From c7ae63201ab193a7dc7fb7e0d81238645111ffac Mon Sep 17 00:00:00 2001
From: Garry Tan <garrytan@gmail.com>
Date: Sun, 14 Jun 2026 11:40:57 -0700
Subject: [PATCH] v1.58.1.0 feat: hermetic local E2E + Conductor prose
 AskUserQuestion (#2004)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* feat: add shared call-time isConductor() helper

Single source of truth for Conductor host detection in TS consumers
(CONDUCTOR_WORKSPACE_PATH / CONDUCTOR_PORT). Reads the passed env at
call time, not a module-load snapshot, so unit tests can pin the env
inline without Bun --preload (esm-hoist-breaks-env-pin-bootstrap).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* test: harden question-preference-hook harness against ambient Conductor env

runHook copied all of process.env into the hook subprocess, so running the
suite inside Conductor (CONDUCTOR_WORKSPACE_PATH/PORT set) would leak those
markers. Strip them so the existing cases deterministically characterize
NON-Conductor behavior before the Conductor branch lands. Baseline: 15 pass.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: PreToolUse hook denies AskUserQuestion in Conductor, redirects to prose

Conductor disables native AskUserQuestion and routes through a flaky MCP
variant that returns '[Tool result missing due to internal error]'. The
hook now denies any AUQ call in a Conductor session and instructs the model
to render a prose decision brief instead (transport avoidance, not preference
enforcement) — firing for one-way doors too, with a typed-confirmation
requirement for destructive paths.

Precedence: never-ask auto-decide still wins (user already settled those);
Conductor prose is the fallback for everything else; non-Conductor behavior
is byte-for-byte unchanged. Restructured the per-question loop to compute
eligibility without early-returning so the Conductor branch can run as the
fallback while preserving memoryContext on every exit.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: Conductor renders AskUserQuestion decisions as prose by default

In Conductor, native AskUserQuestion is disabled and the MCP variant is
flaky, so skills now render every decision as a plain-text prose brief the
user answers by typing a letter — proactively, not as a failure reaction.

- Preamble emits CONDUCTOR_SESSION, gated on != headless so eval/CI inside
  Conductor still BLOCKs instead of rendering prose to nobody.
- AskUserQuestion Format gains a Conductor-default-prose rule (auto-decide
  preferences still apply first; prose decisions log via gstack-question-log
  since PostToolUse never fires), a one-way/destructive typed-confirmation
  rule, and a typed-reply continuation protocol for split chains.
- Regenerated all SKILL.md + ship golden fixtures; bumped affected carve
  skeleton caps to absorb the always-loaded additions.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: deploy the Conductor AskUserQuestion hook (setup + upgrade migration)

The PreToolUse hook only delivers its Conductor-prose guarantee if it's
installed, but setup skips hook registration in non-interactive (conductor/CI)
setups. Two fixes so layer 3 actually deploys:

- setup: treat a Conductor workspace as an implicit opt-in for the PreToolUse
  hook on the silent fall-through (never overriding an explicit opt-out).
- migration v1.58.0.0: re-register the hook for existing Conductor installs on
  /gstack-upgrade, idempotent and respecting plan_tune_hooks=no.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* test: E2E for Conductor prose + fix auto-decide-preserved GSTACK_HOME bug

- New skill-e2e-conductor-prose (periodic): Conductor env + plan-eng-review
  surfaces a prose decision brief, not a silent skip. Header documents this is
  end-to-end behavior coverage; the deterministic Conductor guard is the
  question-preference-hook unit test (the PTY harness can't register the MCP
  variant — Codex #10).
- Fix the pre-existing bug in auto-decide-preserved: it seeded the never-ask
  preference under GSTACK_HOME=tmpHome but never passed GSTACK_HOME into the
  PTY run, so the spawned claude read the real ~/.gstack and the preference
  was inert (Codex #9). Now passes GSTACK_HOME + CONDUCTOR_WORKSPACE_PATH to
  prove auto-decide still wins over the Conductor prose redirect.
- Register both in touchfiles (periodic tier).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* v1.58.0.0 feat: Conductor renders AskUserQuestion decisions as prose

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* test: strip ambient Conductor env in memory-cache-injection hook harness

Same dev-in-Conductor leak fixed for question-preference-hook: this suite's
runHook copies process.env, so running it inside Conductor flipped the
defer-path memoryContext assertions into the [conductor] prose deny. Strip
CONDUCTOR_* so the cases characterize non-Conductor behavior. (CI is headless,
so this only bit local Conductor runs.)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: gstack-detach — run agent eval/bench jobs in their own session

Long agent-run jobs (30-60 min evals, benchmarks) die when the harness sends
SIGTERM to a background task's process group on turn boundaries / monitor
stops / interruptions (observed: 'script test:gate terminated by signal
SIGTERM'). gstack-detach runs the command in a fresh session (python3
os.setsid, or setsid on Linux, nohup fallback) so a group SIGTERM can't reach
it, and wraps it in caffeinate -i on macOS so idle-sleep can't kill it either.
Returns immediately; caller polls the logfile. Secrets stay in env, never argv.

The guard test pins the contract: the command runs in a different process
group than the caller and outlives the launching shell.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: eval:bg* scripts — detached eval runs for agents

Agent-facing convenience scripts that launch the eval suites through
gstack-detach so a harness SIGTERM can't kill a long run. eval:bg (diff-based),
eval:bg:all, eval:bg:gate, eval:bg:periodic — each returns immediately and
streams to /tmp/gstack-evals.log for polling. The plain test:evals / test:e2e
scripts stay foreground for humans.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs: CLAUDE.md — agents must run long evals via gstack-detach

Codifies the detached-execution default: agent-launched eval/benchmark runs go
through bin/gstack-detach (or the eval:bg* scripts) so a harness SIGTERM or
macOS idle-sleep can't kill a 30-60 min run, then poll the log with a
death-aware watcher. Humans keep foreground scripts.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: harden gstack-detach against all four eval-infra killers

The basic bash detach fixed SIGTERM but a real run on a shared dev box hit
three more killers: cross-worktree API saturation (15-way concurrency x a
sibling worktree mass-timed-out the suite), a silent hang (periodic bun died
with no exit marker), and shared-/tmp log contamination (a concurrent
worktree's agent output bled into the log). Rewrite as a portable python3 tool
that bakes in all four fixes:

- fork + setsid: SIGTERM-proof (own session, survives harness polite-quit)
- caffeinate -i on macOS: no idle-sleep death
- --lock NAME (fcntl, machine-wide): concurrent worktrees SERIALIZE instead of
  saturating the shared model API
- run-scoped default log (~/.gstack-dev/eval-runs/<label>-<slug>-<branch>-<ts>-<pid>):
  no cross-worktree collision/contamination
- --timeout watchdog + a guaranteed '### gstack-detach EXIT=<code> ###' sentinel
  on every terminal path: no silent hang, finished-vs-died always detectable

Guard test pins all four: detached pgid differs + outlives launcher, run-scoped
log path, watchdog EXIT=timeout, and lock serialization (second run WAITS).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: eval:bg* use run-scoped logs + machine lock + watchdog

Drop the shared /tmp/gstack-evals.log path (the cross-worktree collision that
contaminated a live run) for gstack-detach's run-scoped default, and add the
machine-wide gstack-evals lock (concurrent worktrees serialize, no API
saturation) plus per-tier watchdog timeouts (60/90/120 min). Each eval:bg*
prints its run-scoped log path to poll.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs: wire detached-eval guidance into /ship + correct CLAUDE.md flags

- /ship eval step (sections/tests.md): long eval suites launch via gstack-detach
  (own session, machine lock, EXIT sentinel) so a turn boundary can't kill a
  30+ min run mid-ship — the exact failure observed during this branch's ship.
- CLAUDE.md: correct the now-stale /tmp reference; document the --lock (serialize
  worktrees, no API saturation), --timeout watchdog, run-scoped log, and the
  guaranteed EXIT sentinel the poller breaks on.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* refactor: extract pure promotedEnv() from conductor-env-shim

Single source of truth for GSTACK_* key promotion semantics. The ambient
promoteConductorEnv() becomes a wrapper; behavior-preserving. Needed by the
hermetic env builder which must not mutate process.env.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: hermetic child-env builder for E2E runners

Allowlist scrub (basics/network/named-auth kept; CONDUCTOR_*, CLAUDE_*,
GSTACK_*, MCP_*, GBRAIN_*, operator credentials dropped), per-runner
extraAllow, overrides merge last, EVALS_HERMETIC=0 byte-identical escape
hatch read at call time (ESM-hoist safe). Sync memoized singleton temp dirs
(<runRoot>/.claude keeps the extractPlanFilePath contract), seeded
.claude.json for non-interactive first run, pid-aware GC of crashed runs.
19 free unit tests.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: session-runner spawns hermetic children + isolation canaries

claude -p children now get the allowlist-scrubbed env and a gated
--strict-mcp-config (EVALS_HERMETIC=0 restores operator env AND args).
Two gate-tier canaries make the clean room falsifiable: hermetic-canary
asserts env redirect + scrub + zero MCP servers + nonzero API-key cost
from the Bash tool_result (never model prose); hermetic-sentinel plants a
poisoned operator config (user CLAUDE.md + MCP server) and proves the
child cannot see it. Empirically verified on claude 2.1.175: print mode
needs no seed config (the seed serves the PTY path); the child CLI sets
CLAUDECODE for its own tools, so that scrub is pinned in unit tests, not
E2E. hermetic-env.ts joins GLOBAL_TOUCHFILES.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: PTY runner spawns hermetic claude sessions

launchClaudePty children get the allowlist-scrubbed env, a gated
--strict-mcp-config, and the session exposes hermeticConfigDir for
forensics (hermetic plan files live under <dir>/plans/ and still match
extractPlanFilePath via the /.claude dir-name contract). Seeded trust
state covers repo-cwd sessions; the 15s trust-watcher stays as fallback.
Verified foreground via the plan-mode-no-op gate test.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: codex/gemini runners spawn hermetic children

Same allowlist scrub as the claude runners, with each provider's auth
surface re-admitted via extraAllow (codex: OPENAI_API_KEY/CODEX_* plus
its tempHome .codex copy; gemini: GEMINI_*/GOOGLE_* with real HOME for
~/.gemini auth). The gemini spawn previously inherited the full operator
env with no env property at all.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: agent-sdk-runner spawns hermetic children via complete Options.env

The historical 'env: breaks SDK auth' failure was partial-env replacement:
Options.env replaces the child's entire environment, so objects lacking
ANTHROPIC_API_KEY killed auth. Passing the complete hermetic env (key +
PATH + redirected CLAUDE_CONFIG_DIR/GSTACK_HOME) works — validated live
via query() with a Bash tool call (success, real cost, Conductor vars
scrubbed). Per-test opts.env merges last; ambient key mutation still
works because the builder reads process.env at call time.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* test: static tripwire pins hermetic wiring in all five runners

Free-tier invariants: every runner builds child env via hermeticChildEnv,
no raw ...process.env spread at any spawn site, --strict-mcp-config gated
on isHermeticEnabled in both claude runners, and no test callsite passes
the operator env into a runner's override parameter (scoped to runner
calls — unit tests spawning gstack bin scripts directly are exempt).
Mirrors the terminal-agent-pid-identity / server-embedder-terminal-port
tripwire idiom.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* test: refresh codex/factory ship goldens with detached-eval block

a38089aa added the gstack-detach guidance to the ship template and
updated the claude golden; the codex and factory goldens missed the same
16-line block. Regenerated via bun run gen:skill-docs.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs: hermetic local E2E is the default; retire stale SDK env warning

CLAUDE.md now documents the hermetic clean room (allowlist scrub, fresh
seeded CLAUDE_CONFIG_DIR, temp GSTACK_HOME, --strict-mcp-config),
EVALS_HERMETIC=0 as the debug escape hatch, and replaces the 'never pass
env: to runAgentSdkTest' rule with the verified mechanism (partial-env
replacement was the failure; complete env is safe).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix: operational-learning fixture copies lib/jsonl-store.ts with the bin

gstack-learnings-log imports $SCRIPT_DIR/../lib/jsonl-store.ts (hasInjection,
v1.57.5.0) — copying only the bin scripts into the temp fixture broke the
script with exit 1 since then. Latent because diff-based selection rarely
runs this test; surfaced when hermetic-env.ts joined GLOBAL_TOUCHFILES and
selected everything. Reproduced outside the hermetic env to confirm blame.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix: ios-qa daemon scenarios use unique pidfiles under --concurrent

All scenarios shared join(workDir, 'daemon.pid') through a module-scope
workDir binding that beforeEach reassigns mid-flight under bun --concurrent.
First daemon claims; siblings get already_running against the test process's
own always-alive pid and fail in milliseconds — the failure mode seen at
15-way gate concurrency. Per-claim unique pidfiles keep the single-instance
semantics under test.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix: workflow judge re-appends body-carved sections after the marker slice

runWorkflowJudge appended sections/*.md before slicing startMarker..endMarker.
That handles skills that moved their MARKERS into sections (plan-eng,
plan-design) but not document-release, which keeps its markers in the
skeleton and carved the workflow BODY (Steps 2-9 -> sections/release-body.md)
AFTER the endMarker — so the slice dropped it and the judge scored
completeness 2 ('Steps 2-9 are in an external file'). Now any carved section
the marker window excluded is re-appended, so the judge sees the full
workflow the agent executes. document-release: completeness 2->5, clarity
3->4. ship/plan-ceo/plan-eng/plan-design judges unchanged (their section
content is already inside the slice, so the head-dedup skips re-append).

Pre-existing since the v1.57.0.0 carve (#1907); surfaced now because
hermetic-env.ts is a global touchfile that selects every llm-judge test.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* harden: hermetic temp-dir GC grace window + half-seed cleanup

Codex adversarial review (ship) flagged two temp-dir lifecycle edges:
- GC deleted any dead-pid dir; PID reuse could delete a freshly-created dir
  whose original pid exited and was recycled to a live process. Now requires
  BOTH a dead pid AND mtime older than a 1h floor.
- A seed-write failure after mkdir left an unseeded dir named with our live
  pid that this process's GC skips, leaking until exit. Now the partial dir
  is torn down before the (still loud) rethrow.

Two findings left as-is by design: HOME stays allowlisted (CLAUDE_CONFIG_DIR
wins for claude; codex/gemini need ~/.codex|~/.gemini auth; FS sandbox is
TODOS.md:454 scope; the hermetic-sentinel canary proves config isolation),
and PTY extraArgs --mcp-config is a deliberate caller opt-in like env overrides.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs: document hermetic-by-default E2E + eval:bg detached runs in CONTRIBUTING

The Testing & evals section now tells contributors that local E2E runners
spawn children through a sealed clean room (allowlist-scrubbed env, seeded
CLAUDE_CONFIG_DIR, temp GSTACK_HOME, --strict-mcp-config) so local signal
matches CI, with EVALS_HERMETIC=0 as the escape hatch. The eval-tools list
gains the eval:bg* detached-run scripts (gstack-detach: SIGTERM-proof,
caffeinate-wrapped, machine-locked, run-scoped logs, EXIT= sentinel).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* chore: sync package.json to 1.58.1.0

The merge took main's package.json (1.58.0.0); gstack-version-bump repair
fixed the working tree but the change was left uncommitted. Without this the
committed tree disagrees with VERSION and CI's version-match test fails.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs: regenerate diagram SKILL.md with Conductor prose preamble

The diagram skill (new from main) was missing the Conductor-session prose
AskUserQuestion blocks that gen-skill-docs propagates to every SKILL.md.
Pure generated output; reproduced by bun run gen:skill-docs.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
---
 CHANGELOG.md                                  | 102 +++++++
 CLAUDE.md                                     |  53 +++-
 CONTRIBUTING.md                               |  31 ++
 SKILL.md                                      |   7 +
 VERSION                                       |   2 +-
 autoplan/SKILL.md                             |  19 +-
 benchmark-models/SKILL.md                     |   7 +
 benchmark/SKILL.md                            |   7 +
 bin/gstack-detach                             | 167 +++++++++++
 browse/SKILL.md                               |   7 +
 canary/SKILL.md                               |  19 +-
 codex/SKILL.md                                |  19 +-
 context-restore/SKILL.md                      |  19 +-
 context-save/SKILL.md                         |  19 +-
 cso/SKILL.md                                  |  19 +-
 design-consultation/SKILL.md                  |  19 +-
 design-html/SKILL.md                          |  19 +-
 design-review/SKILL.md                        |  19 +-
 design-shotgun/SKILL.md                       |  19 +-
 devex-review/SKILL.md                         |  19 +-
 diagram/SKILL.md                              |  19 +-
 document-generate/SKILL.md                    |  19 +-
 document-release/SKILL.md                     |  19 +-
 gstack-upgrade/migrations/v1.58.0.0.sh        |  63 ++++
 health/SKILL.md                               |  19 +-
 .../claude/hooks/question-preference-hook.ts  |  87 +++---
 investigate/SKILL.md                          |  19 +-
 ios-clean/SKILL.md                            |  19 +-
 ios-design-review/SKILL.md                    |  19 +-
 ios-fix/SKILL.md                              |  19 +-
 ios-qa/SKILL.md                               |  19 +-
 ios-sync/SKILL.md                             |  19 +-
 land-and-deploy/SKILL.md                      |  19 +-
 landing-report/SKILL.md                       |  19 +-
 learn/SKILL.md                                |  19 +-
 lib/conductor-env-shim.ts                     |  26 +-
 lib/is-conductor.ts                           |  19 ++
 make-pdf/SKILL.md                             |   7 +
 office-hours/SKILL.md                         |  19 +-
 open-gstack-browser/SKILL.md                  |  19 +-
 package.json                                  |   6 +-
 pair-agent/SKILL.md                           |  19 +-
 plan-ceo-review/SKILL.md                      |  19 +-
 plan-design-review/SKILL.md                   |  19 +-
 plan-devex-review/SKILL.md                    |  19 +-
 plan-eng-review/SKILL.md                      |  19 +-
 plan-tune/SKILL.md                            |  19 +-
 qa-only/SKILL.md                              |  19 +-
 qa/SKILL.md                                   |  19 +-
 retro/SKILL.md                                |  19 +-
 review/SKILL.md                               |  19 +-
 scrape/SKILL.md                               |  19 +-
 .../preamble/generate-ask-user-format.ts      |  12 +-
 .../preamble/generate-preamble-bash.ts        |   7 +
 setup                                         |  20 +-
 setup-browser-cookies/SKILL.md                |   7 +
 setup-deploy/SKILL.md                         |  19 +-
 setup-gbrain/SKILL.md                         |  19 +-
 ship/SKILL.md                                 |  19 +-
 ship/sections/tests.md                        |  16 +
 ship/sections/tests.md.tmpl                   |  16 +
 skillify/SKILL.md                             |  19 +-
 spec/SKILL.md                                 |  38 ++-
 sync-gbrain/SKILL.md                          |  19 +-
 test/agent-sdk-runner.test.ts                 |   8 +-
 test/fixtures/golden/claude-ship-SKILL.md     |  19 +-
 test/fixtures/golden/codex-ship-SKILL.md      |  35 ++-
 test/fixtures/golden/factory-ship-SKILL.md    |  35 ++-
 test/gstack-detach.test.ts                    |  96 ++++++
 test/helpers/agent-sdk-runner.ts              |  20 +-
 test/helpers/carve-guards.ts                  |  20 +-
 test/helpers/claude-pty-runner.ts             |  20 +-
 test/helpers/codex-session-runner.ts          |  14 +-
 test/helpers/gemini-session-runner.ts         |   8 +-
 test/helpers/hermetic-env.test.ts             | 269 +++++++++++++++++
 test/helpers/hermetic-env.ts                  | 276 ++++++++++++++++++
 test/helpers/session-runner.ts                |  10 +-
 test/helpers/touchfiles.ts                    |  19 +-
 test/hermetic-wiring.test.ts                  | 113 +++++++
 test/is-conductor.test.ts                     |  50 ++++
 test/memory-cache-injection.test.ts           |   5 +
 test/preamble-compose.test.ts                 |  10 +
 test/question-preference-hook.test.ts         | 112 ++++++-
 test/resolver-ask-user-format.test.ts         |  21 +-
 test/skill-e2e-auto-decide-preserved.test.ts  |   8 +
 test/skill-e2e-conductor-prose.test.ts        |  69 +++++
 test/skill-e2e-hermetic-canary.test.ts        | 190 ++++++++++++
 test/skill-e2e-ios.test.ts                    |  20 +-
 test/skill-llm-eval.test.ts                   |  16 +-
 89 files changed, 2747 insertions(+), 221 deletions(-)
 create mode 100755 bin/gstack-detach
 create mode 100755 gstack-upgrade/migrations/v1.58.0.0.sh
 create mode 100644 lib/is-conductor.ts
 create mode 100644 test/gstack-detach.test.ts
 create mode 100644 test/helpers/hermetic-env.test.ts
 create mode 100644 test/helpers/hermetic-env.ts
 create mode 100644 test/hermetic-wiring.test.ts
 create mode 100644 test/is-conductor.test.ts
 create mode 100644 test/skill-e2e-conductor-prose.test.ts
 create mode 100644 test/skill-e2e-hermetic-canary.test.ts

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2dd4b64a8..034ab7bcf 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,107 @@
 # Changelog
 
