Garry Tan
7fca42ad8b
* test: delete test-infrastructure dead code (G) - exit-propagation drives the runner's real strict verdict (BunTestOutputClassifier + strictTestExitCode); delete the unused shardRunLooksTruncated predicate. - delete skill-coverage-matrix registry + its gate (nothing reads it; the floor already iterates skillCensus()). - delete touchfiles-facade export-parity tests (Bun fails missing imports at link time) and the duplicated E2E_TIERS tier-value test. - delete brain-cache-spec TRANSPORT_DEFAULT_POLICY, SKILL_RUN_RETENTION_DAYS and the now-unused BrainTrustPolicy type with their literal tests. AUTOPLAN_PREFLIGHT_BUDGET_BYTES stays: skill-preflight-budget enforces it against real resolver output. - delete audit-compliance's JSDoc-comment grep. * test: replace product tests that fake the product with real-boundary tests (F) - design: serve.test.ts drove an inline mirror server; now two tests run the real serve() on an ephemeral port (reload confinement, submit exit 0). - setup-gbrain: rollback + voyage tests execute the template-extracted init blocks (3 sites) instead of drifted local bash copies. - terminal-agent: internalHandler source greps replaced by a behavioral /internal/grant + /internal/revoke auth matrix (no/wrong/valid token). - /health: server-security-surface and the server-auth / security-audit-r2 / sidebar-tabs source greps fold into one liveness-only check on the real body; the L4 sidecar wiring gets a behavioral /pty-inject-scan test. - delete tautologies (browser-manager onDisconnect, memory-command #12), ios swiftui tap fixture self-check, memory-ingest put_page grep, detach source greps, sidebar-agent absence pins, dead-CSS pins + the dead CSS, security-audit-r2 Task 1 + the test-only meta-commands re-export, duplicate generated-SKILL.md checks. - make-pdf coverage-gaps cases move into their owner test files. * test: delete tests of dead eval code (A) - A1: the retired Eng lexical oracle (evaluateEngSeedCoverage, isEngSeedDecisionAUQ), the completion-handoff detector and the retained corpus had no paid caller since v1.87.6; delete their 26 replay files, ~2.6k helper LOC and fixtures, and the dead blocks in 8 mixed files (live hasNativePlanTerminal / batching assertions stay). - A2: dead viewport approvers in autoplan-artifact-permission and their 11 replay files + fixtures; recorder/launcher cases stay. - A3: never-wired oracles and seeders (autoplan-phase-order, eng-finding-fixture, ceo-paired-fixture, design-ui-scope, plan-skill-completion, pty-current-screen, required-reads, transcript-section-logger); plan-seed-submission now decodes through the production createPtyScreen; section manifests name their actual guard. - A4: zero-reference helper exports, plus execGit and invokeAndObserve found by the reachability pass. - 52 fixtures orphaned by the deletions; touchfile and selection-table entries for every deleted path. * test: clean up the paid eval lane (B1-B4, B6, B7) - B1: delete paid files that assert nothing or cannot pass meaningfully: skill-llm-eval-spec and skill-e2e-spec-execute (test.todo), gemini-e2e (+ gemini-session-runner; no gemini CLI in CI), ship-idempotency (red since v1.63), the two opus-4-7 *-sonnet overlay wrappers, conductor-prose (+ its source-evaluation replay), codex-e2e-plan-format; drop their keys, scripts and census rows. - B2: skill-llm-eval grades browse/sections/command-list.md with one union judge that also carries the baseline score pin; regression-vs-baseline deleted (paid run: pass, c4/c4/a4). - B3: memory-pipeline, ios-qa, ios-qa-swift-build and plan-tune-cathedral make no model calls; renamed out of the paid glob so they run on every PR. Swift builds need GSTACK_TEST_SWIFT=1; device stub deleted. - B4: codex-e2e*, outside-voice, aside and ios-device cannot run in the CI image; excluded from the weekly lane with a tracked re-entry condition. - B6: fold opus-47's negative routing controls into skill-routing-e2e journey-negatives (paid run: 3/3 unrouted) and delete the file. - B7: delete the never-green brain-privacy-gate eval; a free gstack-skill-start test now proves consent precedes artifacts egress. * test: retire the finding-count cluster and trim its helpers (C) - C0/C1: the five never-green evals (skill-e2e-autoplan-chain and skill-e2e-plan-{ceo,eng,design,devex}-finding-count) failed on harness and budget, never on skill behavior; delete them, their touchfile/tier ids, AUTOPLAN_CHAIN_BUDGET and the dedicated eighth periodic slice (--slices 7). - C2: delete the helper groups whose only paid consumers were those files (11 modules), trim claude-pty-runner and eng-seeded-coverage to the paid closure, and delete the free replay tests whose assertions exercised only that dead code (89 files, 135 orphaned fixtures). Blocks that used dead code only as input for a live subject keep their assertions: the multiSelect default moved to plan-review-decisions, runner PTY tests use inline caller policies, and the timer-safe budget checks moved to eng-finding-retry-budget. - The eight production-touching files stay except ceo-current-decision-record (its template read only feeds the retired counter). - CARVE_GUARDS.autoplan is behavioral 'none'; TODOS records the lost chain and per-finding cadence coverage with their re-entry tests. * test: fold per-incident replay series into their detector owners (D) Twelve detector families move into one owner test each: 73 incident files become describe blocks in ceo-section-loading-fixture (stale-fill race), model-overlays, coverage-audit-evidence, autoplan-phase-observer, native-auto-decide, outside-voice-evidence, eng-first-review, plan-count-completion, plan-count-file-permission, ceo-mode-option, plan-scope-selection and plan-count-prerequisite. Each block keeps its original code and fixture, so every case still runs; only tests asserting the incident file's own touchfile registration are dropped (41). Touchfile lists that named an incident now name its owner. * test: start the plan-count history PTY on its readiness marker (H) The fake CLI prints a startup marker and the runner waits for it instead of the fixed 8 s startup sleep (8.6 s -> 0.9 s locally). eng-semantic-terminal's sleeping registration cases went with C; plan-count-timeout keeps the fixed wait because it asserts deadline behavior. * test: derive paid touchfiles from each eval's static closure (E) touchfiles.test.ts now checks, per key, that the paid file's static test/helpers and test/fixtures closure (plus fixture paths it names in string literals) is covered, and names the file, path, chain and key to fix when it is not. Free *.test.ts files are no longer touchfiles, so editing a free replay test stops selecting paid evals: 950 entries removed, 653 real closure paths added. The hand-copied inventories go: periodic-fixture-selection, fake-impeccable-touchfiles and 45 per-file selection examples. Selection for the sample edits (plan-eng-review template, claude-pty-runner, plan-count-fixture, gstack-config) loses no case under either profile. CONTRIBUTING documents the rule and its lower bound. * test: skip hollow tier shards and census judges in the paid planner (B5) A paid file is now skipped for a tier lane only when every E2E id it registers is known statically and none has that tier; ids come from the touchfile registrations and literal testName/*IfSelected arguments, so a comment or skill path that quotes another id cannot unschedule it, and computed names keep today's scheduling. --list and the manifest show each skip as "skipped: no E2E_TIERS id has tier <tier>". The weekly gate census drops the LLM judges (--skip-judges); they still run in the periodic census and PR gate lanes. Gate lane 52 -> 42 files, census 41; periodic 77 -> 69. * test: run seven paid evals on the current default capture model (B8) skill-e2e-{auq-matrix,plan-format,qa-bugs,retro,workflow} pinned claude-opus-4-7 and skill-e2e-office-hours plus -brain-writeback pinned claude-sonnet-4-6; none tests a historical model, so they now capture with resolveEvalModel('capture'), and the free harness tests that execute these registrations receive the same resolver. The paid re-pin run passed all of them. skill-e2e-{design,office-hours-phase4,plan-prosons,plan} keep claude-opus-4-7: six of their cases failed on the default model (three timeouts, a missing report file, a format miss and a posture score of 3), so per the plan's fallback they keep their pins with a TODOS entry. The pre-spend estimate and drop threshold are in docs/test-audit-2026-09.md. * test: guard the reduced suite against new test-of-test files - test/test-of-test-ratchet.test.ts records the 228 free tests that import only test/ code and fails on a new one, naming the owner test to extend instead; a stale baseline entry fails with the remove instruction. - test/helpers/resolve-repo-path.ts is the one specifier/literal resolver for the ratchet and the touchfile closure invariant, with its own unit tests. - CONTRIBUTING "Test tiers" describes the paid-failure workflow (fix, then one row in the detector's owner test) and the ratchet; TEST_PORTFOLIO gains the detector -> owner-test table and no longer claims an Autoplan chain eval. - TODOS: automatic exclusion policy for chronically red periodic files (P3), the deferred native-completion table collapse, the unused CEO payment seeder; the PTY readiness item is narrowed to the paid runner. - docs/test-audit-2026-09.md collects the triage, security mapping, inventories, selection proof, behavior-commit decisions and retained false positives. * v1.91.8.0 test: smaller suite, derived paid selection, retired never-green evals Release metadata for the test-reduction branch: VERSION 1.91.8.0 (1.91.7.0 is claimed by #2983), CHANGELOG with the measured before/after table and a contributor section, durations re-recorded on Ubicloud standard-16 (857 files, 0 failures), the agents digest, CONTRIBUTING's after-measurement row, the B8 fallback TODOS entry, and the after metrics, kept-vs-plan notes, B8 run and census estimate in docs/test-audit-2026-09.md. * fix(ubicloud): skip retrieval globs that match nothing instead of reporting a failed pull * test: pin DISABLE_AUTOUPDATER in hermetic env and capture corrupt-seed warning Both EVALS_HERMETIC branches of buildHermeticEnv now carry DISABLE_AUTOUPDATER=1 (the allowlist scrubbed the workflow's copy, so every PTY screen showed the updater's npm-prefix failure). Per-test overrides still win. The corrupt durations-seed test now captures its expected warning and restores the console spy. * style(cso): format lib/cso TypeScript with pinned Prettier Mechanical reformat only. Minified transpile output is byte-identical for 21 of 22 files; witness.ts differs only in three regex flag orders (/mi -> /im), which JavaScript canonicalizes. Source-text assertions over lib/cso now compare whitespace-insensitively with the same tokens. * fix(cso): import join for compiled-launcher assertion witnesses Compiled installs always take the non-Bun branch, which called an unimported join and threw before any runtime-tested assertion could be witnessed. The child command selection is now a pure, platform-aware function; a missing sibling launcher fails with its expected path. * fix(browse): make connect --supervise actually respawn a crashed server The supervisor respawned with a block-scoped env that no longer existed, so every attempt threw and the loop gave up after five tries. The headed env is now one pure helper used by connect and respawn, the loop is an injectable runHeadedSupervisor with behavioral tests, failures name the daemon log and relaunch command, and connect's usage advertises --supervise. * test: one finite PR world for the shared-libs fixture; name dual-voice probe evidence The shared-libs shim served 2 PRs for pulls?state=all and endless full pages for state=open. gh pr list, pulls?state=open|all|closed (per_page/page, short last page, direction) and search/issues now page one deterministic table: PR 7, 600 older open PRs, PR 42 and 3 closed PRs, so five 100-item open-metadata pages still leave older open PRs unchecked. The Contents API lists pinned directories (the captured attempt got 404 for contents/ and contents/src while files resolved, then fell back to a raw host), unknown endpoints return 404 instead of repo metadata, and the read-only detector is unchanged. Free tests cover view agreement, the budget bound, gh/curl agreement and the empty world. Dual-voice outside-voice failures now report probeToolUseId, probeMode and the canonical-match result with the reason the probe output was rejected. * feat: require a zero-error product typecheck and a test type-debt ratchet Adds tsconfig.json (strict) over product code, fixes its remaining 90 diagnostics (type-only, interface corrections, and explicit narrowing), and adds a typecheck job to the required free-tests aggregate running bun run typecheck, the test-code ratchet (identity -> count baseline, fails on new, repeated, or unlocked fixed diagnostics), and the lib/cso format check. Reuses fixes from #2447 where they still applied. * test: follow the headed env helper and the typecheck gate in source-shape checks * fix(test): pin the package.json change kind in shared-input selection tests computePaidCaseSelection read the version-only exemption from git even when changed files were injected, so the shared-input test failed on main and on version-only branches. The exemption is now an optional input; the test pins a real package.json change and covers the version-only case. * test: judge plan-count completion on structured evidence, not wording Replaying run 36385945043's two Design attempts showed the existing routes rejected correct endings: attempt 1 at the typed-completion path field ('- Reviewed plan written to …' is not a 'Plan written to' line), attempt 2 at the leading-fence veto (its final message opens with the dashboard). nativePlanTerminalPreconditions is the structural prefix of hasNativePlanTerminal (behavior unchanged). structuredPlanCompletion adds, inside the existing nativeSummary branch: a complete report (Design binding for Design), a completed review-log row for the expected skill appended during this attempt under the child's GSTACK_HOME/project slug (resolved with bin/gstack-slug) and stamped with the fixture commit, timed between the report/last answer (second resolution) and the final native message, a final message with stop_reason end_turn (now carried on public transcript messages), and no visible question or permission prompt. Timeout summaries add idleFor and lastTerminalCandidate. Terminal and throw captures copy the plan file and review-log rows into the artifact directory; copies are best-effort and recorded in evidence-copy.json. Free regressions: both captured Design endings (trimmed fixture with provenance; report, row and end_turn reconstructed and labelled), the negative controls, and real-PTY completion/timeout runs through the real review logger. * test: structural Design count boundary; TODO proposals are