mirror of https://github.com/garrytan/gstack.git
3 Commits
| Author | SHA1 | Message | Date |
|---|---|---|---|
Garry Tan
|
45cc95d5f4
|
v1.57.5.0 feat: cross-session decision memory + gbrain dream-stage call graph (#1910)
* feat(gbrain-sync): add cycleCompleted() cycle-state probe
Reads `gbrain doctor` cycle_freshness to classify whether a source has
completed a full cycle (completed/never/unknown). A fail naming this source
-> never; a fail naming only other sources -> completed; an absent or
unparseable check -> unknown, so an unrelated doctor failure never masks a
real state. Gates the automatic call-graph build on --full.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(gbrain-sync): --dream call-graph stage with lock-free gate + honest outcome guard
Adds a source-scoped `gbrain dream --source <id>` stage that builds this
worktree's call graph (code-callers/code-callees). Runs lock-free after the
sync lock releases so it never blocks sibling worktrees; a .dream-in-progress
marker dedupes concurrent dreams. --full auto-runs it only when the cycle was
never built; explicit --dream always forces; --no-dream opts out.
The stage parses the cycle's own output and reports the truth, not a flat
"built": a WARN when the schema pack can't extract code symbols, when the
embed phase failed for a missing key, or when 0 edges resolved; OK with the
resolved-edge count otherwise. gbrain exits 0 even when it skips on a held
cycle lock (e.g. autopilot), so that case reports SKIP, not success.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* chore: ignore gbrain .sources/ local staging dir
gbrain writes per-source staging and capability-check artifacts under
.sources/ in the repo root. It's machine-local runtime state, not source.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* docs(gbrain): honest call-graph guidance in /sync-gbrain + pin works on gbrain>=0.41.38
sync-gbrain frames the --dream offer honestly: building a call graph requires a
code-aware schema pack, and the dream stage reports a WARN when it can't. The
verdict's Call graph row mirrors the dream stage's real outcome instead of
assuming a completed cycle means edges exist. The ## GBrain Search Guidance
block written into CLAUDE.md drops the old code-callers --source caveat:
gbrain >=0.41.38.0 honors the .gbrain-source pin for code-callers/code-callees.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(jsonl-store): shared audited JSONL plumbing (injection-reject + atomic append + tolerant read)
Single source of truth extracted for D2A: gstack-learnings-* and the upcoming
gstack-decision-* bins share one injection-pattern list, one atomic single-line
appender, and one tolerant reader. No more drift between stores.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* refactor(learnings-log): use shared hasInjection from lib/jsonl-store (D2A)
Replace the inline injection-pattern copy with the shared list. One audited
write-path rejection across learnings + the upcoming decision store. Behavior
unchanged (35/35 learnings tests green); learnings-search keeps its inline copy
because a structural test pins its bash/bun shape.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(decision): event-sourced decision-memory model (lib/gstack-decision)
decide/supersede/redact events on lib/jsonl-store; active set is computed (no
mutable status), dangling refs tolerated. Free-text is injection-checked and
redact-scanned on write (HIGH secret -> reject). Scope filter (repo/branch/issue)
for relevant resurfacing. File-only + reliable; gbrain not required.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(decision): bounded active snapshot + compaction (redact expunges, supersede archives)
writeSnapshot/readSnapshot/rebuildSnapshot give an O(active) bounded read for the
session-start hot path (D1A). compact() rewrites the log to active, archives
superseded decisions for history, and EXPUNGES redacted ones (dropped, never
archived) so an accidentally-captured secret leaves the store for good.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(decision): gstack-decision-log + gstack-decision-search bins (non-interactive)
Two bins mirroring gstack-learnings-* (D3A). log writes decide/--supersede/--redact/
--compact events + refreshes the bounded snapshot + enqueues for cross-machine sync;
search reads the O(active) snapshot, scope-filtered to current branch, newest-first,
--all to include superseded, --json for machines. Empty store returns silently
(no snapshot write on an empty read).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(memory): surface active decisions at session start + capture nudge (Context Recovery)
Context Recovery now shows recent scope-relevant active decisions (bounded read of
