Codex cross-model doc review (via /document-release, run as part of /ship
Step 18) flagged that the CHANGELOG's contributor bullet about the new
periodic-tier E2E evals didn't scope them to /plan-eng-review — could be
misread as covering /plan-ceo-review too, which still only has static text
coverage (tracked as a P2 follow-up in TODOS.md).
Codex structured review (large-diff gate, Step 11) found 3 more real gaps:
- matchesUnnegated's bare "no" trigger over-fired on common positive-framing
idioms: "No blocker here: create a separate tier model" and "there is no
downside to promoting the payload" both got misread as negated, even
though they're endorsements of the recommendation that follows, not
rejections. Removed bare "no" from the negation list — "not"/"never"/
"cannot"/the "-n't" family are far more reliable negation signals and
don't share this idiom-collision problem.
- SENTENCE_END only recognized period-followed-by-capital as a boundary, so
a bulleted recommendation like "I would not keep this inline.\n- Separate
tier model" (very common LLM output formatting) let the earlier "not"
leak into the bullet below it. Added a second alternative: a newline
followed by a markdown bullet or numbered-list marker now also counts as
a clause boundary. Fixed the boundary-index arithmetic to use the full
match length (em[0].length) instead of a bare +1, since this new
alternative can match more than one character ("\n- ", "\n1. ").
- The minimal-change eval's own hardcoded doesn't-pattern only matched the
ASCII apostrophe (`doesn'?t`), missing the curly one — inconsistent with
matchesUnnegated's own apostrophe handling a few lines away in the same
file.
Added 3 regression tests reproducing each directly.
bun test test/helpers/e2e-helpers.test.ts test/skill-e2e-plan.test.ts
test/skill-validation.test.ts test/parity-suite.test.ts test/touchfiles.test.ts
test/gen-skill-docs.test.ts — 788 pass, 42 skip (paid E2E), 0 fail.
Codex adversarial review (Step 11) flagged that these 4 tests were not
hermetic: the prompt only said "Read plan-eng-review/SKILL.md" (relative,
unpinned), and the copied SKILL.md's own STOP-Read directive references an
ABSOLUTE path (~/.claude/skills/gstack/plan-eng-review/sections/review-sections.md)
for its carved sections file. On any machine with gstack actually installed
at that location — which includes the maintainer's own dev machine — the
agent could read and get graded against the REAL globally-installed skill
instead of this branch's sandboxed copy, silently defeating the point of
the isolated fixture. This is the exact failure mode
test/helpers/auq-sdk-capture.ts's setupSkillDir()/captureSectionReads()
already guards against for other carved-skill tests; these 4 blocks
predated that pattern and didn't use it.
Fixed with a targeted change rather than a full migration to the
auq-sdk-capture harness: extracted planEngReviewDataModelPrompt() into
e2e-helpers.ts, pinning the sandboxed SKILL.md's absolute path and
explicitly forbidding the agent from reading any other SKILL.md ("especially
nothing under ~/.claude or /Users") — matching auq-sdk-capture.ts's own
phrasing. This also resolves the maintainability specialist's earlier DRY
finding (the identical prompt string was duplicated across all 4 blocks) in
the same change, since all 4 now call the one shared function.
Also fixed a second Codex finding: touchfiles.ts only listed
plan-eng-review/** as these tests' dependency, but their actual behavior
lives partly in test/helpers/e2e-helpers.ts (setupPlanEngReviewFixture,
matchesUnnegated) — a future bug there wouldn't have gotten these 4 tests
re-selected by the diff-based selector. Added the file to each test's
touchfile entry (scoped to just these 4, not GLOBAL_TOUCHFILES, since the
new exports aren't used elsewhere yet).
Two other Codex findings reviewed and accepted as-is, not fixed:
- matchesUnnegated can still mis-score idiom-based false negation ("no
doubt", "not only... but also") — a genuine lexical-heuristic limitation,
but these idioms are exceedingly unlikely to co-occur with the specific
verb+object patterns these checks require in real review prose; chasing
every conceivable idiom is diminishing returns for a best-effort matcher.
- The new evals are periodic tier, not gate — consistent with this
repo's existing convention for all Opus-model E2E tests (documented in
CLAUDE.md's two-tier system), not a gap introduced by this branch.
bun test test/helpers/e2e-helpers.test.ts test/skill-e2e-plan.test.ts
test/skill-validation.test.ts test/parity-suite.test.ts test/touchfiles.test.ts
test/gen-skill-docs.test.ts — 786 pass, 42 skip (paid E2E), 0 fail.
Maintainability specialist re-dispatch found the standalone n't alternative
in NEGATION was unreachable: \b n't \b can never match mid-word (no word
boundary exists between the letters immediately before "n" and "n" itself
in "can't", "won't", "aren't", "hasn't"), so only the explicitly-spelled-out
contractions (don't, doesn't, didn't, shouldn't, wouldn't, isn't) were ever
detected — very common ones like "can't"/"won't"/"cannot" fell through
silently. Replaced with a generic [a-z]+n't pattern plus an explicit
"cannot" alternative (which has no apostrophe to match generically).
Also brought the original data-model-bias test's recommendsSeparateModel
check in line with its three newer sibling counterexample tests by routing
it through matchesUnnegated — it was the one E2E block still un-guarded
against a negated false-positive ("I would NOT recommend a separate tier
model" would previously have registered as a pass).
Did not touch rejectsAsUnjustified (measured-denorm test) — it directly
models rejection language ("instead of denormalizing" IS the signal, not a
positive claim needing negation-checking), so wrapping it in
matchesUnnegated would conflict with its own tuned alternatives for no
clear benefit. Did not extract a shared prompt-string constant across the
4 E2E blocks (maintainability finding) — the identical prompt string already
appears verbatim in 3 pre-existing, untouched blocks in this same file;
extracting a constant for only the 4 new blocks would create inconsistency
rather than resolve it, and touching the pre-existing blocks is outside
this PR's scope.
Added unit tests reproducing both contraction gaps directly.
bun test test/helpers/e2e-helpers.test.ts test/skill-e2e-plan.test.ts
test/skill-validation.test.ts test/parity-suite.test.ts test/touchfiles.test.ts
test/gen-skill-docs.test.ts — 783 pass, 42 skip (paid E2E), 0 fail.
Ship-workflow review re-dispatched the testing specialist against the full
updated diff (including the new minimal-change eval) and found 3 more real
bugs in the same family as the earlier fix pass:
- matchesUnnegated's fixed 20-char negation lookback was too narrow: the
patterns it guards have their own internal gaps up to 80 chars, so "I do
not think it's worth extracting the payload into columns" (negation >20
chars before the match) would be misread as an unnegated recommendation.
Now scans back to the start of the current sentence (last ./!/? before the
match) instead of a fixed count — matches the same period-bounded
assumption the calling patterns already make.
- Verified during the fix: neither "not"/"never"/etc. nor the wider window
catches contrastive phrasing ("add this field to the model RATHER THAN
create a separate table") — the rejected alternative matches the pattern
just as strongly as a genuine recommendation, with no negation word
anywhere nearby. Added "rather than"/"instead of" to the shared negation
list, benefiting all four callers, not just the new eval.
- The new minimal-change eval's acceptsInlineAddition check was missing the
matchesUnnegated guard entirely on its first alternative (unlike its
sibling recommendsSeparateModel a few lines above), and
recommendsSeparateModel's verb list missed common recommendation phrasings
(introduce, build, spin out, break out) — deliberately did NOT add "add",
since "add this new field to the existing model" is exactly how the
CORRECT answer gets phrased.
Added 3 unit tests reproducing the sentence-boundary and contrastive-phrasing
fixes directly (not just via the paid E2E path).
bun test test/helpers/e2e-helpers.test.ts test/skill-e2e-plan.test.ts
test/skill-validation.test.ts test/parity-suite.test.ts test/touchfiles.test.ts
test/gen-skill-docs.test.ts — 781 pass, 42 skip (paid E2E), 0 fail.
garrytan/gstack#1048 comment asked for concrete evaluations covering three
cases: normalized models, justified JSON fields, and minimal-change cases.