not findings Replaying run 36385945043 through the Design count predicates: routing, focus and learnings setup was not recognized as setup, Issue 1 was counted pre-review in both attempts (the boundary fired on it), and attempt 2 counted the Font TODO proposal as a finding (review=4 and review=5 for five issues). The paid caller now starts review at the first answered native decision that is not setup (recognized packet, or setup header/question ID), a completion handoff, artifact rendering or a TODO proposal (the review's Add to TODOS.md / Skip / Build it now menu). TODO proposals are recorded as administrative extra decisions. The replay asserts each counted call: both attempts review=5 (Issues 1-5). isDesignCountFirstReview and its controls are unchanged. * test: CEO classifier throws name the question and matched predicates Replaying run 36385945043's FAN-1 and ERR-1 throws (ledger rows reconstructed from rendered diffs) through ceoPaymentFinding: the email obligation's row, subject, option and proposal predicates pass and the ELI10 explanation-defect predicate fails first ('lets that exception fly out', 'the error bubbles up'). Binding the defect to the named ledger row instead (the planned fix) was tried and reverted: scoped to the email seed it flips 30+ existing cf74 still-rejects replays, which require a vocabulary-free, ledger-bound email question to earn credit only through a complete saved comparison. With FAN-1's rendered currentDecision payload reconstructed, the recorded- decision path counts it, so the real saved plan (not uploaded) must have differed; failure artifacts now retain it. The classifier stays fail-closed and unchanged. Its throw now prints the header, the first 200 question characters and each obligation's predicate results. Free regressions with provenance and negative controls: an unrelated question, an email question whose row says it is already rescued, and a ledger ID whose row belongs to another seed. * chore: regenerate the test type-debt baseline on top of #2994 * fix(typecheck): strip the checkout root from ratchet diagnostic identities * fix(test): recognize ledger row-ID split candidates so collection stops at the last ACK Run 36385945043's split-overflow case asked all five candidate decisions by 8m55s, but the live candidate check required the question to open with "E1:" and every option to be a known disposition. The skill cited ledger row IDs ("D2.1 — R-E1: …") and offered "Hold, discuss first", so no candidate was recognized and the attempt ran the whole review (1302s). Identity now comes from the native header; the question must open with that candidate's ledger reference, name only that candidate, and offer exactly one include, defer and cut disposition. The selected answer must still be one of those three. The semantic evaluator and every existing negative control are unchanged; a trimmed capture from the run adds the positive case and four row-ID negative controls. * fix(test): stop the eng batching eval once its floor is proven The case's only verdict is reviewCount >= FLOOR (3). Run 36385945043 had three distinct acknowledged review decisions at 6m41s but kept answering until the ceiling (7) at 12m13s. The registration now passes the runner's existing isCollectionComplete stop once FLOOR non-setup, non-administrative review decisions are acknowledged; the floor check, ceiling, budget and counter are unchanged. A child-process registration test proves the stop predicate and that below-floor and timeout outcomes still fail. * test: add the non-blocking 'marathon' E2E tier Full start-to-finish flows move out of the blocking lanes. E2E_TIERS and E2ETier gain 'marathon'; describeE2ETier('marathon') is enabled only when EVALS_TIER=marathon, so the gate/PR and periodic lanes (and the gate census) never run those cases. The PR profile accepts marathon ids as scheduled elsewhere and defers them with their own reason, even on full fallback. * test: move the full office-hours workflow to marathon; add a periodic design-draft checkpoint The full startup workflow runs 1–3 real spec-review rounds (~280s each) and hit its 1200s capture in run 36385945043 at finalize. Review depth is the product's loop, so the case cannot fit a blocking lane without cutting rounds. It is now marathon tier with every assertion unchanged. skill-e2e-office-hours-design-draft.test.ts (periodic) runs the same fixed interview only through the Write that creates the design (269s in that run) and applies the full validator's design-draft checks, the required section reads and the launch/foreign-skill-read guards. validateOfficeHoursDesignDraft is extracted from validateOfficeHoursCompletion, which still applies it. Selection: office-hours-design-draft is registered periodic; the marathon-only file is already excluded from the gate and periodic plans by the B5 planner rule. Tier-alignment regexes and the valid-tier check accept 'marathon'. A type-only cast in plan-scope-selection.test.ts removes a diagnostic whose union print order made the ratchet identity unstable; baseline tightened. * test: supply the split-overflow fixture's HOLD SCOPE mode as a prerequisite The split actor always answered 0E's mode question with HOLD SCOPE. The skill skips that question on an explicit choice, so the fixture now states it and the attempt starts at the five candidate decisions (about 1.5 min earlier in run 36385945043). Candidates, actor policy, floor and semantic evaluation are unchanged; the fixture test pins the supplied choice. * test: start the eng batching eval with its setup prerequisites supplied Routing setup and cross-project learnings (D1/D2 in run 36385945043) are never counted and are not what the case measures. The registration now uses the runner's existing preconfiguredReviewActor so the attempt starts at the review; engSetupAUQ still vetoes any late setup question. The registration test pins the option. * test: count the design-draft paid file and defer marathon ids in PR selection pins The discovered paid-file census grows by one (skill-e2e-office-hours-design-draft). Full-fallback PR selection defers every non-gate id; the shared-input pins now expect periodic and marathon ids there. * fix(review): resolve the judged revalidation, setup-authority, plan-gate and findings-record ambiguities The census review workflow judge scored clarity/actionability 3 on both attempts: smoke-clock limits appeared to forbid post-repair revalidation, the caller deadline was undefined, 'ask for setup' conflicted with the report-only browser rule, fallback-sourced HIGH discrepancies had no gate decision, and the Step 5.8 record omitted adversarial findings. * fix(office-hours): load the builder section for every builder-mode reply Both census builder-wildness attempts answered a direct request for adjacent unlocks without reading phase-2b-builder-brainstorm.md, whose trigger read as applying only to the generative questions. * fix(sync-gbrain): define Step 4 helper args and one atomic write path Both census read-ready attempts spent turns reading the helper source to resolve <user-args>, inspecting fixture internals kept inside the repo, and reconciling 'Read + Edit' with the tmp+mv atomic write, then hit max turns before the verdict. * refactor(evals): share the import-closure walker and add the E2E shard reuse identity sourceDependencyClosure moves from the workflow-judge adapter into scripts/eval-input-cache.ts unchanged, so judge keys stay byte-identical. scripts/e2e-shard-reuse.ts builds the consumed-input identity of one PR-lane E2E shard (test import closure, every registered case's touchfiles, globals, runner/workflow/setup actions, child env pins, CI image, Claude CLI) and fails closed on anything unknown. Marathon joins the always-fresh purposes. * feat(evals): ~12-minute blocking paid lanes and a non-blocking marathon lane - Planner budget mode (--slice-budget S --jobs J): recorded per-tier wall times pack into as many ~9-minute executors as the work needs; the plan records per-slice estimates and the CI job timeout (supervised worst case + 20 min). evals.yml and evals-periodic.yml derive matrix size and timeout-minutes from it; max-parallel covers every slice at once. - Case shards: plan/design/review-army/shared-libs(-paths) run one registered case per process (<file>#<case id>, exact name pattern, exactly one case). - Retry rule: a timed-out attempt is a verdict. Only files whose every case budget is CAPTURE tier or shorter keep one retry; walls shrink to match. - Marathon tier: positive selection, excluded from gate/periodic planners, run by the new evals-marathon.yml (weekly + dispatch, fresh, own report). - PR-lane E2E reuse of verified first-attempt passes on identical inputs; the report rejects reuse outside the fast PR profile. - Duration seed from census run 36385945043, per tier and per case shard. * docs: blocking lane budget, marathon lane, retry policy and E2E reuse * chore(typecheck): lock in two fixed test diagnostics * fix(ci): drop a duplicated env/jobs block in evals-marathon.yml * test(ship-docsync): shard the doc-sync lifecycle by case and drop the duplicate dispatch-only case ship-docsync ran the same fixture and prompt as ship-docsync-completion and asserted a subset of it. The file now runs one case per process, so its lane wall is its longest case instead of half the sum of thirteen. * fix(evals): plan CI-unrunnable cases as excluded entries, not empty case shards design-review-fix drives the Aside browser and registers test.skip on Linux runners, so its case shard executed zero cases and failed the exact-one-case check in proof census 36597762183 (eval-slices 6). CASE_CI_EXCLUDE (reason + tracking, beside PERIODIC_CI_EXCLUDE) now turns such cases into excluded manifest entries that --list and the manifest surface; every planned case shard still must execute exactly its case. * docs(todos): list the case-level Aside exclusion with the CI-unrunnable evals * fix(plan-ceo-review): restore experience-first expansion framing, require the mode handoff, skip pacing menus Census 36597762183: both mode-routing runs logged provenance and moved on without the mandated handoff chat; the EXPANSION run asked an unauthorized batch/narrow pacing menu instead of the first per-addition question; the expansion-energy proposals led with the spec because v1.87.6.0 dropped 'lead with the felt experience'. The HOLD review detector also rejected a decision whose grounding line named no plan file although the owned source Read binds it. * test(outside-plan-disabled): bind quoted prior-record values by their sentence, not phrase order The parent obeyed the off switch and twice named the seeded completed record as pre-existing, once with the quotation after its owner and once with slash separators; the order-specific stripper counted both as current completion. Timestamp, location, current-claim and value-match controls still reject. * test(outside-plan-disabled): compare named record timestamps as instants; negated authorship is not a current claim The repair rerun named the seeded record by its ISO second (2026-09-29T16:58:52Z vs .727Z) and said 'I did not write'; both were misread as a foreign timestamp and a current write. * test(ceo-section-loading): recognize an arrow-ordered stale-fill execution by event roles The census review traced the seeded race as 'R1 miss -> R1 store read (v1) -> W commit v2 -> W cache.delete -> W fulfills -> R1 cache.set(v1) -> R2 (begun after W) hits v1', but the in-flight gate only accepted race vocabulary or fixed sentence shapes. Order, actor, version and dismissal mutations still fail. * test(design-floor): answer the seed-declared all-seven 0D focus menu while it is pending The actor declares 'Design: review all seven dimensions', but its picker reused designReviewSetupAUQ, which only matches already-answered calls (and a narrower header/label set), so the pending D1 focus menu was never answered and the case waited out its 609 s deadline. The skill's Step 0D requires asking; the fixture now answers it. * test(ceo-mode-routing): accept the skill-mandated Note form and Recommendation reason as HOLD posture HOLD Defer/Keep briefs must use 'Note: options differ in kind' (preamble), but the answered-HOLD path demanded a Completeness score, rejected a one-line Net with a semicolon, and read posture only from ELI10. The rerun's brief applied HOLD SCOPE in its Recommendation reason. Revert the ineffective 'always'/'handoff chat' wording: two runs still skipped the mode handoff. * test(qa-bugs): keep claude-opus-4-7 after qa-b6-static stalled on the default model qa-b6-static timed out on claude-fable-5-1 in census 36597762183 and in one of two targeted reruns. Both times the stream stopped mid-message with no pending tool, right after the model found the disabled submit button, and stayed silent until the 300 s deadline. Per the B8 fallback, re-pin with a TODOS entry; budgets and retries are unchanged. A rerun on opus-4-7 passed (125 s, 5/5 detected). * test(evals): add E2E_KINDS, BEHAVIOR_WHY, EVAL_POLICY and CASE_QUARANTINE skeletons Every E2E_TIERS and LLM_JUDGE_TOUCHFILES key starts as 'rule'; BEHAVIOR_WHY and CASE_QUARANTINE start empty. EVAL_POLICY pre-registers the approved panel (3, majority 2), quarantine entry 0.95/10 and exit 0.97/10, 10% cap, 8-weekly-run expiry, Fisher drift alarm and one INFRA re-dispatch. * test(evals): add trial records, panelVerdict, expectContract and trial-outcomes JSONL EvalTestEntry gains case_id, kind, trial, panel, failure_class and policy_version, stamped from the runner's TRIAL_ENV on isolated trial shards. panelVerdict() is the single verdict function (INCOMPLETE on missing or duplicate trials, contract veto at any count, quarantine hard-break rule, INFRA/INCOMPLETE machine classification). expectContract() records failure_class 'contract' on the collector entry and a sidecar before throwing. trial-outcomes JSONL has a fail-closed writer and a data-only reader. * test(evals): pin the fail-closed rule-shard gate through the real --report path Synthetic slice artifacts for rule fail, timeout, missing slice, unreported entry, hollow, never-started, collector failure and wrong-slice reports all exit red before the panel-verdict gate change lands. * test(evals): retire every paid automatic retry Paid evals never retry (approved 2026-09-29): delete SHORT_CASE_RETRY_FILES and retriesWithinCaseCap, drop the retry fields from the registered wall rows (walls now cover one run plus reserve), make retriesForFiles return 0, pass --retry 0 explicitly, and drop --retry 1 from the package.json paid scripts. Add the eval:pass-rates alias. Tests that pinned the old retry allowance are updated as a policy change; review-finalization-budget now proves late-result recording under the production zero-retry arguments. * test(llm-judge): sample every judge as a pre-registered 3-sample panel Each of the 24 skill-llm-eval judges now draws EVAL_POLICY.judge.samples independent samples of the same prompt concurrently inside the