decisions.active.json via gstack-decision-search) and instructs the agent to treat
them as settled calls and to log durable decisions/reversals. Closes the Phase-1
capture->curate->resurface loop, reliable + file-only. Regen across all hosts folded
in (squash-with-regen); parity 10/10, freshness green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test: refresh ship golden baselines for the memory-loop preamble change
Context Recovery now emits the cross-session-decisions block, so ship's preamble
(all hosts) changed. Golden baselines are hand-maintained copies (gen does not
write them); refresh them from the fresh gen so golden-file regression passes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* docs(memory): document the cross-session decision-memory loop in CLAUDE.md
Adds a '## Cross-session decision memory' section: how to resurface
(gstack-decision-search) and capture (gstack-decision-log) durable decisions,
the supersede/redact/compact verbs, and a crisp durable-vs-trivial definition
so the store stays signal. Reliable file-only path; gbrain not required.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(memory): emit durable decisions from ship/ceo/eng/spec at structured points
Wires the four skills that finalize real decisions to capture them in the
cross-session decision store, from their STRUCTURED outputs (never free-text
scraping):
- ship: the version bump (level + why) at write time
- plan-ceo-review: accepted scope + verdict (branch-scoped)
- plan-eng-review: the architecture verdict + key call (branch-scoped)
- spec: the filed issue's core approach (issue-scoped)
All emits are non-interactive, schema-correct (content in decision/rationale,
source=skill, confidence 1-10), and best-effort (|| true) so a decision-log
failure never blocks the workflow. Includes regen across hosts + refreshed ship
golden baselines.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(memory): optional gbrain --semantic recall for decision search
Adds gstack-decision-search --semantic (with --query): appends a 'Related from
memory' block from gbrain semantic search, scoped to the curated-memory source.
Pure enhancement, reliability-first: a new lib/gstack-decision-semantic.ts is the
ONLY decision module that touches gbrain and is imported lazily only on --semantic,
so the reliable file path never loads gbrain code. Every path degrades to the
reliable file results when gbrain is off, unconfigured, empty, or errors (never
throws, 10s timeout).
Built against the verified gbrain 0.42.x surface (text output [score] slug --
snippet, NOT JSON; curated-memory source resolved by worktree path, not a
gstack-brain-<user> id). Deterministic-contract tests only: parser units,
degrade-to-null when gbrain absent, and a fake-gbrain shim proving scope+search
end-to-end. find-contradictions deferred (no verifiable CLI surface yet + curated
memory not indexed).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(gbrain-sync): self-heal stale autopilot lock (dead-pid)
detectAutopilot treated a lock FILE as proof of life, so a crashed gbrain daemon
left a stale lock that wedged every sync forever (observed: a dead pid refused
--full indefinitely). Now read the holder pid (bare or JSON body) and check
liveness via signal-0: ESRCH=dead → ignore the stale signal and keep checking;
EPERM=alive (other user) → active. A stale lock never masks a live autopilot
process. Pure decision function — does not delete the file; the caller may clean it.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* docs(review): drop stray trailing code fence in TODOS-format
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(test): align section-loading E2E testNames with their TOUCHFILES keys
Pre-existing on main (v1.56.x): the two section-loading E2E tests used
human-label testNames ('/ship section-loading') that don't match their slug
keys ('ship-section-loading') in E2E_TOUCHFILES/E2E_TIERS. Every other E2E test
uses the slug as its testName, and the TOUCHFILES completeness gate requires
testName to be a registered key — so the gate was red. Align both testNames to
their slug keys (also fixes tier lookup for these two periodic tests).
Verified failing on a clean origin/main checkout before the fix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix: pre-landing review fixes (datamark, DRY, compact, coverage)
Addresses the pre-landing review findings (all INFORMATIONAL, no criticals):
- security: datamark resurfaced decision text at the render boundary
(lib/gstack-decision.ts datamark() — neutralizes code fences, --- banners,
<|role|>/</system> markers, control chars, newlines). Applied in
gstack-decision-search human output so stored text can't masquerade as
instructions in Context Recovery (codex hardening #3 / AC #7). --json stays raw.