The prior commits covered the first two (data-model-bias, legitimate-json,
measured-denorm) but not the third — a plan where keeping a single trivial
field inline (not extracting a new model) is the correct call. This was
independently flagged by the ship-workflow coverage audit as the highest-
severity gap in this branch's test coverage.
Adds plan-eng-review-data-model-minimal-change: a synthetic plan adding one
nullable timestamp field to an existing model, no polymorphism, no JSON, no
independent query pattern or consumer. Asserts the skill does NOT reflexively
recommend extracting a separate model for it — exercising the "unless the
extra job is a single trivial field..." exception added to cognitive pattern
#12 and the SRP checklist item in 5374987c.
Also adds test/helpers/e2e-helpers.test.ts: free, deterministic unit tests
for matchesUnnegated() and setupPlanEngReviewFixture(), which were previously
only exercised indirectly by the paid, EVALS=1-gated E2E tests (also flagged
by the coverage audit — a bug in either could silently flip an eval's
pass/fail verdict and look like ordinary LLM wording variance). Along the
way, found and documented a real limitation: the negation-window check
correctly handles multiple sentences (period-bounded), but a single
comma-spliced sentence containing both a negated and a positive match can
still be absorbed into one greedy match — accepted as a known edge case,
since real review prose reliably separates points with periods/bullets.
bun test test/skill-validation.test.ts test/parity-suite.test.ts
test/touchfiles.test.ts test/skill-e2e-plan.test.ts test/helpers/e2e-helpers.test.ts
test/gen-skill-docs.test.ts — all pass, 0 fail.
Ran /review on this branch's diff against main. The always-on testing
specialist found real correctness bugs in the two eval blocks added in
33fbb58b:
- pushesToSplitTheField (legitimate-JSON eval) was negation-blind: a
CORRECT "I would NOT extract the payload into columns" answer contains
the same verb+payload+column proximity as an incorrect one, so it could
fail the test on a passing review. Also missing "normalize" from its verb
list, so a real regression phrased as "normalize this payload into
columns" would slip through undetected.
- rejectsAsUnjustified (measured-denorm eval) missed common rejection
phrasings ("instead of denormalizing", "rather than snapshotting"), and
its normali[sz]e-instead/back/it clause was loose enough to false-positive
on approving language ("stays normalized elsewhere; keep it that way").
- recommendsSeparateModel (pre-existing data-model-bias eval, touched by
this branch to fix its fixture) broke on backtick-wrapped model names, a
common LLM markdown convention.
Added matchesUnnegated() to test/helpers/e2e-helpers.ts — walks every regex
match and checks the ~20 chars before it for a negation word before counting
it as a positive signal — and used it for the split-detector. Broadened the
rejection regex and tightened its false-positive-prone clause. Stripped
markdown formatting before the backtick-affected check. Verified all four
fixes against representative sentences (correct-negative, actual-positive,
previously-missed-regression, and previously-false-positive cases) before
committing — all passed as expected.
Also extracted the maintainability specialist's DRY finding: the three
data-model E2E blocks (bias, legitimate-json, measured-denorm) each
duplicated the same ~20-line mkdtemp/git-init/write-plan/copy-skill fixture
setup. Extracted into setupPlanEngReviewFixture() in e2e-helpers.ts.
bun test test/skill-validation.test.ts test/parity-suite.test.ts
test/touchfiles.test.ts test/gen-skill-docs.test.ts test/skill-e2e-plan.test.ts
— 768 pass, 39 skip (paid E2E), 0 fail.
PR review feedback (garrytan/gstack#1071) flagged this as release metadata
that doesn't belong in a feature PR, and Copilot separately flagged that the
bump had already drifted (VERSION at 1.1.2.0, package.json still at
1.1.1.0). Per this repo's own CHANGELOG convention, the version bump and
changelog entry belong at /ship time — covering every commit that actually
landed on the branch — not mid-review while the content is still being
revised. Reverts VERSION and package.json to match main's current
1.60.1.0 and drops the premature CHANGELOG entry; /ship will add the real
one once this branch is ready to merge.
PR review feedback (garrytan/gstack#1071), resolving the three Copilot
review threads plus the "add positive and counterexample evaluations" ask:
- test/skill-validation.test.ts: renamed the two tests still named "...to
minimal diff..." to "...to right-sized diff..." to match the actual bullet
text the assertions check (Copilot review comment).
- test/skill-e2e-plan.test.ts: updated the Data-Model Bias regression test's
header comment from "minimal diff" to "right-sized diff" (Copilot review
comment); fixed the same test's fixture setup to also copy plan-eng-review's
sections/ directory — main's carve refactor moved the Architecture Review
and Data Model Review Checklist out of SKILL.md into
sections/review-sections.md after this test was written, so the sandbox
was missing the very file the skill's STOP-Read directive points at.
- Added two new periodic-tier E2E cases exercising the counterexamples the
templates now document: "plan-eng-review-data-model-legitimate-json"
(a Stripe-webhook-style JSONField the skill should NOT push back on) and
"plan-eng-review-data-model-measured-denorm" (a scoped denormalization
backed by a stated APM/load-test measurement the skill should NOT insist
on normalizing away). Both registered in test/helpers/touchfiles.ts
(periodic tier, same classification as the existing bias regression test).
Verified: bun test test/skill-validation.test.ts test/parity-suite.test.ts
test/touchfiles.test.ts test/skill-e2e-plan.test.ts — 768 pass, 39 skip
(paid E2E, requires EVALS=1), 0 fail.
PR review feedback (garrytan/gstack#1071): the data-model bias guidance was
overly absolute ("diff size is NOT a goal", "three clean models... in all
respects", "JSONField is not an escape hatch") and lacked counterexamples for
cases where JSON columns or denormalization are the right call.
- "Data model exception to right-sized diff" bullet (both skills): reframes
from an absolute rule to "don't let diff size alone decide the shape",
explicitly allows denormalization backed by a measurement, and adds the
counterexample of a field that would only ever be read through one FK join
with no independent consumer.
- JSONField bullet (both skills): keeps the core diagnostic but adds concrete
legitimate uses — third-party webhook/OAuth payload caching, open-ended
preference bags, event payloads shaped by an external producer, and
schemas still being discovered.
- plan-eng-review cognitive patterns #11 (normalize first) and #13
(structure beats blobs) get the same measured-reason / known-polymorphism
counterexamples.
- Architecture Review's "Data model honesty" bullet and the "Single
Responsibility" and "JSONField as polymorphism escape hatch" checklist
items (now living in sections/review-sections.md.tmpl per main's carve
refactor) get matching nuance.
Bumps carve-guards.ts skeleton-size budgets for plan-eng-review (69_000 ->
70_000) and plan-ceo-review (91_000 -> 92_000) to cover the added prose;
verified against test/parity-suite.test.ts (13/13 pass) and
test/skill-validation.test.ts's load-bearing-bullet guardrails (still pass —
none of the anchor phrases the static tests check were removed).
Merges origin/main (a3259400) into fix-schema-consolidation-bias. Resolves
conflicts in CHANGELOG.md, VERSION, and plan-eng-review/SKILL.md(.tmpl):
main refactored plan-eng-review's review body into an on-demand section
file (sections/review-sections.md), so the schema-normalization Data Model
Review Checklist added by this branch's fix (abafb381) was re-homed there
instead of staying inline. Bumped VERSION/package.json/CHANGELOG to
1.60.2.0 on top of main's 1.60.1.0. Updated two stale assertions in
test/skill-validation.test.ts that still checked SKILL.md directly instead
of the carved sections file, and widened two parity-suite size budgets in
test/helpers/carve-guards.ts to account for the legitimate content growth.
urlBlocklistFilter compared URLs against the exfiltration blocklist with
case-sensitive substring checks, so an uppercased sink (https://WEBHOOK.SITE/x)
bypassed the guard and reached the scoped browser agent. Normalize the page URL
and extracted content URLs to lowercase before comparing, and make URL
extraction scheme-insensitive so HTTPS:// links are still caught.