unchanged JUDGE_MS budget. Numeric dimensions gate on the per-dimension panel mean against the unchanged threshold; booleans (would_browse, consistent) on a strict majority. An erroring sample fails the whole panel and is never resampled; a refusal is an unscored panel only when every sample refused. callJudge's 429 backoff stays: it is transport before any model output. The workflow-judge cache stores and validates only complete panels, and its identity now records the panel and zero file retries. Harness tests that pinned one provider call per case now pin the panel size. * test(evals): classify every live case and re-select a case when its kind changes E2E_KINDS: rule by default (191 E2E ids), 22 behavior cases whose verdict is a live model choice with an acceptable sub-100% per-trial rate, each with a BEHAVIOR_WHY tolerance, and 25 judge entries (the 24 workflow judges plus the fixed-fixture llm-judge-recommendation rubric check). Contract-shaped cases (ask-before-decide, plan-mode no-writes, mandated steps, secrets, the batching floor) stay rule. Behavior requires a known literal registration and an exact Bun test name so the case runs as its own trial shard. Map-diff selection now diffs E2E_KINDS and BEHAVIOR_WHY per key, and a base revision without them selects every key, so a kind flip runs the panel it introduces. test/eval-kinds.test.ts enforces coverage, tolerances, isolatability and the reviewed counts, printing the literal to add. * feat(evals): per-case pass rates with Wilson intervals, identity series and quarantine policy scripts/eval-flake-rank.ts becomes eval:pass-rates (eval:flake-rank stays an alias, and the legacy aggregate stays exported). It reads eval-store's trial-outcomes JSONL from the last N completed evals-periodic runs on this branch and main (gh, downloading only the trial-outcomes artifact, cached and size-capped, parsed as data), plus local eval dirs, and prints per-case per-trial pass rates with 95% Wilson intervals. A series is a case's own touchfiles minus GLOBAL_TOUCHFILES (caseSeriesIdentities, for the report job to stamp), per model, CLI version and policy version. Labels: INCONCLUSIVE, BROKEN, FLAKY, FAILING, PASSING. --backfill imports legacy slice artifacts as pre-policy trials (first attempt only, attributed by registry id, never guessed) for display only. --gate fails with ACTION REQUIRED on post-policy evidence only: drift below the quarantine entry rule, a rule case behaving like behavior, a one-sided Fisher drop against the previous identity (Holm-controlled), and quarantine entries that met their exit rule, expired after 8 weekly runs, broke the 10% tier cap or are invalid. CASE_QUARANTINE entries now carry a failureClass (detector, harness or model-latency); a product defect has no class and is never quarantined. The policy test pins EVAL_POLICY's approved constants. * feat(eval-pass-rates): attribute legacy records by the exact slug of their display name * ci(image): pin Claude Code 2.1.284 so the eval model is recognized 2.1.251 logs [claude-code:unrecognized_model] for claude-fable-5-1, the eval capture/judge default. 2.1.284 does not. The gate PTY smoke subset (plan-ceo/plan-devex plan-mode, plan-mode-no-op) parses on the new TUI; plan-design-review-plan-mode passed at 293 s on 2.1.284 and timed out at 300 s on 2.1.251 on the same tree. * test(eng-batching): grade the floor once the review report is complete A completed GSTACK REVIEW REPORT ends the review, so the review-question count is final there. Run 36606688266 wrote its report at 1,248 s and closed the session at 1,318 s; the case now stops collection and applies the unchanged floor at the report instead of waiting out the session. No budget changes. * test(eng-batching): bind unsourced native briefs through the report's target Run 36606688266 asked ten separate native review questions (D1-D9 bound to ledger records R1-R9) and failed reviewCount=0 < FLOOR=3: its briefs named the plan by title instead of citing PLAN.md, its report declared 'Review target (fixed): PLAN.md' under '# Engineering review: <plan>', and it kept an unfenced copy of the plan's own H1. The named-source route now accepts those spellings and non-inline ledger briefs. The same replay rejects a foreign, mixed, duplicate or missing target, another plan's title or copied H1, a brief naming another plan or file, a mismatched saved brief, and re-asks. The run-36597762183 capture still counts 3. * fix(plan-design-review): treat a designer with no API key as unavailable Both proof runs (36597762183, 36606688266) printed DESIGN_READY, hit 'No OpenAI API key found' on the first $D variants call, then hand-built HTML/CSS wireframes, screenshots and a comparison board for ~195-245 s before the first review question; the second run timed out at 600 s. A failed first generation now takes the existing text-only path, and the skill forbids substituting hand-built mockups. * fix(deslop-shared-libs): read related sources together within the turn limit Run 36606688266's opportunity audit read sixteen sources one per turn and stopped at error_max_turns; the passing run 36597762183 read the same files in three batched commands. The skill now says turns are bounded and asks for parallel reads or one read-only command per step. * test(ceo-mode-routing): submit a mode review that scrolled past the viewport Run 36606688266 bundled routing, learnings and the mode choice into one native call. Its review panel was taller than the terminal, so the tab bar scrolled off, ceoModeSubmissionInput returned null for 240 s and HOLD SCOPE was never submitted ('no posture match'). With no bar on screen the viewport must still end at the focused Submit prompt, and the accumulated screen text supplies the one complete panel, authenticated exactly as before. Replay controls reject another mode, an unoffered answer, an altered question, a quoted panel, trailing output, a moved cursor and an answered or changed call. * docs(evals): document the pre-registered verdict policy, quarantine, pass-rate history and arithmetic AGENTS.md replaces the retry rule with the approved policy text (no retries; kind fixes trials; no added trials, samples or dispatches after a result; quarantine by CASE_QUARANTINE only; one INFRA/INCOMPLETE re-dispatch) and notes that a pre-registered fixed panel is not rejudging. CONTRIBUTING gains the kind rules, the judge panel, eval:pass-rates and an 'Add a paid eval' checklist. TESTING_INTERNALS describes verdicts, quarantine, history and the arithmetic, including the rule term: 1 trial vs 2-of-3 red rates at p = 0.99/0.95/0.90/0.70/0.30 and lane all-green probabilities for the current 191 rule / 22 behavior / 25 judge registry. * feat(evals): trial planner, slice exit split and panel-verdict report Planner: behavior and quarantined cases become panels of isolated trial shards (<file>#<id>~t<N>) bound by EVALS_SELECTION_JSON=[id] and the exact test name; the file shard excludes them by name. Trials of one case never share a slice, result slugs are unique, panels are validated whole, unknown registrations throw, and the planner prints a capacity preflight. Executor: each trial shard gets its TRIAL_ENV identity and a trial record (outcome, failure class, cause, cost); every shard writes a JUnit report. The slice exit now means execution completeness: a failed rule shard or a trial without a record reds the runner, a failed trial does not. Report: panelVerdict() decides every panel of the first run attempt (later attempts are reported, never replacing it); rule shards keep the unchanged fail-closed checks; collector records all count (no last-attempt wins); census runs enforce the quarantine cap and expiry. It writes collector-outcomes v2, trial-outcomes.jsonl (trials plus JUnit rule/judge cases), report-summary.md, and one headline + failure block with rerun commands, and flags INFRA/INCOMPLETE-only reds for the one re-dispatch. The fail-open suite gains the panel cases: behavior 1/3 red, 2/3 green with its failed trial shown, missing trial INCOMPLETE, contract at 2/3 red, quarantined 1/3 green, 0/3 and contract red, missing slice red, and a later attempt never replacing the first. * chore(evals): refresh paid duration seeds from proof runs 36597762183 and 36606688266 Both tiers, merged in run order (the later run wins). Notable: split-overflow 1332s -> 504s, section-loading 604s -> 342s, mode-routing 575s -> 444s; multi-finding-batching 734s -> 1318s (its red path in run 36606688266). * feat(evals): stamp trial series identities and fit panels to the live registry - scripts/eval-trial-series.ts stamps series_identity (eval-flake-rank's caseSeriesIdentities) on a report's trial-outcomes JSONL as its own step, keeping the history tool out of the paid runner's closure; TrialOutcomeRecord gains the optional series_identity field. - Slice-count plans let a registered trial spill into an ordinary lane when its siblings hold every long lane, so panels never share a runner. - Re-audited test-selection.ts (Stream B added the E2E_KINDS/BEHAVIOR_WHY map-diff; no new module loading) and repinned its hash. - Detach and release floors now count trial shards (66 periodic trials in 22 panels): periodic floor 33,821s, still under eval:bg:periodic's 67,380s. - Coordination fixtures supply the executor's trial records. * ci(evals): attempt-scoped artifacts, verdict-v2 PR comment, weekly pass-rate gate and one INFRA re-dispatch - Slice, census and marathon artifacts carry -a<run_attempt>; reports download them per artifact (no merge), so records never overwrite and a re-run never replaces the first attempt's verdict. - Planners pass --max-parallel for the capacity preflight (24/16 unchanged: the refreshed periodic plan needs 24 slices, the gate census 12). - PR comment: jq-only job reads collector-outcomes v2 (headline, sanitized failure block); the group_by(.name)|last recomputation is gone. - Reports stamp series identities, upload trial-outcomes-* for history, and shard logs upload always (a failed trial no longer reds its runner). - Weekly report: headline + failure block of both lanes in the issue body, the eval:pass-rates --gate step (fails closed without history), close the issue on a green run, and UC-E1: when every red is machine-classified INFRA/INCOMPLETE, one re-dispatch as a new run in its own concurrency group (redispatch_of), both runs reported. * feat(evals): planner-side whole-panel reuse and negative receipts The planner job restores this PR's receipt store once and ships a single filtered set with the plan: a pass or panel receipt with a same-or-newer FAIL for its input identity is dropped, and a panel receipt ships only as a whole PASS panel (re-verified with panelVerdict) from one run. Executors read only that set (no per-slice cache restore or save), so every trial of a panel sees the same receipts; a trial reuses its own record from the panel receipt, keeping a split PASS's failed trial. Trial identities drop the trial index (run-scoped) and bind the panel policy. Executed shards carry their input identity; the report turns a whole fresh PASS panel into a panel receipt and a FAIL panel or failed rule shard into a negative receipt, and marks a panel that mixes reused and fresh trials INCOMPLETE. The report job merges plan, slice and report receipts (newest per file) and saves one store per run. Also fixes two TS2352 casts in browse/test/dia-macos-qualification.test.ts whose diagnostic text drifted with program order (baseline locked, fix only). * feat(evals): --case/--trials local diagnosis and panels in local sharded runs bun run scripts/test-paid-shards.ts --case <id> [--trials N] runs N independent trials of one case through the CI panel runner (trial shards, TRIAL_ENV identity, name-pattern isolation) and prints its panelVerdict(); N defaults to the case's policy panel and CI never reads it. The local sharded path (test:gate:sharded, test:periodic:sharded) now plans the same trial shards and exclusions as CI and exits on execution completeness plus panel verdicts. * test(pty): grant an owned Create pane whose title row is cropped The targeted batching rerun on Claude Code 2.1.284 left its first report Write unanswered for 1,372 s and timed out: the viewport began at the pane's relative file row and rule, with the 'Create file' title cropped above, so the preview parser rejected the file row as foreign. That row must now resolve to the owned path and is skipped before the unchanged line-by-line preview match. Replay controls reject another file, another directory and an edited preview row. * fix(evals): tsx-safe generics in eval-flake-rank, legacy artifact names, no-retry wall docs * test(evals): record the read-only and detector-row invariants as contracts shared-libs-opportunity-judgment and review-design-lite are behavior cases: their recommendation and checklist judgments may vary, but the read-only invariant (commands, provider requests, fixture bytes, hooks, state) and the deterministic fake-engine detector rows are contracts. Both now go through expectContract, so any failure vetoes the panel. * test(judges): sample the recommendation rubric as a panel; never re-ask armJudge llm-judge-recommendation is a judge case: each fixture now draws a 3-sample judgePanel, gates reason_substance on the panel mean and the present/commits/has_because checks on a 2-of-3 majority, thresholds unchanged. armJudge no longer re-asks on a malformed verdict; it is a failed sample, as the judge policy requires. * test(evals): record a pre-turn API or CLI failure as infra recordE2E sets failure_class 'infra' on a failed session whose runner reports error_api, timeout_startup, error_output_stream or a non-zero CLI exit with zero turns and no assistant event. A model refusal, a timeout after model work, max turns, or an explicit caller pass/class keeps its ordinary classification. * test: pin every-record outcome counts and the twelve doc-sync callbacks * test(eng-batching): read the report target as a field, not a spelling The next targeted rerun (Claude Code 2.1.284) again asked eleven separate native questions and again counted zero: its briefs named no plan and its report declared '- **Review target (fixed):** `/abs/PLAN.md`' under '# Eng Review — PLAN.md: <plan>'. An unsourced brief now inherits the one current target field that names a PLAN.md file, whatever its list or emphasis markup; its ledger record still supplies the cited finding and must reproduce the brief exactly. A brief that names its plan must still match the report title. Replays of all three captures count 9, 9 and 3; controls reject a foreign, duplicate or missing target and an archived title. * fix(evals): --case list mode and name precheck; case-shard qa-callers; refresh batching and design-with-ui seeds * chore(release): v1.91.9.0 * test: settle the post-response composer before seeding; give the TPA recorder adapter its infra helper submitPlanSeed accepted a stale empty composer when the transcript recorded end_turn before the CLI repainted (late-repaint-typed-current