+## [1.58.1.0] - 2026-06-14
+
+## **Local evals stop lying. Spawned `claude` test children run in a sealed clean room,**
+## **and in Conductor every decision is a plain-text brief you answer with a letter.**
+
+Two things shipped here. First, the local E2E harness is now hermetic by default:
+every spawned agent (claude -p, the real-PTY plan-mode runner, the Agent SDK
+runner, plus the codex and gemini runners) gets an allowlist-scrubbed environment,
+a fresh seeded `CLAUDE_CONFIG_DIR`, a temp `GSTACK_HOME`, and `--strict-mcp-config`.
+Before this, a dev machine leaked the operator's `~/.claude` config, MCP servers
+(gbrain, Conductor), skills, `~/.gstack` decision logs, and `CONDUCTOR_*`/`CLAUDECODE`
+env into every child, so local eval results disagreed with CI for reasons that had
+nothing to do with the code under test. Now local signal matches CI. Set
+`EVALS_HERMETIC=0` to debug against real operator state.
+
+Second, in a Conductor session gstack no longer fights Conductor's flaky
+AskUserQuestion tool. It detects the session and renders every decision as a prose
+brief, a labeled question with a recommendation, per-option completeness scores, and
+"reply with a letter," enforced by a PreToolUse hook that denies the tool and
+redirects to prose. Destructive confirmations demand an explicit typed answer.
+
+Agents that launch long eval runs get `gstack-detach`: a SIGTERM-proof, idle-sleep-proof
+wrapper (fresh session + `caffeinate`) with a machine-wide lock so concurrent
+worktrees serialize instead of saturating the model API, run-scoped logs, and a
+guaranteed `EXIT=` sentinel so a poller never mistakes silence for success.
+
+### The numbers that matter
+
+Measured against the gate eval suite on a contaminated dev box (gbrain MCP up, live
+Conductor session, sibling worktrees). Reproduce: `bun test` (free unit + wiring
+tripwire) and `EVALS=1 EVALS_TIER=gate bun test test/skill-e2e-hermetic-canary.test.ts`.
+
+| Metric | Before | After | Δ |
+|--------|--------|-------|---|
+| Spawned-child env | full operator `process.env` | allowlist-scrubbed | sealed |
+| Runners hermeticized | 0 of 5 | 5 of 5 | +5 |
+| Operator MCP servers visible to child | all (gbrain, Conductor) | 0 (`--strict-mcp-config`) | isolated |
+| Config isolation proof | none | poisoned-operator sentinel canary | falsifiable |
+| Long eval runs surviving a turn-boundary SIGTERM | no | yes (`gstack-detach`) | survives |
+
+The clean room is falsifiable, not asserted: a `hermetic-sentinel` gate canary
+plants a poisoned operator config (a user `CLAUDE.md` + an MCP server) and fails if
+the child can see any of it, and a free static tripwire fails CI if any runner
+reverts to a raw `process.env` spread.
+
+### What this means for contributors
+
+Run evals locally and trust the result. You no longer have to push to CI to find
+out whether a failure was real or just your machine bleeding context into the agent.
+Three latent bugs the old harness hid surfaced the moment the suite ran clean and
+are fixed: a coverage-judge that scored carved skills against half a document, an
+ios-qa daemon test that collided on a shared pidfile under concurrency, and an
+operational-learning fixture missing a lib it imports. Start a run with
+`bun run eval:bg:gate`; flip `EVALS_HERMETIC=0` only when you deliberately want your
+real `~/.claude` in the loop.
+
+### Itemized changes
+
+#### Added
+- **Hermetic E2E environment** (`test/helpers/hermetic-env.ts`): allowlist env
+  builder (process basics, network/proxy vars, named `ANTHROPIC_*` auth, per-runner
+  `extraAllow`), pure `promotedEnv()` shared with `lib/conductor-env-shim.ts`, a
+  sync-memoized singleton temp dir (`<runRoot>/.claude` keeps the plan-file path
+  contract), a seeded `.claude.json` for non-interactive first run, and pid-aware GC
+  of crashed runs. Default-on; `EVALS_HERMETIC=0` restores the legacy env AND drops
+  `--strict-mcp-config`.
+- **Two gate-tier isolation canaries** (`test/skill-e2e-hermetic-canary.test.ts`):
+  `hermetic-canary` asserts env redirect + scrub + zero MCP servers + nonzero
+  API-key cost from the Bash tool_result (not model prose); `hermetic-sentinel`
+  proves the child cannot see a planted poisoned operator config.
+- **Static wiring tripwire** (`test/hermetic-wiring.test.ts`): free-tier invariants
+  that fail CI if any of the five runners drops `hermeticChildEnv()`, the gated
+  `--strict-mcp-config`, or leaks `process.env` through a callsite override.
+- **`gstack-detach`** + `eval:bg` / `eval:bg:all` / `eval:bg:gate` / `eval:bg:periodic`
+  scripts: detached, SIGTERM-proof, `caffeinate`-wrapped eval runs with a machine-wide
+  lock, per-run logs under `~/.gstack-dev/eval-runs/`, a watchdog, and an `EXIT=`
+  sentinel.
+- **Conductor prose AskUserQuestion**: when a Conductor session is detected, every
+  decision renders as a prose brief (labeled question, recommendation, per-option
+  completeness, reply-with-a-letter), enforced by a PreToolUse hook that denies the
+  tool and redirects. Auto-decide preferences still apply first; destructive
+  confirmations require an explicit typed answer. Installed for Conductor even in
+  non-interactive setup, with an upgrade migration for existing installs.
+
+#### Changed
+- All five E2E runners (`session-runner`, `claude-pty-runner`, `agent-sdk-runner`,
+  `codex-session-runner`, `gemini-session-runner`) spawn children through
+  `hermeticChildEnv()`. The Agent SDK runner now receives a COMPLETE hermetic env
+  via `Options.env` (the old "never pass env: to the SDK" rule was partial-env
+  replacement; a complete env is safe).
+- `hermetic-env.ts` is a global touchfile, so any change to it selects every E2E +
+  judge test.
+- CLAUDE.md documents hermetic-by-default local evals and retires the stale SDK env
+  warning.
+
+#### Fixed
+- The workflow LLM-judge now re-appends body-carved `sections/*.md` after the marker
+  slice, so carved skills (document-release) are judged on the full workflow the
+  agent executes instead of a half-document.
+- ios-qa daemon scenarios use unique pidfiles, fixing `already_running` collisions
+  under `bun test --concurrent`.
+
 ## [1.58.0.0] - 2026-06-12
 
 ## **Your documents grow diagrams. Mermaid and excalidraw fences render as real pictures,**
diff --git a/CLAUDE.md b/CLAUDE.md
index 03384ae79..984844902 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -31,11 +31,26 @@ use Codex's own auth from `~/.codex/` config — no `OPENAI_API_KEY` env var nee
 `lib/conductor-env-shim.ts`) promotes `GSTACK_ANTHROPIC_API_KEY` /
 `GSTACK_OPENAI_API_KEY` to their canonical names inside gstack's TS binaries.
 Tests run through gstack entrypoints inherit this promotion automatically.
-Don't echo the key value to stdout, logs, or shell history. When passing to a
-test's Agent SDK, do NOT pass `env: {...}` to `runAgentSdkTest` — the SDK's
-auth pipeline doesn't pick up the key the same way when env is supplied as an
-object (confirmed failure mode). Mutate `process.env.ANTHROPIC_API_KEY`
-ambiently before the call and restore in `finally`.
+Don't echo the key value to stdout, logs, or shell history. The historical
+"never pass `env:` to `runAgentSdkTest`" rule is retired: the failure was
+partial-env replacement (the SDK's `Options.env` REPLACES the child's entire
+environment, so an object without the key broke auth). The runner now always
+passes a COMPLETE hermetic env with per-test `env:` merged last, so per-test
+overrides are safe; ambient `process.env.ANTHROPIC_API_KEY` mutation also
+still works (the env builder reads process.env at call time).
+
+**Hermetic local E2E (default).** Every E2E runner (claude -p, PTY, Agent
+SDK, codex, gemini) spawns children through `test/helpers/hermetic-env.ts`:
+allowlist-scrubbed env (operator `CONDUCTOR_*`, `CLAUDE_*`, `GSTACK_*`,
+`MCP_*`, `GBRAIN_*`, and credentials like `GH_TOKEN` never reach children),
+a fresh seeded `CLAUDE_CONFIG_DIR` (no operator `~/.claude` CLAUDE.md /
+MCP servers / skills), a temp `GSTACK_HOME`, and `--strict-mcp-config`.
+Local eval signal matches CI. Debug against real operator state with
+`EVALS_HERMETIC=0` (restores the legacy env AND drops the strict-MCP flag).
+Per-test `env:` overrides merge last, so deliberate contamination
+(`CONDUCTOR_WORKSPACE_PATH`, per-test `GSTACK_HOME`) keeps working. Wiring
+is pinned by `test/hermetic-wiring.test.ts` (static tripwire) and two
+gate-tier canaries in `test/skill-e2e-hermetic-canary.test.ts`.
 
 E2E tests stream progress in real-time (tool-by-tool via `--output-format stream-json
 --verbose`). Results are persisted to `~/.gstack-dev/evals/` with auto-comparison
@@ -828,6 +843,34 @@ them. Report progress at each check (which tests passed, which are running, any
 failures so far). The user wants to see the run complete, not a promise that
 you'll check later.
 
+## Running evals as an agent: always detach (SIGTERM-proof)
+
+When **you (an agent/harness)** launch a long eval/benchmark run, run it through
+`bin/gstack-detach` — NEVER as a plain backgrounded Bash task. A plain background
+task lives in the harness's process group, so a SIGTERM ("polite quit") on a turn
+boundary, a stopped Monitor, or an interruption kills the run mid-flight (observed:
+`script "test:gate" was terminated by signal SIGTERM` ~40 min into a run). On macOS
+the run can also die to idle-sleep. `gstack-detach` fixes both: a fresh session
+(escapes the group SIGTERM) wrapped in `caffeinate -i` (blocks idle-sleep).
+
+- Use the `eval:bg*` scripts (`eval:bg`, `eval:bg:all`, `eval:bg:gate`,
+  `eval:bg:periodic`) — they wrap the eval command in `gstack-detach` with the
+  machine-wide `gstack-evals` lock (concurrent worktrees serialize instead of
+  saturating the shared model API), a per-tier watchdog, and a **run-scoped** log
+  under `~/.gstack-dev/eval-runs/` (no shared-`/tmp` collision). Each prints its
+  log path. Or call `gstack-detach [--lock NAME] [--timeout SECS] [--label LBL] --
+  <cmd>` directly for any long agent job. Export `ANTHROPIC_API_KEY` first (never
+  pass keys in argv).
+- Then **poll the printed logfile** with a death-aware watcher: break on the
+  guaranteed `### gstack-detach EXIT=<code> ###` sentinel (success AND failure are
+  both marked, so silence is never mistaken for success). The detached run survives
+  even if your watcher gets reaped, so re-checking the log always works.
+- Why the lock: a shared dev box with several Conductor worktrees will rate-limit
+  the model API if two eval suites run at once (15-way concurrency each), which
+  mass-times-out E2E tests. The lock makes the second run WAIT, not collide.
+- Humans running `bun run test:evals` foreground in their own terminal don't need
+  this — Ctrl-C is intended there. Detachment is for agent-launched runs only.
+
 ## E2E test fixtures: extract, don't copy
 
 **NEVER copy a full SKILL.md file into an E2E test fixture.** SKILL.md files are
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 5a56ef5d3..b75d4a898 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -176,6 +176,18 @@ EVALS=1 bun test test/skill-e2e-*.test.ts
 - Saves full NDJSON transcripts and failure JSON for debugging
 - Tests live in `test/skill-e2e-*.test.ts` (split by category), runner logic in `test/helpers/session-runner.ts`
 
+**Hermetic by default.** Every E2E runner (claude -p, the real-PTY plan-mode
+runner, the Agent SDK runner, plus the codex and gemini runners) spawns its child
+through `test/helpers/hermetic-env.ts`: an allowlist-scrubbed environment, a fresh
+seeded `CLAUDE_CONFIG_DIR`, a temp `GSTACK_HOME`, and `--strict-mcp-config`. Your
+operator `~/.claude` config, MCP servers (gbrain, Conductor), skills, `~/.gstack`
+decision logs, and `CONDUCTOR_*` env never leak into the child, so local eval
+signal matches CI instead of disagreeing for reasons unrelated to the code under
+test. Set `EVALS_HERMETIC=0` to debug against your real operator state (this also
+drops `--strict-mcp-config`). The wiring is pinned by `test/hermetic-wiring.test.ts`
+(a free static tripwire) and two gate-tier isolation canaries in
+`test/skill-e2e-hermetic-canary.test.ts`.
+
 ### E2E observability
 