Fixes#2190
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Parses package.json overrides + every resolved basic-ftp specifier in bun.lock
(including nested paths like get-uri/basic-ftp) and fails if any is below 5.3.1.
Deterministic and offline. Fires on the pre-fix tree (basic-ftp@5.2.0) and would
also catch a direct-dependency-only bump that leaves a nested vulnerable copy.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
basic-ftp reaches the tree only transitively:
puppeteer-core > @puppeteer/browsers > proxy-agent > pac-proxy-agent > get-uri > basic-ftp
Versions <= 5.3.0 carry four HIGH advisories, all fixed in 5.3.1:
- GHSA-chqc-8p9q-pq6q (CVE-2026-39983) FTP command injection via CRLF
- GHSA-6v7q-wjvx-w8wg incomplete CRLF protection (USER/PASS + MKD bypass)
- GHSA-rpmf-866q-6p89 DoS via unbounded multiline control-response buffering
- GHSA-rp42-5vxx-qpwr DoS via unbounded memory in Client.list()
Pin via a bun `overrides` entry rather than a phantom direct dependency.
Overriding forces every basic-ftp in the tree to 5.3.1, including get-uri's
nested copy; `bun audit` then reports zero basic-ftp advisories (total 37 -> 33,
HIGH 13 -> 9). A direct-dependency bump leaves get-uri/basic-ftp at the
vulnerable version and audit still flags all four.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
endpoint-hash returns "local" for stdio/PGLite, but validators only
allowed hex suffixes, so setup-gbrain could not persist trust policy.
Co-authored-by: Cursor <cursoragent@cursor.com>
Exercise resultFromGeminiStream against current stream-json fixtures
(including empty-success hardening). Recognize GEMINI_API_KEY and map
IneligibleTierError to auth — personal OAuth free-tier is no longer
supported by gemini CLI.
Co-authored-by: Cursor <cursoragent@cursor.com>
GeminiAdapter was reading message.text and result.usage, so current CLI
content/stats events produced empty $0 success rows. Accept content with
an assistant role guard, stats token fallbacks, init model, and treat
empty exit-0 output as an error (#2159).
Co-authored-by: Cursor <cursoragent@cursor.com>
Root cause of months of silent local failure: the sandbox copied skill dirs to
the repo root, but claude >= 2.x resolves slash commands strictly from
registered skills, so /autoplan short-circuited with 'Unknown command' (0
turns, ~1s) on every attempt. Install /autoplan + review skills at
project-level .claude/skills/ (same pattern as skill-routing-e2e).
Also: the transcript filter matched entry.type === 'tool_use', a shape that
never appears at the top level of raw stream-json, so assertions only ever saw
the final result text; filter on assistant/user events instead. Hang
protection accepts the Phase 1 review dispatch (Agent/Task tool call carrying
review instructions) as progress evidence, since full Phase 1 completion is
15+ min of subagent work. Budget raised to 10 min / 40 turns.
Invisible in CI: the file is in neither evals.yml nor evals-periodic.yml
matrices (coverage decision filed in TODOS.md).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
proc.kill() only signals the sh -c wrapper; the claude child survives as an
orphan that inherited our stdout/stderr pipes, blocking the stream drain until
it exits (observed: a 600s spawn timeout stretching to 1431s and tripping bun's
per-test timeout with no result). On timeout, cancel the stdout reader; race
the stderr drain against child exit + 5s grace. Streamed transcript lines
survive the cancel, so callers still get their evidence.
Regression test: test/session-runner-timeout.test.ts (fake claude spawns a
pipe-holding orphan; fails in 30s without the fix, passes in 8s with it).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: first-run activation — project-aware scaffold, router front door, onboarding nudges
Adds the activation system that drives a new install toward a concrete first move:
- bin/gstack-first-task-detect: local-git+filesystem repo classifier emitting one
validated enum bucket (greenfield/code_<lang>/branch_ahead/dirty_default/clean_default),
portable timeouts, fail-safe empty output.
- generate-first-run-guidance.ts: unified preamble section — first-run project-aware
scaffold + returning-session plan->review->ship tip, gated on a persistent .activated
marker and never run in headless. Detection wired lazily in generate-preamble-bash.ts.
- SKILL.md.tmpl: top-level gstack skill is now a pure router (browse body removed; it
lives in /browse), routing any request and sending browser/QA work to /browse.
- setup: first-move nudge on first install. office-hours: closing handoff that launches
the next review via the Skill tool.
- telemetry-ingest: accept onboarding/first_task_scaffold_shown/handoff/route event types.
* test: cover first-run detection + repoint browse-content assertions to /browse
- New unit tests for every detection bucket, the eval-safe enum contract, and the
first-run gating (test/preamble-first-task-scaffold.test.ts); periodic E2E that runs
the detector through the real harness (test/skill-e2e-first-task-scaffold.test.ts).
- Repoint browse-content assertions (gen-skill-docs, audit-compliance, skill-validation,
LLM-judge eval) from the root skill to browse/SKILL.md following the router split;
add a regression pinning that the router carries no browse body.
- Register first-task-scaffold touchfiles + periodic tier; bump parity/carve size caps
~1-2KB per skill for the shared first-run-guidance preamble section.
- Refresh ship golden fixtures for the preamble addition.
* chore: regenerate SKILL.md + llms.txt for first-run activation
* chore: bump version and changelog (v1.58.5.0)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(test): repoint bws skillmd-* setup-block assertions to browse/SKILL.md
The skillmd-setup-discovery / -no-local-binary / -outside-git E2E tests extracted
the `## SETUP`→`## IMPORTANT` browse binary-discovery block from the root SKILL.md.
P2 moved that block to browse/SKILL.md (end anchor is now `## Core QA Patterns`),
so the slice came back empty and the `browse/dist/browse` guard failed. Repoint to
browse/SKILL.md. Verified: 7/7 e2e-browse pass locally.
* fix(test): tolerate skill-discovery race in PTY plan-mode smoke
The e2e-pty-plan-smoke suite (office-hours / plan-mode-no-op) failed in CI with
`Unknown command: /office-hours` (claude exited ~10s) while passing locally. Root
cause: a cold CI container's overlay-FS scan of the symlinked ~/.claude/skills
registry finishes AFTER the runner's 8s boot grace, so the first `/skill` send
reaches claude before the skill is indexed and is rejected as unknown. The runner
gave up on the first "Unknown command:" line.
runPlanSkillObservation now re-sends the skill command up to 3x (6s apart),
re-marking the buffer each time so stale scrollback can't re-trip the check,
before concluding the skill is genuinely unregistered. A real dangling-symlink /
missing-skill still surfaces as 'exited' (after retries), preserving the original
diagnostic. Pure-helper contract unchanged: 95/95 unit tests pass.
This is a pre-existing harness bug (fails identically on #2077's own branch, which
introduced the suite) surfaced while shipping the activation feature.
* debug(ci): temporarily instrument pty-smoke skill discovery
Capture claude version, env, registry tree, and a claude -p discovery probe to
pin why /office-hours isn't discovered in CI (retries proved it's not a race).
Temporary — revert once the registry fix is identified.
* chore: revert pty-smoke harness experiments (race-retry + CI debug step)
Diagnosis is conclusive and the experiments aren't the fix, so restore the
harness to its original state (net-zero diff vs main for both files).