fails 5/5 on the old helper, passes 5/5 now). The TPA recording fixture extracted recordE2E without isPreTurnInfraFailure, so every failed case threw before recording. * test(autoplan-dual-voice): unwrap Claude Code 2.1.284 subagent hand-back frames; accept read-only probe diagnostics; record before asserting Census run 36626737820: the native CEO report arrived framed and indented, so its INPUT line never matched, and the model's exact probe plus two variable echoes was not canonical. A column-zero line inside a frame, command substitution, backticks, redirects, assignments, CODEX_MODE echoes and output line-count mismatches stay rejected. The failure now records before asserting. * ci(image): keep Claude Code 2.1.251; test(ceo-mode-routing): keep HOLD's own deferrals in scope before assessing its rigor decision 2.1.284 enables per-turn effort for claude-fable-5-1: in gate census 36626737820, 66 of 84 sessions ran longer than on 2.1.251 (+20% session time, +32% thinking tokens) and 11 cases timed out on unchanged budgets. HOLD SCOPE's 0G step asks its own defer/keep menu; the actor answered it Defer and the assessment then judged that scope question as the rigor decision. The actor now answers that menu Keep and assesses the next one. * test: attribute quoted prior-record field lists, state the judge reason bound in its schema, move split-overflow to marathon Census 36629958451 reds: - outside-plan-disabled-no-fallback: the model quoted the pre-existing record as a parenthesized field list with its exact timestamp; attribution now requires that exact timestamp and the record's own field values. - plan-devex-peer-comparison-classification: the judge correctly returned missing but wrote a 1069-character reason, voiding the judgment; structured outputs cannot enforce maxLength, so the bound is stated on the field. - plan-ceo-split-overflow ran 504-1188 s as one PTY flow and set the periodic lane's wall clock; it now runs weekly in the marathon lane. * test: supply holdDeferKeepIndex to the CEO routing mocks and follow split-overflow into the marathon lane The registered-callback fixtures mock ceo-mode-option and lacked the new export; the split fixtures asserted the periodic tier; the registered-budget check looked for split-overflow only in the periodic manifest. * fix(qa): checkpoint receipts print the report link for their exploration file qa-functional-webhook-report failed in two of three censuses because the report linked .qa-evidence/NNN capture folders as "checkpoints" and never linked exploration-NNN.json. The checkpoint receipt now prints link: [checkpoint NNN](exploration-NNN.json), and the functional report template says capture folders are not checkpoints. * docs: final census numbers in the v1.91.9.0 entry; file the paid-eval follow-ups * ci(evals): name the PR-comment loop's unused fields so shellcheck passes (SC2034) * fix(plan-ceo-review): tighten expansion pacing wording to fit the skeleton cap after the main merge The merged skeleton measured 80,166 bytes against its unchanged 80,150 cap. Same instructions: ask separately for each addition, in turn, with no pacing menu; lead each proposal with the felt experience, then shape, effort and impact. * fix(eval-pass-rates): match trial-outcome files by basename so Windows backslash paths are read * fix(evals): repair proof-run reds in design-consultation, document-release, design and QA fixtures - design-consultation Phase 1 asks one brief that confirms context and decides research; the confirm-only first question scored substance 2. - document-release defines ship-owned inputs, exact steps and the JSON result, and drops stale spawned-from-/ship text (judge actionability 3.67 -> 4/4/4). - plan-design-with-ui accepts the Step 0D focus menu the same way the shared picker does ("focus on specific ones?"). - plan-design-review plan-mode saves in three Edits instead of one final Write. - QA functional annotations ask for the full 40-character revision. - Outside-disabled attribution judges quoted prior-record data by its exact timestamp or a dated, pre-existing-record sentence; four captured phrasings replay clean and current claims still fail. - --case can select autoplan-dual-voice by its literal test name. * test(design): revert the three-Edit plan-mode flow A focused paid run still timed out at 300 s: the first three passes alone took 150 s of thinking. The case stays a named timeout red rather than cutting review depth. * test: accept 'review mode = X' auto-decide declarations and parenthetical scope exclusions in the shared-libs actor auto-decide-preserved: the product auto-decided HOLD SCOPE and said "Decision: review mode = HOLD SCOPE"; the grammar knew only "is" and ":". shared-libs-plan-callers: the recommended option said "(no hardening)" and the actor read "hardening" as an expansion. Both replay the captured text, keep negative controls, and passed focused paid runs. * fix(review): pass Review Army checklists by path, run research alongside dispatch, always probe the design detector; state review-log invocation and statuses in the caller fixture - review-army-perf-n-plus-one: the parent copied full checklists into agent prompts and ran web research before dispatch (290 s on a 12-line diff); 212 s now. - review-design-lite: 5 of 6 captured trials reported the detector absent without probing; the probe is mandatory and its first line is reported, and the contract credits only fake-engine rule ids the checklist never names. - review-exploratory-small-cli: the fixture never gave review-log's direct invocation or status vocabulary; the model ran it through bun and wrote status "blocked". The prompt states both and the validator rejects out-of-vocabulary review statuses. Each case passed a focused paid run after repair. * docs(changelog): proof-run product fixes * fix(ship): always run the design-lite detector probe; test(shared-libs): credit a failed first file view and deferred-reuse Skip wording - /ship design-lite: the probe is mandatory and any non-ready first line is stated, matching /review (5 of 6 captured /review trials had skipped it). - shared-libs-pr-coverage: the first PR 42 page-1 read printed only a jq error, so the one refetch is a legitimate recovery, charged to the same budget. - shared-libs-review-prior-coverage: the Skip option said a future review can "reuse it once snapshot coverage holds"; a conditional tail on the recorded decision is not product work. Captured-text regressions and negative controls. * fix(ship,qa,document-release): repair proof-run regressions and fixture gaps - ship-docsync-completion: yesterday's audit-scope result dropped the section's status, so /ship spliced one in; the section now opens with **Status:**. - ship-docsync-missing-asset: a missing section or old Ship-owned mode blocks before launch. - ship-docsync-late-result: the invocation record says prepare already saves the candidate selection (no extra Read; budget unchanged). - qa exploratory: await the method Reads before the first probe. - qa-callers fixture: quote the real review-log record template; allow the git log command plan-completion prescribes. - qa functional observer: a receipt caught mid-link(2) is checked at stop instead of failing with ENOENT (reproduced from CI). Each repaired case passed a focused paid run. * ci(image): pin Claude Code 2.1.284, the version users run Request-body capture shows both 2.1.251 and 2.1.284 send effort "high" to claude-fable-5-1; 2.1.284 adds the model's own profile. The slower 2.1.284 census was mostly API latency: its SDK-only judges were 25% slower too. Nine previously slow cases pass on 2.1.284 within unchanged budgets. * test: one owner per case id, a structural devex 0B setup rule, and correct design/gbrain actors - plan-design-review-plan-mode was registered by two files; the PTY smoke is now plan-design-review-plan-mode-smoke, and a registry test requires one owner per case in case-sharded files. - plan-devex-finding-floor: the template's 0B narrative-confirmation question is classified as setup structurally instead of timing out a Haiku assessor. - setup-gbrain-remote: the actor accepted 'skip' on the MCP-registration question the test asserts; it now accepts that question and declines others. - design-review-plugin-handoff: the fake engine cited a file absent from the fixture repo and index.html linked a missing styles.css. Captured-question regressions with negative controls; each case passed a focused paid run. * test: PTY harness handles clipped reviews and bundled setup tabs; AUQ judge uses structured output; design-consultation carve declines optional outside voices - ceo mode routing: a Submit review taller than the viewport, a setup tab bundled after the mode tab, and a clip through the mode question each hung or misread the run; the native answer is still verified after Submit. - judgeRecommendation requests a 1-5 enum schema; a malformed Haiku reply had scored substance 0 for a 4/5 brief. Judge failures now propagate. - carve section-loading for design-consultation declines the optional outside voices (a supported path) and treats DESIGN.md as the report; timeout unchanged. The Step 0E handoff defect is not fixed (0/15 samples across four wordings, none shipped) and is filed in TODOS. * test: fold the design-consultation completion replay into carve-section-sharding (test-of-test ratchet) * docs(todos): record the pre-push hook shard-order hang * test(qa-callers): disable git auto maintenance in the fixture repo (same guard as shared-libs; from #3002) * test(office-hours-attempt): the fake judge SDK response carries stop_reason like the real API (structured judge requires end_turn) * fix(qa): the caller STOP line says to await the method Reads before any probe ship-exploratory-plan-checks: the model read exploratory.md and sent a capture in the same response, before seeing the section's own await rule. * fix(qa): number the qa value-bar questions from 1 and say reproduced bugs already answer the first two * fix(qa): define evidence.json where it is built, point the preparation gate at the next section, name measured command durations in the report template Recurring qa/qa-only workflow-judge complaints in CI (clarity/actionability 3.33). * fix(plan-eng-review,review): a disallowed question tool is not headless; report kept tests only when some were skipped * fix(plan-eng-review): keep the headless-rule contract phrases adjacent * fix(evals): cut path variance at its measured sources - gstack-qa-evidence capture prints startedAt/completedAt/durationMs and, for --deadline captures, remainingMs; the functional report takes durations from them. The section clock notice asks for one clock read up front instead of one after every checkpoint (QA runs spent 7-14% of tool calls on date -u). - ship plan-completion: skip the audit dispatch when discovery already found no plan (the dispatch-vs-skip conflict produced an optional 60-100 s subagent). - materialize/checkpoint validation errors state the expected schema, so a rejected annotations file is fixable in one call instead of blocking the phase. - session-runner counts turns from the transcript when a run times out, so timeouts stop reporting 'turn 0'. * fix(evals): count timeout turns only from object transcript events * test(qa-callers): deterministic child transport, completion-time handoff reads, compact phase report The exploratory caller cases exist to prove the caller starts and bounds exploratory QA. Their native adversarial reviewer (review) and plan audit (ship plan-checks) now come from recorded child outputs instead of a live subagent, handoff freshness reads are required before completion records rather than every bookkeeping log, and the phase report is compact. Measured: 194-257 s per case against 208-284 s before, no subagent calls. * test(ship-docsync): seed fault cases at their gate instead of replaying attempt 1 The post-dispatch fault cases (missing-marker, launch-failure, timeout-unsettled, late-result, stale-before, stale-after, recovery) now start from a fixture-owned attempt 1: the real actor prepares and dispatches it, its verbatim output is saved once, and the invocation journal carries its pre-dispatch entry with the child asset hashes. The model resumes at Parent processing with a trimmed read list, inspect named as the authoritative repository observation, and recovery's intermediate checkpoint folded into the next attempt's pre-dispatch entry. Assertions count only parent-issued transport events and require a read of the saved attempt-1 output; missing-asset and the legacy failure case keep the full model-driven first attempt, and their prompts are byte-identical. * test(ship-docsync): name the seeded read list and cap journal/report length The first seeded stale-before run spent calls locating documentation.md (two ls sweeps), reading through cat and re-Reading the record before Edit, and ~40 s composing 1.5-2.2 KB entries and report. Name every seeded read path, ask for native Read, and bound entry/report length. * test(ship-docsync): trim the seeded parent's measured model time Measured on the seeded runs: one read the 78 KB ship/SKILL.md, the post-child freshness comparison spent 18-32 s of thinking over full inspect contents, and the final response restated the report (~1.1 KB). Say the phase excerpt stands in for ship/SKILL.md, compare hashes first and read content only for changed paths, and end with one status line. * feat(qa-evidence): enforce the checkpoint sequence and fill report bookkeeping in code - capture refuses to run another probe until a checkpoint anchored on the latest complete capture names this capture as its next command, and every complete capture prints that requirement. - materialize fills revision, runtime, cwd and learning (checkpoints whose next native command differs) when omitted and prints the reportLinks the report must include; the QA section shrinks accordingly. * test(qa-callers): hand the caller phase its invocation-start observations and review token; fix(next-version): fetch without auto maintenance - Every caller case receives the diff, status, log, untracked list, HEAD and an already-captured review start token, so the phase spends its budget on the contract under test instead of re-running setup reads. - gstack-next-version's fetches pass --no-auto-maintenance. On git 2.55 a completed fetch forks detached maintenance in the caller's repository; the free suite's live smoke test ran it inside the CI checkout, and every shard-12 pre-push hook hang so far followed a completed smoke fetch. * feat(deslop-shared-libs): route every Git read through bin/gstack-safe-git The skill made the model retype a long safe-Git prefix on each call and a dropped flag failed shared-libs-read-only. bin/gstack-safe-git applies the fixed env + flag prefix, adds --no-ext-diff --no-textconv to log/show/diff, allows diff only between two explicit object IDs and ls-files only in the NUL-delimited overlay form, and refuses every other shape with one line naming the