- DRY: extract resolveSlug/gitBranch/flagValue to lib/bin-context.ts; both
decision bins use it instead of duplicating the helpers.
- compact(): batch the archive append (one write, not N) and shrink the
mid-compact crash window; simplify the opaque branch/issue ternary.
- coverage: learnings-log injection rejection (D2A wiring), search --recent/
--scope + NaN-safe --recent, datamark-applied, unparseable lock body,
compact-empty, corrupt-snapshot degrade.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(security): close adversarial-review findings in decision memory
Adversarial review (Claude subagent) found a CRITICAL the specialist pass missed:
- F1 (CRITICAL): 'Human:'/'Assistant:' turn-prefixes bypassed BOTH the write-time
denylist AND datamark(), landing verbatim in agent context inside the trusted
ACTIVE DECISIONS fence. Add 'human:' (+ 'disregard previous', 'from now on') to
the shared denylist, and have datamark() neutralize Human:/Assistant:/System:/User:
turn-prefixes (ZWSP) at the render boundary.
- F2: datamark() only stripped ASCII C0; extend to Unicode line terminators
(U+0085/2028/2029) and U+007F so 'strip newlines' actually holds.
- F3: validateDecide blocked only HIGH secrets; MEDIUM-tier PII (e.g. SSN) persisted
silently and synced cross-machine. The store is non-interactive (no confirm path),
so fail closed on MEDIUM too.
- F4: compact() was a lock-free read-modify-rewrite that could clobber a concurrent
append (lost decision). Add an O_EXCL compact lock + a pre-rename size recheck that
aborts untouched (skipped=true) if an append landed; caller re-runs.
- F7: filterByScope unknown/garbage scope fell through to 'return true' (leaked into
every context); fail conservative (false).
F5 (pid reuse) and F6 (pgrep over-match) are intentionally left as-is: both fail SAFE
(over-refuse sync); making them precise would introduce a fail-DANGEROUS path
(allowing sync during a real autopilot). True disambiguation needs gbrain to stamp the
lock with a start-time, which gstack doesn't own. F8 (compact moves history to archive)
is by design.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(security): close cross-model (Codex) adversarial findings
Codex adversarial review found a HIGH the Claude pass missed plus 3 mediums:
- C1 (HIGH): gstack-decision-search --all returned every decide and IGNORED redact
events, so a redacted secret still resurfaced via --all until compact ran. --all
now excludes redacted (redact = expunge from every read path), still showing
superseded history.
- C-med: semantic (external gbrain) slug/snippet were printed raw — datamark them too
so a gbrain hit can't spoof role markers / fences into agent context.
- C4: semanticRecall fell back to an UNSCOPED gbrain search when no curated-memory
source resolved, pulling code/doc corpora mislabeled as 'related decisions'. Now
returns null (degrade) when there's no worktree-backed memory source.
- C5: validateDecide scanned only decision/rationale/alternatives; branch and issue
are stored + surfaced (raw via --json), so include them in the injection+secret scan.