What the CI debug step proved: `claude -p` returns READY — claude v2.1.187 fully
DISCOVERS /office-hours from the symlinked registry. Only the interactive PTY TUI
rejects it as "Unknown command" (and it received the full command text). So the
e2e-pty-plan-smoke failure is a claude 2.1.187 interactive-TUI regression (skills
discovered by `claude -p` aren't exposed as TUI slash commands), pre-existing in
the #2077 harness and failing identically on its own origin branch — unrelated to
this activation PR. The race-retry can't help (the TUI genuinely lacks the
command); the debug step also tripped actionlint (shellcheck SC2012). Both reverted.
* fix(ci): copy SKILL.md as real files in pty-smoke registry (cross-mount symlink)
The e2e-pty-plan-smoke suite failed with "Unknown command: /office-hours" in CI
while passing locally. Root cause (proven, not guessed): claude 2.1.187's
interactive-TUI skill scanner does not follow the /github/home -> /__w cross-mount
symlink the registry used for per-skill SKILL.md. Evidence: a CI debug step showed
`claude -p` discovered the skill (printed READY), and a local macOS repro with the
identical symlinked registry recognized /office-hours — isolating the failure to
the container's cross-mount symlink, not registration content, claude version,
duplicate names, or a race.
Fix: register the per-skill SKILL.md + sections as REAL copies (same mount as
$HOME) so the TUI reads them directly. The gstack root stays a symlink — the
preamble's runtime bash resolves bin/* and sections/* through it and bash follows
cross-mount symlinks fine.
* fix(ci): guard rm expansion in pty-smoke registry (shellcheck SC2115)
* fix(ci): also register pty-smoke skills project-scoped (cwd/.claude/skills)
The real-file user-dir registration still left the TUI rejecting /office-hours in
the container. claude's interactive TUI surfaces /slash commands from the PROJECT
dir (<cwd>/.claude/skills); the smokes run with cwd=$REPO whose .claude/skills is
gitignored (absent on a fresh CI checkout), so the user-dir registry feeds
`claude -p` (READY) but not the TUI. Populate $REPO/.claude/skills with real
SKILL.md + sections copies (no gstack symlink there — it would point at its own
parent; runtime paths use the user-dir gstack symlink).
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(gbrain): stop forcing GBRAIN_PREPARE on transaction-mode poolers (#1965)
buildGbrainEnv auto-set GBRAIN_PREPARE=true whenever DATABASE_URL targeted
port 6543, and the /sync-gbrain capability check exported it for the rest
of the skill run. Both had the semantics inverted: gbrain auto-disables
prepared statements on transaction-mode poolers because they break every
write there ("prepared statement does not exist"); GBRAIN_PREPARE=true is
gbrain's documented override for SESSION-mode poolers on 6543, not a
requirement for transaction mode. The #1435 search symptom the auto-set
worked around was fixed gbrain-side.
Remove both force-sets. A caller-set GBRAIN_PREPARE (either value) still
passes through untouched, preserving the session-mode-on-6543 escape hatch.
isTransactionModePooler stays exported.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(gbrain): classify probe timeout as its own status; sync proceeds instead of skipping (#1964)
The 5s engine probe misclassified healthy-but-slow engines (cold Supabase
pooler connections measured at 6.9-10.7s) as broken-config, so /sync-gbrain
silently skipped code+memory and told the user their config was malformed.
- New "timeout" status: probe killed at the deadline with no recognized
stderr pattern. Default deadline is now 15s, overridable via
GSTACK_GBRAIN_PROBE_TIMEOUT_MS (tests set 300ms against a fake that
sleeps 2s).
- Sync stages PROCEED on timeout with a stderr warning naming the env knob;
a genuinely-dead engine surfaces its real error at the first operation
instead of a false config diagnosis.
- Consistency everywhere "ok" gated behavior: gstack-gbrain-detect --is-ok
exits 0 on timeout, and gen-skill-docs' detection gate accepts it, so a
slow engine no longer silently suppresses brain-aware features.
- Status cache: key now includes the effective probe timeout (raising it
invalidates a cached timeout) and GBRAIN_HOME; config detection honors
GBRAIN_HOME so relocated-home users stop being misclassified as
missing-config.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(bins): cygpath-normalize SCRIPT_DIR for bun imports; surface learnings-log errors (#1950)
Under Windows git-bash, pwd yields a POSIX path (/c/Users/...) that Bun on
Windows cannot resolve as an ES module specifier. gstack-learnings-log
interpolates SCRIPT_DIR into a bun -e import, so every invocation died with
"Cannot find module" — and 2>/dev/null swallowed the error, silently
dropping every AI-logged learning for Windows users.
- 3-line cygpath -m guard in gstack-learnings-log and gstack-question-log
(which gains the same import shape in the next commit). Matches the
duplicated IS_WINDOWS convention in setup; no shared shell lib exists.
- learnings-log adopts question-log's set +e / TMPERR capture pattern
wholesale: validation errors now print to stderr. The old
`if [ $? -ne 0 ]` check was dead code under set -euo pipefail — the
script exited at the failing assignment before reaching it.
- New test/bin-windows-bun-import-paths.test.ts: static invariant (any
bash bin interpolating $SCRIPT_DIR into a bun -e import must carry the
guard) + behavioral end-to-end run invoked via `bash <bin>` — added to
the windows-free-tests workflow list so the conversion is proven on the
only platform where the bug exists.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(question-log): dedupe INJECTION_PATTERNS via lib/jsonl-store (#1934)
bin/gstack-question-log carried a local copy of the injection-pattern list,
so pattern fixes to lib/jsonl-store.ts never propagated — including the
/override[:\s]/i false-positive fix arriving via community PR #1940.
Import the shared hasInjection instead (enabled by the previous commit's
cygpath guard). question-log also gets the lib's stricter superset
(human:, disregard, from-now-on, approve-all patterns).
Tests pin the contract in a #1940-order-independent way: an "Override:
ignore all previous instructions" header is rejected, "prose overrides the
deterministic table" is accepted, and a static invariant keeps local
INJECTION_PATTERNS duplicates out of the bin.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(security): community-pulse + both dashboards never report fake zeros (#1947)
The security-signaling surface failed open at three layers — every failure
mode read as a reassuring "0 attacks" / "0 installs":
- community-pulse edge function: supabase-js returns {data,error} without
throwing, and all five queries discarded `error` — a DB outage produced
real-looking zeros via the SUCCESS path, and the catch (also returning
zeros with HTTP 200) was unreachable for query failures. Every query now
destructures and throws; the catch serves the stale cache (marked
"stale": true) when one exists, else 503 {"error":"pulse_unavailable"}.
Success responses carry "status":"ok" so clients can distinguish
authoritative data from legacy backends. NOTE: the edge function deploys
out-of-band (supabase functions deploy community-pulse).
- gstack-security-dashboard: captures the HTTP status; non-200 / network
failure / error body / missing section → "unknown — backend error";
jq missing → "unknown — install jq" (the lossy grep fallback broke on
nested arrays and under-reported attacks as zero — removed); a 200
without the new marker shows figures with an "unverified (legacy
backend)" note. Also fixes a latent display bug: the TOTAL grep matched
the digit 7 inside "attacks_last_7_days" and misreported every count.
- gstack-community-dashboard: same class — curl || echo "{}" plus
grep || echo "0" printed "Weekly active installs: 0" on any failure.
Now "unknown — backend error (HTTP N)".
test/security-dashboard-fallback.test.ts pins the matrix (200+marker,
200-legacy, 503, network failure) x (jq present, jq absent) for both bins:
"unknown" states never render as 0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(telemetry): redact error_message spans before they leave the machine (#1947)
error_message was uploaded with only quote/newline escaping — stack traces
and failed-API errors can embed credentials, private paths, and hostnames,
and the sync path strips only _repo_slug/_branch.