allowed forms. The template now points at the installed helper (host global runtime via {{SAFE_GIT}}) and drops the prose it enforces. Fixtures resolve the helper to this checkout, the git shim records the safety environment, and isGuardedGitRequest requires the complete prefix (env included) for every repository read. * test(shared-libs): tee to a discard device is not a file write Paid shared-libs-opportunity-judgment t1 on 1213b01 failed read-only on '... | tee /dev/null | sha256sum'. The detector flagged any tee operand while the same devices are allowed for redirection. tee now fails only when an operand is a real file; tee to a file, -a file and -- -a stay violations. * fix(qa-evidence,observer): reject placeholder metadata and replay-only learning; declare the docs atomic-write target - materialize measures revision, runtime and cwd itself and rejects supplied values that differ (CI run wrote revision "HEAD" and runtime "bun"), and refuses learning checkpoints that replay the same probe, naming the fix. - The docs write observer treats Claude Code's atomic temp for the authorized doc target as transient, so a temp renamed before its per-file watch no longer marks the observation incomplete (ship-docsync-completion flake). Per-file monitoring outside declared targets stays fail-closed. * test(qa-functional): fix mode requires only the happy scenario from the model (carried byte-identical from #3002 183b01f4..3e6074b4) verifyQANativeRegression already reruns all eight webhook scenarios on the repaired source, so the model-side eight-scenario requirement in fix mode duplicated harness coverage and pushed qa-functional-webhook-fix past its budget. qa-only still requires every scenario. * fix(deslop-shared-libs): probe the audited repository with -C <repo> A CI run probed safe-git from the session directory above the target repo, so the capability probe never touched the repository and the run fell back to the API without a local attempt. The probe (and any call from elsewhere) now names the audited repository. * test(qa-deadline): never attach a reader to the full-pipe fixture's stdout The full-pipe receipt test attached a 'data' listener (flowing mode) and then paused; on CI the reader could drain the 2 MB write before the pause, so the receipt write never blocked and the helper exited 0 in ~126 ms. The stdout pipe now stays unread until the assertion, which is what the test means to model. * feat(qa): helpers answer --help, and the QA eval interfaces declare it Approved by Garry: asking gstack-qa-evidence or gstack-qa-deadline for usage is read-only, so both helpers print usage and exit 0 on --help (the evidence usage now names the annotation shape), and the functional and caller command allowlists accept exactly 'bun <path>/bin/gstack-qa-{evidence,deadline} --help'. Two CI runs failed only on that call. * fix(qa): after an input change, a probe is affected unless shown otherwise CI late-input run finished in time but revalidated only the happy probe after the locale input changed and reported the stale adverse probe green. The revalidation step now treats any probe not shown to be unaffected as affected. * test(shared-libs): seed the lifecycle replay's first Step 3 pass instead of replaying it shared-libs-review-lifecycle ran ~88% of its 300 s session budget (12-run census median 265 s, 4/24 sessions timed out). The fixture now executes pass 1's Step 3 once with the real logger and Git: a real unused REVIEW_START, then the diff, inventories, attributes/config/index flags, gstack-review-read output and every file's bytes and sha256, saved to one observation. The model resumes at Step 4 with an exact four-file first read, the observation named as the authoritative pass-1 repository read, one post-fix verification, an explicit pass-2 read list and a twelve-line summary. Pass 2 still runs its own --start, diff, reads, fingerprint and stage actor before --finish. The actor scope now states that a current settled final-pass actor result supplies the replaced QA/adversarial prerequisites and that the no-credit disclosure is a reporting label: one r1 session persisted completed:false from that ambiguity. New assertions: the final binding never uses the seeded token's start or tree, and the observation was read; free controls finish the seeded token (binding changed) and omit the observation read, and both fail. * test(shared-libs): trim the resumed review replays' setup and report Every sibling review session (revalidation, path-eligibility, index-flags, prior-coverage) loaded qa/sections/exploratory.md and often scope.md although its QA and native adversarial results are supplied synthetic inputs, then spent a second request on shared-code-reuse.md and base metadata. The resumed scope now states that the supplied results replace Step 4's QA method loading; the revalidation contract names one first response (workflow, checklist, finding, prerequisites, shared-code-reuse.md, base metadata) and caps the summary at twelve lines. Receipt order, direct source reads, the checker, the question and final persistence are unchanged. * fix(review): define what a Step 5c Skip option says Step 5c named "B) Skip" without saying what its description may claim. Two CI captures (path-eligibility on131d43be, index-flags on4643cb85) offered a Skip whose description added effects beyond declining: "The extraction can be applied in a later editing review pass" and "replacing the invalidated prior Skip". Those read as change commitments, so the no-change actor refused both. Step 5c now says to describe Skip only as no code/index change with the Skip recorded; adjacent lines are compacted so the review parity caps hold unchanged. Both exact packets are kept as a free regression: still refused, and accepted once Skip follows the rule. The actor's classifier is unchanged. * fix(qa-evidence): every complete capture needs an evidence row; test(tpa): accept the hyphenated app-specific-password spelling - materialize refuses when a complete capture has no evidence row and is not named in limits (CI cli-report omitted capture 004), naming the missing IDs. - tpa-apple-ban's detector required 'app-specific password' with a space; the CI answer said 'app-specific-password path' and was otherwise correct. * test(qa-observer): fix mode treats atomic temps of authorized src/test writes as transient CI webhook-fix failed with 'Could not watch test/worker.regression-1.test.ts.tmp...': Claude Code's Write renamed its temp before the per-file watch was added. The functional eval now tells the observer its mode, and a temp whose target that mode may write is observed through its directory watch. Report-only mode and undeclared paths keep failing closed. * feat(qa-evidence): refuse evidence observed on an older input snapshot than the latest capture When native probe output declares a top-level input snapshot, materialize compares each evidence row with the latest capture's snapshot and refuses stale rows unless they are classified superseded, naming the captures to rerun. ship-exploratory-late-input kept reporting a pre-change adverse probe green after the input changed. * test(qa-functional): point the fixture at the helper's --help instead of its source A CI webhook-fix run spent three turns reading lib/qa-evidence.ts to learn the interface and timed out just before materialize (agreed with #3002's owner). * feat(qa-evidence): captures list the caller's declared-but-unrun required probes GSTACK_QA_REQUIRED_PROBES (a JSON array of native child commands) makes every capture print requiredRemaining; it never judges pass or fail. The functional eval passes the webhook list from QA_WEBHOOK_REQUIRED_SCENARIOS, which the verdict now reads too, so the nudge and the verdict share one source (agreed with #3002's owner). CI webhook-report kept stopping with scenarios unrun. * test(review-army): record N+1's pre-dispatch stages and scope the session to Step 4.5 review-army-perf-n-plus-one timed out in 7 of 13 CI runs on this branch (passing 245-280 s of 300). Each session spent ~95 s on setup (the full extracted SKILL, checklist, section greps, exploratory.md, diff-scope/stats/learnings, tooling checks), ran Step 4's core pass, a search-before-recommending WebSearch, and wrote a 10-16 KB report (~100 s after the Red Team returned). The fixture now stages only review/sections/review-army.md plus the performance and red-team checklists, and hands the session the recorded detect-scope, specialist-stats and learnings outputs and the diff. The caller passes --performance (every CI parent already treated the prompt as that force flag against the <50-line skip), declares the core pass, QA, adversarial review, web research, Fix-First and persistence out of scope, and caps the report at the selection line, the SPECIALIST REVIEW block and the Red Team result (30 lines). The Performance specialist and the conditional Red Team are still real foreground subagents, and the report still has to surface the N+1. New assertion: a foreground Performance specialist dispatch precedes the Red Team dispatch. Free controls omit the Performance dispatch or background it, and both fail; the budget lifecycle adapter supplies the current result shape. Touchfiles now include the .rb fixture the case reads. * test(review-army): share the recorded Step 4.5 staging with consensus and supply its Red Team review-army-consensus (periodic) timed out in 2 of 13 census sessions; passing runs took 213-297 s of 300. Like N+1 it spent ~30-50 s reading the whole extracted SKILL, checklist and every specialist file, sometimes dispatched an unrequested Maintainability specialist, then ran a Red Team (60-70 s) and a second merge before writing a 9-15 KB report. The N+1 staging and scope text move into stageReviewArmySession / reviewArmyScope / reviewArmyChecklists (the N+1 prompt renders byte-identical). Consensus now records its detect-scope, stats, learnings and diff, stages the Review Army section with the security and testing checklists, forces --security --testing, and caps the report like N+1. Its Red Team is outside the multi-specialist contract, so the fixture supplies a labeled synthetic NO FINDINGS result instead of a dispatch. The existing SQL-finding and browser-error assertions are unchanged; the lifecycle adapter's spawnSync now returns the git output the staging reads. * docs(changelog): v1.91.10.0 records the flake census and its repairs * test(strict-output): give the spool-prefix child time to finish before the pending stream times out windows-free-tests failed on9a7a7e54: the 150 ms shared deadline raced Bun startup on Windows, so the child was killed mid-write and the spool held a partial payload. Only the never-released extra stream should time out; the child now has 3 s. * fix(qa-evidence): accept a single limits string; test(qa-callers): read the handoff first when a probe snapshot changes CI late-input spent a turn rewriting limits as an array after materialize refused a string, and a ten-read sweep hunting for the changed input before it read reports/HANDOFF.md, then timed out at 300 s. * test(autoplan-dual-voice): unwrap the framed native report before Claude Code 2.1.284's agentId/usage trailer * test(section-loading): credit a Bash print that contains every line of the carved section * test(auto-decide): ask for the selected mode in the skill's mode handoff line, not a separate public decision * test(plan-ceo floor): scope preservation approves no premise, approach or remedy * test(autoplan-dual-voice): the fixture declares that delivered bash blocks run alone, diagnostics separately * test(coverage-audit): a fenced plain-word caption in a successful && read chain is display only Census 36776104571 plan-eng capture read both owned files with cat -n in one successful && chain; the caption 'echo "=== git diff main --stat ==="' fell outside the two-token caption grammar, so both reads lost credit. Accept a fenced caption of plain words; unfenced command strings, expansions, redirection, -e escapes and ; / || tails stay rejected. * test(office-hours): a fork whose outer options are the seeded shapes is the Phase 4 question Census trials 1-2 captured complete Phase 4 forks (A) Server-side B) Client-side C) Hybrid, recommendation with because) whose prose used none of the vocabulary words. Accept two seeded shapes as outer options as Phase 4 specificity; the earlier-phase, nested, fenced and single-shape controls still fail. * fix(review): design-lite rows keep the detector's [rule-id]; the e2e detector rows point at the diff The output template had no rule-id slot, so rows merged with checklist items dropped the detector id (census t2, local t1). Rows now carry [rule-id]. The fake engine's sample rows named a foreign fixture path at line 0; the e2e remaps them to landing.html/styles.css so trials stop spending turns reconciling it. * test(shared-libs): the plan actor reads scheduler parity and unchanged-scope lists Census 36776104571's question preserved the contract ('behaving exactly like the scheduler', 'scheduler parity holds by construction') and excluded work with 'Existing copies and helper hardening stay unchanged'. Accept exactly/parity as preservation (negated forms refuse) and a bare noun list that stays unchanged as an exclusion for the expansion scan only; verb-led clauses still refuse. * fix(qa-only,qa): name the exploratory read point and finalization order; judge qa with its browser assets qa-only judges cited 'next section' pointing at the wrong heading, an exploratory trigger that contradicted its read point, clock ownership in mixed runs and the unstated order of exploratory section 4 vs reporting. The qa judge penalized the absent qa-report-template and issue-taxonomy that qa-patterns loads; with them in, it found issue-taxonomy's dangling 'rule 13' (the consent rule is browser rule 3). * test(ship-docsync): seeded attempt 1 counts toward the limit; transport counts ignore calls that never reached the state file - CI launch-failure retried after the seeded attempt 1 as if that attempt were the fixture's; the seeded prompt now says attempt 1 is this invocation's and a further attempt needs what Blocked recovery requires. - A late-result run typo'd the state path once (ENOENT, the actor never ran), then repeated the call correctly; the per-action count compared both calls with one actor event. Only calls naming the real state file are counted. * fix(plan-eng-review): show the accepted dedicated read form for coverage-diagram sources CI plan-eng-coverage-audit mixed package/config and git diff into the source read; the review variant, whose prompt shows the && display form, does not. The plan trace step now shows it too, within the unchanged size cap. * test(sync-gbrain-readiness): a negation earlier in the claim clause is not a search/write readiness claim The census unknown actor wrote 'nothing about read, search, or write capability is confirmed either way' after a YELLOW/WARN verdict. The claim window started at 'write', so the leading 'nothing' was outside it. Check the clause subject for nothing/neither/none/no; keep the original in-claim negations. Replay of the captured output passes; positive controls still flag an unnegated claim. * fix(office-hours): a forcing question's recommendation takes the position the founder's words support auq-matrix office-hours asked D1 Demand as options about the founder's own evidence and, with no rule for that shape, recommended 'answer whichever is TRUE — A is marked recommended only because it is the strongest position' (substance 2). Say what such a recommendation is: the option the founder's own words support, why it matters for the next step, and what would change it. * fix(plan-ceo-review): name the mode preference command and the exact handoff line auto-decide-preserved at6fcb0981: the model never ran the preference check, read 'check ... through the preamble' as already done, auto-selected 'per your preference setting', and wrote 'Selected mode: HOLD SCOPE, auto-decided from your tuned preference' instead of the AUTO_DECIDE handoff line. At9a7a7e54it ran the check but wrote 'Decision: HOLD SCOPE is the review mode for ...'