 When E2E tests run, they produce machine-readable artifacts in `~/.gstack-dev/`:
@@ -198,6 +210,25 @@ bun run eval:compare         # compare two runs — shows per-test deltas + Take
 bun run eval:summary         # aggregate stats + per-test efficiency averages across runs
 ```
 
+**Detached runs for agents and long suites.** When an agent (or you, for a run
+you don't want to babysit) launches a long eval, use the `eval:bg*` scripts. They
+wrap the eval command in `bin/gstack-detach`: a fresh session that escapes a
+turn-boundary SIGTERM, a `caffeinate` wrapper that blocks idle-sleep, a machine-wide
+`gstack-evals` lock so concurrent worktrees serialize instead of saturating the
+model API, a run-scoped log under `~/.gstack-dev/eval-runs/`, a per-tier watchdog,
+and a guaranteed `### gstack-detach EXIT=<code> ###` sentinel so a poller never
+mistakes silence for success.
+
+```bash
+bun run eval:bg              # detached test:evals (diff-based)
+bun run eval:bg:all          # detached test:evals:all
+bun run eval:bg:gate         # detached gate-tier suite
+bun run eval:bg:periodic     # detached periodic-tier suite
+```
+
+Each prints its log path. Humans running `bun run test:evals` foreground in their
+own terminal don't need this — Ctrl-C is intended there.
+
 **Eval comparison commentary:** `eval:compare` generates natural-language Takeaway sections interpreting what changed between runs — flagging regressions, noting improvements, calling out efficiency gains (fewer turns, faster, cheaper), and producing an overall summary. This is driven by `generateCommentary()` in `eval-store.ts`.
 
 Artifacts are never cleaned up — they accumulate in `~/.gstack-dev/` for post-mortem debugging and trend analysis.
diff --git a/SKILL.md b/SKILL.md
index 8711ae7f3..90774950e 100644
--- a/SKILL.md
+++ b/SKILL.md
@@ -48,6 +48,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
diff --git a/VERSION b/VERSION
index 3a62339b5..eb4d8b4b5 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-1.58.0.0
+1.58.1.0
\ No newline at end of file
diff --git a/autoplan/SKILL.md b/autoplan/SKILL.md
index de7174b03..49db38ff9 100644
--- a/autoplan/SKILL.md
+++ b/autoplan/SKILL.md
@@ -57,6 +57,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -306,7 +313,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -328,7 +337,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -412,7 +425,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/benchmark-models/SKILL.md b/benchmark-models/SKILL.md
index 06fbd6a7b..9ea455ad7 100644
--- a/benchmark-models/SKILL.md
+++ b/benchmark-models/SKILL.md
@@ -51,6 +51,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
diff --git a/benchmark/SKILL.md b/benchmark/SKILL.md
index 20b15e0d6..6aacca4b4 100644
--- a/benchmark/SKILL.md
+++ b/benchmark/SKILL.md
@@ -51,6 +51,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
diff --git a/bin/gstack-detach b/bin/gstack-detach
new file mode 100755
index 000000000..101e86eee
--- /dev/null
+++ b/bin/gstack-detach
@@ -0,0 +1,167 @@
+#!/usr/bin/env python3
+"""gstack-detach — run a long agent job (evals, benchmarks, syncs) robustly.
+
+Agent-launched long jobs on a shared dev box keep dying to environmental
+killers. This tool bakes in the fixes so gstack (and every gstack user) runs
+them properly:
+
+  * SIGTERM-proof: fork + setsid puts the job in its OWN session, so the
+    harness's "polite quit" SIGTERM to the launching process group can't reach
+    it (observed: `script "test:gate" was terminated by signal SIGTERM`).
+  * No idle-sleep death (macOS): wraps the command in `caffeinate -i`.
+  * No cross-worktree API saturation: `--lock NAME` takes a machine-wide
+    advisory lock so concurrent Conductor worktrees SERIALIZE their eval runs
+    instead of saturating the shared model API (which mass-times-out E2E suites).
+  * No shared-/tmp collision: a run-scoped log path by default
+    (~/.gstack-dev/eval-runs/<label>-<slug>-<branch>-<ts>-<pid>.log), so
+    concurrent runs never clobber or contaminate each other's logs.
+  * No silent hang: `--timeout SECS` watchdog kills a stalled run, and a
+    `### gstack-detach EXIT=<code> ###` sentinel is ALWAYS appended on a
+    terminal path so a poller can tell finished-vs-died (silence != success).
+
+Usage:
+  gstack-detach [--log PATH] [--lock NAME] [--timeout SECS] [--label LBL] -- CMD [ARGS...]
+
+Prints `gstack-detach LOG <path>` and returns immediately. Poll the log; break
+on `### gstack-detach EXIT=` (both success and failure are marked).
+
+Secrets are inherited from the environment ONLY — never pass an API key in argv.
+"""
+import argparse
+import os
+import shutil
+import signal
+import subprocess
+import sys
+import time
+from datetime import datetime, timezone
+
+
+def _now():
+    return datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
+
+
+def _git(*args):
+    try:
+        return subprocess.check_output(["git", *args], stderr=subprocess.DEVNULL, text=True).strip()
+    except Exception:
+        return ""
+
+
+def run_scoped_log(label):
+    base = os.path.expanduser("~/.gstack-dev/eval-runs")
+    os.makedirs(base, exist_ok=True)
+    root = _git("rev-parse", "--show-toplevel")
+    slug = os.path.basename(root) if root else "unknown"
+    branch = (_git("branch", "--show-current") or "nobranch").replace("/", "-")
+    stamp = datetime.now(timezone.utc).strftime("%Y%m%d-%H%M%S")
+    return os.path.join(base, f"{label}-{slug}-{branch}-{stamp}-{os.getpid()}.log")
+
+
+def log_line(path, msg):
+    with open(path, "ab", buffering=0) as f:
+        f.write((msg + "\n").encode("utf-8", "replace"))
+
+
+def acquire_lock(name, log):
+    """Machine-wide advisory lock via fcntl (portable on macOS + Linux). Blocks
+    until free so concurrent worktrees serialize rather than saturate the API.
+    Returns the held fd (kept open for the process lifetime)."""
+    import fcntl
+
+    d = os.path.expanduser("~/.gstack/locks")
+    os.makedirs(d, exist_ok=True)
+    fd = open(os.path.join(d, f"{name}.lock"), "w")
+    try:
+        fcntl.flock(fd, fcntl.LOCK_EX | fcntl.LOCK_NB)
+    except OSError:
+        log_line(log, f"### gstack-detach WAITING for lock '{name}' (another run holds it) ### {_now()}")
+        fcntl.flock(fd, fcntl.LOCK_EX)  # block until released
+    fd.write(f"{os.getpid()} {_now()}\n")
+    fd.flush()
+    log_line(log, f"### gstack-detach LOCK '{name}' ACQUIRED ### {_now()}")
+    return fd
+
+
+def child_run(args, log):
+    lock_fd = acquire_lock(args.lock, log) if args.lock else None
+    cmd = args.cmd
+    if shutil.which("caffeinate"):  # macOS: block idle-sleep for the run
+        cmd = ["caffeinate", "-i", *cmd]
+    log_line(log, f"### gstack-detach START label={args.label} pgid={os.getpgid(0)} ### {_now()}")
+    with open(log, "ab", buffering=0) as f:
+        # start_new_session: the command runs in its OWN process group so the
+        # watchdog can killpg() it without also killing this supervisor (which
+        # must survive to write the EXIT sentinel).
+        proc = subprocess.Popen(
+            cmd, stdout=f, stderr=subprocess.STDOUT, stdin=subprocess.DEVNULL, start_new_session=True
+        )
+        if args.timeout and args.timeout > 0:
+            try:
+                code = proc.wait(timeout=args.timeout)
+            except subprocess.TimeoutExpired:
+                log_line(log, f"### gstack-detach WATCHDOG fired after {args.timeout}s — killing ### {_now()}")
+                try:
+                    os.killpg(os.getpgid(proc.pid), signal.SIGTERM)
+                except Exception:
+                    pass
+                time.sleep(5)
+                try:
+                    proc.kill()
+                except Exception:
+                    pass
+                code = "timeout"
+        else:
+            code = proc.wait()
+    log_line(log, f"### gstack-detach EXIT={code} ### {_now()}")
+    if lock_fd:
+        try:
+            lock_fd.close()
+        except Exception:
+            pass
+
+
+def main():
+    ap = argparse.ArgumentParser(add_help=True)
+    ap.add_argument("--log")
+    ap.add_argument("--lock")
+    ap.add_argument("--timeout", type=int, default=0)
+    ap.add_argument("--label", default="job")
+    ap.add_argument("cmd", nargs=argparse.REMAINDER)
+    args = ap.parse_args()
+
+    cmd = args.cmd
+    if cmd and cmd[0] == "--":
+        cmd = cmd[1:]
+    if not cmd:
+        print("gstack-detach: no command given (usage: gstack-detach [opts] -- CMD...)", file=sys.stderr)
+        sys.exit(2)
+    args.cmd = cmd
+
+    log = args.log or run_scoped_log(args.label)
+    os.makedirs(os.path.dirname(log) or ".", exist_ok=True)
+    open(log, "ab").close()
+
+    # Detach: fork so the launching shell returns immediately, then setsid in the
+    # child to escape the harness's process group / controlling terminal.
+    if os.fork() > 0:
+        # flush BEFORE os._exit — os._exit skips stdio buffer flush, which would
+        # otherwise drop this line and leave the caller without the log path.
+        print(f"gstack-detach LOG {log}", flush=True)
+        os._exit(0)
+    os.setsid()
+    devnull = os.open(os.devnull, os.O_RDWR)
+    os.dup2(devnull, 0)
+    lf = os.open(log, os.O_WRONLY | os.O_APPEND | os.O_CREAT, 0o644)
+    os.dup2(lf, 1)
+    os.dup2(lf, 2)
+    try:
+        child_run(args, log)
+    except Exception as e:  # never leave the log without a terminal marker
+        log_line(log, f"### gstack-detach ERROR {e!r} ### {_now()}")
+        log_line(log, f"### gstack-detach EXIT=error ### {_now()}")
+    os._exit(0)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/browse/SKILL.md b/browse/SKILL.md
index 0f670d47a..26816be4b 100644
--- a/browse/SKILL.md
+++ b/browse/SKILL.md
@@ -49,6 +49,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
diff --git a/canary/SKILL.md b/canary/SKILL.md
index d2412d881..b72b99364 100644
--- a/canary/SKILL.md
+++ b/canary/SKILL.md
@@ -49,6 +49,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -298,7 +305,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -320,7 +329,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -404,7 +417,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/codex/SKILL.md b/codex/SKILL.md
index e15c16ec2..cd40e075d 100644
--- a/codex/SKILL.md
+++ b/codex/SKILL.md
@@ -52,6 +52,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -301,7 +308,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -323,7 +332,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -407,7 +420,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/context-restore/SKILL.md b/context-restore/SKILL.md
index ce286184d..11b95b4e2 100644
--- a/context-restore/SKILL.md
+++ b/context-restore/SKILL.md
@@ -53,6 +53,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -302,7 +309,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -324,7 +333,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -408,7 +421,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/context-save/SKILL.md b/context-save/SKILL.md
index d9602de4f..c135352ed 100644
--- a/context-save/SKILL.md
+++ b/context-save/SKILL.md
@@ -52,6 +52,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -301,7 +308,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -323,7 +332,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -407,7 +420,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/cso/SKILL.md b/cso/SKILL.md
index 13f564808..9843d26ca 100644
--- a/cso/SKILL.md
+++ b/cso/SKILL.md
@@ -55,6 +55,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -304,7 +311,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -326,7 +335,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -410,7 +423,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/design-consultation/SKILL.md b/design-consultation/SKILL.md
index 7759e7f02..2ecca91c0 100644
--- a/design-consultation/SKILL.md
+++ b/design-consultation/SKILL.md
@@ -75,6 +75,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -324,7 +331,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -346,7 +355,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -430,7 +443,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/design-html/SKILL.md b/design-html/SKILL.md
index f58f13800..78a0bc926 100644
--- a/design-html/SKILL.md
+++ b/design-html/SKILL.md
@@ -56,6 +56,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -305,7 +312,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -327,7 +336,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -411,7 +424,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/design-review/SKILL.md b/design-review/SKILL.md
index 8c36d8dab..6aa627271 100644
--- a/design-review/SKILL.md
+++ b/design-review/SKILL.md
@@ -53,6 +53,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -302,7 +309,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -324,7 +333,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -408,7 +421,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/design-shotgun/SKILL.md b/design-shotgun/SKILL.md
index 68fdce7d6..18506be7d 100644
--- a/design-shotgun/SKILL.md
+++ b/design-shotgun/SKILL.md
@@ -70,6 +70,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -319,7 +326,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -341,7 +350,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -425,7 +438,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/devex-review/SKILL.md b/devex-review/SKILL.md
index 791db192f..990755b50 100644
--- a/devex-review/SKILL.md
+++ b/devex-review/SKILL.md
@@ -55,6 +55,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -304,7 +311,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -326,7 +335,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -410,7 +423,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/diagram/SKILL.md b/diagram/SKILL.md
index 1ea8fd1a2..345a81140 100644
--- a/diagram/SKILL.md
+++ b/diagram/SKILL.md
@@ -50,6 +50,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -299,7 +306,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -321,7 +330,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -405,7 +418,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/document-generate/SKILL.md b/document-generate/SKILL.md
index 70e71bdb0..b257ad76d 100644
--- a/document-generate/SKILL.md
+++ b/document-generate/SKILL.md
@@ -55,6 +55,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -304,7 +311,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -326,7 +335,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -410,7 +423,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/document-release/SKILL.md b/document-release/SKILL.md
index 5ebf20002..f6aef206c 100644
--- a/document-release/SKILL.md
+++ b/document-release/SKILL.md
@@ -53,6 +53,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -302,7 +309,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -324,7 +333,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -408,7 +421,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/gstack-upgrade/migrations/v1.58.0.0.sh b/gstack-upgrade/migrations/v1.58.0.0.sh
new file mode 100755
index 000000000..da2252286
--- /dev/null
+++ b/gstack-upgrade/migrations/v1.58.0.0.sh
@@ -0,0 +1,63 @@
+#!/usr/bin/env bash
+# Migration: v1.58.0.0 — register the PreToolUse AskUserQuestion hook for
+# existing Conductor installs.
+#
+# Why a migration: v1.58 makes the PreToolUse question-preference-hook also
+# deny the flaky Conductor AskUserQuestion and redirect to a prose decision
+# brief. But setup's hook-install block skips silently in non-interactive
+# (conductor/CI) setups, and existing users who previously declined plan-tune
+# hooks would never pick up the new Conductor backstop. This re-registers the
+# hook for Conductor users so layer 3 actually deploys.
+#
+# Affected: users who run gstack inside Conductor and don't already have the
+# PreToolUse hook installed.
+#
+# Scope guard: only acts inside a Conductor session (CONDUCTOR_* present) and
+# never overrides an explicit `plan_tune_hooks` opt-out.
+#
+# Idempotent: gstack-settings-hook dedupes by (event, matcher, source), and a
+# .done touchfile gates re-runs.
+
+set -u
+
+GSTACK_HOME="${HOME}/.gstack"
+MIGRATION_DIR="${GSTACK_HOME}/.migrations"
+DONE="${MIGRATION_DIR}/v1.58.0.0.done"
+mkdir -p "${MIGRATION_DIR}" 2>/dev/null || true
+[ -f "${DONE}" ] && exit 0
+
+# Only relevant inside Conductor — the prose-default behavior is Conductor-scoped.
+if [ -z "${CONDUCTOR_WORKSPACE_PATH:-}" ] && [ -z "${CONDUCTOR_PORT:-}" ]; then
+  touch "${DONE}"
+  exit 0
+fi
+
+SCRIPT_DIR="$(cd "$(dirname "$0")/../.." && pwd)"
+SETTINGS_HOOK="${SCRIPT_DIR}/bin/gstack-settings-hook"
+PREF_HOOK="${SCRIPT_DIR}/hosts/claude/hooks/question-preference-hook"
+CONFIG_BIN="${SCRIPT_DIR}/bin/gstack-config"
+
+# Respect an explicit opt-out — don't force a hook on a user who said no.
+_PT=$("${CONFIG_BIN}" get plan_tune_hooks 2>/dev/null || echo "")
+_PT=$(printf '%s' "${_PT}" | tr '[:upper:]' '[:lower:]' | tr -d '[:space:]')
+case "${_PT}" in
+  n|no|false|skip|off|0)
+    echo "  [v1.58.0.0] plan_tune_hooks opted out — leaving Conductor on guidance-only prose." >&2
+    touch "${DONE}"
+    exit 0
+    ;;
+esac
+
+if [ -x "${SETTINGS_HOOK}" ] && [ -x "${PREF_HOOK}" ]; then
+  "${SETTINGS_HOOK}" add-event \
+    --event PreToolUse \
+    --matcher '(AskUserQuestion|mcp__.*__AskUserQuestion)' \
+    --command "${PREF_HOOK}" \
+    --source plan-tune-cathedral \
+    --timeout 5 2>/dev/null \
+    && echo "  [v1.58.0.0] Conductor AskUserQuestion prose hook registered (PreToolUse)." >&2 \
+    || echo "  [v1.58.0.0] WARN: could not register the PreToolUse hook; run ./setup --plan-tune-hooks." >&2
+fi
+
+touch "${DONE}"
+exit 0
diff --git a/health/SKILL.md b/health/SKILL.md
index df0210492..9ba9bbbb3 100644
--- a/health/SKILL.md
+++ b/health/SKILL.md
@@ -51,6 +51,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -300,7 +307,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -322,7 +331,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -406,7 +419,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/hosts/claude/hooks/question-preference-hook.ts b/hosts/claude/hooks/question-preference-hook.ts
index dde1bda0c..12cbd5ea2 100644
--- a/hosts/claude/hooks/question-preference-hook.ts
+++ b/hosts/claude/hooks/question-preference-hook.ts
@@ -40,6 +40,7 @@ import * as fs from 'fs';
 import * as path from 'path';
 import * as os from 'os';
 import { spawnSync } from 'child_process';
+import { isConductor } from '../../../lib/is-conductor';
 
 interface HookStdin {
   session_id?: string;
@@ -400,57 +401,77 @@ async function main(): Promise<void> {
     ? '[plan-tune memory] Past answers suggest: ' + contextNuggets.join(' | ')
     : undefined;
 
+  // Determine whether EVERY question is eligible for never-ask auto-decide.
+  // We deliberately do NOT early-return defer on the first ineligible question:
+  // a Conductor session still needs the [conductor] prose deny as a fallback,
+  // so we compute eligibility, then branch. memoryContext is preserved on every
+  // non-enforcing exit. (All-or-nothing per-call semantics are unchanged: any
+  // ineligible question makes the whole call not auto-decidable.)
   const autoDecisions: Array<{ id: string; recommended: string }> = [];
+  let fullyAutoDecidable = true;
   for (const q of questions) {
     const qText = q.question || '';
     const marker = qText.match(MARKER_RE);
-    if (!marker) {
-      defer(memoryContext);
-      return;
-    }
+    if (!marker) { fullyAutoDecidable = false; break; }
     const questionId = marker[1];
     const pref = lookupPreference(slug, questionId);
-    if (!pref.preference || pref.preference === 'always-ask') {
-      defer(memoryContext);
-      return;
-    }
+    if (!pref.preference || pref.preference === 'always-ask') { fullyAutoDecidable = false; break; }
 
     const entry = registry[questionId];
     const doorType = entry?.door_type || 'two-way';
-    if (doorType === 'one-way') {
-      // Safety override — even never-ask doesn't bypass one-way doors.
-      defer(memoryContext);
-      return;
-    }
+    // Safety override — even never-ask doesn't bypass one-way doors.
+    if (doorType === 'one-way') { fullyAutoDecidable = false; break; }
 
     const opts = optionLabels(q.options || []);
     const { recommended, ambiguous } = extractRecommended(qText, opts);
-    if (!recommended || ambiguous) {
-      // Refuse-on-ambiguous per D2 — fail safe, ask normally.
-      defer(memoryContext);
-      return;
-    }
+    // Refuse-on-ambiguous per D2 — fail safe.
+    if (!recommended || ambiguous) { fullyAutoDecidable = false; break; }
     autoDecisions.push({ id: questionId, recommended });
   }
 
-  // All questions were eligible for enforcement.
-  markAutoDecided(stdin.session_id, stdin.tool_use_id);
+  if (fullyAutoDecidable && autoDecisions.length > 0) {
+    // All questions were eligible for enforcement.
+    markAutoDecided(stdin.session_id, stdin.tool_use_id);
 
-  // Log each auto-decided question now, since deny prevents PostToolUse from
-  // firing. /plan-tune Recent auto-decisions reads source=auto-decided events.
-  for (let i = 0; i < autoDecisions.length; i++) {
-    const d = autoDecisions[i];
-    const q = questions[i];
-    const qText = (q.question || '').replace(MARKER_RE, '').trim();
-    const opts = optionLabels(q.options || []);
-    logAutoDecided(d.id, qText, d.recommended, opts.length, stdin.session_id, stdin.tool_use_id, stdin.cwd);
+    // Log each auto-decided question now, since deny prevents PostToolUse from
+    // firing. /plan-tune Recent auto-decisions reads source=auto-decided events.
+    for (let i = 0; i < autoDecisions.length; i++) {
+      const d = autoDecisions[i];
+      const q = questions[i];
+      const qText = (q.question || '').replace(MARKER_RE, '').trim();
+      const opts = optionLabels(q.options || []);
+      logAutoDecided(d.id, qText, d.recommended, opts.length, stdin.session_id, stdin.tool_use_id, stdin.cwd);
+    }
+
+    const reasonLines = autoDecisions.map(
+      (d) =>
+        `[plan-tune auto-decide] ${d.id} → ${d.recommended} (your never-ask preference). Proceed with that option without re-prompting. Change with /plan-tune.`,
+    );
+    deny(reasonLines.join('\n'));
+    return;
   }
 
-  const reasonLines = autoDecisions.map(
-    (d) =>
-      `[plan-tune auto-decide] ${d.id} → ${d.recommended} (your never-ask preference). Proceed with that option without re-prompting. Change with /plan-tune.`,
-  );
-  deny(reasonLines.join('\n'));
+  // Not fully auto-decidable. In Conductor, AskUserQuestion is unreliable
+  // (native is disabled, the mcp__conductor__AskUserQuestion variant is flaky),
+  // so deny the tool and redirect to a prose decision brief. This is TRANSPORT
+  // AVOIDANCE, not preference enforcement: it fires regardless of marker,
+  // preference, or door type — including one-way doors, which must reach the
+  // human via prose rather than the unreliable tool.
+  if (isConductor()) {
+    const conductorReason =
+      '[conductor] AskUserQuestion is unreliable in Conductor (native disabled, MCP variant flaky). ' +