New lib/redact-engine.ts export redactFindingSpans(): replaces EVERY
finding's span with <REDACTED-{id}> regardless of tier (applyRedactions is
the interactive PII-only path and exits nonzero on credential findings, so
it can't serve machine egress). Returns null when a span can't be located —
callers drop the whole payload rather than risk a leak.
gstack-telemetry-log pipes error_message through it at LOG time, so the
local JSONL at rest is clean too; surrounding text survives for crash
triage. FAIL CLOSED: bun missing, engine error, or non-JSON-string output
all null the field. Tests pin: embedded ghp_ token → <REDACTED-github.pat>
with context intact; redactor unavailable → null; raw bytes on disk never
contain the token.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(redact): prepush guard fails closed on git failure; /ship owns hook install (#1946)
Two gaps closed:
1. Fail closed. The git() helper returned "" on ANY non-zero exit or
maxBuffer overflow (status null), addedLinesFor produced an empty
string, and the push sailed through unscanned — fail-open on exactly
the oversized-diff case where a large secret-bearing blob is most
likely. The diff call now uses a strict variant that throws; main
blocks with a clear message naming the GSTACK_REDACT_PREPUSH=skip
escape valve. Probe calls (symbolic-ref, rev-parse, merge-base) keep
the permissive helper — their failures are normal control flow.
2. Install path. The hook was installed by nothing ("opt-in, installed by
nothing" was the issue's words). ./setup runs in the gstack checkout —
the wrong repo for a per-project hook — so it gets a one-line hint
only. /ship owns per-repo install: config redact_prepush_hook=true +
hook missing → silent install (consent already given); config unset +
no ~/.gstack/.redact-prepush-prompted marker → one-time machine-wide
AskUserQuestion offer, answer persisted. ship/SKILL.md regenerated in
this same commit (check-freshness bisect discipline).
Tests: unscannable diff (bogus SHAs) → exit 1 + valve named; empty-but-
successful diff → exit 0; static asserts pin setup as hint-only and the
ship template as the installer surface.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(redact): six new credential patterns — GitLab, HuggingFace, npm, DigitalOcean, Bearer, GCP SA (#1946)
Coverage gaps from the #1946 security review, including token types for
tooling gstack itself drives (glab):
HIGH (block): gitlab.token (glpat-/glptt-/gldt-), huggingface.token (hf_),
npm.token (npm_), digitalocean.token (dop_v1_), gcp.service_account (the
JSON-escaped "private_key" form that dodges pem.private_key's literal-block
match when minified, confirmed by "private_key_id" proximity).
MEDIUM (warn): auth.bearer — the most FP-prone shape in the set (docs are
full of "Authorization: Bearer <token>"), so it requires header-context
proximity and the same entropy>=3.0 + placeholder validator recipe as
env.kv. "Bearer YOUR_TOKEN_HERE" never fires; calibration over coverage,
per the cries-wolf principle.
All shapes are linear-time; test/redact-pattern-lint.test.ts covers them
automatically. Engine tests add positive + placeholder-negative cases per
pattern.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: coverage-audit additions for the fix wave
Ship Step 7 gap-fill (all passing, 248 tests across the touched suites):
memory + dream stage probe-timeout proceeds, gbrain-detect override paths,
stale-flag passthrough, 200-body-missing-.security fail-closed case,
telemetry redaction edges, and credential-pattern edge cases.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: pre-landing review fixes
Review army findings (1 critical, auto-fixed with regression tests):
- CRITICAL (security specialist, verified live): redactFindingSpans spliced
only the regex capture span, and pem.private_key / gcp.service_account
capture just the BEGIN-header — the key body survived "redaction" and
shipped via telemetry. Marker-only patterns now drop the whole payload
(null, fail closed). Overlapping spans (Bearer+JWT on the same bytes) are
coalesced before splicing so stale offsets can't leave partial secret
bytes behind.
- gitStrict: drop the dead `|| r.status === null` disjunct (null !== 0
already covers it); add the signal-kill/null-status regression test the
docstring promised.
- security-dashboard human mode flags stale snapshots ("figures may be out
of date") instead of presenting frozen counts as current.
- community-dashboard marker check uses jq when available — the grep-only
variant misclassified whitespaced/reserialized bodies as legacy.
- telemetry fail-closed test now shadows bun with a failing stub
(deterministic on any host layout); stale "five status cases" describe
title renamed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: adversarial review fixes (Claude + Codex cross-model passes)
Both adversarial passes ran against the wave; every FIXABLE finding landed
with a regression test:
- probeTimeoutMs clamps to >=1ms: a fractional override floored to 0, and
execFileSync treats timeout:0 as NO timeout — the probe that exists to
bound hangs could hang forever (found by both models independently).
- /ship silent hook install now requires the hooks dir to live inside
.git: with core.hooksPath (husky's COMMITTED .husky/), the chaining
installer would have renamed the team's committed pre-push and written a
machine-local wrapper into the working tree (found by both models).
- gstack-config gbrain-refresh accepts the "timeout" status — the last
consumer still gating on literal "ok" (Codex); gstack-gbrain-detect's
config-derived fields honor GBRAIN_HOME so the detection JSON can't
report status ok alongside config_exists false (Codex).
- prepush: a remote sha absent locally (shallow clone / stale fetch) falls
back to the merge-base/empty-tree range — scans MORE, never blocks a
legitimate push into training users toward --no-verify.
- dashboards: curl's own 000 no longer doubles to "HTTP 000000"; the
community dashboard flags stale snapshots like the security one; array
sections parse via jq (the sed/grep loops truncated at the first ']');
the no-jq marker grep tolerates whitespace.
- telemetry: multi-line redactor output nulls the field instead of
corrupting the JSONL record; setup's hint fires only when the config key
is genuinely unset (an explicit false is a recorded decline); the /ship
prompt marker honors GSTACK_HOME.
Kept as designed (cross-model tension noted): Bearer stays MEDIUM in the
prepush gate — a HIGH Bearer would block every docs example; the entropy
validator can't eliminate that FP class, and MEDIUM warns visibly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore: bump version and changelog (v1.57.11.0)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs: P1 TODO — eval harness live progress + incremental persistence
Root-caused during this ship: a killed eval run was indistinguishable from a
healthy one for hours (per-file output buffering across mega test files, no
incremental eval-store writes, no honest liveness signal). Full context and
starting points in the entry.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: fix operational-learning E2E fixture — copy lib/jsonl-store.ts
Pre-existing breakage, proven on main: gstack-learnings-log has imported
lib/jsonl-store.ts (shared injection patterns) since v1.57.5.0 / #1910, but
the fixture copies only the bin scripts — the bin exits 1 before writing
anything, on main silently (stderr swallowed) and on this branch loudly
(the #1950 error-surfacing made the four-day-old failure visible). A real
install always ships bin/ and lib/ together; the fixture now does too.
Verified: the fixture-shaped invocation writes the learning (exit 0) with
lib present, exits 1 on both main and this branch without it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ios-qa): isolate E2E tests under --concurrent (3 real races)
The ios-qa E2E file failed intermittently under `bun test --concurrent`
(the eval harness default). Three distinct shared-state races, all fixed:
1. Shared pidfile: a module-level `workDir` reassigned in beforeEach was
clobbered by parallel tests, so concurrent daemons collided on the same
pidfile and the loser returned `already_running`. Each test now gets its
own dir via makeWorkDir().
2. process.env path globals: tests set GSTACK_IOS_AUDIT_PATH /
_ATTEMPTS_PATH / _ALLOWLIST_PATH on the shared process env; concurrent
tests stomped each other's audit/attempts destinations. Threaded
auditPath/attemptsPath/allowlistPath through DaemonOptions (and
mintForCaller) as explicit args — env is no longer load-bearing.
3. afterEach cleanup race: the per-test cleanup drained a shared dir array,
so the first test to finish deleted still-running tests' workDirs
mid-assertion. Moved to afterAll (cleans once, after all settle).