. Neither matched the handoff template the observer recognizes. Name gstack-question-preference --check at the point of use and say the handoff begins with the exact matching line. Collapse the audit block's comment padding to stay within the unchanged 80150-byte skeleton cap. * test(section-loading): record the CEO capture's report and transcript The6fcb0981census failed hasStaleFillRaceFinding (line 98), but the case records nothing beyond junit, so the report the detector judged is gone. Return the SkillTestResult from captureSectionReads and record it, with the full saved report, through the eval collector on pass and fail. * test(design): plan-mode names its read list and caps its additions and summary At6fcb0981plan-design-review-plan-mode timed out at 300 s (9 turns): 22 cat/sed chunk reads (~50 s), then a 28 KB plan Write (~150 s), before the read-back finished. The9a7a7e54pass took 240 s with a 24.6 KB Write. Read SKILL.md, review-sections.md and plan.md natively in one response, keep additions under 14,000 characters and the summary within ten lines. Budgets unchanged. * test(plan-mode-no-op): require prose evidence before a waiting verdict ends eng/design runs (carried byte-identical from #3002) With the prose fallback forced, the gate renders as a lettered menu; a judge 'waiting' verdict on a spinner-only frame ended the run as 'asked' before the menu rendered, so the scope-gate check failed on unchanged behavior. * feat(qa-evidence): materialize computes the phase verdict; callers must report it Approved by Garry: the helper, not the model, decides whether evidence can pass. materialize writes verdict {status, open} into evidence.json and prints it: fail or blocked from row classifications, inconclusive while any row is superseded, a complete capture is withheld, a declared required probe is unrun or there is no evidence, else pass. The caller fixture requires receipt.status to equal that verdict. CI late-input kept reporting pass with a superseded happy probe. * test(qa-callers): compare the receipt with the helper verdict only when evidence.json was materialized The producer free tests run captures without materialize; evidence.json is optional for callers, so its absence is not a verdict mismatch. * test(llm-judge): run the ship workflow judge at medium effort so its panel fits JUDGE_MS claude-fable-5-1 accepts only adaptive thinking (thinking.type.enabled with budget_tokens returns 400), so effort is the available thinking control. Measured on the exact ship judge request (105,301 input tokens): - default effort, 18 samples: thinking 5,086-10,881 tokens, 75.9-144.7 s; 3 of 18 passed the 120 s deadline (about 42% of 3-sample panels). - medium effort, 18 samples: thinking 2,749-5,762, output at most 6,144 tokens, 43.1-77.9 s; scores 4/4/4 in 16 of 18 (clarity 3 in two), versus 14 of 18 at default. callJudge gains an effort option sent as output_config.effort; only the ship judge sets it. Rubric, floors, panel size, deadline, model and max_tokens are unchanged. The cache identity records effort. * test(llm-judge): ask frontier workflow judges for 120-word reasoning under the unchanged 150-word check Told "under 150 words", the ship judge's reasoning landed at 130-156 words (3 of 18 probe samples at 152-156), so the structured-response check failed about one panel in three independent of effort. The prompt's frontier block and the response schema description now say under 120 words; the validator still rejects 150 words or more. The changed prompt bytes reach only the two frontier judges: ship/SKILL.md workflow (prompt and schema) and review/SKILL.md workflow (prompt). * test(llm-judge): type the stream transport mock call * test(plan-ceo floor): the request answers only the questions it names PR lane 36794871032 (head20d6e98f): the CEO floor ran 608 s without a question. Its Step 0 recorded the premise gap and approach choice as unresolved ledger rows, then said "this session supplies all answers up front, so no decision brief was dispatched" and wrote Sections 1-11.2734e203stopped scope preservation from approving the premise; this time the actor block (declined setup, recall, outside voices, HOLD SCOPE) and the fixture's "complete user request is available from the start" were read as pre-answering every review question. The CEO actor now states that the request answers only the routing, recall, outside-reviewer and review-mode questions it names. * test(plan-devex floor): a 'Partly wrong' 0B answer is the narrative confirmation PR lane 36794871032: the DX floor asked its D1 narrative confirmation (Accurate, proceed / Partly wrong, correct it / Way off, actual is...). The deterministic setup rule accepted only 'Some ... wrong', so the question went to the LLM assessor, which hit its 30 s spawnSync cap (ETIMEDOUT) and ended the case as assessment_error at 141 s, the same failure as census 36641820398. The rule now accepts 'partly' beside 'some'; the captured question is a free regression and the remedy-option controls still go to the assessor. * test(design-review plugin handoff): quoted report text is not an install command PR lane 36794871032: every behavioral check passed except noInstallOrOverride, which matched "no `npx impeccable`" inside the quoted heredoc that wrote detector-output.md. Nothing was installed or downloaded. The check now drops quoted-delimiter heredoc bodies (literal data) before matching; unquoted bodies, which can expand $(...), and unterminated bodies stay checked. Free controls cover the captured write, bare npx, an IMPECCABLE_BIN override, an unquoted $(npx ...), npx after the delimiter and an unterminated body. * test(review-army delivery audit): stage only the plan-completion section and record its git reads PR lane 36794871032: the case timed out at its 120 s budget after 7 turns (previous lane passed in 45 s). The session read the 46 KB extracted SKILL in three passes (cat to persisted output, grep, sed), ran its own git reads, wrote a 74-line report, then inspected and ran gstack-learnings-log and rewrote the report's Learnings section. As in the Step 4.5 cases (17ee2e54/2bd4651c), the fixture now stages only review/sections/plan-completion.md, hands the session the recorded git log and diff, declares the HIGH-impact question, its Scope Check, learnings logging and later steps outside the capture, and caps the report at the audit block and its DISCREPANCY entries (30 lines). The NOT DONE and email assertions are unchanged. * feat(qa-evidence): one capture call records the causal note for the previous capture capture R NNN [--public] (--deadline D|--timeout-ms MS) --after PREV --hypothesis 'TEXT' -- CMD publishes exploration-NNN.json {observationCapture, observationArgv, observed, hypothesis, nextCapture, nextArgv} before running CMD, refusing unless PREV is the latest complete capture. The receipt carries checkpoint/checkpointSha256; validators bind the note to the transcript's capture calls by capture ID and receipt hash instead of exact command strings. The separate checkpoint command and the capture guard keep working; materialize learning accepts both note shapes and still rejects same-probe replays. Prose and eval fixture prompts teach the merged form. * fix(qa-evidence): a superseded row stops holding the verdict open once its probe is rerun on current inputs materialize requires an old-snapshot row to be classified superseded, and its verdict kept every superseded row open, so rerunning the probe (what its own error tells the model to do) could never reach pass; late-input reran 3 and 9 on the new snapshot and still got inconclusive. A superseded row now closes only when a non-superseded row with the same captured argv observed the current snapshot. Re-materializing an already-published evidence.json names the cause instead of failing generically. * test(plan-eng batching): count saved decisions whose label drops the (recommended) marker or whose report is titled 'Eng Review Report — <plan>' * fix(qa): browser-only runs skip annotations/materialize; only Q captures can anchor evidence rows * test(design): plan-mode length is a drafting target, not a check to measure and trim * test(llm-judge): structured output for doc, outcome and posture judges so reasoning quotes cannot break JSON * test(ship-docsync): steer skill file reads to Read; large cat output becomes an unpageable preview * docs(changelog): browser-only QA evidence and structured judge output * test(qa-only cleanup): refusal scenarios get a 1 s budget and an absolute worker deadline; 300 ms starved under parallel load * fix(office-hours, design-consultation): ask the goal question and read the mode section first; ask the memorable-thing question on its own * test(outside-disabled): a record named by the retained record's own clock and then disowned owns its completed status * test(context-skills): install gstack-paths in the fixture bin; without it the model guessed the checkpoint root * test(ceo mode routing): SCOPE EXPANSION posture credits plural 'expansions' * test(ship-docsync): name the unmet atomic-replacement check on a forbidden temp-file write * fix(qa): browser-only runs materialize an empty evidence list with checkpoints in limits, matching /qa-only * test(qa callers): an accepted review-log record may cite checkpoints as finding evidence * fix(plan-eng-review): state that a disallowed question tool never qualifies as headless before the headless action * merge follow-up: re-record paid CLI parity for #2999's flags; trim merged review, qa-only and plan-eng wording toward the size caps * test(golden): refresh codex/factory ship goldens for the trimmed caller QA wording * test(coverage-audit fixture): disable git auto maintenance so cleanup is not racing a detached git writer * test(parity): raise review, qa and plan-eng caps to the measured merged size of #2999 and #3002 (each fit alone), documented per cap * fix(qa-evidence): materialize rejects an unrecognized classification before publishing, so the one-shot verdict cannot be locked inconclusive by a descriptive label
gstack
"I don't think I've typed like a line of code probably since December, basically, which is an extremely large change." — Andrej Karpathy, No Priors podcast, March 2026
When I heard Karpathy say this, I wanted to find out how. How does one person ship like a team of twenty? Peter Steinberger built OpenClaw — 247K GitHub stars — essentially solo with AI agents. The revolution is here. A single builder with the right tooling can move faster than a traditional team.
I'm Garry Tan, President & CEO of Y Combinator. I've worked with thousands of startups — Coinbase, Instacart, Rippling — when they were one or two people in a garage. Before YC, I was one of the first eng/PM/designers at Palantir, cofounded Posterous (sold to Twitter), and built Bookface, YC's internal social network.
gstack is my answer. I've been building products for twenty years, and right now I'm shipping more products than I ever have. In the last 60 days: 3 production services, 40+ shipped features, part-time, while running YC full-time. On logical code change — not raw LOC, which AI inflates — my 2026 run rate is ~810× my 2013 pace (11,417 vs 14 logical lines/day). Year-to-date (through April 18), 2026 has already produced 240× the entire 2013 year. Measured across 40 public + private garrytan/* repos including Bookface, after excluding one demo repo. AI wrote most of it. The point isn't who typed it, it's what shipped.
The LOC critics aren't wrong that raw line counts inflate with AI. They are wrong that normalized-for-inflation, I'm less productive. I'm more productive, by a lot. Full methodology, caveats, and reproduction script: On the LOC Controversy.
2026 — 1,237 contributions and counting:
2013 — when I built Bookface at YC (772 contributions):
Same person. Different era. The difference is the tooling.
gstack is how I do it. It turns Claude Code into a virtual engineering team — a CEO who rethinks the product, an eng manager who locks architecture, a designer who catches AI slop, a reviewer who finds production bugs, a QA lead who opens a real browser, a security officer who runs OWASP + STRIDE audits, and a release engineer who ships the PR. Twenty-three specialists and eight power tools, all slash commands, all Markdown, all free, MIT license.
This is my open source software factory. I use it every day. I'm sharing it because these tools should be available to everyone.
Fork it. Improve it. Make it yours. And if you want to hate on free open source software — you're welcome to, but I'd rather you just try it first.
Who this is for:
- Founders and CEOs — especially technical ones who still want to ship
- First-time Claude Code users — structured roles instead of a blank prompt
- Tech leads and staff engineers — rigorous review, QA, and release automation on every PR
Quick start
- Install gstack (30 seconds — see below)
- Run
/office-hours— describe what you're building - Run
/plan-ceo-reviewon any feature idea - Run
/reviewon any branch with changes - Run
/qaon your staging URL or an isolated local API, CLI, job or webhook - Stop there. You'll know if this is for you.
Install — 30 seconds
Requirements: Claude Code, Git, Bun v1.0+, Node.js (Windows only). Recommended on macOS: the Aside browser (macOS 15+) — browser skills, /make-pdf, and /diagram drive it first, with your real logged-in sessions. Without it, ./setup builds gstack's own bundled browser and the same skills use that. /cso additionally needs a Bun release with all four --no-compile-autoload-* build flags plus a native toolchain: a static-capable C compiler on Linux, Xcode command-line tools on macOS, or Visual Studio 2022 Build Tools with Desktop development with C++ on Windows. If those are absent, setup installs everything else, removes stale CSO helpers, and /cso reports not assessed with the prerequisite.