+      'Do NOT call AskUserQuestion (native or any mcp__*__AskUserQuestion). Render this decision as a ' +
+      'PROSE decision brief now: a D<N> label, an ELI10 of the issue, a Recommendation line, then one ' +
+      'paragraph per choice carrying its `(recommended)` marker and `Completeness: X/10`; tell the user ' +
+      'to reply with a letter, then STOP. For a one-way/destructive confirmation, require an explicit ' +
+      'typed confirmation and do NOT proceed on a vague reply. Capture the decision with gstack-question-log ' +
+      '(PostToolUse will not fire on a prose path).' +
+      (memoryContext ? `\n${memoryContext}` : '');
+    deny(conductorReason);
+    return;
+  }
+
+  defer(memoryContext);
 }
 
 main().catch((e) => {
diff --git a/investigate/SKILL.md b/investigate/SKILL.md
index ab3b3dee4..0e27f4d34 100644
--- a/investigate/SKILL.md
+++ b/investigate/SKILL.md
@@ -90,6 +90,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -339,7 +346,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -361,7 +370,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -445,7 +458,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/ios-clean/SKILL.md b/ios-clean/SKILL.md
index db14d0f9c..d969640a2 100644
--- a/ios-clean/SKILL.md
+++ b/ios-clean/SKILL.md
@@ -53,6 +53,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -302,7 +309,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -324,7 +333,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -408,7 +421,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/ios-design-review/SKILL.md b/ios-design-review/SKILL.md
index 8c32707e4..6a890f26a 100644
--- a/ios-design-review/SKILL.md
+++ b/ios-design-review/SKILL.md
@@ -55,6 +55,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -304,7 +311,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -326,7 +335,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -410,7 +423,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/ios-fix/SKILL.md b/ios-fix/SKILL.md
index 68b5cef22..464691bc8 100644
--- a/ios-fix/SKILL.md
+++ b/ios-fix/SKILL.md
@@ -56,6 +56,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -305,7 +312,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -327,7 +336,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -411,7 +424,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/ios-qa/SKILL.md b/ios-qa/SKILL.md
index 96453be06..a57421d3c 100644
--- a/ios-qa/SKILL.md
+++ b/ios-qa/SKILL.md
@@ -59,6 +59,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -308,7 +315,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -330,7 +339,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -414,7 +427,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/ios-sync/SKILL.md b/ios-sync/SKILL.md
index 9026a8670..a55b2a426 100644
--- a/ios-sync/SKILL.md
+++ b/ios-sync/SKILL.md
@@ -53,6 +53,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -302,7 +309,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -324,7 +333,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -408,7 +421,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/land-and-deploy/SKILL.md b/land-and-deploy/SKILL.md
index 2f1179573..f2c3e75c7 100644
--- a/land-and-deploy/SKILL.md
+++ b/land-and-deploy/SKILL.md
@@ -48,6 +48,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -297,7 +304,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -319,7 +328,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -403,7 +416,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/landing-report/SKILL.md b/landing-report/SKILL.md
index 713971163..3d33ab3ad 100644
--- a/landing-report/SKILL.md
+++ b/landing-report/SKILL.md
@@ -49,6 +49,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -298,7 +305,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -320,7 +329,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -404,7 +417,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/learn/SKILL.md b/learn/SKILL.md
index 5221f126e..4266c1e36 100644
--- a/learn/SKILL.md
+++ b/learn/SKILL.md
@@ -51,6 +51,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -300,7 +307,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -322,7 +331,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -406,7 +419,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/lib/conductor-env-shim.ts b/lib/conductor-env-shim.ts
index f8804a1b9..6f4ae08d9 100644
--- a/lib/conductor-env-shim.ts
+++ b/lib/conductor-env-shim.ts
@@ -7,12 +7,30 @@
  *
  * Import this for its side effect: `import "../lib/conductor-env-shim";`
  */
-export function promoteConductorEnv(): void {
-  for (const key of ["ANTHROPIC_API_KEY", "OPENAI_API_KEY"] as const) {
-    if (!process.env[key] && process.env[`GSTACK_${key}`]) {
-      process.env[key] = process.env[`GSTACK_${key}`];
+const PROMOTED_KEYS = ["ANTHROPIC_API_KEY", "OPENAI_API_KEY"] as const;
+
+/**
+ * Pure form: returns a copy of `base` with each GSTACK_-prefixed key promoted
+ * to its canonical name when the canonical is empty. Single source of truth
+ * for promotion semantics — used by the ambient mutator below and by the
+ * hermetic env builder (test/helpers/hermetic-env.ts), which must not mutate
+ * process.env.
+ */
+export function promotedEnv(base: NodeJS.ProcessEnv): NodeJS.ProcessEnv {
+  const out: NodeJS.ProcessEnv = { ...base };
+  for (const key of PROMOTED_KEYS) {
+    if (!out[key] && out[`GSTACK_${key}`]) {
+      out[key] = out[`GSTACK_${key}`];
     }
   }
+  return out;
+}
+
+export function promoteConductorEnv(): void {
+  const promoted = promotedEnv(process.env);
+  for (const key of PROMOTED_KEYS) {
+    if (promoted[key]) process.env[key] = promoted[key];
+  }
 }
 
 promoteConductorEnv();
diff --git a/lib/is-conductor.ts b/lib/is-conductor.ts
new file mode 100644
index 000000000..3f31c88af
--- /dev/null
+++ b/lib/is-conductor.ts
@@ -0,0 +1,19 @@
+/**
+ * Conductor host detection — single source of truth for TS consumers.
+ *
+ * Conductor (the Mac app that runs many coding agents in parallel) sets
+ * CONDUCTOR_WORKSPACE_PATH / CONDUCTOR_PORT in the session env. The same two
+ * vars are what `bin/gstack-session-kind` keys on (it collapses Conductor into
+ * `interactive`, so it can't be reused to distinguish Conductor specifically —
+ * hence this dedicated helper).
+ *
+ * IMPORTANT: detection is a CALL-TIME read of the passed-in env (default
+ * `process.env`), never a module-load-time snapshot. ESM hoists static imports
+ * above any in-file `process.env.X = ...`, so a load-time read can't be pinned
+ * by a test without Bun --preload. Reading at call time lets unit tests set
+ * `process.env.CONDUCTOR_WORKSPACE_PATH` inline before invoking. See the
+ * `esm-hoist-breaks-env-pin-bootstrap` learning.
+ */
+export function isConductor(env: NodeJS.ProcessEnv = process.env): boolean {
+  return !!(env.CONDUCTOR_WORKSPACE_PATH || env.CONDUCTOR_PORT);
+}
diff --git a/make-pdf/SKILL.md b/make-pdf/SKILL.md
index 252336d7e..832c24a19 100644
--- a/make-pdf/SKILL.md
+++ b/make-pdf/SKILL.md
@@ -50,6 +50,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
diff --git a/office-hours/SKILL.md b/office-hours/SKILL.md
index 3562f17f6..b02ec8fa6 100644
--- a/office-hours/SKILL.md
+++ b/office-hours/SKILL.md
@@ -86,6 +86,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -335,7 +342,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -357,7 +366,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -441,7 +454,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/open-gstack-browser/SKILL.md b/open-gstack-browser/SKILL.md
index 36e87b6f0..09ed56123 100644
--- a/open-gstack-browser/SKILL.md
+++ b/open-gstack-browser/SKILL.md
@@ -48,6 +48,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -297,7 +304,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -319,7 +328,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -403,7 +416,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/package.json b/package.json
index b2f92dd5a..454e85569 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "gstack",
-  "version": "1.58.0.0",
+  "version": "1.58.1.0",
   "description": "Garry's Stack — Claude Code skills + fast headless browser. One repo, one install, entire AI engineering workflow.",
   "license": "MIT",
   "type": "module",
@@ -34,6 +34,10 @@
     "skill:check": "bun run scripts/skill-check.ts",
     "dev:skill": "bun run scripts/dev-skill.ts",
     "start": "bun run browse/src/server.ts",
+    "eval:bg": "bin/gstack-detach --label evals --lock gstack-evals --timeout 5400 -- bun run test:evals",
+    "eval:bg:all": "bin/gstack-detach --label evals-all --lock gstack-evals --timeout 7200 -- bun run test:evals:all",
+    "eval:bg:gate": "bin/gstack-detach --label evals-gate --lock gstack-evals --timeout 3600 -- bun run test:gate",
+    "eval:bg:periodic": "bin/gstack-detach --label evals-periodic --lock gstack-evals --timeout 5400 -- bun run test:periodic",
     "eval:list": "bun run scripts/eval-list.ts",
     "eval:compare": "bun run scripts/eval-compare.ts",
     "eval:summary": "bun run scripts/eval-summary.ts",
diff --git a/pair-agent/SKILL.md b/pair-agent/SKILL.md
index 2daa26de4..b6e192070 100644
--- a/pair-agent/SKILL.md
+++ b/pair-agent/SKILL.md
@@ -50,6 +50,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -299,7 +306,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -321,7 +330,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -405,7 +418,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/plan-ceo-review/SKILL.md b/plan-ceo-review/SKILL.md
index a3c4107eb..126c9282c 100644
--- a/plan-ceo-review/SKILL.md
+++ b/plan-ceo-review/SKILL.md
@@ -80,6 +80,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -329,7 +336,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -351,7 +360,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -435,7 +448,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/plan-design-review/SKILL.md b/plan-design-review/SKILL.md
index 539175b4a..baaba9568 100644
--- a/plan-design-review/SKILL.md
+++ b/plan-design-review/SKILL.md
@@ -52,6 +52,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -301,7 +308,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -323,7 +332,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -407,7 +420,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/plan-devex-review/SKILL.md b/plan-devex-review/SKILL.md
index 7f75f1023..3492fd51d 100644
--- a/plan-devex-review/SKILL.md
+++ b/plan-devex-review/SKILL.md
@@ -58,6 +58,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -307,7 +314,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -329,7 +338,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -413,7 +426,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/plan-eng-review/SKILL.md b/plan-eng-review/SKILL.md
index 58c5cc9c4..5a20a2295 100644
--- a/plan-eng-review/SKILL.md
+++ b/plan-eng-review/SKILL.md
@@ -56,6 +56,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -305,7 +312,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -327,7 +336,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -411,7 +424,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/plan-tune/SKILL.md b/plan-tune/SKILL.md
index 20264af84..ecd2c8d71 100644
--- a/plan-tune/SKILL.md
+++ b/plan-tune/SKILL.md
@@ -61,6 +61,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -310,7 +317,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -332,7 +341,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -416,7 +429,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/qa-only/SKILL.md b/qa-only/SKILL.md
index aef82bd88..ec2cea40d 100644
--- a/qa-only/SKILL.md
+++ b/qa-only/SKILL.md
@@ -51,6 +51,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -300,7 +307,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -322,7 +331,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -406,7 +419,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/qa/SKILL.md b/qa/SKILL.md
index dfce8cb02..4108212ea 100644
--- a/qa/SKILL.md
+++ b/qa/SKILL.md
@@ -57,6 +57,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -306,7 +313,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -328,7 +337,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -412,7 +425,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/retro/SKILL.md b/retro/SKILL.md
index f0684165e..cb65e98cf 100644
--- a/retro/SKILL.md
+++ b/retro/SKILL.md
@@ -68,6 +68,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -317,7 +324,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -339,7 +348,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -423,7 +436,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/review/SKILL.md b/review/SKILL.md
index 16e90389f..9ce97c052 100644
--- a/review/SKILL.md
+++ b/review/SKILL.md
@@ -53,6 +53,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -302,7 +309,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -324,7 +333,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -408,7 +421,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/scrape/SKILL.md b/scrape/SKILL.md
index 23545e725..54e1187c6 100644
--- a/scrape/SKILL.md
+++ b/scrape/SKILL.md
@@ -49,6 +49,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -298,7 +305,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -320,7 +329,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -404,7 +417,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/scripts/resolvers/preamble/generate-ask-user-format.ts b/scripts/resolvers/preamble/generate-ask-user-format.ts
index 12090dc2c..533c002c0 100644
--- a/scripts/resolvers/preamble/generate-ask-user-format.ts
+++ b/scripts/resolvers/preamble/generate-ask-user-format.ts
@@ -7,7 +7,9 @@ export function generateAskUserFormat(_ctx: TemplateContext): string {
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. \`mcp__conductor__AskUserQuestion\` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any \`mcp__*__AskUserQuestion\` variant is in your tool list, prefer it. Hosts may disable native AUQ via \`--disallowedTools AskUserQuestion\` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if \`CONDUCTOR_SESSION: true\` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any \`mcp__*__AskUserQuestion\` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns \`[Tool result missing due to internal error]\`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a \`[plan-tune auto-decide] <id> → <option>\` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with \`bin/gstack-question-log\` (the PostToolUse capture hook never fires on a prose path, so \`/plan-tune\` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any \`mcp__*__AskUserQuestion\` variant is in your tool list, prefer it. Hosts may disable native AUQ via \`--disallowedTools AskUserQuestion\` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -29,7 +31,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit \`Completeness: X/10\` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a \`Recommendation: <choice> because <reason>\` line plus the \`(recommended)\` marker on that choice.
 
-Layout: a \`D<N>\` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its \`(recommended)\` marker, its \`Completeness: X/10\`, and 2-4 sentences of reasoning — never a bare bullet list; a closing \`Net:\` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a \`D<N>\` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its \`(recommended)\` marker, its \`Completeness: X/10\`, and 2-4 sentences of reasoning — never a bare bullet list; a closing \`Net:\` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (\`D<N>\`, or \`D<N>.k\` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which \`D<N>.k\` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -113,7 +119,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + \`(recommended)\` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless \`CONDUCTOR_SESSION: true\` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + \`(recommended)\` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \\u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/scripts/resolvers/preamble/generate-preamble-bash.ts b/scripts/resolvers/preamble/generate-preamble-bash.ts
index ecc30cef7..f82318f0c 100644
--- a/scripts/resolvers/preamble/generate-preamble-bash.ts
+++ b/scripts/resolvers/preamble/generate-preamble-bash.ts
@@ -36,6 +36,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(${ctx.paths.binDir}/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "\${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "\${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(${ctx.paths.binDir}/gstack-config get telemetry 2>/dev/null || true)
diff --git a/setup b/setup
index ec1db22b7..e4d31f881 100755
--- a/setup
+++ b/setup
@@ -1371,6 +1371,17 @@ if [ "$NO_TEAM_MODE" -ne 1 ] \
     *)                       PT_DECISION="prompt" ;;
   esac
 
+  # Conductor host reliability: the PreToolUse preference hook also carries the
+  # Conductor-prose enforcement (deny the flaky mcp__conductor__AskUserQuestion,
+  # redirect to a prose decision brief). A Conductor workspace setup otherwise
+  # falls through to "prompt" → the non-interactive skip below, leaving Conductor
+  # users without that backstop. Treat Conductor as an implicit opt-in — but
+  # only on the silent fall-through, never overriding an explicit --no-plan-tune-hooks.
+  if [ "$PT_DECISION" = "prompt" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+    PT_DECISION="yes"
+    _PT_CONDUCTOR_AUTO=1
+  fi
+
   _install_plan_tune_hooks() {
     "$SETTINGS_HOOK" add-event \
       --event PostToolUse \
@@ -1405,10 +1416,15 @@ if [ "$NO_TEAM_MODE" -ne 1 ] \
     log ""
     log "Plan-tune hooks already installed. Run \`$SETTINGS_HOOK list-sources\` to inspect."
   elif [ "$PT_DECISION" = "yes" ]; then
-    # Explicit opt-in (flag / env / config). Non-interactive.
+    # Explicit opt-in (flag / env / config) or Conductor implicit opt-in. Non-interactive.
     _install_plan_tune_hooks
     log ""
-    log "Plan-tune hooks installed. Run /plan-tune anytime to inspect."
+    if [ "${_PT_CONDUCTOR_AUTO:-0}" -eq 1 ]; then
+      log "AskUserQuestion reliability hooks installed (Conductor detected): decisions"
+      log "render as a prose brief instead of the flaky AskUserQuestion tool. Inspect with /plan-tune."
+    else
+      log "Plan-tune hooks installed. Run /plan-tune anytime to inspect."
+    fi
     touch "$PLAN_TUNE_INSTALL_MARKER"
   elif [ "$PT_DECISION" = "no" ]; then
     # Explicit opt-out (flag / env / config). Non-interactive.
diff --git a/setup-browser-cookies/SKILL.md b/setup-browser-cookies/SKILL.md
index 0d0a4f005..53f24d05c 100644
--- a/setup-browser-cookies/SKILL.md
+++ b/setup-browser-cookies/SKILL.md
@@ -45,6 +45,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
diff --git a/setup-deploy/SKILL.md b/setup-deploy/SKILL.md
index 2873903ae..9765c3798 100644
--- a/setup-deploy/SKILL.md
+++ b/setup-deploy/SKILL.md
@@ -52,6 +52,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -301,7 +308,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -323,7 +332,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -407,7 +420,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/setup-gbrain/SKILL.md b/setup-gbrain/SKILL.md
index 2b6c3a666..e6a041710 100644
--- a/setup-gbrain/SKILL.md
+++ b/setup-gbrain/SKILL.md
@@ -51,6 +51,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -300,7 +307,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -322,7 +331,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -406,7 +419,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/ship/SKILL.md b/ship/SKILL.md
index f78dcd5a0..a8ebb77d7 100644
--- a/ship/SKILL.md
+++ b/ship/SKILL.md
@@ -53,6 +53,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -302,7 +309,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -324,7 +333,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -408,7 +421,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/ship/sections/tests.md b/ship/sections/tests.md
index 2ba17a96b..e9d4dd819 100644
--- a/ship/sections/tests.md
+++ b/ship/sections/tests.md
@@ -332,6 +332,22 @@ EVAL_JUDGE_TIER=full EVAL_VERBOSE=1 bin/test-lane --eval test/evals/<suite>_eval
 
 If multiple suites need to run, run them sequentially (each needs a test lane). If the first suite fails, stop immediately — don't burn API cost on remaining suites.
 
+**Long eval suites (30+ min): launch detached so a turn boundary can't kill them.**
+A plain backgrounded eval lives in the harness's process group and dies to a
+SIGTERM ("polite quit") on a turn boundary, a stopped monitor, or an interruption
+(observed mid-`/ship`: `script terminated by signal SIGTERM`). Run it through
+`~/.claude/skills/gstack/bin/gstack-detach` instead — it survives in its own
+session, serializes against other worktrees via a machine lock (no API
+saturation), and writes a guaranteed `### gstack-detach EXIT=<code> ###` sentinel:
+
+```bash
+~/.claude/skills/gstack/bin/gstack-detach --label ship-evals --lock gstack-evals --timeout 5400 -- <project eval command>
+```
+
+Then poll the printed log path; break on the `EXIT=` sentinel (covers both pass
+and crash — silence is never success). The detached run survives even if your
+poller is reaped.
+
 **4. Check results:**
 
 - **If any eval fails:** Show the failures, the cost dashboard, and **STOP**. Do not proceed.
diff --git a/ship/sections/tests.md.tmpl b/ship/sections/tests.md.tmpl
index c9ba9ed6f..e6d53d495 100644
--- a/ship/sections/tests.md.tmpl
+++ b/ship/sections/tests.md.tmpl
@@ -76,6 +76,22 @@ EVAL_JUDGE_TIER=full EVAL_VERBOSE=1 bin/test-lane --eval test/evals/<suite>_eval
 
 If multiple suites need to run, run them sequentially (each needs a test lane). If the first suite fails, stop immediately — don't burn API cost on remaining suites.
 
+**Long eval suites (30+ min): launch detached so a turn boundary can't kill them.**
+A plain backgrounded eval lives in the harness's process group and dies to a
+SIGTERM ("polite quit") on a turn boundary, a stopped monitor, or an interruption
+(observed mid-`/ship`: `script terminated by signal SIGTERM`). Run it through
+`~/.claude/skills/gstack/bin/gstack-detach` instead — it survives in its own
+session, serializes against other worktrees via a machine lock (no API
+saturation), and writes a guaranteed `### gstack-detach EXIT=<code> ###` sentinel:
+
+```bash
+~/.claude/skills/gstack/bin/gstack-detach --label ship-evals --lock gstack-evals --timeout 5400 -- <project eval command>
+```
+
+Then poll the printed log path; break on the `EXIT=` sentinel (covers both pass
+and crash — silence is never success). The detached run survives even if your
+poller is reaped.
+
 **4. Check results:**
 
 - **If any eval fails:** Show the failures, the cost dashboard, and **STOP**. Do not proceed.
diff --git a/skillify/SKILL.md b/skillify/SKILL.md
index 6daf635bd..c3db38b4d 100644
--- a/skillify/SKILL.md
+++ b/skillify/SKILL.md
@@ -49,6 +49,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -298,7 +305,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -320,7 +329,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -404,7 +417,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/spec/SKILL.md b/spec/SKILL.md