Verified: 5/5 clean full-suite runs at --max-concurrency 15 (was
intermittent); daemon unit suite 91/91; daemon source compiles. The paths
default to the env-derived locations when options are omitted, so the
production CLI path is unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(pty): pin spawned claude to EVALS model chain (default claude-sonnet-4-6)
launchClaudePty spawned the interactive `claude` TUI with no --model flag, so
the child inherited the operator's ~/.claude/settings.json model. On a
slow-thinking model that meant 5+ min of extended thinking on empty plan-mode
context, timing out the plan-mode smoke tests regardless of contention. Pin the
model via opts.model ?? EVALS_MODEL ?? 'claude-sonnet-4-6' — byte-identical to
session-runner.ts:144, so PTY and `claude -p` evals always agree.
Pushed before extraArgs (last flag wins, so a per-test --model still overrides).
Placement leaves the spawn region byte-stable for a clean merge with the
in-flight hermetic-env branch. Plumbed model through the three plan-skill
wrappers. Static-grep tripwires guard the pin, its fallback chain, the
before-extraArgs ordering, and all three wrapper forwards.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(pty): detect markdown bold-bullet prose AUQs (fixes office-hours smoke)
office-hours auto-mode renders its mode question as `- **Building a startup**`
markdown bullets (office-hours/SKILL.md.tmpl:102) with no letter/number marker.
isProseAUQVisible only matched `A)`-style lettered or `1.`-style numbered
options, so the question went undetected: the model surfaced it at ~2m19s
(well under the 300s budget) but the harness kept scoring the run "working"
off the spinner glyphs and timed out — a false timeout on a question that was
already on screen.
Add Pattern 3: when an interrogative line ('?') is present AND 3+ bold-bullet
markers (`- **`) appear in the 4KB tail, classify as a prose AUQ. Bold is the
discriminator vs incidental prose bullets; the line anchor is dropped (stripAnsi
can collapse option lines) and the existing `❯ 1.` cursor gate still defers to a
live native list. Wires through the existing classifyVisible 'asked' path and the
timeout high-water-mark, so office-hours now classifies 'asked' instead of
'timeout'. Five unit cases: the office-hours render passes; no-'?', <3-bullet,
plain-bullet, and native-cursor cases stay false.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(pty): detect stripAnsi-collapsed prose AUQs + judge spinner-precedence
The plan-eng/plan-design plan-mode + finding-floor smokes timed out even when
the skill HAD rendered a complete prose AskUserQuestion and was waiting: the PTY
strips cursor-positioning escapes, collapsing the option newlines/spaces so
"A) ..." arrives as "A(recommended)" / "-B:" and "Reply with A, B, or C" as
"ReplywithA,B,orC". Every line-anchored detector (Patterns 1-3) returns false on
those bytes, so proseAUQEverObserved never latched and the run timed out on a
question that was already on screen.
Add Pattern 4/5: a two-signal collapsed-form detector — a reply/recommendation
marker (space-insensitive "reply with [A-D]", "Recommendation:", or
"(recommended)") AND 2+ distinct A-D letters each punctuated by ) : or (. The
conjunction is what separates a real AUQ from incidental report prose; verified
true on the verbatim failing-run buffers where Patterns 1-3 return false.
Also fix the Haiku judge spinner bias: of 614 verdicts, 569 were 'working' and
95 of those noted a question was visible — Claude Code keeps the spinner
animating at an idle prose decision, so the judge coin-flipped. Add a precedence
override: when an option list AND a Recommendation/Reply instruction are both
visible, classify WAITING even with spinner glyphs. Kept the strict dual-signal
gate (never option-list-alone) so auto-decide-preserved doesn't flip.
5 unit tests pin the two-signal contract (2 true on real collapsed bytes, 3
false guards). 90 -> 95 pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(plan-review): ask-first scope gate for plan-eng + plan-design review
On an empty/cold invocation, plan-eng-review and plan-design-review would dive
straight into repo exploration (plan-eng) or a 7-pass mockup+audit (plan-design)
and only ask the user much later, if at all. plan-ceo-review already asks first
via an unconditional Step-0 gate and behaves well; these two did not.
Add a hard-STOP scope gate as the FIRST operational instruction in each skill
(above the design-doc check / pre-review audit / mockup defaults it explicitly
overrides): the first tool call must be AskUserQuestion confirming the review
target, before any git/Read/Grep/Glob/Bash or mockup generation. Under
--disallowedTools the options render as plain column-0 lettered prose with a
Recommendation + "Reply with A, B, or C" line so the answer is detectable.
This is correct cold-start UX (confirm what to review before grinding a full
review on nothing) and it is the product half of the plan-mode smoke fix; the
harness collapsed-form detector is the deterministic half that catches the ask
however it renders. Templates + regenerated SKILL.md (default variant).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(tiers): reclassify stochastic plan-eng/plan-design ask-first smokes as periodic
plan-eng-review and plan-design-review run a long explore/audit before their
first AskUserQuestion, so whether the plan-mode + finding-floor smokes reach a
terminal outcome within the 300s/600s budget depends on stochastic ask-first
compliance (measured ~50-67%/run even with the hardened gate). Per the
"non-deterministic -> periodic" tiering rule, move the four affected smokes
(plan-eng/plan-design review-plan-mode + finding-floor) to periodic.
The deterministic harness fix (collapsed-form detector + judge precedence) and
the ask-first gate lift these from always-failing to mostly-passing and are the
real product+harness improvements; periodic monitoring tracks the rate weekly
without blocking PRs on an LLM coin-flip. plan-ceo/plan-devex ask-first reliably
and stay gate-tier.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* ci(evals): gate the deterministic PTY plan-mode smokes in CI
The real-PTY plan-mode smokes never ran in CI — the gate was local-only. Add an
e2e-pty-plan-smoke matrix suite running the two deterministically-reliable ones
(office-hours-auto-mode, plan-mode-no-op) so a regression there blocks PRs. The
stochastic plan-eng/plan-design ask-first smokes stay periodic (touchfiles
E2E_TIERS) and are not CI-gated.
A fresh CI container has no ~/.claude.json, so the spawned interactive `claude`
would wedge on the onboarding + API-key-approval dialog. Add a scoped seed step
(hasCompletedOnboarding + key approval, its own ANTHROPIC_API_KEY env) before the
run — mirrors what the hermetic E2E child env seeds. Per-suite timeout override
(35 min) via matrix.suite.timeout so the PTY suite has headroom for --retry 2
without bumping the other 12 suites. Report runner count 12 -> 13.
Validate via workflow_dispatch before relying on the gate (PTY-in-CI is new).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* ci(evals): install gstack skill registry for the PTY smoke suite
The first dry-run of e2e-pty-plan-smoke failed: the spawned interactive `claude`
printed "Unknown command: /plan-ceo-review". .claude/skills is gitignored, so a
fresh CI checkout has no gstack skill registry and the TUI can't resolve
/office-hours or /plan-ceo-review.
Add a Register step (scoped to the suite, after Seed, before Run) that mirrors
setup's --no-prefix user-scoped registry minimally: $HOME/.claude/skills/gstack
-> repo (resolves the preambles' absolute ~/.claude/skills/gstack/bin/* and
<skill>/sections/* paths) + per-skill SKILL.md/sections symlinks for the two
skills these tests invoke. HOME is /github/home in this container and the runner
adds no HOME/CLAUDE_CONFIG_DIR override (no hermetic mode), so $HOME is the right
anchor — the Seed step already proved claude reads it. No ./setup (binary build
+ Chromium + fonts + /dev/tty prompt); SKILL.md + bin/ + sections/ are committed.