C2 (snapshot staleness) / C3 (compact TOCTOU residual): accepted for a single-user
store — atomic appends never lose the event, rebuilds self-heal, and the compact
size-recheck leaves only a sub-ms window; full append-locking would break the
lock-free append design.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* chore: bump version and changelog (v1.57.5.0)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
|
Garry Tan
|
b9371d716e
|
v1.34.2.0 fix wave: /codex review on CLI 0.130+, /investigate learnings, /sync-gbrain on Supabase (3 community-reported bugs) (#1478)
* fix(learnings): accept type:"investigation" in gstack-learnings-log The /investigate skill instructed agents to log learnings with type:"investigation", but bin/gstack-learnings-log:22 rejected anything not in [pattern, pitfall, preference, architecture, tool, operational]. Every investigation run exited 1 to stderr and the learning was dropped, silently to the user. Fix: add 'investigation' to ALLOWED_TYPES. Regression test: round-trips a learning with type:"investigation" and asserts exit 0 + file write; second test reads investigate/SKILL.md.tmpl and asserts it emits the literal type:"investigation" string, guarding the template/validator contract at both ends. Fixes #1423. Reported by diogolealassis. * fix(gbrain): engine detection survives gbrain ≥0.25 schema + non-zero doctor exit freshDetectEngineTier() in lib/gstack-memory-helpers.ts returned engine: "unknown" for every Supabase user on gbrain ≥0.25. Two stacking bugs: 1. execSync("gbrain doctor --json --fast 2>/dev/null") threw on non-zero exit. gbrain doctor exits 1 whenever health_score < 100, which is essentially every fresh install due to resolver_health warnings. The JSON output never reached the parser. 2. gbrain ≥0.25 shipped schema_version:2 doctor output that dropped the top-level 'engine' field entirely. Result: every /sync-gbrain on Supabase logged 'engine=unknown' and skipped all sync stages silently. Fix: - Replace execSync with execFileSync (no shell, no bash-specific 2>/dev/null redirect; portable to Windows). - Recover stdout from the thrown error object so non-zero exits still parse. - Fall back to reading gbrain's config.json (respecting GBRAIN_HOME env var, defaulting to ~/.gbrain/config.json) when doctor output doesn't surface an engine field. - Add logGbrainError() helper that appends one-line JSONL to ~/.gstack/.gbrain-errors.jsonl on parse failure, so future regressions leave a forensic trail. The "supabase" tier here means "remote postgres" in practice — gbrain config uses engine:"postgres" for both real Supabase and any other remote postgres (e.g. local-postgres-for-testing). Downstream sync code treats them identically, so the label compression is intentional and documented inline. Regression test: existing detectEngineTier suite now isolates HOME + GBRAIN_HOME + PATH to temp dirs (closes a flake source where the prior tests would read whatever was on the reviewer's machine). New test forces gbrain off PATH, writes a synthetic config.json with engine:"postgres", asserts detectEngineTier() returns engine:"supabase". Fixes #1415. Patch shape contributed by Shiv @shivasymbl (tested on gstack v1.31.0.0 + gbrain v0.31.3 + Supabase). * fix(codex): /codex review works on Codex CLI ≥0.130.0 Codex CLI 0.130.0 made [PROMPT] and --base <BRANCH> mutually exclusive at argv level. Step 2A of codex/SKILL.md.tmpl had always passed both (the filesystem boundary prefix as the prompt argument + the base branch), so every /codex review call died with: error: the argument '[PROMPT]' cannot be used with '--base <BRANCH>' Fix: split Step 2A into two paths. Default (no custom user instructions): bare 'codex review --base <base>'. Codex's review prompt is internally diff-scoped, so the model focuses on the changes against base. The filesystem boundary prefix is dropped here because Codex 0.130 has no documented system-prompt config key (probed -c 'system_prompt="..."' against 0.130 — the flag is silently accepted but the value isn't applied). Skill files under .claude/ and agents/ are public, so this is a token-efficiency concern, not a safety one. Custom instructions (/codex review <focus>): route through codex exec with the diff written to a tempfile, inlined into the prompt between explicit DIFF_START / DIFF_END markers. The boundary is preserved here because codex exec isn't auto-scoped to the diff. The DIFF_START/END delimiters tell the model where data ends and instructions resume, which materially reduces prompt-injection hijack rates when the diff contains adversarial content. Note on bash semantics: codex's earlier review flagged the exec route as "command injection via $_DIFF interpolation." That framing is wrong — bash parameter expansion does not re-evaluate $(...) or backticks inside the expanded value, so a diff containing $(rm -rf /) is plain string data to codex exec. The real risk is prompt injection (model-side, not shell-side), which the DIFF_START/END