When qualified CSO runtime images are published, setup gives each automatic preload a 30-second window plus a bounded setup allowance for the declared catalog. For slower registries, set an integer such as GSTACK_CSO_IMAGE_PULL_TIMEOUT_SECONDS=120 (accepted range: 5–300 seconds). One image timing out does not consume the remaining images' windows; setup reports partial progress and a later run resumes from exact digests already present in local Docker. The complete preload is capped at one hour.
Step 1: Install on your machine
Open Claude Code and paste this. Claude does the rest.
Install gstack: run
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setupthen add a "gstack" section to CLAUDE.md that says to use the /browse skill from gstack for all web browsing, never use mcp__claude-in-chrome__* tools, and lists the available skills: /office-hours, /plan-ceo-review, /plan-eng-review, /plan-design-review, /design-consultation, /design-shotgun, /design-html, /review, /deslop-shared-libs, /test-audit, /ship, /land-and-deploy, /canary, /benchmark, /browse, /connect-chrome, /qa, /qa-only, /design-review, /scrape, /setup-browser-cookies, /setup-deploy, /setup-gbrain, /retro, /investigate, /document-release, /document-generate, /codex, /cso, /autoplan, /plan-devex-review, /devex-review, /careful, /freeze, /guard, /unfreeze, /gstack-upgrade, /learn. Then ask the user if they also want to add gstack to the current project so teammates get it.
Step 2: Team mode — auto-update for shared repos (recommended)
From inside your repo, paste this. Switches you to team mode, bootstraps the repo so teammates get gstack automatically, and commits the change:
(cd ~/.claude/skills/gstack && ./setup --team) && ~/.claude/skills/gstack/bin/gstack-team-init required && git add .claude/ CLAUDE.md && git commit -m "require gstack for AI-assisted work"
No vendored files in your repo, no version drift, no manual upgrades. Every Claude Code session starts with a fast auto-update check (throttled to once/hour, network-failure-safe, completely silent).
Swap required for optional if you'd rather nudge teammates than block them.
OpenClaw
OpenClaw spawns Claude Code sessions via ACP, so every gstack skill just works when Claude Code has gstack installed. Paste this to your OpenClaw agent:
Install gstack: run
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setupto install gstack for Claude Code. Then add a "Coding Tasks" section to AGENTS.md that says: when spawning Claude Code sessions for coding work, tell the session to use gstack skills. Include these examples — security audit: "Load gstack. Run /cso", code review: "Load gstack. Run /review", QA test a URL: "Load gstack. Run /qa https://...", build a feature end-to-end: "Load gstack. Run /autoplan, implement the plan, then run /ship", plan before building: "Load gstack. Run /office-hours then /autoplan. Save the plan, don't implement."
After setup, just talk to your OpenClaw agent naturally:
| You say | What happens |
|---|---|
| "Fix the typo in README" | Simple — Claude Code session, no gstack needed |
| "Run a security audit on this repo" | Spawns Claude Code with Run /cso |
| "Build me a notifications feature" | Spawns Claude Code with /autoplan → implement → /ship |
| "Help me plan the v2 API redesign" | Spawns Claude Code with /office-hours → /autoplan, saves plan |
See docs/OPENCLAW.md for advanced dispatch routing and the gstack-lite/gstack-full prompt templates.
Native OpenClaw Skills (via ClawHub)
Four methodology skills that work directly in your OpenClaw agent, no Claude Code session needed. Install from ClawHub:
clawhub install gstack-openclaw-office-hours gstack-openclaw-ceo-review gstack-openclaw-investigate gstack-openclaw-retro
| Skill | What it does |
|---|---|
gstack-openclaw-office-hours |
Product interrogation with 6 forcing questions |
gstack-openclaw-ceo-review |
Strategic challenge with 4 scope modes |
gstack-openclaw-investigate |
Root cause debugging methodology |
gstack-openclaw-retro |
Weekly engineering retrospective |
These are conversational skills. Your OpenClaw agent runs them directly via chat.
Other AI Agents
gstack works on 10 AI coding agents, not just Claude. Setup auto-detects which agents you have installed:
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/gstack
cd ~/gstack && ./setup
Or target a specific agent with ./setup --host <name>:
| Agent | Flag | What you get |
|---|---|---|
| OpenAI Codex CLI | --host codex |
Full install → ${CODEX_HOME:-~/.codex}/skills/gstack-*/ |
| OpenCode | --host opencode |
Full install → ~/.config/opencode/skills/gstack-*/ |
| Cursor | --host cursor |
Full install → ~/.cursor/skills/gstack-*/ |
| Factory Droid | --host factory |
Full install → ~/.factory/skills/gstack-*/ |
| Kiro | --host kiro |
Full install → ~/.kiro/skills/gstack-*/ |
| Slate | --host slate |
Pointer to the Claude install (Slate reads .claude/skills as a fallback) |
| OpenClaw | --host openclaw |
ACP spawn pointers + methodology artifacts via gen:skill-docs --host openclaw + the instruction-only digest below (full guide: docs/OPENCLAW.md) |
| Hermes | --host hermes |
Methodology artifacts via gen:skill-docs --host hermes + the instruction-only digest below |
| GBrain (mod) | --host gbrain |
Brain-aware skill variants, shipped from the GBrain repo |
Outside reviews require the selected CLI to be installed and authenticated: Claude Code when using gstack in Codex, or Codex on other harnesses. External harnesses discover these commands as /gstack-claude-code and /gstack-codex; each harness omits its own wrapper. Explicit provider requests keep that provider. The existing codex_reviews setting controls automatic outside reviews where supported, regardless of the provider selected.
/claude has been renamed to /claude-code. Re-run ./setup --host <name> to migrate managed installations, including other harnesses sharing the checkout. Setup preserves the previous installation if replacement generation or installation fails and prints repair instructions.
Instruction-only tier (any rules-reading agent — Zed, Amp, Jules, side projects):
copy the 2KB digest at agents-digest/gstack-AGENTS.md
into a location your agent reads (for example, append it to your project's AGENTS.md).
It carries gstack's ethos, reuse ladder, and voice rules — no install required. The
digest's first line shows its gstack version; re-copy it after upgrading.
For Codex, setup reads the top-level model from
${CODEX_HOME:-~/.codex}/config.toml and generates the matching behavioral
profile, falling back to gpt-6-astra when no usable model is configured.
gpt-5.6-sol automatically receives bounded-scope instructions that
finish the requested lake without expanding into adjacent cleanup or speculative
hardening. The Sol profile is exact-match only: dated snapshots and other 5.6
variants get the generic GPT profile, and setup warns on near-misses like
gpt-5.6-sol-2026-08-01. Override detection with ./setup --host codex --model <id> — the
override applies to that run only; set model in your Codex config.toml to
make it stick across upgrades. After changing your Codex model, rerun
./setup --host codex to regenerate the skills.
gstack-owned Codex invocations and evals default to gpt-6-astra. Set
GSTACK_CODEX_MODEL=<model> to override that runtime default; an explicitly
requested model takes precedence. Runtime model selection is separate from
the setup-time behavioral profile above. /claude-code (gstack-claude-code
on Codex) preserves Claude's configured model. Set GSTACK_CLAUDE_MODEL=<model>
or name a model in your request to override it for the invocation, including
resumed consultations. See eval defaults and overrides
for capture, judge, and benchmark model selection.
Want to add support for another agent? See docs/ADDING_A_HOST.md. It's one TypeScript config file, zero code changes.
See it work
You: I want to build a daily briefing app for my calendar.
You: /office-hours
Claude: [asks about the pain — specific examples, not hypotheticals]
You: Multiple Google calendars, events with stale info, wrong locations.
Prep takes forever and the results aren't good enough...
Claude: I'm going to push back on the framing. You said "daily briefing
app." But what you actually described is a personal chief of
staff AI.
[extracts 5 capabilities you didn't realize you were describing]
[challenges 4 premises — you agree, disagree, or adjust]
[generates 3 implementation approaches with effort estimates]
RECOMMENDATION: Ship the narrowest wedge tomorrow, learn from
real usage. The full vision is a 3-month project — start with
the daily briefing that actually works.
[writes design doc → feeds into downstream skills automatically]
You: /plan-ceo-review
[reads the design doc, challenges scope, runs 10-section review]
You: /plan-eng-review
[ASCII diagrams for data flow, state machines, error paths]
[test matrix, failure modes, security concerns]
You: Approve plan. Exit plan mode.
[writes 2,400 lines across 11 files. ~8 minutes.]
You: /review
[AUTO-FIXED] 2 issues. [ASK] Race condition → you approve fix.
You: /qa https://staging.myapp.com
[opens real browser, clicks through flows, finds and fixes a bug]
You: /ship
Tests: 42 → 51 (+9 new). PR: github.com/you/app/pull/42
You said "daily briefing app." The agent said "you're building a chief of staff AI" — because it listened to your pain, not your feature request. Eight commands, end to end. That is not a copilot. That is a team.
The sprint
gstack is a process, not a collection of tools. The skills run in the order a sprint runs:
Think → Plan → Build → Review → Test → Ship → Reflect
Each skill feeds into the next. /office-hours writes a design doc that /plan-ceo-review reads. /plan-eng-review writes a test plan that /qa picks up. /review catches bugs that /ship verifies are fixed. Nothing falls through the cracks because every step knows what came before it.