Self-validating: fails the step loudly on a dangling symlink or missing
`name:` frontmatter, so a moved target surfaces here instead of as a silent
35-min "Unknown command" timeout.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* chore: bump version and changelog (v1.58.4.0)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* feat: Layer C stealth — chrome.*, Notification, per-install hardware, toString Proxy (gbrowser T1+T3+D6)
Three additions stacked into the existing applyStealth() init script
to close the visible automation tells that today push GBrowser users
into Google's /sorry/index captcha and similar:
T1 — Strip Playwright's automation default args:
--enable-automation (kills "Chrome is being
controlled" infobar)
--disable-popup-blocking, --disable-component-update,
--disable-default-apps (Patchright's list — each
is a documented tell)
Now centralized in STEALTH_IGNORE_DEFAULT_ARGS export, used by BOTH
launchHeaded() and handoff() (the headless → headed re-launch path).
D6 — Drop "GStackBrowser" UA branding suffix:
Real Chrome's UA ends `Safari/537.36`, not `Safari/537.36 GStackBrowser`.
The branded suffix was a high-entropy classifier for any vendor that
grep'd UA for known automation/test-browser strings. Branding still
lives in the wrapper .app name + Dock icon + tray — does not need
to leak via the UA string for the product to be "GBrowser." Resolves
the "looks like Chrome but identifies as GStackBrowser" contradiction
codex review #18 flagged.
T3 — Layer C init-script additions in stealth.ts:
1. Function.prototype.toString Proxy (must run first). Wraps every
patched getter / function in a WeakSet so they report
`function NAME() { [native code] }` at every recursion depth,
defeating the depth-3+ integrity check
(fn.toString.toString.toString().includes('[native code]')).
2. window.chrome.runtime / chrome.app / chrome.csi / chrome.loadTimes
restoration with full enum shape (OnInstalledReason, PlatformArch,
PlatformOs, etc.) + method bodies. Real Chrome ships these; their
absence is universally checked. Vendor research (gbrowser plan
deep-dive on Cloudflare + DataDome) confirmed both vendors probe
this shape directly.
3. Notification.permission aligned to 'default'. The existing inline
addInitScript already spoofs permissions.query({name:'notifications'})
to return 'prompt' — Notification.permission being 'denied' while
Permissions returns 'prompt' is a cross-source inconsistency that
detectors flag specifically.
4. Per-install hardware values via GSTACK_HW_CONCURRENCY /
GSTACK_DEVICE_MEMORY env vars (set by gbd's host_profile.go from
system_profiler + sysctl). Reporting real host values within the
Chrome shape avoids the cross-user GBrowser fingerprint cluster
that hardcoded defaults would create. Codex review #10 flagged
hardcoding as creating contradictions across Apple Silicon / Intel
/ UA-CH architecture.
5. Selenium 25-global cleanup + PhantomJS + NightmareJS + Watir +
Playwright (__pwInitScripts, __playwright__binding__) static-name
deletion. The inline block continues to handle the dynamic
cdc_/__webdriver/__selenium/__driver prefixes.
D7 (codex correction) kept: still do NOT fake navigator.plugins or
navigator.languages. Synthesizing those triggers MORE consistency
flags from modern fingerprinters than letting Chromium surface them
natively.
Test coverage:
- 15 new tests in stealth-layer-c.test.ts covering: launch-flag
exports, script structure, toString-Proxy installs first, every
spoof present, hardware values interpolated from input (not
hardcoded), Selenium global cleanup spot-check, no GStackBrowser
leak in stealth payload, backwards-compat exports preserved.
- All 8 existing stealth-webdriver tests still pass.
- All 2 existing browser-manager-unit tests still pass.
For GBrowser specifically: this is the gstack-side half of Phase 1 / T1
+ T3 + D6 in the anti-detection plan. The gbrowser repo's submodule
pointer bump will land alongside this.
* feat: buildGStackLaunchArgs — Pack 1 cmdline-switch construction for gbrowser
New stealth.ts export that turns the GSTACK_* env vars (already populated
by gbrowser's gbd from host_profile.go) into the --gstack-* cmdline
switches the Pack 1 Chromium patches read at WebGL getParameter,
NavigatorUA::userAgentData, NavigatorConcurrentHardware::hardwareConcurrency,
and NavigatorDeviceMemory::deviceMemory time.
Wired into all three launchArgs sites: launch() (headless), launchHeaded()
(real product path), and handoff() (headless → headed re-launch).
Mapping:
GSTACK_GPU_VENDOR → --gstack-gpu-vendor
GSTACK_GPU_RENDERER → --gstack-gpu-renderer
GSTACK_PLATFORM → --gstack-ua-platform (with mapping:
MacARM/MacIntel → macOS, Win32 → Windows,
Linux x86_64 → Linux)
GSTACK_GPU_CHIPSET → --gstack-ua-model
GSTACK_HW_CONCURRENCY → --gstack-hw-concurrency
GSTACK_DEVICE_MEMORY → --gstack-device-memory
Each switch is emitted only when its env var is non-empty — empty
values fall through to the patch's "no override" path, which returns
the real Chromium native value. Safe to ship on Chromium builds
without the Pack 1 patches applied (zero behavior change).
The patches themselves live in the gbrowser repo at chromium/patches/
{webgl-vendor-spoof,ua-client-hints-stealth,worker-navigator-stealth}.patch.
Both halves (gstack arg construction + gbrowser C++ patches) must
land + Chromium rebuild before the spoof reaches the WebGL/UA-CH/
hardware accessors. Currently dormant until then.
Tests (browse/test/stealth-layer-c.test.ts):
7 new buildGStackLaunchArgs cases — empty env, all-populated, partial,
platform mapping (MacARM/MacIntel/Win32/Linux), unrecognized platform
fallthrough, vendor-with-spaces escape-safety.
All 32 stealth/browser-manager tests pass.
For GBrowser specifically: gstack-side half of the Pack 1 flag plumbing.
gbrowser repo will bump the submodule pointer to this commit, then re-run
bun run test/anti-bot/evidence-run.ts to verify creepjs's "33% headless"
score drops after Pack 1 + Chromium rebuild.
* feat: buildGStackLaunchArgs adds --gstack-suppress-prepare-stack-trace
Pack 2 / B11 flag plumbing for the new
error-preparestacktrace-stealth.patch in gbrowser/chromium/patches/.
Always emit --gstack-suppress-prepare-stack-trace unless the caller
explicitly sets GSTACK_CDP_STEALTH=off in the environment. Off by
default in patch behavior (no-op without the C++ patch), so this is
safe on stock Playwright Chromium too.
Closes the Cloudflare canary trick where a page sets
Error.prepareStackTrace and watches for it to fire during CDP
serialization of a logged Error object.
Tests:
All 33 stealth/browser-manager tests pass. New cases:
- GSTACK_CDP_STEALTH=off disables suppression
- empty env still emits the always-on flag (count=1)
- all-populated env now emits 7 flags (was 6)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(browse): enable Chromium sandbox on headed launchPersistentContext
Mirrors v1.40.0.1 from main lineage (PR #1617). Cherry-picked onto
gbrowser-anti-detection so the GBrowser submodule can consume the fix
without waiting for main to merge.
Playwright auto-adds --no-sandbox whenever chromiumSandbox !== true
(playwright-core/lib/server/chromium/chromium.js:291-292). The headless
chromium.launch() site set the option; the two headed sites
(launchHeaded() and handoff()) did not. Every headed launch on macOS
and Linux showed Chromium's yellow "unsupported command-line flag:
--no-sandbox" infobar.
shouldEnableChromiumSandbox() centralizes the Win32 / CI / CONTAINER /
root heuristic that previously lived only in the headless path's
explicit --no-sandbox push at :225. All three launch sites now use the
helper, and six unit tests pin the policy across darwin, linux, win32,
CI, CONTAINER, and root.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* v1.40.0.2 fix(browse): Cmd+Q on managed Chromium stops triggering supervisor respawn
Three browser.on('disconnected') handlers in browse/src/browser-manager.ts
(launch, launchHeaded, handoff) each exited with a non-zero code on every
disconnect, regardless of cause. Process supervisors that consume our exit
code (gbrowser's gbd HealthMonitor in cmd/gbd/health.go) treated user
Cmd+Q identical to a Chromium crash and respawned with exponential
backoff, so the visible browser kept reappearing after the user closed it.