index 2c3b02f9f..89198cb91 100644
--- a/spec/SKILL.md
+++ b/spec/SKILL.md
@@ -50,6 +50,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -299,7 +306,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -321,7 +330,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -405,7 +418,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
@@ -1078,6 +1091,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -1327,7 +1347,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -1349,7 +1371,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -1433,7 +1459,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/sync-gbrain/SKILL.md b/sync-gbrain/SKILL.md
index a2abc2141..eba49d403 100644
--- a/sync-gbrain/SKILL.md
+++ b/sync-gbrain/SKILL.md
@@ -51,6 +51,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -300,7 +307,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -322,7 +331,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -406,7 +419,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/test/agent-sdk-runner.test.ts b/test/agent-sdk-runner.test.ts
index 39c5db812..b760d6833 100644
--- a/test/agent-sdk-runner.test.ts
+++ b/test/agent-sdk-runner.test.ts
@@ -347,7 +347,13 @@ describe('runAgentSdkTest — options propagation', () => {
     expect(opts.permissionMode).toBe('bypassPermissions');
     expect(opts.allowDangerouslySkipPermissions).toBe(true);
     expect(opts.settingSources).toEqual([]);
-    expect(opts.env).toEqual({ ANTHROPIC_API_KEY: 'fake' });
+    // env is the COMPLETE hermetic env with the per-test override merged
+    // last — partial pass-through was the documented SDK auth-breaker
+    // (Options.env replaces the child's entire environment).
+    expect(opts.env?.ANTHROPIC_API_KEY).toBe('fake');
+    expect(opts.env?.PATH).toBeTruthy();
+    expect(opts.env?.CLAUDE_CONFIG_DIR).toMatch(/\/\.claude$/);
+    expect(opts.env?.GSTACK_HOME).toContain('gstack-home');
     expect(opts.pathToClaudeCodeExecutable).toBe('/fake/path/claude');
   });
 
diff --git a/test/fixtures/golden/claude-ship-SKILL.md b/test/fixtures/golden/claude-ship-SKILL.md
index f78dcd5a0..a8ebb77d7 100644
--- a/test/fixtures/golden/claude-ship-SKILL.md
+++ b/test/fixtures/golden/claude-ship-SKILL.md
@@ -53,6 +53,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$(~/.claude/skills/gstack/bin/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$(~/.claude/skills/gstack/bin/gstack-config get telemetry 2>/dev/null || true)
@@ -302,7 +309,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -324,7 +333,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -408,7 +421,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
diff --git a/test/fixtures/golden/codex-ship-SKILL.md b/test/fixtures/golden/codex-ship-SKILL.md
index 95de61439..347e1ff81 100644
--- a/test/fixtures/golden/codex-ship-SKILL.md
+++ b/test/fixtures/golden/codex-ship-SKILL.md
@@ -39,6 +39,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$($GSTACK_BIN/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$($GSTACK_BIN/gstack-config get telemetry 2>/dev/null || true)
@@ -288,7 +295,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -310,7 +319,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -394,7 +407,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
@@ -1277,6 +1290,22 @@ EVAL_JUDGE_TIER=full EVAL_VERBOSE=1 bin/test-lane --eval test/evals/<suite>_eval
 
 If multiple suites need to run, run them sequentially (each needs a test lane). If the first suite fails, stop immediately — don't burn API cost on remaining suites.
 
+**Long eval suites (30+ min): launch detached so a turn boundary can't kill them.**
+A plain backgrounded eval lives in the harness's process group and dies to a
+SIGTERM ("polite quit") on a turn boundary, a stopped monitor, or an interruption
+(observed mid-`/ship`: `script terminated by signal SIGTERM`). Run it through
+`$GSTACK_ROOT/bin/gstack-detach` instead — it survives in its own
+session, serializes against other worktrees via a machine lock (no API
+saturation), and writes a guaranteed `### gstack-detach EXIT=<code> ###` sentinel:
+
+```bash
+$GSTACK_ROOT/bin/gstack-detach --label ship-evals --lock gstack-evals --timeout 5400 -- <project eval command>
+```
+
+Then poll the printed log path; break on the `EXIT=` sentinel (covers both pass
+and crash — silence is never success). The detached run survives even if your
+poller is reaped.
+
 **4. Check results:**
 
 - **If any eval fails:** Show the failures, the cost dashboard, and **STOP**. Do not proceed.
diff --git a/test/fixtures/golden/factory-ship-SKILL.md b/test/fixtures/golden/factory-ship-SKILL.md
index dd8d0e01d..ee0ee83a2 100644
--- a/test/fixtures/golden/factory-ship-SKILL.md
+++ b/test/fixtures/golden/factory-ship-SKILL.md
@@ -41,6 +41,13 @@ echo "REPO_MODE: $REPO_MODE"
 _SESSION_KIND=$($GSTACK_BIN/gstack-session-kind 2>/dev/null || echo "interactive")
 case "$_SESSION_KIND" in spawned|headless|interactive) ;; *) _SESSION_KIND="interactive" ;; esac
 echo "SESSION_KIND: $_SESSION_KIND"
+# Conductor host: AskUserQuestion is unreliable here (native disabled, MCP
+# variant flaky), so skills render decisions as prose instead of calling the
+# tool. Gated on !headless so an eval/CI run INSIDE Conductor (GSTACK_HEADLESS)
+# still BLOCKs rather than rendering prose to nobody.
+if [ "$_SESSION_KIND" != "headless" ] && { [ -n "${CONDUCTOR_WORKSPACE_PATH:-}" ] || [ -n "${CONDUCTOR_PORT:-}" ]; }; then
+  echo "CONDUCTOR_SESSION: true"
+fi
 _LAKE_SEEN=$([ -f ~/.gstack/.completeness-intro-seen ] && echo "yes" || echo "no")
 echo "LAKE_INTRO: $_LAKE_SEEN"
 _TEL=$($GSTACK_BIN/gstack-config get telemetry 2>/dev/null || true)
@@ -290,7 +297,9 @@ AI orchestrator (e.g., OpenClaw). In spawned sessions:
 
 "AskUserQuestion" can resolve to two tools at runtime: the **host MCP variant** (e.g. `mcp__conductor__AskUserQuestion` — appears in your tool list when the host registers it) or the **native** Claude Code tool.
 
-**Rule:** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
+**Conductor rule (read before the MCP rule):** if `CONDUCTOR_SESSION: true` was echoed by the preamble, do NOT call AskUserQuestion at all — neither native nor any `mcp__*__AskUserQuestion` variant. Render EVERY decision brief as the **prose form** below and STOP. This is proactive, not a reaction to a failure: Conductor disables native AUQ and its MCP variant is flaky (it returns `[Tool result missing due to internal error]`), so prose is the reliable path. **Auto-decide preferences still apply first:** if a `[plan-tune auto-decide] <id> → <option>` result has already surfaced for a question, proceed with that option (no prose). Because in Conductor you go straight to prose without ever calling the tool, this auto-decide-first ordering is enforced HERE, not only by the PreToolUse hook. When you render a Conductor prose brief, also capture it with `bin/gstack-question-log` (the PostToolUse capture hook never fires on a prose path, so `/plan-tune` history/learning depends on this call).
+
+**Rule (non-Conductor):** if any `mcp__*__AskUserQuestion` variant is in your tool list, prefer it. Hosts may disable native AUQ via `--disallowedTools AskUserQuestion` (Conductor does, by default) and route through their MCP variant; calling native there silently fails. Same questions/options shape; same decision-brief format applies.
 
 If AskUserQuestion is unavailable (no variant in your tool list) OR a call to it fails, do NOT silently auto-decide or write the decision to the plan file as a substitute. Follow the **failure fallback** below.
 
@@ -312,7 +321,11 @@ Tell three outcomes apart:
 2. **Completeness scores per choice** — explicit `Completeness: X/10` on EACH choice (10 complete, 7 happy-path, 3 shortcut); use the kind-note when options differ in kind not coverage, but never silently drop the score.
 3. **The recommendation and why** — a `Recommendation: <choice> because <reason>` line plus the `(recommended)` marker on that choice.
 
-Layout: a `D<N>` title + a one-line note that AskUserQuestion failed and to reply with a letter; the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+Layout: a `D<N>` title + a one-line note to reply with a letter (in Conductor this is the normal path; elsewhere it means AskUserQuestion was unavailable or errored); the issue ELI10; the Recommendation line; then ONE paragraph per choice carrying its `(recommended)` marker, its `Completeness: X/10`, and 2-4 sentences of reasoning — never a bare bullet list; a closing `Net:` line. Split chains / 5+ options: one prose block per per-option call, in sequence. Then STOP and wait — the user's typed answer is the decision. In plan mode this satisfies end-of-turn like a tool call.
+
+**Continuation — mapping a typed reply back to a brief.** Each brief carries a stable label (`D<N>`, or `D<N>.k` in a split chain). The user references it (e.g. "3.2: B"). A bare letter maps to the single most-recent UNANSWERED brief; if more than one is open (a split chain), do NOT guess — ask which `D<N>.k` it answers. Never apply a bare letter ambiguously across a chain.
+
+**One-way / destructive confirmations in prose.** When the decision is a one-way door (irreversible or destructive — delete, force-push, drop, overwrite), prose is a WEAKER gate than the tool, so make it stronger: require an explicit typed confirmation (the exact option letter or word), state plainly what is irreversible, and NEVER proceed on a vague, partial, or ambiguous reply — re-ask instead. Treat silence or "ok"/"sure" without the explicit choice as not-yet-confirmed.
 
 ### Format
 
@@ -396,7 +409,7 @@ Before calling AskUserQuestion, verify:
 - [ ] (recommended) label on one option (even for neutral-posture)
 - [ ] Dual-scale effort labels on effort-bearing options (human / CC)
 - [ ] Net line closes the decision
-- [ ] You are calling the tool, not writing prose — unless the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
+- [ ] You are calling the tool, not writing prose — unless `CONDUCTOR_SESSION: true` (then prose is the DEFAULT, not the tool) OR the documented failure fallback applies (then: prose with the mandatory triad — issue ELI10, per-choice Completeness, Recommendation + `(recommended)` — and a "reply with a letter" instruction, then STOP)
 - [ ] Non-ASCII characters (CJK / accents) written directly, NOT \u-escaped
 - [ ] If you had 5+ options, you split (or batched into ≤4-groups) — did NOT drop any
 - [ ] If you split, you checked dependencies between options before firing the chain
@@ -1279,6 +1292,22 @@ EVAL_JUDGE_TIER=full EVAL_VERBOSE=1 bin/test-lane --eval test/evals/<suite>_eval
 
 If multiple suites need to run, run them sequentially (each needs a test lane). If the first suite fails, stop immediately — don't burn API cost on remaining suites.
 
+**Long eval suites (30+ min): launch detached so a turn boundary can't kill them.**
+A plain backgrounded eval lives in the harness's process group and dies to a
+SIGTERM ("polite quit") on a turn boundary, a stopped monitor, or an interruption
+(observed mid-`/ship`: `script terminated by signal SIGTERM`). Run it through
+`$GSTACK_ROOT/bin/gstack-detach` instead — it survives in its own
+session, serializes against other worktrees via a machine lock (no API
+saturation), and writes a guaranteed `### gstack-detach EXIT=<code> ###` sentinel:
+
+```bash
+$GSTACK_ROOT/bin/gstack-detach --label ship-evals --lock gstack-evals --timeout 5400 -- <project eval command>
+```
+
+Then poll the printed log path; break on the `EXIT=` sentinel (covers both pass
+and crash — silence is never success). The detached run survives even if your
+poller is reaped.
+
 **4. Check results:**
 
 - **If any eval fails:** Show the failures, the cost dashboard, and **STOP**. Do not proceed.
diff --git a/test/gstack-detach.test.ts b/test/gstack-detach.test.ts
new file mode 100644
index 000000000..926bf253a
--- /dev/null
+++ b/test/gstack-detach.test.ts
@@ -0,0 +1,96 @@
+/**
+ * gstack-detach — the eval-infra robustness guard. Pins the four killer fixes:
+ *   1. SIGTERM-proof detachment (runs in a different process group, outlives the launcher)
+ *   2. run-scoped default log path (no shared-/tmp collision between worktrees)
+ *   3. watchdog --timeout (no silent hang) + guaranteed EXIT sentinel
+ *   4. machine-wide --lock serialization (no cross-worktree API saturation)
+ */
+import { describe, test, expect } from 'bun:test';
+import { spawnSync, spawn } from 'child_process';
+import * as fs from 'fs';
+import * as os from 'os';
+import * as path from 'path';
+
+const ROOT = path.resolve(import.meta.dir, '..');
+const DETACH = path.join(ROOT, 'bin', 'gstack-detach');
+
+function ownPgid(): string {
+  return (spawnSync('ps', ['-o', 'pgid=', '-p', String(process.pid)], { encoding: 'utf-8' }).stdout || '').trim();
+}
+function waitFor(pred: () => boolean, ms: number): boolean {
+  const end = Date.now() + ms;
+  while (Date.now() < end) {
+    if (pred()) return true;
+    spawnSync('sleep', ['0.2']);
+  }
+  return pred();
+}
+function logHas(p: string, needle: string): boolean {
+  try { return fs.readFileSync(p, 'utf-8').includes(needle); } catch { return false; }
+}
+
+describe('gstack-detach', () => {
+  test('detaches (different pgid), returns immediately, completes, writes EXIT sentinel', () => {
+    const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'gd-'));
+    const log = path.join(dir, 'run.log');
+    try {
+      const t0 = Date.now();
+      const r = spawnSync(DETACH, ['--log', log, '--', 'bash', '-c', 'sleep 2; echo body-ran'], { encoding: 'utf-8', timeout: 10000 });
+      const elapsed = Date.now() - t0;
+      expect(r.status).toBe(0);
+      expect(r.stdout).toContain(`gstack-detach LOG ${log}`);
+      expect(elapsed).toBeLessThan(1500);                         // non-blocking
+      expect(waitFor(() => logHas(log, '### gstack-detach EXIT=0 ###'), 8000)).toBe(true);
+      expect(logHas(log, 'body-ran')).toBe(true);                 // ran to completion after launcher returned
+      const m = fs.readFileSync(log, 'utf-8').match(/pgid=(\d+)/);
+      expect(m).not.toBeNull();
+      expect(m![1]).not.toBe(ownPgid());                          // detached into its own group
+    } finally { fs.rmSync(dir, { recursive: true, force: true }); }
+  }, 15000);
+
+  test('default log is run-scoped under ~/.gstack-dev/eval-runs (no shared /tmp)', () => {
+    const r = spawnSync(DETACH, ['--label', 'unittest', '--', 'true'], { encoding: 'utf-8', timeout: 10000 });
+    const log = (r.stdout.match(/gstack-detach LOG (\S+)/) || [])[1];
+    try {
+      expect(log).toContain('/.gstack-dev/eval-runs/');
+      expect(path.basename(log)).toContain('unittest-');
+      expect(path.basename(log)).toMatch(/-\d+\.log$/);            // pid-unique
+      waitFor(() => logHas(log, '### gstack-detach EXIT=0 ###'), 6000);
+    } finally { if (log) fs.rmSync(log, { force: true }); }
+  }, 12000);
+
+  test('watchdog kills a stalled run and records EXIT=timeout (no silent hang)', () => {
+    const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'gd-'));
+    const log = path.join(dir, 'run.log');
+    try {
+      spawnSync(DETACH, ['--log', log, '--timeout', '1', '--', 'sleep', '60'], { encoding: 'utf-8', timeout: 10000 });
+      expect(waitFor(() => logHas(log, '### gstack-detach EXIT=timeout ###'), 12000)).toBe(true);
+      expect(logHas(log, 'WATCHDOG fired')).toBe(true);
+    } finally { fs.rmSync(dir, { recursive: true, force: true }); }
+  }, 16000);
+
+  test('machine --lock serializes concurrent runs (second WAITS for the first)', () => {
+    const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'gd-'));
+    const lock = `gstack-detach-test-${process.pid}`;
+    const logA = path.join(dir, 'a.log');
+    const logB = path.join(dir, 'b.log');
+    try {
+      // First holds the lock for ~3s; second must wait then acquire.
+      spawnSync(DETACH, ['--log', logA, '--lock', lock, '--', 'sleep', '3'], { encoding: 'utf-8', timeout: 10000 });
+      waitFor(() => logHas(logA, "ACQUIRED"), 4000);
+      spawnSync(DETACH, ['--log', logB, '--lock', lock, '--', 'echo', 'second-ran'], { encoding: 'utf-8', timeout: 10000 });
+      // Second should report WAITING (first still holds it) then ACQUIRE after release.
+      expect(waitFor(() => logHas(logB, 'WAITING for lock'), 4000)).toBe(true);
+      expect(waitFor(() => logHas(logB, '### gstack-detach EXIT=0 ###'), 12000)).toBe(true);
+      expect(logHas(logB, 'second-ran')).toBe(true);
+    } finally {
+      fs.rmSync(dir, { recursive: true, force: true });
+      fs.rmSync(path.join(os.homedir(), '.gstack', 'locks', `${lock}.lock`), { force: true });
+    }
+  }, 20000);
+
+  test('rejects missing command (exit 2)', () => {
+    const r = spawnSync(DETACH, ['--label', 'x'], { encoding: 'utf-8' });
+    expect(r.status).toBe(2);
+  });
+});
diff --git a/test/helpers/agent-sdk-runner.ts b/test/helpers/agent-sdk-runner.ts
index fc50f7884..9bd5eb8f8 100644
--- a/test/helpers/agent-sdk-runner.ts
+++ b/test/helpers/agent-sdk-runner.ts
@@ -36,6 +36,7 @@ import {
 import * as fs from 'fs';
 import * as path from 'path';
 import { resolveClaudeBinary as resolveClaudeBinaryShared } from '../../browse/src/claude-bin';
+import { hermeticChildEnv } from './hermetic-env';
 import type { SkillTestResult } from './session-runner';
 
 // ---------------------------------------------------------------------------
@@ -300,12 +301,17 @@ export async function runAgentSdkTest(
   const queryImpl: QueryProvider = opts.queryProvider ?? query;
   const model = opts.model ?? 'claude-opus-4-7';
 
-  // NOTE on GSTACK_HEADLESS: the SDK child inherits process.env, so headless
-  // classification for eval/E2E runs is set by the `test:gate` / `test:evals`
-  // package.json scripts (scoped to that invocation), NOT mutated here. We must not
-  // pass sdkOpts.env (it breaks the SDK auth pipeline — see CLAUDE.md) and must not
-  // mutate process.env ambiently (it would leak headless into later interactive-path
-  // tests in the same Bun process — Codex review finding).
+  // NOTE on env: the SDK child gets the COMPLETE hermetic env (allowlist
+  // scrub + ANTHROPIC_API_KEY + hermetic CLAUDE_CONFIG_DIR/GSTACK_HOME), with
+  // per-test opts.env merging last. The historical "passing env: breaks SDK
+  // auth" failure (old CLAUDE.md warning) was partial-env replacement —
+  // Options.env REPLACES the child's entire environment, so an object without
+  // the key killed auth. A complete env is safe (validated 2026-06-12 via
+  // query() with hermeticChildEnv(): success, real cost, Bash tool working).
+  // Do not mutate process.env ambiently here (it would leak into later
+  // interactive-path tests in the same Bun process — Codex review finding);
+  // ambient ANTHROPIC_API_KEY mutation by tests still works because the
+  // builder reads process.env at call time.
 
   let attempt = 0;
   let lastErr: unknown = null;
@@ -356,7 +362,7 @@ export async function runAgentSdkTest(
         permissionMode: resolvedPermissionMode,
         allowDangerouslySkipPermissions: resolvedPermissionMode === 'bypassPermissions',
         settingSources: opts.settingSources ?? [],
-        env: opts.env,
+        env: hermeticChildEnv(opts.env),
         pathToClaudeCodeExecutable: opts.pathToClaudeCodeExecutable,
         ...(hasCanUseTool ? { canUseTool: opts.canUseTool } : {}),
       };
diff --git a/test/helpers/carve-guards.ts b/test/helpers/carve-guards.ts
index 6a06ef086..11e559d8b 100644
--- a/test/helpers/carve-guards.ts
+++ b/test/helpers/carve-guards.ts
@@ -183,7 +183,9 @@ export const CARVE_GUARDS: Record<string, CarveGuard> = {
       gateAfterStop: 'EXIT PLAN MODE GATE',
     },
     behavioral: 'plan',
-    maxSkeletonBytes: 82_000,
+    // +Conductor AUQ-default-prose rule + one-way/continuation safety in the
+    // always-loaded AskUserQuestion Format section.
+    maxSkeletonBytes: 84_000,
     minUnionBytes: 70_000,
     mustContain: ['design', 'visual'],
   },
@@ -199,7 +201,9 @@ export const CARVE_GUARDS: Record<string, CarveGuard> = {
       gateAfterStop: 'EXIT PLAN MODE GATE',
     },
     behavioral: 'plan',
-    maxSkeletonBytes: 76_000,
+    // +Conductor AUQ-default-prose rule + one-way/destructive prose safety +
+    // continuation protocol in the always-loaded AskUserQuestion Format section.
+    maxSkeletonBytes: 78_000,
     minUnionBytes: 70_000,
     mustContain: ['developer experience', 'Getting Started'],
     // Default-on Codex outside-voice (codexPreflight block + CODEX_MODE branch
@@ -237,7 +241,9 @@ export const CARVE_GUARDS: Record<string, CarveGuard> = {
       gateAfterStop: undefined,
     },
     behavioral: 'prompt',
-    maxSkeletonBytes: 50_000,
+    // +Conductor AUQ-default-prose rule + one-way/continuation safety in the
+    // always-loaded AskUserQuestion Format section.
+    maxSkeletonBytes: 53_000,
     minUnionBytes: 55_000,
     mustContain: ['CHANGELOG', 'Diataxis', 'coverage'],
     // Two intentional additions stack on this small skill: the AUQ-failure prose
@@ -262,7 +268,9 @@ export const CARVE_GUARDS: Record<string, CarveGuard> = {
       gateAfterStop: undefined,
     },
     behavioral: 'prompt',
-    maxSkeletonBytes: 64_000,
+    // +Conductor AUQ-default-prose rule + one-way/continuation safety in the
+    // always-loaded AskUserQuestion Format section.
+    maxSkeletonBytes: 67_000,
     minUnionBytes: 72_000,
     mustContain: ['Typography', 'Color', 'Aesthetic Direction'],
     // Cross-cutting preamble growth (v1.57.2.0 AUQ-failure prose fallback ~2KB +
@@ -298,7 +306,9 @@ export const CARVE_GUARDS: Record<string, CarveGuard> = {
       gateAfterStop: undefined,
     },
     behavioral: 'prompt',
-    maxSkeletonBytes: 70_000,
+    // +Conductor AUQ-default-prose rule + one-way/continuation safety in the
+    // always-loaded AskUserQuestion Format section.
+    maxSkeletonBytes: 73_000,
     minUnionBytes: 72_000,
     mustContain: ['OWASP', 'STRIDE', 'daily', 'comprehensive', 'verif'],
     // cso keeps its mode-dispatch + FP-filtering phases always-loaded, so the
diff --git a/test/helpers/claude-pty-runner.ts b/test/helpers/claude-pty-runner.ts
index f06ba05f3..a371af866 100644
--- a/test/helpers/claude-pty-runner.ts
+++ b/test/helpers/claude-pty-runner.ts
@@ -24,6 +24,7 @@
 import * as fs from 'fs';
 import * as os from 'os';
 import * as path from 'path';
+import { hermeticChildEnv, isHermeticEnabled } from './hermetic-env';
 
 /** Strip ANSI escapes for pattern-matching against visible text. */
 export function stripAnsi(s: string): string {
@@ -120,6 +121,13 @@ export interface ClaudePtySession {
   exited(): boolean;
   /** Exit code, if known. */
   exitCode(): number | null;
+  /**
+   * The hermetic CLAUDE_CONFIG_DIR this session's claude was pointed at, or
+   * null when EVALS_HERMETIC=0. Forensics: hermetic plan files live under
+   * `<hermeticConfigDir>/plans/` (extractPlanFilePath still matches them —
+   * the dir name ends in `/.claude` by contract).
+   */
+  hermeticConfigDir: string | null;
   /**
    * Send SIGINT, then SIGKILL after 1s. Always safe to call multiple times.