pattern mitigates. Regression tests in test/codex-hardening.test.ts assert across BOTH codex/SKILL.md.tmpl AND the generated codex/SKILL.md: 1. No 'codex review' invocation line combines a quoted-string OR variable positional argument with --base. 2. Step 2A still contains either bare 'codex review --base' OR 'codex exec' (guards against accidental deletion of both fix paths). Fixes #1428. Reported by Stashub. * test: raise timeouts for slow integration tests Two test files were timing out at the default 5s on developer machines, both pre-existing on origin/main but unrelated to this branch's bug fixes: - test/gstack-artifacts-init.test.ts: 13 tests spawning real subprocesses via fake gh/glab/git shims in PATH. bun's fork+exec overhead pushed these past 5s consistently. Added a local test-wrapper that aliases test() with a 30s timeout (matches the brain-sync.test.ts pattern already in the repo). - test/gstack-next-version.test.ts: one integration smoke test that spawns 'bun run ./bin/gstack-next-version' and parses the resulting JSON. The subprocess does a 'gh pr list' against the live GitHub API to enumerate claimed version slots. Network latency makes 5s tight; raised this single test to 30s. No production code changed. The tests already passed deterministically once given enough wall-clock time. * chore: bump version and changelog (v1.34.2.0) Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> |
|
|
Garry Tan
|
ae0a9ad195
|
feat: GStack Learns — per-project self-learning infrastructure (v0.13.4.0) (#622)
* feat: learnings + confidence resolvers — cross-skill memory infrastructure
Three new resolvers for the self-learning system:
- LEARNINGS_SEARCH: tells skills to load prior learnings before analysis
- LEARNINGS_LOG: tells skills to capture discoveries after completing work
- CONFIDENCE_CALIBRATION: adds 1-10 confidence scoring to all review findings
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: learnings bin scripts — append-only JSONL read/write
gstack-learnings-log: validates JSON, auto-injects timestamp, appends to
~/.gstack/projects/$SLUG/learnings.jsonl. Append-only (no mutation).
gstack-learnings-search: reads/filters/dedupes learnings with confidence
decay (observed/inferred lose 1pt/30d), cross-project discovery, and
"latest winner" resolution per key+type.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: learnings count in preamble output
Every skill now prints "LEARNINGS: N entries loaded" during preamble,
making the compounding loop visible to the user.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: integrate learnings + confidence into 9 skill templates
Add {{LEARNINGS_SEARCH}}, {{LEARNINGS_LOG}}, and {{CONFIDENCE_CALIBRATION}}
placeholders to review, ship, plan-eng-review, plan-ceo-review, office-hours,
investigate, retro, and cso templates. Regenerated all SKILL.md files.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: /learn skill — manage project learnings
New skill for reviewing, searching, pruning, and exporting what gstack
has learned across sessions. Commands: /learn, /learn search, /learn prune,
/learn export, /learn stats, /learn add.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* docs: self-learning roadmap — 5-release design doc
Covers: R1 GStack Learns (v0.14), R2 Review Army (v0.15), R3 Smart Ceremony
(v0.16), R4 /autoship (v0.17), R5 Studio (v0.18). Inspired by Compound
Engineering, adapted to GStack's architecture.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* test: learnings bin script unit tests — 13 tests, free
Tests gstack-learnings-log (valid/invalid JSON, timestamp injection,
append-only) and gstack-learnings-search (dedup, type/query/limit filters,
confidence decay, user-stated no-decay, malformed JSONL skip).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* chore: bump version and changelog (v0.13.4.0)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* test: learnings resolver + bin script edge case tests — 21 new tests, free
Adds gen-skill-docs coverage for LEARNINGS_SEARCH, LEARNINGS_LOG, and
CONFIDENCE_CALIBRATION resolvers. Adds bin script edge cases: timestamp
preservation, special characters, files array, sort order, type grouping,
combined filtering, missing fields, confidence floor at 0.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: sync package.json version with VERSION file (0.13.4.0)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* chore: gitignore .factory/ — generated output, not source
Same pattern as .claude/skills/ and .agents/. These SKILL.md files are
generated from .tmpl templates by gen:skill-docs --host factory.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* test: /learn E2E — seed 3 learnings, verify agent surfaces them
Seeds N+1 query pattern, stale cache pitfall, and rubocop preference
into learnings.jsonl, then runs /learn and checks that at least 2/3
appear in the agent's output. Gate tier, ~$0.25/run.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|