Add resolveDisconnectCause(browser) that reads the underlying ChildProcess
exitCode + signalCode (waiting up to 1s for the exit event if the
disconnected event fired first). Exit code 0 + no signal = clean user
quit; anything else = crash, signal-kill, or OOM.
Wire the resolver into all three disconnect handlers:
- launch() (headless): clean → exit 0, crash → exit 1 (was always 1)
- launchHeaded() (headed): clean → exit 0, crash → exit 2 (was always 2)
onDisconnect() cleanup callback still runs in both cases.
- handoff() (re-launch): same as launch() via the helper.
Preserve the per-path crash codes (1 vs 2) so any supervisor that
differentiated headed vs headless crashes keeps working.
Seven new unit tests in browse-manager-unit.test.ts cover the resolver
across already-exited, signal-killed (SIGSEGV / SIGKILL), async exits,
and null-browser inputs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(browse): apply stealth on every launch path + share automation-artifact cleanup
handoff() built cmdline args but never called applyStealth, so a handed-off
browser had no JS stealth (no webdriver mask, no chrome.* shape, no toString
proxy). And the cdc_/Permissions cleanup shim lived inline in launchHeaded()
only, so headless launch() reported Notification.permission='default' without
the matching permissions.query='prompt' answer — the exact cross-source
inconsistency the shim exists to prevent.
Move the cleanup into AUTOMATION_ARTIFACT_CLEANUP_SCRIPT inside applyStealth so
all three launch paths (launch, launchHeaded, handoff) get identical stealth,
and call applyStealth(newContext) in handoff() before restoreState() navigates.
A static tripwire in browser-manager-unit.test.ts fails CI if any launch path
drops the applyStealth call again.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(browse): make --gstack-suppress-prepare-stack-trace opt-in, not default-on
buildGStackLaunchArgs() pushed the flag unless GSTACK_CDP_STEALTH=off, i.e.
on-by-default — contradicting its own comment ("off by default, only for
gbrowser builds"). The switch is read by a C++ patch that only exists in
gbrowser; on stock Playwright Chromium it is an unknown switch.
Flip to opt-in: emit only when GSTACK_CDP_STEALTH is on/1/true. gbd opts in by
exporting GSTACK_CDP_STEALTH=on; stock installs leave it unset so the flag
never reaches a Chromium that wouldn't understand it. Comment now matches code.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* docs(browse): correct stale stealth comments
The file-level stealth.ts docstring claimed "we DON'T fake navigator.plugins"
while the same file now ships EXTENDED_STEALTH_SCRIPT, which does fake plugins
when GSTACK_STEALTH=extended. Clarify that Layer C (the always-on default)
doesn't fake plugins and the opt-in extended mode does, as the documented
"actively lies, may break sites" escape hatch.
Also fix the launch()/launchHeaded() comments that said "mask navigator.webdriver
only" — applyStealth (Layer C) also restores window.chrome.*, aligns
Notification.permission, and sets per-install hardware.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(browse): runtime + extended-mode coverage for the stealth blend
The stealth tests were all static string-shape assertions; nothing executed
the script in a real page. Add real-Chromium runtime checks via applyStealth +
page.evaluate:
- Layer C runtime: window.chrome.* rich shape, Notification.permission='default'
paired with permissions.query notifications='prompt' (guards the shim now
running on every path), and patched getters reporting [native code].
- Per-install hardware: navigator.hardwareConcurrency/deviceMemory reflect the
GSTACK_* env profile.
- Extended-mode blend: navigator.plugins is faked when GSTACK_STEALTH=extended,
Layer C still wins window.chrome.runtime, and navigator.webdriver stays false
(own-prop getter survives extended's prototype delete).
- Persistent-context (launchHeaded/handoff) parity now uses a page created
AFTER applyStealth — the old test checked pages()[0], which predates the init
script, so webdriver was false only via the launch arg, not Layer C.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(browse): handoff() + launchHeaded() spread the shared STEALTH_LAUNCH_ARGS
handoff() built its launch args from only ['--hide-crash-restore-bubble',
...buildGStackLaunchArgs()], omitting STEALTH_LAUNCH_ARGS — so a handed-off
browser kept the --disable-blink-features=AutomationControlled tell that
launch() and launchHeaded() strip. launchHeaded() also hardcoded the flag as a
literal. Both now spread the shared constant, so the AutomationControlled flag
lives in one place across all three launch paths.
Tripwires: STEALTH_LAUNCH_ARGS spread into >= 3 sites (no inline literal) and
STEALTH_IGNORE_DEFAULT_ARGS wired into both persistent-context paths.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* refactor(browse): drop dead HostProfile.platform, export test internals
HostProfile.platform was set by readHostProfile but never read by
buildStealthScript — the platform spoof is owned by the UA-CH cmdline switch in
buildGStackLaunchArgs (which reads GSTACK_PLATFORM directly). Remove the dead
field. Export readHostProfile and AUTOMATION_ARTIFACT_CLEANUP_SCRIPT so their
clamp/shape invariants can be unit-tested. Correct the stale "25 Selenium
globals" count comment and note the extended cdc_ scan is redundant-but-retained
for standalone use.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(browse): cover readHostProfile clamp, toString depth-3, chrome.* calls
Pre-landing review coverage gaps:
- readHostProfile clamps 0/negative/NaN/missing env to 8 (a deviceMemory=0 or
NaN would be a glaring bot tell) — now asserted.
- toString proxy survives the depth-3 recursion trick
(fn.toString.toString.toString().includes('[native code]')), the headline
claim that was only tested at depth-1.
- chrome.csi() and chrome.loadTimes() are invoked (not just typeof-checked) and
runtime.connect() throws the native-shaped "No matching signature" error.
- AUTOMATION_ARTIFACT_CLEANUP_SCRIPT static shape (cdc_/__webdriver strip +
notifications->prompt) as a hermetic backup for the live-Chromium pairing test.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(browse): recreateContext() re-applies stealth (closes 4th un-stealth path)
useragent and viewport --scale route through recreateContext(), which rebuilds
the BrowserContext via newContext() — a fresh context with no init scripts. It
never called applyStealth, so a routine useragent/viewport-scale command
silently dropped webdriver masking, window.chrome.* shape, hardware spoof, and
the cdc/Permissions cleanup on every restored page. Caught by the cross-model
adversarial review (Codex) after the Claude pass and eng review missed it.
Both the main and fallback paths now call applyStealth before any page is
created. The launch-path tripwire is raised to >= 4 sites and now asserts the
recreateContext() body specifically, so the regression class can't recur.
Also documents the load-bearing trust assumption on buildGStackLaunchArgs /
readHostProfile (GSTACK_* must be gbd-sourced, never page/remote data — the
injection-safety argument depends on it) and the notifications-permission
spoof tradeoff.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* chore: bump version and changelog (v1.58.3.0)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* docs: sync browser stealth docs to Layer C (v1.58.3.0)
BROWSER.md "Stealth scope" still described the default as navigator.webdriver
masking only; Layer C is now the always-on default across all four
context-creation paths. Update the stealth-scope prose, the "What GStack
Browser means" blurb (stock-Chrome UA, no GStackBrowser suffix, captchas can
still get through at the CDP layer), the stealth.ts source-map line, and the
env-vars table (GSTACK_STEALTH, GSTACK_CDP_STEALTH, GSTACK_GPU_*, GSTACK_PLATFORM,
GSTACK_HW_CONCURRENCY/GSTACK_DEVICE_MEMORY + the explicit --gstack-* switches and
ignoreDefaultArgs stripping). Correct the stale "narrows to navigator.webdriver
masking only" premise on the open CDP-patch TODO (the TODO itself stays open —
the CDP-protocol layer is still unaddressed).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
David Grant
Joshua France
t
Sinabina
Sina
Garry Tan