    * Awaits process exit before resolving.
@@ -1143,8 +1151,17 @@ export async function launchClaudePty(
   if (permissionMode !== null) {
     args.push('--permission-mode', permissionMode);
   }
+  // Hermetic children get zero MCP servers; gated on the same call-time
+  // check as the env scrub so EVALS_HERMETIC=0 restores operator MCP too.
+  // Before opts.extraArgs so a test could theoretically supply --mcp-config.
+  const hermetic = isHermeticEnabled();
+  if (hermetic) args.push('--strict-mcp-config');
   if (opts.extraArgs) args.push(...opts.extraArgs);
 
+  // Hermetic by default (test/helpers/hermetic-env.ts): operator session
+  // context never reaches the child; per-test opts.env merges last.
+  const childEnv = hermeticChildEnv(opts.env);
+
   // eslint-disable-next-line @typescript-eslint/no-explicit-any
   const proc = (Bun as any).spawn([claudePath, ...args], {
     terminal: {
@@ -1155,7 +1172,7 @@ export async function launchClaudePty(
       },
     },
     cwd,
-    env: { ...process.env, ...(opts.env ?? {}) },
+    env: childEnv,
   });
 
   // Track exit so waitForAny can fail fast if claude crashes.
@@ -1307,6 +1324,7 @@ export async function launchClaudePty(
     pid: () => proc.pid as number | undefined,
     exited: () => exited,
     exitCode: () => exitCodeCaptured,
+    hermeticConfigDir: hermetic ? childEnv.CLAUDE_CONFIG_DIR ?? null : null,
     close,
   };
 }
diff --git a/test/helpers/codex-session-runner.ts b/test/helpers/codex-session-runner.ts
index 0be9dd7d6..404aa6bb2 100644
--- a/test/helpers/codex-session-runner.ts
+++ b/test/helpers/codex-session-runner.ts
@@ -15,6 +15,7 @@
 import * as fs from 'fs';
 import * as path from 'path';
 import * as os from 'os';
+import { hermeticChildEnv } from './hermetic-env';
 
 // --- Interfaces ---
 
@@ -201,15 +202,18 @@ export async function runCodexSkill(opts: {
     // Build codex exec command
     const args = ['exec', prompt, '--json', '-s', sandbox];
 
-    // Spawn codex with temp HOME so it discovers our installed skill
+    // Spawn codex with temp HOME so it discovers our installed skill.
+    // Hermetic scrub (test/helpers/hermetic-env.ts) with codex's auth surface
+    // re-admitted: codex auths from $HOME/.codex (copied into tempHome above)
+    // plus OPENAI_API_KEY/CODEX_* when present. HOME override merges last.
     const proc = Bun.spawn(['codex', ...args], {
       cwd: cwd || skillDir,
       stdout: 'pipe',
       stderr: 'pipe',
-      env: {
-        ...process.env,
-        HOME: tempHome,
-      },
+      env: hermeticChildEnv(
+        { HOME: tempHome },
+        { extraAllow: ['OPENAI_API_KEY', 'CODEX_*'] },
+      ),
     });
 
     // Race against timeout
diff --git a/test/helpers/gemini-session-runner.ts b/test/helpers/gemini-session-runner.ts
index 06393c383..4f2f040e3 100644
--- a/test/helpers/gemini-session-runner.ts
+++ b/test/helpers/gemini-session-runner.ts
@@ -14,6 +14,7 @@
  */
 
 import * as path from 'path';
+import { hermeticChildEnv } from './hermetic-env';
 
 // --- Interfaces ---
 
@@ -122,11 +123,16 @@ export async function runGeminiSkill(opts: {
   // Build gemini command
   const args = ['-p', prompt, '--output-format', 'stream-json', '--yolo'];
 
-  // Spawn gemini — uses real HOME for auth, cwd for skill discovery
+  // Spawn gemini — uses real HOME for auth (~/.gemini; HOME is allowlisted),
+  // cwd for skill discovery. Hermetic scrub with gemini's auth surface
+  // re-admitted (previously this spawn inherited the full operator env).
   const proc = Bun.spawn(['gemini', ...args], {
     cwd: cwd || process.cwd(),
     stdout: 'pipe',
     stderr: 'pipe',
+    env: hermeticChildEnv(undefined, {
+      extraAllow: ['GEMINI_API_KEY', 'GOOGLE_API_KEY', 'GOOGLE_APPLICATION_CREDENTIALS', 'GOOGLE_CLOUD_*', 'GEMINI_*'],
+    }),
   });
 
   // Race against timeout
diff --git a/test/helpers/hermetic-env.test.ts b/test/helpers/hermetic-env.test.ts
new file mode 100644
index 000000000..b4503b60c
--- /dev/null
+++ b/test/helpers/hermetic-env.test.ts
@@ -0,0 +1,269 @@
+/**
+ * Unit tests for the hermetic child-env builder. Free tier — no API calls.
+ *
+ * Pins three contracts:
+ * 1. Allowlist semantics: contamination vars dropped, basics/auth/network
+ *    kept, overrides merge last, EVALS_HERMETIC=0 is byte-identical legacy.
+ * 2. Seed-config shape: 20-char key suffix, trusted dirs, undefined-key safe.
+ * 3. Dir lifecycle: /.claude suffix (extractPlanFilePath contract —
+ *    claude-pty-runner.ts:191), sync singleton reuse, pid-aware GC.
+ */
+
+import { describe, test, expect, afterAll } from 'bun:test';
+import * as fs from 'fs';
+import * as path from 'path';
+import * as os from 'os';
+import {
+  buildHermeticEnv,
+  buildSeedConfig,
+  isHermeticEnabled,
+  getHermeticDirs,
+  gcStaleHermeticDirs,
+  hermeticChildEnv,
+} from './hermetic-env';
+
+const CONTAMINATED: NodeJS.ProcessEnv = {
+  PATH: '/usr/bin', HOME: '/Users/op', TMPDIR: '/tmp', TERM: 'xterm',
+  ANTHROPIC_API_KEY: 'sk-ant-0123456789abcdefghijklmn',
+  ANTHROPIC_BASE_URL: 'https://proxy.example/api',
+  ANTHROPIC_MODEL: 'sneaky-model-override',
+  EVALS_MODEL: 'claude-sonnet-4-6',
+  GITHUB_ACTIONS: 'true',
+  HTTPS_PROXY: 'http://corp:3128',
+  NODE_EXTRA_CA_CERTS: '/etc/corp.pem',
+  CONDUCTOR_WORKSPACE_PATH: '/Users/op/conductor/ws',
+  CONDUCTOR_SESSION: '1',
+  CLAUDECODE: '1',
+  CLAUDE_CODE_ENTRYPOINT: 'cli',
+  CLAUDE_CONFIG_DIR: '/Users/op/.claude',
+  GSTACK_HOME: '/Users/op/.gstack',
+  GSTACK_HEADLESS_DEFAULT: 'x',
+  MCP_TIMEOUT: '5000',
+  GBRAIN_ENDPOINT: 'http://localhost:1234',
+  OPENAI_API_KEY: 'sk-openai-secret',
+  VOYAGE_API_KEY: 'vg-secret',
+  GH_TOKEN: 'gho_secret',
+  SSH_AUTH_SOCK: '/tmp/ssh.sock',
+  GIT_AUTHOR_NAME: 'Op',
+};
+
+const HERMETIC_VARS = { CLAUDE_CONFIG_DIR: '/x/.claude', GSTACK_HOME: '/x/gstack-home' };
+
+describe('buildHermeticEnv allowlist', () => {
+  const env = buildHermeticEnv(CONTAMINATED, HERMETIC_VARS);
+
+  test('keeps process basics, network, CI, and eval knobs', () => {
+    expect(env.PATH).toBe('/usr/bin');
+    expect(env.HOME).toBe('/Users/op');
+    expect(env.EVALS_MODEL).toBe('claude-sonnet-4-6');
+    expect(env.GITHUB_ACTIONS).toBe('true');
+    expect(env.HTTPS_PROXY).toBe('http://corp:3128');
+    expect(env.NODE_EXTRA_CA_CERTS).toBe('/etc/corp.pem');
+  });
+
+  test('keeps named auth vars but not the broad ANTHROPIC_ prefix', () => {
+    expect(env.ANTHROPIC_API_KEY).toBe(CONTAMINATED.ANTHROPIC_API_KEY);
+    expect(env.ANTHROPIC_BASE_URL).toBe(CONTAMINATED.ANTHROPIC_BASE_URL);
+    expect(env.ANTHROPIC_MODEL).toBeUndefined(); // behavior knob, not auth
+  });
+
+  test('drops session-context and operator-credential vars', () => {
+    for (const k of [
+      'CONDUCTOR_WORKSPACE_PATH', 'CONDUCTOR_SESSION', 'CLAUDECODE',
+      'CLAUDE_CODE_ENTRYPOINT', 'GSTACK_HEADLESS_DEFAULT', 'MCP_TIMEOUT',
+      'GBRAIN_ENDPOINT', 'OPENAI_API_KEY', 'VOYAGE_API_KEY', 'GH_TOKEN',
+      'SSH_AUTH_SOCK', 'GIT_AUTHOR_NAME',
+    ]) {
+      expect(env[k]).toBeUndefined();
+    }
+  });
+
+  test('redirects CLAUDE_CONFIG_DIR and GSTACK_HOME to hermetic values', () => {
+    expect(env.CLAUDE_CONFIG_DIR).toBe('/x/.claude');
+    expect(env.GSTACK_HOME).toBe('/x/gstack-home');
+  });
+
+  test('overrides merge last — per-test re-contamination is deliberate', () => {
+    const e = buildHermeticEnv(CONTAMINATED, HERMETIC_VARS, {
+      CONDUCTOR_WORKSPACE_PATH: '/tmp/test-ws',
+      GSTACK_HOME: '/tmp/test-home',
+      GSTACK_HEADLESS: '',
+    });
+    expect(e.CONDUCTOR_WORKSPACE_PATH).toBe('/tmp/test-ws');
+    expect(e.GSTACK_HOME).toBe('/tmp/test-home');
+    expect(e.GSTACK_HEADLESS).toBe('');
+  });
+
+  test('promotes GSTACK_ANTHROPIC_API_KEY when canonical absent (shared shim fn)', () => {
+    const base = { ...CONTAMINATED } as NodeJS.ProcessEnv;
+    delete base.ANTHROPIC_API_KEY;
+    base.GSTACK_ANTHROPIC_API_KEY = 'sk-ant-promoted-9876543210';
+    const e = buildHermeticEnv(base, HERMETIC_VARS);
+    expect(e.ANTHROPIC_API_KEY).toBe('sk-ant-promoted-9876543210');
+    expect(e.GSTACK_ANTHROPIC_API_KEY).toBeUndefined(); // GSTACK_* still dropped
+  });
+
+  test('extraAllow re-admits exact names and prefixes per runner', () => {
+    const e = buildHermeticEnv(CONTAMINATED, HERMETIC_VARS, undefined, {
+      extraAllow: ['OPENAI_API_KEY', 'GIT_*'],
+    });
+    expect(e.OPENAI_API_KEY).toBe('sk-openai-secret');
+    expect(e.GIT_AUTHOR_NAME).toBe('Op');
+    expect(e.GH_TOKEN).toBeUndefined(); // not in extraAllow
+  });
+
+  test('TERM falls back when base omits it', () => {
+    const base = { ...CONTAMINATED } as NodeJS.ProcessEnv;
+    delete base.TERM;
+    expect(buildHermeticEnv(base, HERMETIC_VARS).TERM).toBe('xterm-256color');
+  });
+});
+
+describe('EVALS_HERMETIC=0 escape hatch', () => {
+  test('returns byte-identical legacy env, overrides still last', () => {
+    const base = { ...CONTAMINATED, EVALS_HERMETIC: '0' } as NodeJS.ProcessEnv;
+    const e = buildHermeticEnv(base, HERMETIC_VARS, { GSTACK_HEADLESS: '1' });
+    // Legacy spread: every base var survives, hermeticVars NOT applied.
+    expect(e.CONDUCTOR_WORKSPACE_PATH).toBe(CONTAMINATED.CONDUCTOR_WORKSPACE_PATH);
+    expect(e.CLAUDE_CONFIG_DIR).toBe('/Users/op/.claude');
+    expect(e.GSTACK_HOME).toBe('/Users/op/.gstack');
+    expect(e.GSTACK_HEADLESS).toBe('1');
+    expect(e).toEqual({ ...(base as Record<string, string>), GSTACK_HEADLESS: '1' });
+  });
+
+  test('isHermeticEnabled reads at call time (ESM-hoist safety)', () => {
+    const prev = process.env.EVALS_HERMETIC;
+    try {
+      process.env.EVALS_HERMETIC = '0';
+      expect(isHermeticEnabled()).toBe(false);
+      process.env.EVALS_HERMETIC = '1';
+      expect(isHermeticEnabled()).toBe(true);
+      delete process.env.EVALS_HERMETIC;
+      expect(isHermeticEnabled()).toBe(true);
+    } finally {
+      if (prev === undefined) delete process.env.EVALS_HERMETIC;
+      else process.env.EVALS_HERMETIC = prev;
+    }
+  });
+});
+
+describe('buildSeedConfig', () => {
+  test('stores only the 20-char key suffix and trusts the given dirs', () => {
+    const seed = buildSeedConfig({
+      apiKey: 'sk-ant-0123456789abcdefghijklmn',
+      trustedDirs: ['/repo/root'],
+    }) as any;
+    expect(seed.hasCompletedOnboarding).toBe(true);
+    const approved = seed.customApiKeyResponses.approved;
+    expect(approved).toHaveLength(1);
+    expect(approved[0]).toHaveLength(20);
+    expect('sk-ant-0123456789abcdefghijklmn'.endsWith(approved[0])).toBe(true);
+    expect(seed.projects['/repo/root'].hasTrustDialogAccepted).toBe(true);
+    expect(seed.projects['/repo/root'].hasCompletedProjectOnboarding).toBe(true);
+  });
+
+  test('apiKey undefined → omits customApiKeyResponses, does not throw', () => {
+    const seed = buildSeedConfig({ apiKey: undefined, trustedDirs: [] }) as any;
+    expect(seed.customApiKeyResponses).toBeUndefined();
+    expect(seed.hasCompletedOnboarding).toBe(true);
+  });
+
+  test('no full key material anywhere in the seed', () => {
+    const key = 'sk-ant-0123456789abcdefghijklmn';
+    const json = JSON.stringify(buildSeedConfig({ apiKey: key, trustedDirs: [] }));
+    expect(json.includes(key)).toBe(false);
+  });
+});
+
+describe('getHermeticDirs lifecycle', () => {
+  test('configDir ends in /.claude — extractPlanFilePath contract', () => {
+    // claude-pty-runner.ts:191 anchors plan paths on `.claude/plans/` under
+    // /var|/tmp prefixes; the dir-name suffix is what keeps PTY plan-mode
+    // tests extracting hermetic plan files with zero extractor changes.
+    const dirs = getHermeticDirs();
+    expect(dirs.configDir.endsWith(`${path.sep}.claude`)).toBe(true);
+    expect(dirs.configDir.startsWith(os.tmpdir())).toBe(true);
+  });
+
+  test('sync singleton: repeat calls return the same dirs', () => {
+    expect(getHermeticDirs()).toBe(getHermeticDirs());
+  });
+
+  test('seeds .claude.json in the config dir', () => {
+    const dirs = getHermeticDirs();
+    const seed = JSON.parse(fs.readFileSync(path.join(dirs.configDir, '.claude.json'), 'utf-8'));
+    expect(seed.hasCompletedOnboarding).toBe(true);
+    const root = path.resolve(__dirname, '..', '..');
+    expect(seed.projects[root].hasTrustDialogAccepted).toBe(true);
+  });
+});
+
+describe('gcStaleHermeticDirs', () => {
+  test('removes dead-pid dirs, keeps live-pid and foreign dirs', () => {
+    const tmp = fs.mkdtempSync(path.join(os.tmpdir(), 'hermetic-gc-test-'));
+    // Find a pid that is definitely dead: spawn-and-reap is overkill; use a
+    // huge pid beyond pid_max on macOS/Linux defaults.
+    const deadPid = 99999999;
+    const dead = path.join(tmp, `gstack-hermetic-${deadPid}-abc`);
+    const live = path.join(tmp, `gstack-hermetic-${process.pid}-abc`);
+    const foreign = path.join(tmp, 'unrelated-dir');
+    const malformed = path.join(tmp, 'gstack-hermetic-notapid-abc');
+    for (const d of [dead, live, foreign, malformed]) fs.mkdirSync(d);
+    // GC only reclaims dirs older than its 1h age floor (PID-reuse guard);
+    // backdate the dead-pid dir's mtime so it qualifies.
+    const old = new Date(Date.now() - 2 * 60 * 60 * 1000);
+    fs.utimesSync(dead, old, old);
+
+    gcStaleHermeticDirs(tmp);
+
+    expect(fs.existsSync(dead)).toBe(false);
+    expect(fs.existsSync(live)).toBe(true);
+    expect(fs.existsSync(foreign)).toBe(true);
+    expect(fs.existsSync(malformed)).toBe(true); // never guess on malformed names
+    fs.rmSync(tmp, { recursive: true, force: true });
+  });
+
+  test('keeps a fresh dead-pid dir (PID-reuse grace window)', () => {
+    const tmp = fs.mkdtempSync(path.join(os.tmpdir(), 'hermetic-gc-fresh-'));
+    // Dead pid but just created — must survive GC, else PID reuse could delete
+    // a dir whose original pid exited and got recycled to a live process.
+    const freshDead = path.join(tmp, 'gstack-hermetic-99999999-xyz');
+    fs.mkdirSync(freshDead);
+    gcStaleHermeticDirs(tmp);
+    expect(fs.existsSync(freshDead)).toBe(true);
+    fs.rmSync(tmp, { recursive: true, force: true });
+  });
+});
+
+describe('hermeticChildEnv composition', () => {
+  test('hermetic by default: redirects config dirs, drops contamination', () => {
+    // process.env in a real test run may carry CONDUCTOR_*/CLAUDECODE — the
+    // composition must scrub them and point at the singleton dirs.
+    const e = hermeticChildEnv({ GSTACK_HEADLESS: '1' });
+    const dirs = getHermeticDirs();
+    expect(e.CLAUDE_CONFIG_DIR).toBe(dirs.configDir);
+    expect(e.GSTACK_HOME).toBe(dirs.gstackHome);
+    expect(e.GSTACK_HEADLESS).toBe('1');
+    expect(e.CLAUDECODE).toBeUndefined();
+    expect(e.CONDUCTOR_WORKSPACE_PATH).toBeUndefined();
+  });
+
+  test('EVALS_HERMETIC=0: legacy passthrough of live process.env', () => {
+    const prev = process.env.EVALS_HERMETIC;
+    try {
+      process.env.EVALS_HERMETIC = '0';
+      const e = hermeticChildEnv({ EXTRA: 'x' });
+      expect(e.PATH).toBe(process.env.PATH as string);
+      expect(e.EXTRA).toBe('x');
+      // No hermetic redirection in legacy mode.
+      expect(e.CLAUDE_CONFIG_DIR).toBe(process.env.CLAUDE_CONFIG_DIR as any);
+    } finally {
+      if (prev === undefined) delete process.env.EVALS_HERMETIC;
+      else process.env.EVALS_HERMETIC = prev;
+    }
+  });
+});
+
+afterAll(() => {
+  // The singleton's own exit hook handles runRoot; nothing else to clean.
+});
diff --git a/test/helpers/hermetic-env.ts b/test/helpers/hermetic-env.ts
new file mode 100644
index 000000000..def6dd9bc
--- /dev/null
+++ b/test/helpers/hermetic-env.ts
@@ -0,0 +1,276 @@
+/**
+ * Hermetic child environment for E2E test runners.
+ *
+ * Local E2E runs spawn `claude` (and codex/gemini/SDK) children that, until
+ * this module, inherited the operator's full session context: ~/.claude
+ * (user CLAUDE.md, .claude.json MCP servers incl. gbrain + Conductor,
+ * skills), ~/.gstack decision logs, and CONDUCTOR_-/CLAUDECODE-style env vars.
+ * CI was hermetic only by accident (fresh Docker /home/runner). This module
+ * makes local children see a CI-equivalent clean room by default.
+ *
+ *   operator shell (contaminated)            hermetic child env
+ *   ┌─────────────────────────────┐  buildHermeticEnv()
+ *   │ PATH, HOME, TMPDIR, ...     │── allowlist ─────────► kept
+ *   │ HTTP(S)_PROXY, SSL_CERT_*   │── allowlist ─────────► kept (network)
+ *   │ ANTHROPIC_API_KEY/BASE_URL/ │── named list ────────► kept (auth)
+ *   │   AUTH_TOKEN                │
+ *   │ GSTACK_ANTHROPIC_API_KEY    │── promotedEnv() ─────► ANTHROPIC_API_KEY
+ *   │ CONDUCTOR_*, CLAUDECODE,    │
+ *   │ CLAUDE_*, GSTACK_*, MCP_*,  │── dropped ───────────► ∅
+ *   │ GBRAIN_*, GH_TOKEN, ...     │
+ *   └─────────────────────────────┘
+ *      + per-runner extraAllow (codex: OpenAI vars; gemini: Google vars)
+ *      + CLAUDE_CONFIG_DIR=<runRoot>/.claude  GSTACK_HOME=<runRoot>/gstack-home
+ *      + per-test overrides spread LAST
+ *
+ * Escape hatch: EVALS_HERMETIC=0 restores the legacy contaminated env
+ * byte-identically (runners must also gate --strict-mcp-config on
+ * isHermeticEnabled() so the escape hatch restores args too).
+ *
+ * isHermeticEnabled() is evaluated at CALL time, never at module load —
+ * ESM hoists imports above any in-file `process.env.EVALS_HERMETIC = '0'`
+ * assignment, so a module-load-time read would silently ignore test pins.
+ */
+
+import * as fs from 'fs';
+import * as path from 'path';
+import * as os from 'os';
+import { promotedEnv } from '../../lib/conductor-env-shim';
+import { isProcessAlive } from '../../browse/src/error-handling';
+
+/** Exact env names a hermetic child keeps. Everything not listed (or matched
+ * by a prefix rule below) is dropped. */
+const ALLOW_EXACT = new Set([
+  // Process basics
+  'PATH', 'HOME', 'TMPDIR', 'TERM', 'COLORTERM', 'LANG', 'LC_ALL', 'SHELL',
+  'USER', 'LOGNAME', 'TZ', 'NODE_ENV', 'CI',
+  // Browser/runtime caches the child legitimately shares with the operator
+  'PLAYWRIGHT_BROWSERS_PATH',
+  // Network reachability — without these, children on proxied networks can't
+  // reach the Anthropic API at all
+  'HTTP_PROXY', 'HTTPS_PROXY', 'NO_PROXY',
+  'http_proxy', 'https_proxy', 'no_proxy',
+  'SSL_CERT_FILE', 'SSL_CERT_DIR', 'NODE_EXTRA_CA_CERTS',
+  // Auth — named, NOT the broad ANTHROPIC_* prefix: a prefix rule would
+  // smuggle model/beta/debug knobs that change eval behavior
+  'ANTHROPIC_API_KEY',   // the auth credential evals require
+  'ANTHROPIC_BASE_URL',  // API endpoint override (corp proxies)
+  'ANTHROPIC_AUTH_TOKEN', // bearer-token auth variant
+]);
+
+/** Prefix rules: eval-harness knobs + CI metadata. Deliberately NOT here:
+ * CONDUCTOR_* / CLAUDE_* (incl. CLAUDECODE, CLAUDE_CODE_ENTRYPOINT) /
+ * GSTACK_* / MCP_* / GBRAIN_* — session-context contamination; and operator
+ * credentials (GH_TOKEN, SSH_AUTH_SOCK, GIT_*, OPENAI_API_KEY,
+ * VOYAGE_API_KEY) — CI doesn't have them and eval children have no business
+ * using them. A test that legitimately needs one opts in via its own env
+ * override; a provider runner (codex/gemini) re-admits its auth vars via
+ * opts.extraAllow. */
+const ALLOW_PREFIXES = ['EVALS_', 'GITHUB_'];
+
+export interface HermeticEnvOpts {
+  /** Per-runner additional allowed names (exact match) or prefixes (entries
+   * ending in '*'). Example: codex runner passes ['OPENAI_API_KEY', 'CODEX_*']. */
+  extraAllow?: string[];
+}
+
+/** EVALS_HERMETIC !== '0'. Read at call time (see module doc — ESM hoist). */
+export function isHermeticEnabled(env: NodeJS.ProcessEnv = process.env): boolean {
+  return env.EVALS_HERMETIC !== '0';
+}
+
+/**
+ * Pure allowlist scrub. No I/O. Overrides spread LAST so per-test env
+ * (GSTACK_HOME, CONDUCTOR_WORKSPACE_PATH, GSTACK_HEADLESS opt-out) always
+ * wins over the scrub — that is the documented re-contamination escape and
+ * the wiring tripwire forbids passing raw process.env through it.
+ */
+export function buildHermeticEnv(
+  base: NodeJS.ProcessEnv,
+  hermeticVars: Record<string, string>,
+  overrides?: Record<string, string | undefined>,
+  opts?: HermeticEnvOpts,
+): Record<string, string> {
+  if (!isHermeticEnabled(base)) {
+    // Escape hatch: byte-identical to the legacy spread.
+    const legacy: Record<string, string> = {};
+    for (const [k, v] of Object.entries(base)) if (v !== undefined) legacy[k] = v;
+    for (const [k, v] of Object.entries(overrides ?? {})) if (v !== undefined) legacy[k] = v;
+    return legacy;
+  }
+
+  const promoted = promotedEnv(base);
+  const extraExact = new Set<string>();
+  const extraPrefixes: string[] = [];
+  for (const entry of opts?.extraAllow ?? []) {
+    if (entry.endsWith('*')) extraPrefixes.push(entry.slice(0, -1));
+    else extraExact.add(entry);
+  }
+
+  const out: Record<string, string> = {};
+  for (const [k, v] of Object.entries(promoted)) {
+    if (v === undefined) continue;
+    const allowed =
+      ALLOW_EXACT.has(k) ||
+      extraExact.has(k) ||
+      ALLOW_PREFIXES.some((p) => k.startsWith(p)) ||
+      extraPrefixes.some((p) => k.startsWith(p));
+    if (allowed) out[k] = v;
+  }
+  if (!out.TERM) out.TERM = 'xterm-256color';
+  Object.assign(out, hermeticVars);
+  for (const [k, v] of Object.entries(overrides ?? {})) if (v !== undefined) out[k] = v;
+  return out;
+}
+
+export interface SeedConfigOpts {
+  /** When undefined (operator has no key exported), customApiKeyResponses is
+   * omitted — the child fails auth exactly as it would today, no throw here. */
+  apiKey: string | undefined;
+  trustedDirs: string[];
+}
+
+/**
+ * Minimal $CLAUDE_CONFIG_DIR/.claude.json for fresh-config children.
+ *
+ * Empirically verified 2026-06-12 on claude 2.1.175: PRINT MODE (`claude -p`)
+ * with ANTHROPIC_API_KEY needs NO seed at all — a fresh empty config dir ran
+ * non-interactively (exit 0, real cost billed to the key). The seed exists
+ * for the PTY path, where first-run TUI prompts DO appear:
+ * - hasCompletedOnboarding: suppresses the onboarding flow
+ * - customApiKeyResponses.approved: suppresses the "use this API key?"
+ *   prompt; entries are the key's LAST 20 CHARS (shape verified against a
+ *   real ~/.claude.json)
+ * - projects[dir].hasTrustDialogAccepted: pre-trusts repo-cwd PTY sessions
+ *   (the pty-runner's 15s trust-watcher remains as fallback for temp cwds)
+ * bypassPermissionsModeAccepted was considered and dropped: absent from a
+ * real config even though --dangerously-skip-permissions is in daily use.
+ */
+export function buildSeedConfig(opts: SeedConfigOpts): Record<string, unknown> {
+  const seed: Record<string, unknown> = {
+    hasCompletedOnboarding: true,
+    projects: Object.fromEntries(
+      opts.trustedDirs.map((dir) => [
+        dir,
+        { hasTrustDialogAccepted: true, hasCompletedProjectOnboarding: true },
+      ]),
+    ),
+  };
+  if (opts.apiKey) {
+    seed.customApiKeyResponses = { approved: [opts.apiKey.slice(-20)] };
+  }
+  return seed;
+}
+
+export interface HermeticDirs {
+  /** Ends in `/.claude` — load-bearing: extractPlanFilePath in
+   * claude-pty-runner.ts:191 anchors plan-file paths on `.claude/plans/`
+   * under a /var|/tmp prefix. Renaming this segment breaks PTY plan tests. */
+  configDir: string;
+  gstackHome: string;
+  runRoot: string;
+}
+
+const DIR_PREFIX = 'gstack-hermetic-';
+
+let cachedDirs: HermeticDirs | null = null;
+
+/** Repo root for the trusted-dir seed: test files live in <root>/test/helpers. */
+function repoRoot(): string {
+  return path.resolve(__dirname, '..', '..');
+}
+
+/**
+ * Sync memoized per-process singleton — intentionally NO async gap between
+ * the cache check and create+seed, so concurrent first calls under
+ * `bun test --concurrent` cannot double-create or observe a half-seeded dir.
+ * Shared across all tests in the process: that matches CI's within-job
+ * shared /home/runner (operator isolation, not per-test isolation).
+ */
+export function getHermeticDirs(): HermeticDirs {
+  if (cachedDirs) return cachedDirs;
+
+  gcStaleHermeticDirs();
+
+  // Embed our pid so the GC of future processes can check liveness.
+  const runRoot = fs.mkdtempSync(path.join(os.tmpdir(), `${DIR_PREFIX}${process.pid}-`));
+  const configDir = path.join(runRoot, '.claude');
+  const gstackHome = path.join(runRoot, 'gstack-home');
+
+  // A half-seeded config dir means children hang on first-run prompts until
+  // the test timeout — far worse than failing loudly here. So we throw on
+  // failure, but tear down the partial dir first: an unseeded runRoot named
+  // with our (alive) pid would be skipped by this process's GC and leak until
+  // process exit, so remove it before rethrowing.
+  try {
+    fs.mkdirSync(configDir, { recursive: true });
+    fs.mkdirSync(gstackHome, { recursive: true });
+    const seed = buildSeedConfig({
+      apiKey: process.env.ANTHROPIC_API_KEY ?? process.env.GSTACK_ANTHROPIC_API_KEY,
+      trustedDirs: [repoRoot()],
+    });
+    fs.writeFileSync(path.join(configDir, '.claude.json'), JSON.stringify(seed, null, 2));
+  } catch (err) {
+    try { fs.rmSync(runRoot, { recursive: true, force: true }); } catch { /* best-effort */ }
+    throw err;
+  }
+
+  process.on('exit', () => {
+    // Exit handlers cannot await: sync best-effort removal only. Anything
+    // left behind is reclaimed by the next process's pid-aware GC.
+    try { fs.rmSync(runRoot, { recursive: true, force: true }); } catch { /* GC reclaims */ }
+  });
+
+  cachedDirs = { configDir, gstackHome, runRoot };
+  return cachedDirs;
+}
+
+/** A dir younger than this is never GC'd even if its pid looks dead — guards
+ * against PID reuse deleting a freshly-created dir whose original pid exited
+ * and was recycled to an unrelated live process between create and GC. */
+const GC_MIN_AGE_MS = 60 * 60 * 1000; // 1h
+
+/**
+ * Reclaim leftovers from crashed runs. Two signals, both required: the
+ * embedded pid is dead AND the dir is older than GC_MIN_AGE_MS. Pid-alone
+ * would risk PID-reuse false-deletes of live dirs; age-alone would delete a
+ * live >24h eval run's config out from under it. Exported for tests.
+ */
+export function gcStaleHermeticDirs(tmpDir: string = os.tmpdir()): void {
+  let entries: string[];
+  try { entries = fs.readdirSync(tmpDir); } catch { return; }
+  const now = Date.now();
+  for (const name of entries) {
+    if (!name.startsWith(DIR_PREFIX)) continue;
+    const pidStr = name.slice(DIR_PREFIX.length).split('-')[0];
+    const pid = Number(pidStr);
+    if (!Number.isInteger(pid) || pid <= 0) continue;
+    if (pid === process.pid || isProcessAlive(pid)) continue;
+    const full = path.join(tmpDir, name);
+    try {
+      if (now - fs.statSync(full).mtimeMs < GC_MIN_AGE_MS) continue; // too fresh
+    } catch { continue; } // vanished or unreadable — leave it
+    try { fs.rmSync(full, { recursive: true, force: true }); } catch { /* best-effort */ }
+  }
+}
+
+/**
+ * The composition runners use: scrub process.env, point the child at the
+ * singleton hermetic dirs, apply per-test overrides last. Returns the legacy
+ * env untouched when EVALS_HERMETIC=0 (and skips dir creation entirely).
+ */
+export function hermeticChildEnv(
+  overrides?: Record<string, string | undefined>,
+  opts?: HermeticEnvOpts,
+): Record<string, string> {
+  if (!isHermeticEnabled()) {
+    return buildHermeticEnv(process.env, {}, overrides, opts);
+  }
+  const dirs = getHermeticDirs();
+  return buildHermeticEnv(
+    process.env,
+    { CLAUDE_CONFIG_DIR: dirs.configDir, GSTACK_HOME: dirs.gstackHome },
+    overrides,
+    opts,
+  );
+}
diff --git a/test/helpers/session-runner.ts b/test/helpers/session-runner.ts
index 675255cf2..c29c59022 100644
--- a/test/helpers/session-runner.ts
+++ b/test/helpers/session-runner.ts
@@ -10,6 +10,7 @@ import * as fs from 'fs';
 import * as path from 'path';
 import * as os from 'os';
 import { getProjectEvalDir } from './eval-store';
+import { hermeticChildEnv, isHermeticEnabled } from './hermetic-env';
 
 const GSTACK_DEV_DIR = path.join(os.homedir(), '.gstack-dev');
 const HEARTBEAT_PATH = path.join(GSTACK_DEV_DIR, 'e2e-live.json'); // heartbeat stays global
@@ -167,6 +168,10 @@ export async function runSkillTest(options: {
     '--max-turns', String(maxTurns),
     '--allowed-tools', ...allowedTools,
   ];
+  // Hermetic children get zero MCP servers (no --mcp-config is passed).
+  // Gated on the same call-time check as the env scrub so EVALS_HERMETIC=0
+  // restores operator MCP along with the operator env.
+  if (isHermeticEnabled()) args.push('--strict-mcp-config');
 
   // Write prompt to a temp file OUTSIDE workingDirectory to avoid race conditions
   // where afterAll cleanup deletes the dir before cat reads the file (especially
@@ -176,11 +181,14 @@ export async function runSkillTest(options: {
 
   const proc = Bun.spawn(['sh', '-c', `cat "${promptFile}" | claude ${args.map(a => `"${a}"`).join(' ')}`], {
     cwd: workingDirectory,
+    // Hermetic by default (see test/helpers/hermetic-env.ts): operator
+    // session context (CONDUCTOR_*, CLAUDECODE, ~/.claude config, ~/.gstack)
+    // never reaches the child; EVALS_HERMETIC=0 restores the legacy env.
     // Default GSTACK_HEADLESS=1 so eval/E2E runs classify as headless (BLOCK on an
     // AskUserQuestion failure rather than emit a prose question no human reads). A
     // suite exercising the INTERACTIVE prose-fallback path opts out by passing
     // `env: { GSTACK_HEADLESS: '' }` — extraEnv wins because it spreads last.
-    env: { ...process.env, GSTACK_HEADLESS: '1', ...extraEnv },
+    env: hermeticChildEnv({ GSTACK_HEADLESS: '1', ...extraEnv }),
     stdout: 'pipe',
     stderr: 'pipe',
   });
diff --git a/test/helpers/touchfiles.ts b/test/helpers/touchfiles.ts
index 68bc2062e..946a38007 100644
--- a/test/helpers/touchfiles.ts
+++ b/test/helpers/touchfiles.ts
@@ -36,6 +36,11 @@ export const E2E_TOUCHFILES: Record<string, string[]> = {
   'browse-basic':    ['browse/src/**', 'browse/test/test-server.ts'],
   'browse-snapshot': ['browse/src/**', 'browse/test/test-server.ts'],
 
+  // Hermetic isolation canaries (hermetic-env.ts is also a GLOBAL touchfile;
+  // these entries exist so the canaries themselves stay tier-classified)
+  'hermetic-canary':   ['test/helpers/hermetic-env.ts', 'test/helpers/session-runner.ts', 'test/skill-e2e-hermetic-canary.test.ts', 'lib/conductor-env-shim.ts'],
+  'hermetic-sentinel': ['test/helpers/hermetic-env.ts', 'test/helpers/session-runner.ts', 'test/skill-e2e-hermetic-canary.test.ts', 'lib/conductor-env-shim.ts'],
+
   // SKILL.md setup + preamble (depend on ROOT SKILL.md + gen-skill-docs)
   'skillmd-setup-discovery':  ['SKILL.md', 'SKILL.md.tmpl', 'scripts/gen-skill-docs.ts'],
   'skillmd-no-local-binary':  ['SKILL.md', 'SKILL.md.tmpl', 'scripts/gen-skill-docs.ts'],
@@ -111,7 +116,12 @@ export const E2E_TOUCHFILES: Record<string, string[]> = {
   // written a never-ask preference, AUQ should still auto-decide rather than
   // surfacing the question. Touches the question-tuning + preference
   // infrastructure plus the resolvers that own the AUTO_DECIDE preamble.
-  'auto-decide-preserved':        ['scripts/resolvers/question-tuning.ts', 'scripts/resolvers/preamble/generate-ask-user-format.ts', 'scripts/resolvers/preamble/generate-completion-status.ts', 'plan-ceo-review/**', 'bin/gstack-question-preference', 'bin/gstack-config', 'bin/gstack-slug', 'test/helpers/claude-pty-runner.ts'],
+  'auto-decide-preserved':        ['scripts/resolvers/question-tuning.ts', 'scripts/resolvers/preamble/generate-ask-user-format.ts', 'scripts/resolvers/preamble/generate-preamble-bash.ts', 'scripts/resolvers/preamble/generate-completion-status.ts', 'plan-ceo-review/**', 'bin/gstack-question-preference', 'bin/gstack-config', 'bin/gstack-slug', 'hosts/claude/hooks/question-preference-hook.ts', 'lib/is-conductor.ts', 'test/helpers/claude-pty-runner.ts'],
+
+  // Conductor → prose decision brief (Conductor signal makes prose the default;
+  // the PreToolUse hook denies the flaky tool). Touches the resolver that owns
+  // the Conductor rule, the preamble signal, the hook, and the detection helper.
+  'conductor-prose':              ['scripts/resolvers/preamble/generate-ask-user-format.ts', 'scripts/resolvers/preamble/generate-preamble-bash.ts', 'scripts/resolvers/preamble.ts', 'plan-eng-review/**', 'hosts/claude/hooks/question-preference-hook.ts', 'lib/is-conductor.ts', 'test/helpers/claude-pty-runner.ts', 'test/skill-e2e-conductor-prose.test.ts'],
 
   // Real-PTY E2E batch (#6 new tests on the harness).
   // Each one tests behavior the SDK harness can't observe (rendered TTY,
@@ -437,6 +447,11 @@ export const E2E_TIERS: Record<string, 'gate' | 'periodic'> = {
   'browse-basic': 'gate',
   'browse-snapshot': 'gate',
 
+  // Hermetic isolation — gate (deterministic env/config assertions; if the
+  // clean room breaks, every other eval's signal is contaminated)
+  'hermetic-canary': 'gate',
+  'hermetic-sentinel': 'gate',
+
   // SKILL.md setup — gate (if setup breaks, no skill works)
   'skillmd-setup-discovery': 'gate',
   'skillmd-no-local-binary': 'gate',
@@ -510,6 +525,7 @@ export const E2E_TIERS: Record<string, 'gate' | 'periodic'> = {
   // v1.21+ auto-mode regression tests
   'office-hours-auto-mode': 'gate',
   'auto-decide-preserved': 'periodic',
+  'conductor-prose': 'periodic',
   'e2e-harness-audit': 'gate',
 
   // Real-PTY E2E batch — tier classification:
@@ -785,6 +801,7 @@ export const LLM_JUDGE_TOUCHFILES: Record<string, string[]> = {
  */
 export const GLOBAL_TOUCHFILES = [
   'test/helpers/session-runner.ts',  // All E2E tests use this runner
+  'test/helpers/hermetic-env.ts',    // Changes every E2E child's environment
   'test/helpers/eval-store.ts',      // All E2E tests store results here
   'test/helpers/touchfiles.ts',      // Self-referential — reclassifying wrong is dangerous
 ];
diff --git a/test/hermetic-wiring.test.ts b/test/hermetic-wiring.test.ts
new file mode 100644
index 000000000..08528586d
--- /dev/null
+++ b/test/hermetic-wiring.test.ts
@@ -0,0 +1,113 @@
+/**
+ * Static-grep tripwire for the hermetic E2E wiring. Free tier — no API.
+ *
+ * Every E2E runner spawns its child through hermeticChildEnv(); if a refactor
+ * reverts any spawn site to a raw `...process.env` spread (or a callsite
+ * smuggles the operator env back in through the overrides parameter), local
+ * evals silently re-contaminate and nothing fails until a human notices
+ * weird results again — which took three burned suites last time.
+ *
+ * Pattern mirrors browse/test/terminal-agent-pid-identity.test.ts and
+ * browse/test/server-embedder-terminal-port.test.ts: read source files as
+ * text, assert invariants on their contents. Brittle by design — renaming
+ * the helper must force the author to look here.
+ */
+
+import { describe, test, expect } from 'bun:test';
+import * as fs from 'fs';
+import * as path from 'path';
+
+const ROOT = path.resolve(new URL(import.meta.url).pathname, '..', '..');
+
+const RUNNERS = [
+  'test/helpers/session-runner.ts',
+  'test/helpers/claude-pty-runner.ts',
+  'test/helpers/codex-session-runner.ts',
+  'test/helpers/gemini-session-runner.ts',
+  'test/helpers/agent-sdk-runner.ts',
+];
+
+function read(rel: string): string {
+  return fs.readFileSync(path.join(ROOT, rel), 'utf-8');
+}
+
+describe('hermetic wiring tripwire', () => {
+  test('every runner builds its child env via hermeticChildEnv()', () => {
+    for (const rel of RUNNERS) {
+      const src = read(rel);
+      expect(src.includes('hermeticChildEnv(') ).toBe(true);
+      expect(src.includes("from './hermetic-env'")).toBe(true);
+    }
+  });
+
+  test('no runner spawns a child with a raw process.env spread', () => {
+    // `...process.env` inside an env object is the exact pre-hermetic leak.
+    // hermetic-env.ts itself legitimately READS process.env (call-time
+    // snapshot); the runners must not SPREAD it into a child env.
+    for (const rel of RUNNERS) {
+      const offenders = read(rel)
+        .split('\n')
+        .map((line, i) => ({ line, n: i + 1 }))
+        .filter(({ line }) => line.includes('...process.env'));
+      expect(
+        offenders,
+        `${rel} spreads raw process.env into a child env at line(s) ` +
+          offenders.map((o) => o.n).join(', ') +
+          ' — route through hermeticChildEnv() instead',
+      ).toEqual([]);
+    }
+  });
+
+  test('claude runners gate --strict-mcp-config on isHermeticEnabled()', () => {
+    // Zero MCP servers for hermetic children; EVALS_HERMETIC=0 must restore
+    // operator MCP along with the operator env (the flag may not be
+    // unconditional, or the escape hatch lies).
+    for (const rel of ['test/helpers/session-runner.ts', 'test/helpers/claude-pty-runner.ts']) {
+      const src = read(rel);
+      expect(src.includes('--strict-mcp-config')).toBe(true);
+      const gated =
+        /if\s*\(\s*isHermeticEnabled\(\)\s*\)\s*(args\.push\(\s*)?['"]--strict-mcp-config['"]/.test(src) ||
+        /const hermetic = isHermeticEnabled\(\);[\s\S]{0,200}if\s*\(hermetic\)\s*args\.push\(\s*['"]--strict-mcp-config['"]/.test(src);
+      expect(gated, `${rel}: --strict-mcp-config must be gated on isHermeticEnabled()`).toBe(true);
+    }
+  });
+
+  test('no test callsite passes the whole operator env as a RUNNER override', () => {
+    // Overrides merge last by design (per-test GSTACK_HOME etc.) — passing
+    // process.env itself through that hole defeats the entire scrub. Scoped
+    // to OUR runner calls: unit tests that spawnSync gstack bin scripts with
+    // `...process.env` are test-process spawns, not eval children, and are
+    // legitimately the test's own business.
+    const RUNNER_CALL =
+      /\b(runSkillTest|launchClaudePty|runPlanSkillObservation|runPlanSkillCounting|runPlanSkillFloorCheck|runAgentSdkTest|runCodexSkillTest|runGeminiSkillTest)\s*\(/;
+    const DIRECT_SPAWN = /\b(spawnSync|spawn|execSync|exec|Bun\.spawn|Bun\.spawnSync)\s*\(/;
+    const testDir = path.join(ROOT, 'test');
+    const offenders: string[] = [];
+    const walk = (dir: string) => {
+      for (const entry of fs.readdirSync(dir, { withFileTypes: true })) {
+        const full = path.join(dir, entry.name);
+        if (entry.isDirectory()) { walk(full); continue; }
+        if (!entry.name.endsWith('.test.ts')) continue;
+        if (entry.name === 'hermetic-wiring.test.ts') continue;
+        const lines = fs.readFileSync(full, 'utf-8').split('\n');
+        for (let i = 0; i < lines.length; i++) {
+          if (!/env:\s*(\{\s*\.\.\.\s*process\.env|process\.env\b(?!\.))/.test(lines[i])) continue;
+          // Walk backwards to the nearest enclosing call: runner vs direct spawn.
+          for (let j = i; j >= Math.max(0, i - 25); j--) {
+            if (DIRECT_SPAWN.test(lines[j])) break; // test's own spawn — fine
+            if (RUNNER_CALL.test(lines[j])) {
+              offenders.push(`${path.relative(ROOT, full)}:${i + 1}`);
+              break;
+            }
+          }
+        }
+      }
+    };
+    walk(testDir);
+    expect(
+      offenders,
+      'These callsites pass the operator env into an eval child, defeating the hermetic scrub: ' +
+        offenders.join(', '),
+    ).toEqual([]);
+  });
+});
diff --git a/test/is-conductor.test.ts b/test/is-conductor.test.ts
new file mode 100644
index 000000000..be06eabe4
--- /dev/null
+++ b/test/is-conductor.test.ts
@@ -0,0 +1,50 @@
+import { describe, test, expect } from 'bun:test';
+import { isConductor } from '../lib/is-conductor';
+
+describe('is-conductor', () => {
+  test('true when CONDUCTOR_WORKSPACE_PATH is set', () => {
+    expect(isConductor({ CONDUCTOR_WORKSPACE_PATH: '/Users/x/conductor/ws' })).toBe(true);
+  });
+
+  test('true when CONDUCTOR_PORT is set', () => {
+    expect(isConductor({ CONDUCTOR_PORT: '55070' })).toBe(true);
+  });
+
+  test('true when both are set', () => {
+    expect(isConductor({ CONDUCTOR_WORKSPACE_PATH: '/ws', CONDUCTOR_PORT: '55070' })).toBe(true);
+  });
+
+  test('false when neither is set', () => {
+    expect(isConductor({ HOME: '/Users/x', PATH: '/usr/bin' })).toBe(false);
+  });
+
+  test('false on an empty env', () => {
+    expect(isConductor({})).toBe(false);
+  });
+
+  test('false when the vars are present but empty (Codex #1 hardening — empty != set)', () => {
+    expect(isConductor({ CONDUCTOR_WORKSPACE_PATH: '', CONDUCTOR_PORT: '' })).toBe(false);
+  });
+
+  test('reads the passed env at call time, not a module-load snapshot', () => {
+    const env: NodeJS.ProcessEnv = {};
+    expect(isConductor(env)).toBe(false);
+    // mutate AFTER the first call — a call-time read must see the new value
+    env.CONDUCTOR_PORT = '55070';
+    expect(isConductor(env)).toBe(true);
+  });
+
+  test('defaults to process.env when no arg is passed', () => {
+    const saved = process.env.CONDUCTOR_PORT;
+    try {
+      process.env.CONDUCTOR_PORT = '12345';
+      expect(isConductor()).toBe(true);
+      delete process.env.CONDUCTOR_PORT;
+      // CONDUCTOR_WORKSPACE_PATH may be set in a real Conductor session; guard the assertion
+      if (!process.env.CONDUCTOR_WORKSPACE_PATH) expect(isConductor()).toBe(false);
+    } finally {
+      if (saved === undefined) delete process.env.CONDUCTOR_PORT;
+      else process.env.CONDUCTOR_PORT = saved;
+    }
+  });
+});
diff --git a/test/memory-cache-injection.test.ts b/test/memory-cache-injection.test.ts
index 3330f8d2a..3ab6a2144 100644
--- a/test/memory-cache-injection.test.ts
+++ b/test/memory-cache-injection.test.ts
@@ -43,6 +43,11 @@ function runHook(stdin: object): { stdout: string; stderr: string; status: numbe
   env.GSTACK_STATE_ROOT = stateRoot;
   env.GSTACK_QUESTION_LOG_NO_DERIVE = '1';
   delete env.GSTACK_HOME;
+  // These cases assert the defer-path memoryContext injection. Strip ambient
+  // Conductor markers so running inside Conductor (CONDUCTOR_WORKSPACE_PATH/PORT
+  // set) doesn't flip the hook into the [conductor] prose deny instead of defer.
+  delete env.CONDUCTOR_WORKSPACE_PATH;
+  delete env.CONDUCTOR_PORT;
   const res = spawnSync(HOOK, [], {
     env,
     input: JSON.stringify({ ...stdin, cwd: fixtureCwd }),
diff --git a/test/preamble-compose.test.ts b/test/preamble-compose.test.ts
index 22fdfd7c8..f2bcf8bfc 100644
--- a/test/preamble-compose.test.ts
+++ b/test/preamble-compose.test.ts
@@ -70,3 +70,13 @@ describe('Preamble composition order', () => {
     expect(out).not.toContain('## AskUserQuestion Format');
   });
 });
+
+describe('Conductor signal (preamble bash)', () => {
+  test('claude preamble emits CONDUCTOR_SESSION, gated on != headless (Issue 8)', () => {
+    const out = generatePreamble(makeCtx('claude', 2, 'claude'));
+    expect(out).toContain('echo "CONDUCTOR_SESSION: true"');
+    // The emission must be suppressed when the session is headless (eval/CI
+    // inside Conductor must BLOCK, not render prose to nobody).
+    expect(out).toMatch(/"\$_SESSION_KIND" != "headless"[\s\S]*CONDUCTOR_WORKSPACE_PATH[\s\S]*CONDUCTOR_PORT[\s\S]*CONDUCTOR_SESSION: true/);
+  });
+});
diff --git a/test/question-preference-hook.test.ts b/test/question-preference-hook.test.ts
index 6b06d22f4..39de02f4e 100644
--- a/test/question-preference-hook.test.ts
+++ b/test/question-preference-hook.test.ts
@@ -60,7 +60,7 @@ function writeGlobalPref(questionId: string, preference: string): void {
   fs.writeFileSync(f, JSON.stringify(prefs, null, 2));
 }
 
-function runHook(stdin: object, cwd?: string): {
+function runHook(stdin: object, cwd?: string, extraEnv?: Record<string, string>): {
   stdout: string;
   stderr: string;
   status: number;
@@ -72,7 +72,15 @@ function runHook(stdin: object, cwd?: string): {
   }
   env.GSTACK_STATE_ROOT = stateRoot;
   delete env.GSTACK_HOME;
+  // Strip ambient Conductor markers so these cases characterize NON-Conductor
+  // behavior deterministically — otherwise running the suite inside Conductor
+  // (CONDUCTOR_WORKSPACE_PATH/PORT set) would flip every defer into the
+  // [conductor] prose deny. The Conductor cases below opt back in explicitly
+  // via extraEnv.
+  delete env.CONDUCTOR_WORKSPACE_PATH;
+  delete env.CONDUCTOR_PORT;
   env.GSTACK_QUESTION_LOG_NO_DERIVE = '1';
+  if (extraEnv) Object.assign(env, extraEnv);
   const res = spawnSync(HOOK, [], {
     env,
     input: JSON.stringify({ ...stdin, cwd: cwd || fixtureCwd }),
@@ -337,6 +345,108 @@ describe('MCP variant', () => {
   });
 });
 
+// ----------------------------------------------------------------------
+// Conductor: deny + prose redirect (transport avoidance, not preference)
+// ----------------------------------------------------------------------
+
+describe('Conductor prose redirect', () => {
+  const CONDUCTOR = { CONDUCTOR_PORT: '55070' };
+
+  test('two-way, no preference → deny with [conductor] prose directive', () => {
+    const r = runHook({
+      session_id: 'c1',
+      tool_name: 'AskUserQuestion',
+      tool_use_id: 'tu-c1',
+      tool_input: {
+        questions: [
+          { question: '<gstack-qid:test-q> Need approval?', options: ['A) Yes (recommended)', 'B) No'] },
+        ],
+      },
+    }, undefined, CONDUCTOR);
+    expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('deny');
+    expect(r.parsed?.hookSpecificOutput?.permissionDecisionReason).toContain('[conductor]');
+    expect(r.parsed?.hookSpecificOutput?.permissionDecisionReason).toMatch(/do not call askuserquestion/i);
+    expect(r.parsed?.hookSpecificOutput?.permissionDecisionReason).toMatch(/reply with a letter/i);
+  });
+
+  test('UNMARKED question (modal path) → deny with prose directive', () => {
+    const r = runHook({
+      session_id: 'c2',
+      tool_name: 'AskUserQuestion',
+      tool_use_id: 'tu-c2',
+      tool_input: {
+        questions: [
+          { question: 'No marker — an ad-hoc question', options: ['A) Yes (recommended)', 'B) No'] },
+        ],
+      },
+    }, undefined, CONDUCTOR);
+    expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('deny');
+    expect(r.parsed?.hookSpecificOutput?.permissionDecisionReason).toContain('[conductor]');
+  });
+
+  test('one-way door → deny with prose directive (NOT defer — destructive must reach human via prose)', () => {
+    const r = runHook({
+      session_id: 'c3',
+      tool_name: 'AskUserQuestion',
+      tool_use_id: 'tu-c3',
+      tool_input: {
+        questions: [
+          {
+            question: '<gstack-qid:ship-test-failure-triage> Tests failed.',
+            options: ['A) Fix now (recommended)', 'B) Investigate', 'C) Ack and ship'],
+          },
+        ],
+      },
+    }, undefined, CONDUCTOR);
+    expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('deny');
+    expect(r.parsed?.hookSpecificOutput?.permissionDecisionReason).toContain('[conductor]');
+    expect(r.parsed?.hookSpecificOutput?.permissionDecisionReason).toMatch(/typed confirmation/i);
+  });
+
+  test('CONDUCTOR_WORKSPACE_PATH alone also triggers the redirect', () => {
+    const r = runHook({
+      session_id: 'c4',
+      tool_name: 'mcp__conductor__AskUserQuestion',
+      tool_use_id: 'tu-c4',
+      tool_input: {
+        questions: [{ question: '<gstack-qid:test-q> Pick?', options: ['A) X (recommended)', 'B) Y'] }],
+      },
+    }, undefined, { CONDUCTOR_WORKSPACE_PATH: '/Users/x/conductor/ws' });
+    expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('deny');
+    expect(r.parsed?.hookSpecificOutput?.permissionDecisionReason).toContain('[conductor]');
+  });
+
+  test('PRECEDENCE: full never-ask auto-decide still wins over Conductor prose', () => {
+    writeProjectPref('ship-pre-landing-review-fix', 'never-ask');
+    const r = runHook({
+      session_id: 'c5',
+      tool_name: 'AskUserQuestion',
+      tool_use_id: 'tu-c5',
+      tool_input: {
+        questions: [
+          {
+            question: '<gstack-qid:ship-pre-landing-review-fix> Pre-landing review flagged issue.',
+            options: ['A) Fix now (recommended)', 'B) Skip'],
+          },
+        ],
+      },
+    }, undefined, CONDUCTOR);
+    expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('deny');
+    // auto-decide reason, NOT the conductor prose reason
+    expect(r.parsed?.hookSpecificOutput?.permissionDecisionReason).toContain('plan-tune auto-decide');
+    expect(r.parsed?.hookSpecificOutput?.permissionDecisionReason).not.toContain('[conductor]');
+  });
+
+  test('non-AUQ tool in Conductor → still defer (no redirect on unrelated tools)', () => {
+    const r = runHook(
+      { session_id: 'c6', tool_name: 'Bash', tool_use_id: 'tu-c6', tool_input: {} },
+      undefined,
+      CONDUCTOR,
+    );
+    expect(r.parsed?.hookSpecificOutput?.permissionDecision).toBe('defer');
+  });
+});
+
 // ----------------------------------------------------------------------
 // Auto-decided event logging (since PostToolUse never fires on deny)
 // ----------------------------------------------------------------------
diff --git a/test/resolver-ask-user-format.test.ts b/test/resolver-ask-user-format.test.ts
index 69e66a055..0b498af0a 100644
--- a/test/resolver-ask-user-format.test.ts
+++ b/test/resolver-ask-user-format.test.ts
@@ -225,8 +225,25 @@ describe('generateAskUserFormat — runtime-failure prose fallback', () => {
     expect(out).toMatch(/must be sent as tool_use, not prose — unless the documented failure fallback/);
   });
 
-  test('OV2: the self-check "not writing prose" line carries the fallback qualifier', () => {
-    expect(out).toMatch(/not writing prose — unless the documented failure fallback applies/);
+  test('OV2: the self-check "not writing prose" line carries the Conductor + fallback qualifiers', () => {
+    // After the Conductor-default-prose change, the exception is two-pronged:
+    // CONDUCTOR_SESSION makes prose the default, OR the documented failure fallback.
+    expect(out).toMatch(/not writing prose — unless `CONDUCTOR_SESSION: true`[\s\S]*OR the documented failure fallback applies/);
+  });
+
+  // Conductor-default-prose contract (the proactive path, distinct from the
+  // failure fallback). Guards the Tool-resolution rule + self-check wording.
+  test('Conductor: do-not-call rule present in Tool resolution', () => {
+    expect(out).toMatch(/CONDUCTOR_SESSION: true/);
+    expect(out).toMatch(/do NOT call AskUserQuestion at all/);
+    expect(out).toMatch(/Auto-decide preferences still apply first/);
+    expect(out).toMatch(/gstack-question-log/);
+  });
+
+  test('Conductor: one-way prose rule + continuation protocol present', () => {
+    expect(out).toMatch(/one-way\b[\s\S]*typed confirmation/i);
+    expect(out).toMatch(/never proceed on a vague/i);
+    expect(out).toMatch(/Continuation — mapping a typed reply/);
   });
 });
 
diff --git a/test/skill-e2e-auto-decide-preserved.test.ts b/test/skill-e2e-auto-decide-preserved.test.ts
index 8b773d5fc..7e74d8438 100644
--- a/test/skill-e2e-auto-decide-preserved.test.ts
+++ b/test/skill-e2e-auto-decide-preserved.test.ts
@@ -100,11 +100,19 @@ describeE2E('AUTO_DECIDE opt-in preserved under Conductor flags (periodic)', ()
       }
 
       // 4. Run /plan-ceo-review with the Conductor flag set + isolated state.
+      //    GSTACK_HOME=tmpHome is REQUIRED: the preference + question_tuning were
+      //    seeded there. Without it the spawned claude reads the real ~/.gstack,
+      //    never sees the never-ask preference, and the test silently exercises
+      //    the wrong state root (pre-existing bug, Codex #9 / Issue 13).
+      //    CONDUCTOR_WORKSPACE_PATH additionally proves auto-decide still WINS
+      //    over the Conductor prose redirect (precedence: settled preference
+      //    beats transport-avoidance).
       const obs = await runPlanSkillObservation({
         skillName: 'plan-ceo-review',
         inPlanMode: true,
         extraArgs: ['--disallowedTools', 'AskUserQuestion'],
         timeoutMs: 300_000,
+        env: { GSTACK_HOME: tmpHome, CONDUCTOR_WORKSPACE_PATH: tmpHome },
       });
 
       // 5. Pass: 'auto_decided' (the strongest signal) or 'plan_ready' with
diff --git a/test/skill-e2e-conductor-prose.test.ts b/test/skill-e2e-conductor-prose.test.ts
new file mode 100644
index 000000000..a130dcab5
--- /dev/null
+++ b/test/skill-e2e-conductor-prose.test.ts
@@ -0,0 +1,69 @@
+/**
+ * Conductor → prose decision brief (periodic-tier, paid, real-PTY).
+ *
+ * Proves the end-to-end behavior: when CONDUCTOR_SESSION is signalled, a skill
+ * that hits a decision renders a PROSE decision brief and waits, instead of
+ * silently skipping the user.
+ *
+ * SCOPE — read before trusting this as the Conductor guard. This is END-TO-END
+ * BEHAVIOR coverage, NOT the discriminating Conductor guarantee:
+ *   - The deterministic guard is test/question-preference-hook.test.ts
+ *     ("Conductor prose redirect") — it sets process.env.CONDUCTOR_* and asserts
+ *     the PreToolUse hook denies + redirects. That test CAN fail on unfixed code.
+ *   - The PTY harness here cannot register `mcp__conductor__AskUserQuestion`, so
+ *     it tests "native AUQ unavailable + Conductor signal → prose," NOT "the MCP
+ *     variant exists and must not be called" (Codex #10). Under --disallowedTools
+ *     a present-human interactive session already prose-falls-back, so this test
+ *     is a smoke check that the Conductor path still produces a prose brief, not
+ *     a proof that the Conductor signal (vs the generic fallback) drove it.
+ *
+ * Periodic tier: model-behavior, non-deterministic.
+ */
+
+import { describe, test, expect } from 'bun:test';
+import { runPlanSkillObservation } from './helpers/claude-pty-runner';
+
+const shouldRun = !!process.env.EVALS && process.env.EVALS_TIER === 'periodic';
+const describeE2E = shouldRun ? describe : describe.skip;
+
+const FLAWED_PLAN = `# Plan: add a "developer-friendly" pricing tier
+
+## Goal
+Increase developer adoption.
+
+## Premise
+No tests mentioned, no rollout plan, no auth check on the upgrade endpoint.
+Adds a Stripe tier, a React pricing page, a Postgres entitlements table, and a
+Redis cache. The team "feels like" it should be cheaper; no developer was asked.
+`;
+
+describeE2E('Conductor renders decisions as prose (periodic)', () => {
+  test('plan-eng-review in a Conductor session surfaces a PROSE decision brief, not a silent skip', async () => {
+    const obs = await runPlanSkillObservation({
+      skillName: 'plan-eng-review',
+      inPlanMode: true,
+      // Mimic Conductor: native AUQ disabled + the Conductor env signal present.
+      extraArgs: ['--disallowedTools', 'AskUserQuestion'],
+      env: { CONDUCTOR_WORKSPACE_PATH: '/tmp/conductor-prose-e2e' },
+      initialPlanContent: FLAWED_PLAN,
+      timeoutMs: 300_000,
+    });
+
+    // The decision must reach the human as prose. 'silent_write' (wrote findings
+    // to the plan without asking) is the precise failure we guard against.
+    if (obs.outcome === 'silent_write') {
+      throw new Error(
+        `Conductor prose regression: skill wrote findings without surfacing a decision.\n` +
+          `summary: ${obs.summary}\n--- evidence ---\n${obs.evidence}`,
+      );
+    }
+    if (obs.outcome === 'exited' || obs.outcome === 'timeout') {
+      throw new Error(
+        `Conductor prose test inconclusive: outcome=${obs.outcome}\n` +
+          `summary: ${obs.summary}\n--- evidence ---\n${obs.evidence}`,
+      );
+    }
+    // A prose-rendered decision brief was observed at some point in the run.
+    expect(obs.proseAUQEverObserved).toBe(true);
+  }, 360_000);
+});
diff --git a/test/skill-e2e-hermetic-canary.test.ts b/test/skill-e2e-hermetic-canary.test.ts
new file mode 100644
index 000000000..1356b7dce
--- /dev/null
+++ b/test/skill-e2e-hermetic-canary.test.ts
@@ -0,0 +1,190 @@
+/**
+ * Hermetic-isolation canaries (gate tier, ~$0.02 each, deterministic).
+ *
+ * Two tests that make the hermeticity claim FALSIFIABLE instead of asserted:
+ *
+ * 1. `hermetic-canary` — env + auth isolation. Plants contamination vars in
+ *    the TEST process env, spawns a child through the real runner, and
+ *    asserts from the Bash tool_result in the stream-json transcript (never
+ *    the model's prose — prose can hallucinate) that the child saw a temp
+ *    `/.claude` config dir, a temp GSTACK_HOME, and none of the planted
+ *    contamination. Auth hermeticity: hard-fails when ANTHROPIC_API_KEY is
+ *    absent (a skip here would be a silent hole), and asserts
+ *    total_cost_usd > 0 — subscription/keychain OAuth reports cost 0, so
+ *    nonzero cost is the discriminator that the API key actually paid
+ *    (verified empirically 2026-06-12; the result record exposes no
+ *    auth-source field, so cost is the best available signal — residual
+ *    gap documented in the plan).
+ *
+ * 2. `hermetic-sentinel` — config isolation, the poisoned-operator probe.
+ *    Builds a FAKE operator config tree (user CLAUDE.md + an mcpServers
+ *    entry) and points the test process's CLAUDE_CONFIG_DIR at it. If the
+ *    hermetic redirect ever breaks, the child loads that poisoned tree and
+ *    the probes fire: init.mcp_servers would list the planted server
+ *    (semantic proof that --strict-mcp-config + the redirect yield ZERO MCP
+ *    servers, not an assumption), and the child's config dir would contain
+ *    the poisoned CLAUDE.md.
+ *
+ * Both canaries double as the seed-schema / CLI version-skew tripwire: a
+ * claude release that changes first-run behavior or config discovery fails
+ * here first, loudly, in the gate tier.
+ */
+
+import { expect, afterAll } from 'bun:test';
+import * as fs from 'fs';
+import * as path from 'path';
+import * as os from 'os';
+import { runSkillTest } from './helpers/session-runner';
+import {
+  describeIfSelected, testIfSelected, createEvalCollector, finalizeEvalCollector,
+  recordE2E, runId, logCost,
+} from './helpers/e2e-helpers';
+
+const evalCollector = createEvalCollector('e2e-hermetic');
+
+// Cheap + deterministic: the canaries assert environment facts, not model
+// quality, so the smallest model is the right tool.
+const CANARY_MODEL = 'claude-haiku-4-5-20251001';
+
+/** Extract concatenated tool_result text from the stream-json transcript. */
+function toolResultText(transcript: any[]): string {
+  const chunks: string[] = [];
+  for (const event of transcript) {
+    if (event.type !== 'user') continue;
+    for (const item of event.message?.content ?? []) {
+      if (item.type !== 'tool_result') continue;
+      if (typeof item.content === 'string') chunks.push(item.content);
+      else for (const c of item.content ?? []) if (c.type === 'text') chunks.push(c.text);
+    }
+  }
+  return chunks.join('\n');
+}
+
+function initEvent(transcript: any[]): any {
+  return transcript.find((e) => e.type === 'system' && e.subtype === 'init');
+}
+
+describeIfSelected('hermetic isolation canaries', ['hermetic-canary', 'hermetic-sentinel'], () => {
+  testIfSelected('hermetic-canary', async () => {
+    // Auth hermeticity is part of the contract: a missing key must FAIL the
+    // gate, not skip it — a skipped canary is a silent hole.
+    if (!process.env.ANTHROPIC_API_KEY) {
+      throw new Error('hermetic-canary requires ANTHROPIC_API_KEY (source ~/.zshrc); refusing to skip');
+    }
+
+    const workDir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermetic-canary-'));
+    // Plant contamination deterministically — the operator env may or may not
+    // carry these, so set them ourselves and restore after.
+    const planted: Record<string, string> = {
+      CONDUCTOR_WORKSPACE_PATH: '/tmp/poison-conductor-ws',
+      GBRAIN_POISON_PROBE: 'leaked',
+    };
+    const prev: Record<string, string | undefined> = {};
+    for (const [k, v] of Object.entries(planted)) { prev[k] = process.env[k]; process.env[k] = v; }
+
+    try {
+      const result = await runSkillTest({
+        prompt: 'Run exactly this bash command and then stop: ' +
+          'echo "CFG=$CLAUDE_CONFIG_DIR"; echo "GH=$GSTACK_HOME"; ' +
+          'echo "CW=$CONDUCTOR_WORKSPACE_PATH"; echo "GP=$GBRAIN_POISON_PROBE"',
+        workingDirectory: workDir,
+        maxTurns: 3,
+        allowedTools: ['Bash'],
+        timeout: 120_000,
+        testName: 'hermetic-canary',
+        runId,
+        model: CANARY_MODEL,
+      });
+      logCost('hermetic-canary', result);
+      recordE2E(evalCollector, 'hermetic-canary', 'e2e-hermetic', result);
+
+      expect(result.exitReason).toBe('success');
+
+      // Deterministic: assert the Bash tool OUTPUT, not the model's prose.
+      const bashOut = toolResultText(result.transcript);
+      const cfg = bashOut.match(/CFG=(\S*)/)?.[1] ?? '';
+      expect(cfg).toMatch(/gstack-hermetic-.*\/\.claude$/);
+      expect(bashOut).toMatch(/GH=\S*gstack-home/);
+      // Planted contamination must not reach the child. CLAUDECODE is NOT
+      // probed here: the child claude CLI sets CLAUDECODE=1 for its own tool
+      // subprocesses (verified empirically — CI behaves identically), so the
+      // Bash tool can't observe our scrub of it; the unit test pins that.
+      expect(bashOut).toMatch(/(^|\n)CW=\s*($|\n)/); // planted Conductor var scrubbed
+      expect(bashOut).toMatch(/(^|\n)GP=\s*($|\n)/); // GBRAIN_* scrubbed
+
+      // Zero MCP servers — semantic, from the init event, not a flag grep.
+      const init = initEvent(result.transcript);
+      expect(init).toBeTruthy();
+      expect(init.mcp_servers ?? []).toHaveLength(0);
+
+      // Auth: nonzero cost = the API key paid (OAuth/keychain reports 0).
+      expect(result.transcript.find((e) => e.type === 'result')?.total_cost_usd).toBeGreaterThan(0);
+    } finally {
+      for (const [k, v] of Object.entries(prev)) {
+        if (v === undefined) delete process.env[k]; else process.env[k] = v;
+      }
+      fs.rmSync(workDir, { recursive: true, force: true });
+    }
+  }, 180_000);
+
+  testIfSelected('hermetic-sentinel', async () => {
+    if (!process.env.ANTHROPIC_API_KEY) {
+      throw new Error('hermetic-sentinel requires ANTHROPIC_API_KEY (source ~/.zshrc); refusing to skip');
+    }
+
+    const workDir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermetic-sentinel-'));
+    // Poisoned operator config tree: if the hermetic redirect breaks, the
+    // child discovers this dir and both probes below fire.
+    const poisonRoot = fs.mkdtempSync(path.join(os.tmpdir(), 'hermetic-poison-'));
+    const poisonCfg = path.join(poisonRoot, '.claude');
+    fs.mkdirSync(poisonCfg, { recursive: true });
+    fs.writeFileSync(path.join(poisonCfg, 'CLAUDE.md'), 'POISONED OPERATOR MEMORY — must never load\n');
+    fs.writeFileSync(path.join(poisonCfg, '.claude.json'), JSON.stringify({
+      hasCompletedOnboarding: true,
+      mcpServers: { 'sentinel-mcp': { command: '/usr/bin/true', args: [] } },
+    }));
+    const prevCfgDir = process.env.CLAUDE_CONFIG_DIR;
+    process.env.CLAUDE_CONFIG_DIR = poisonCfg;
+
+    try {
+      const result = await runSkillTest({
+        prompt: 'Run exactly this bash command and then stop: ' +
+          'echo "CFG=$CLAUDE_CONFIG_DIR"; ' +
+          'if [ -f "$CLAUDE_CONFIG_DIR/CLAUDE.md" ]; then echo "USER_MD=present"; else echo "USER_MD=absent"; fi',
+        workingDirectory: workDir,
+        maxTurns: 3,
+        allowedTools: ['Bash'],
+        timeout: 120_000,
+        testName: 'hermetic-sentinel',
+        runId,
+        model: CANARY_MODEL,
+      });
+      logCost('hermetic-sentinel', result);
+      recordE2E(evalCollector, 'hermetic-sentinel', 'e2e-hermetic', result);
+
+      expect(result.exitReason).toBe('success');
+
+      const bashOut = toolResultText(result.transcript);
+      const cfg = bashOut.match(/CFG=(\S*)/)?.[1] ?? '';
+      // The redirect must beat the poisoned operator value...
+      expect(cfg).not.toBe(poisonCfg);
+      expect(cfg).toMatch(/gstack-hermetic-.*\/\.claude$/);
+      // ...and the active config dir must not carry the poisoned user memory.
+      expect(bashOut).toContain('USER_MD=absent');
+
+      // The planted MCP server must be invisible: zero servers in init.
+      const init = initEvent(result.transcript);
+      expect(init).toBeTruthy();
+      const servers = (init.mcp_servers ?? []).map((s: any) => s?.name ?? s);
+      expect(servers).toHaveLength(0);
+      expect(JSON.stringify(servers)).not.toContain('sentinel-mcp');
+    } finally {
+      if (prevCfgDir === undefined) delete process.env.CLAUDE_CONFIG_DIR;
+      else process.env.CLAUDE_CONFIG_DIR = prevCfgDir;
+      fs.rmSync(workDir, { recursive: true, force: true });
+      fs.rmSync(poisonRoot, { recursive: true, force: true });
+    }
+  }, 180_000);
+});
+
+afterAll(() => finalizeEvalCollector(evalCollector));
diff --git a/test/skill-e2e-ios.test.ts b/test/skill-e2e-ios.test.ts
index 70676d8ab..0f6c3b721 100644
--- a/test/skill-e2e-ios.test.ts
+++ b/test/skill-e2e-ios.test.ts
@@ -36,6 +36,16 @@ afterEach(() => {
   rmSync(workDir, { recursive: true, force: true });
 });
 
+// Under `bun test --concurrent`, overlapping tests read the SAME shared
+// `workDir` binding (beforeEach reassigns it mid-flight), so a fixed
+// 'daemon.pid' name collides: the first daemon claims it and every sibling
+// gets already_running against the test process's own (always-alive) pid —
+// the exact failure seen in full gate runs at 15-way concurrency. Unique
+// per-claim pidfiles keep the single-instance semantics under test while
+// removing the cross-test collision.
+let pidfileSeq = 0;
+const uniquePidfile = () => join(workDir, `daemon-${++pidfileSeq}.pid`);
+
 interface StubState {
   loggedIn: boolean;
   username: string;
@@ -205,7 +215,7 @@ class AppState {
       const daemon = await startDaemon({
         loopbackPort: 0,
         tailnetEnabled: false,
-        pidfilePath: join(workDir, 'daemon.pid'),
+        pidfilePath: uniquePidfile(),
         tunnelProvider: async () => tunnel,
       });
       if ('error' in daemon) throw new Error(daemon.error);
@@ -249,7 +259,7 @@ describe('ios-qa E2E (agent-flow simulation)', () => {
       const daemon = await startDaemon({
         loopbackPort: 0,
         tailnetEnabled: false,
-        pidfilePath: join(workDir, 'daemon.pid'),
+        pidfilePath: uniquePidfile(),
         tunnelProvider: async () => tunnel,
       });
       if ('error' in daemon) throw new Error(daemon.error);
@@ -314,7 +324,7 @@ describe('ios-qa E2E (agent-flow simulation)', () => {
       const daemon = await startDaemon({
         loopbackPort: 0,
         tailnetEnabled: false,
-        pidfilePath: join(workDir, 'daemon.pid'),
+        pidfilePath: uniquePidfile(),
         tunnelProvider: async () => tunnel,
       });
       if ('error' in daemon) throw new Error(daemon.error);
@@ -352,7 +362,7 @@ describe('ios-qa E2E (agent-flow simulation)', () => {
       const daemon = await startDaemon({
         loopbackPort: 0,
         tailnetEnabled: true,
-        pidfilePath: join(workDir, 'daemon.pid'),
+        pidfilePath: uniquePidfile(),
         tunnelProvider: async () => tunnel,
         probeImpl: async () => ({ ok: true, ownIdentity: 'mac@e2e' }),
         whoIsImpl: async () => ({ identity: 'agent@e2e', raw: {} }),
@@ -430,7 +440,7 @@ describe('ios-qa E2E (agent-flow simulation)', () => {
       const daemon = await startDaemon({
         loopbackPort: 0,
         tailnetEnabled: true,
-        pidfilePath: join(workDir, 'daemon.pid'),
+        pidfilePath: uniquePidfile(),
         tunnelProvider: async () => tunnel,
         probeImpl: async () => ({ ok: true, ownIdentity: 'mac@e2e' }),
         whoIsImpl: async () => ({ identity: 'readonly@e2e', raw: {} }),
diff --git a/test/skill-llm-eval.test.ts b/test/skill-llm-eval.test.ts
index eb47526b8..3b0d8661c 100644
--- a/test/skill-llm-eval.test.ts
+++ b/test/skill-llm-eval.test.ts
@@ -546,10 +546,13 @@ async function runWorkflowJudge(opts: {
   // slice markers vanish from the skeleton and the judge scores empty content.
   let content = fs.readFileSync(path.join(ROOT, opts.skillPath), 'utf-8');
   const secDir = path.join(ROOT, path.dirname(opts.skillPath), 'sections');
+  const sectionBodies: string[] = [];
   if (fs.existsSync(secDir)) {
     for (const f of fs.readdirSync(secDir).sort()) {
       if (f.endsWith('.md') && !f.endsWith('.md.tmpl')) {
-        content += '\n' + fs.readFileSync(path.join(secDir, f), 'utf-8');
+        const body = fs.readFileSync(path.join(secDir, f), 'utf-8');
+        sectionBodies.push(body);
+        content += '\n' + body;
       }
     }
   }
@@ -565,6 +568,17 @@ async function runWorkflowJudge(opts: {
     section = content.slice(startIdx);
   }
 
+  // Two carve shapes exist. plan-eng/plan-design moved the MARKERS into the
+  // section files, so the slice above already reaches the carved content.
+  // document-release instead keeps its markers in the skeleton and carves the
+  // workflow BODY (Steps 2-9 → sections/release-body.md) AFTER the endMarker,
+  // so the marker slice drops it. Re-append any carved section the window
+  // excluded, so the judge always sees the full workflow the agent executes.
+  for (const body of sectionBodies) {
+    const head = body.trim().slice(0, 120);
+    if (head && !section.includes(head)) section += '\n' + body;
+  }
+
   const scores = await callJudge<JudgeScore>(`You are evaluating the quality of ${opts.judgeContext} for an AI coding agent.
 
 The agent reads this document to learn ${opts.judgeGoal}. It references external tools and files