| Skill | Your specialist | What they do |
|---|---|---|
/office-hours |
YC Office Hours | Start here. Six forcing questions that reframe your product before you write code. Pushes back on your framing, challenges premises, generates implementation alternatives. Design doc feeds into every downstream skill. |
/plan-ceo-review |
CEO / Founder | Rethink the problem. Find the 10-star product hiding inside the request. Four modes: Expansion, Selective Expansion, Hold Scope, Reduction. |
/plan-eng-review |
Eng Manager | Lock in architecture, data flow, diagrams, edge cases, and tests. Forces hidden assumptions into the open. |
/plan-design-review |
Senior Designer | Rates each design dimension 0-10, explains what a 10 looks like, then edits the plan to get there. AI Slop detection. Interactive — one AskUserQuestion per design choice. |
/plan-devex-review |
Developer Experience Lead | Interactive DX review: explores developer personas, benchmarks against competitors' TTHW, designs your magical moment, traces friction points step by step. Three modes: DX EXPANSION, DX POLISH, DX TRIAGE. 20-45 forcing questions. |
/design-consultation |
Design Partner | Build a complete design system from scratch. Researches the landscape, proposes creative risks, generates realistic product mockups. Writes DESIGN.md in the open DESIGN.md format, so impeccable, Google Stitch, and any tool that reads it share one file. |
/review |
Staff Engineer | Find the bugs that pass CI but blow up in production. Auto-fixes the obvious ones. Flags completeness gaps. Advisory simplification lens flags over-built code — never blocks, never auto-applies. |
/deslop-shared-libs |
Shared Code Reviewer | Find worthwhile shared-code extractions in recent work. Compares up to five opportunities and recommends the best three, with source evidence, reliability gains, and total code savings. Recommendations only. |
/test-audit |
Test Auditor | Sweep existing tests for ones that cost more than they protect: source greps, duplicates, assertion-free probes, test-only exports. Every candidate carries an evidence card; report-only unless you approve a batch. |
/investigate |
Debugger | Systematic root-cause debugging. Iron Law: no fixes without investigation. Traces data flow, tests hypotheses, stops after 3 failed fixes. |
/design-review |
Designer Who Codes | Same audit as /plan-design-review, then fixes what it finds. Atomic commits, before/after screenshots. If you have impeccable installed, its engine runs first and every mechanical finding arrives tagged with its rule id. |
/devex-review |
DX Tester | Live developer experience audit. Actually tests your onboarding: navigates docs, tries the getting started flow, times TTHW, screenshots errors. Compares against /plan-devex-review scores — the boomerang that shows if your plan matched reality. |
/design-shotgun |
Design Explorer | "Show me options." Generates 4-6 AI mockup variants, opens a comparison board in your browser, collects your feedback, and iterates. Taste memory learns what you like. Repeat until you love something, then hand it to /design-html. |
/design-html |
Design Engineer | Turn a mockup into production HTML that actually works. Pretext computed layout: text reflows, heights adjust, layouts are dynamic. 30KB, zero deps. Detects React/Svelte/Vue. Smart API routing per design type (landing page vs dashboard vs form). One slop-gate pass through the impeccable engine when you have it. The output is shippable, not a demo. |
/qa |
QA Lead | Explore browser, API, CLI, job and webhook behavior. Reproduce bugs, write failing regressions, fix the cause and re-verify before committing. |
/qa-only |
QA Reporter | Explore and report with replayable evidence. Suggest regression cases without changing product code or tests. |
/pair-agent |
Multi-Agent Coordinator | Share gstack's own browser with any AI agent. One command, one paste, connected. Works with OpenClaw, Hermes, Codex, Cursor, or anything that can curl. Each agent gets its own tab. Auto-launches headed mode so you watch everything. Auto-starts ngrok tunnel for remote agents. Scoped tokens, tab isolation, rate limiting, activity attribution. (Runs on the bundled browser — the fallback engine; agents driving Aside just open their own tabs.) |
/cso |
Chief Security Officer | Security audit with an application model, supported findings, independent challenge, and explicit coverage. Static assessment remains available without catalog profiles. With matching qualified profiles, comprehensive mode adds contained runtime/scanner execution and reviewable repair candidates for Node/Bun, Python, and Rails. Runtime-tested bundles authenticate separate external assertions. Project-test completion remains self_reported because target code controls the test process; tested is reserved for a future target-independent completion witness. |
/ship |
Release Engineer | Sync main, run tests, explore changed behavior, audit coverage and docs, then verify, push and open a PR. |
/land-and-deploy |
Release Engineer | Merge the PR, wait for CI and deploy, verify production health. One command from "approved" to "verified in production." |
/canary |
SRE | Post-deploy monitoring loop. Watches for console errors, performance regressions, and page failures. |
/benchmark |
Performance Engineer | Baseline page load times, Core Web Vitals, and resource sizes. Compare before/after on every PR. |
/document-release |
Technical Writer | Audit changed behavior against project docs on every ship, before final verification and publication. Also runs standalone. Shows updated, reviewed/current or blocked docs and any remaining gaps. |
/document-generate |
Documentation Author | Generate missing docs from scratch using the Diataxis framework. Researches the codebase first, then writes reference / how-to / tutorial / explanation docs that actually match the code. Invokable standalone or chained from /document-release when the coverage map finds gaps. Learn more: tutorial • how-to • why Diataxis. |
/retro |
Eng Manager | Team-aware weekly retro. Per-person breakdowns, shipping streaks, test health trends, growth opportunities. /retro global runs across all your projects and AI tools (Claude Code, Codex, Gemini). |
/browse |
QA Engineer | Give the agent eyes. Drives your Aside browser first — your real sessions, real clicks, real screenshots — through deterministic aside repl scripts. No Aside? It falls back to gstack's own Chromium: real clicks, ~100ms per command, and /open-gstack-browser shows it headed with sidebar, anti-bot stealth, and auto model routing. Every other browser skill stands on it. |
/scrape |
Data Extractor | Pull structured data off a web page — tables, lists, prices — in your Aside browser with the page's real logged-in state. On the fallback browser, /skillify turns the flow into a permanent browser-skill that runs in ~200ms next time. |
/setup-browser-cookies |
Session Manager | Copy selected cookies from Chrome, Chromium, Brave, Edge, Windows-only Opera and Opera GX, or macOS-only Comet, Arc, and Dia into gstack's bundled browser. Choose your profile and domains; copying and sign-in verification are separate. Only needed on the fallback path — Aside already has your sessions. |
/autoplan |
Review Pipeline | One command, fully reviewed plan. Runs CEO → design → DX → eng review automatically (eng always last, so the shipping gate reviews the final amended plan) with encoded decision principles. Surfaces only taste decisions for your approval. |
/spec |
Spec Author | Turn vague intent into a precise, executable spec in five phases (why, scope, technical with mandatory code-reading, draft, file). Outside-review quality gate before filing (Claude Code on Codex; Codex on other harnesses; blocks below 7/10), fail-closed secret redaction, dedupe against existing issues, archive to $GSTACK_STATE_ROOT/projects/$SLUG/specs/ for team-corpus recall. --execute spawns claude -p in a fresh worktree; /ship auto-closes the source issue on merge. Plan-mode aware. |
/learn |
Memory | Manage what gstack learned across sessions. Review, search, prune, and export project-specific patterns, pitfalls, and preferences. Learnings compound across sessions so gstack gets smarter on your codebase over time. |
/make-pdf |
Publisher | Markdown in, publication-quality document out. Mermaid and excalidraw fences render as vector diagrams, fully offline. Images scale to the page and never truncate; wide diagrams get their own landscape page. --to html emits one self-contained file, --to docx a Word doc. |
/diagram |
Diagram Maker | English in, editable diagram out. Emits a triplet: mermaid source, .excalidraw you can open and edit on excalidraw.com (hand-drawn style), and rendered SVG/PNG. Zero network. Embed the source in markdown and /make-pdf renders it. |
QA without a webpage
Use the same commands for browser and non-browser software. Start in a repository with its documented native command and an isolated local fixture; name the target and the behavior you want checked. For example:
/qa-only Test this repo's CLI using its documented local fixture. Check valid and invalid input, exit codes, stdout/stderr, and cancellation. Report only; do not change code or tests. For each finding, include the exact command, expected and actual results, and which checks remain untested. Keep requests inside the fixture; ask before contacting an external service.
/qa Test this repo's local webhook and worker fixture. Explore duplicate deliveries and recovery after a partial failure. Keep all effects inside the fixture; preserve reproduced bugs in native regression tests before repairing them.
QA first tells you which surface, tools and write permissions it will use. A CLI or API does not need a browser. Browser targets keep real browser testing; developer experience audits load only when onboarding, installation or ergonomics are in scope. If the native tools or safe fixture are unavailable, the report names the blocker and untested contracts instead of inventing a pass or installing another framework.
Exploration means learning from each result and choosing the next useful challenge,
not running random commands. Before the next discovery probe, QA saves a short
exploration-NNN.json evidence note with the previous result, the assumption being
tested and the next command. The final report links these notes; they do not require
an extra chat message between probes. A discovered bug must be reproducible, and its new test
must fail for the bug before the repair and pass afterward. Unit tests protect logic;
integration and end-to-end tests protect real boundaries that mocks would hide.
/qa-only proposes those tests without writing them.
Normal /review and /ship run a bounded version on changed behavior and nearby
risks automatically, including small diffs without a plan or web server. Existing
fix/test approval rules still apply. Missing dependencies, denied actions and time
limits remain visible coverage gaps; a short smoke pass never means exhaustive QA.
Production access and destructive or external effects require specific permission.
Bounded exploration uses an executable deadline guard, not an estimated clock: it refuses late probes and stops owned foreground work at the limit. Unfinished checks stay visible in the report. Required plan checks remain outside the review/ship smoke budget. See QA deadlines for command, platform and cleanup limits.
Every ship also runs the existing documentation audit, including repeat ships and existing PR updates. Clear factual corrections join the final checked change; risky rewrites need approval. A failed audit stops for recovery or explicit acceptance of the named risk rather than silently dropping its result. Ship owns versioning, Git and PR publication; the docs helper does not commit or push independently.
Which review should I use?
| Building for... | Plan stage (before code) | Live audit (after shipping) |
|---|---|---|
| End users (UI, web app, mobile) | /plan-design-review |
/design-review |
| Developers (API, CLI, SDK, docs) | /plan-devex-review |
/devex-review |
| Architecture (data flow, perf, tests) | /plan-eng-review |
/review |
| All of the above | /autoplan (runs CEO → design → DX → eng, auto-detects which apply; eng always last) |
— |
Power tools
| Skill | What it does |
|---|---|
/codex |
Second Opinion — independent code review from OpenAI Codex CLI. Review, challenge, and consult modes. Available on every harness except Codex. |
/claude-code |
Second Opinion — independent code review from Claude Code. Review, challenge, and consult modes, with session continuity for consultation. Available on every harness except Claude Code. |
/careful |
Safety Guardrails — warns before destructive commands (rm -rf, DROP TABLE, force-push). Say "be careful" to activate. Override any MEDIUM warning; root/home recursive deletes and default-branch force-pushes are hard-denied. |
/freeze |
Edit Lock — restrict file edits to one directory. Prevents accidental changes outside scope while debugging. |
/guard |
Full Safety — /careful + /freeze in one command. Maximum safety for prod work. |
/unfreeze |
Unlock — remove the /freeze boundary. |
/open-gstack-browser |
GStack Browser — launch gstack's own browser headed, with sidebar, anti-bot stealth, auto model routing (Sonnet for actions, Opus for analysis), one-click cookie import, and Claude Code integration. Clean up pages, take smart screenshots, edit CSS, and pass info back to your terminal. The visible face of the fallback engine; with Aside open you watch the agent's tabs there instead. |
/setup-deploy |
Deploy Configurator — one-time setup for /land-and-deploy. Detects your platform, production URL, and deploy commands. |
/setup-gbrain |
GBrain Onboarding — from zero to running gbrain in under 5 minutes. PGLite local, Supabase existing URL, or auto-provision a new Supabase project via Management API. MCP registration for Claude Code + per-repo trust triad (read-write/read-only/deny). Full guide. |
/sync-gbrain |
Keep Brain Current — re-index this repo's code into gbrain via gbrain sources add + gbrain sync --strategy code, refresh the ## GBrain Search Guidance block in CLAUDE.md, and auto-remove guidance when the capability check fails. --incremental (default), --full, --dry-run. Idempotent; safe to re-run. |
/gstack-upgrade |
Self-Updater — upgrade gstack to latest. Detects global vs vendored install, syncs both, shows what changed. |
/ios-qa |
iOS Live-Device QA (v1.43.0.0+) — drive a real iPhone over USB CoreDevice via an embedded StateServer in the app. Read Swift source, codegen typed @Observable accessors, run the agent loop. Optional --tailnet flag exposes the device to OpenClaw or any HTTP-capable agent on your Tailscale tailnet so remote agents can run iOS QA without ever touching the hardware. Capability-tier allowlist (observe/interact/mutate/restore), per-device session lock, audit log. |
/ios-fix, /ios-design-review, /ios-clean, /ios-sync |
iOS bug-fix loop, designer's-eye HIG audit, debug-bridge cleanup, and accessor resync. See docs/skills.md. End-to-end walkthrough: docs/howto-ios-testing-with-gstack.md. |
Standalone binaries
Beyond the slash-command skills, gstack ships standalone CLIs for workflows that don't belong inside a session:
| Command | What it does |
|---|---|
gstack-model-benchmark |
Cross-model benchmark — run the same prompt through Claude, GPT (via Codex CLI), and Gemini; compare latency, tokens, cost, and (optionally) LLM-judge quality score. Auth detected per provider, unavailable providers skip cleanly. Output as table, JSON, or markdown. --dry-run validates flags + auth without spending API calls. |
gstack-taste-update |
Design taste learning — writes approvals and rejections from /design-shotgun into a persistent per-project taste profile. Decays 5%/week. Feeds back into future variant generation so the system learns what you actually pick. |
gstack-egress |
Egress receipt auditor — every gstack-initiated off-machine send writes a tamper-evident, hash-chained receipt to ~/.gstack/security/egress.jsonl before the send. list shows what gstack attempted to send and to which host, grants shows the standing consent settings plus the exact command that revokes each, verify recomputes the hash chain and exits 3 on tamper (catches edits, reordering, and mid-chain deletion; truncating or deleting the ledger itself is out of scope — it's a forensic log, not tamper-proof storage). |
gstack-context-bill |
Token bill-of-materials — read-only, offline audit of what an installed skills tree costs in tokens: always-on frontmatter every session pays vs per-invocation SKILL.md + forced references. --diff compares two trees, --budget enforces a ceiling, --exact opts into Anthropic count_tokens (sends file text off-machine; writes an egress receipt first, degrades to the offline estimate if the receipt can't be written). |
gstack-code-intelligence |
Code-intelligence provider picker — wraps GBrain, Sourcebot, and Graphify behind one interface: options/status to see what's available, select to pick one, index/search to use it, suggest to check whether the one-time indexing offer should fire here. The offer triggers on large repos (1,000+ tracked files; a decline is persisted). Non-local providers refuse to index or search until you record per-repo consent (consent <repo> yes|no — the query text is repo-derived content), the per-repo trust policy's deny and read-only tiers veto write-class operations regardless of consent, and every off-machine send writes an egress receipt. Fully optional — with nothing selected, gstack falls back to grep. |
gstack-verify-gate |
Verification stop hook (opt-in) — blocks a Claude Code turn from ending until the project's declared verify command passes (after 3 blocked re-entries it yields with a loud still-RED warning instead of looping forever). Declare it on one line in CLAUDE.md: <!-- gstack:verify: bun test -->. Hooks bypass the permission system, so a declared command never runs until you trust it once per repo (gstack-verify-gate --trust); editing the command invalidates trust until re-granted, and every grant is audit-logged. ./setup never registers it for you — opt in with gstack-settings-hook add-event --event Stop --command ~/.claude/skills/gstack/bin/gstack-verify-gate --source verify-gate, remove with gstack-settings-hook remove-source --source verify-gate. |
gstack-memorable |
Memorable recall bridge (opt-in, third party, Claude Code only) — connects Claude Code to the external Memorable CLI through gstack instead of the vendor's own installer, so the hook gets gstack's guarantees: an explicit consent key (memorable_recall, off by default, listed by gstack-egress grants), a fail-closed egress receipt for every prompt handed over (gstack-egress list --sink memorable-recall), a HIGH-tier secret pre-scan, a trust envelope and 8 KiB cap on whatever comes back, an allowlisted environment and process-group containment for the vendor process, and clean removal. enable registers the hook at the stable install with a 5 s timeout and never runs the vendor's own consent command; disable revokes the gate first and removes the entry by identity even after Claude Code strips the tag; status is read-only. gstack never installs Memorable, and what its binary sends is the vendor's claim, not gstack's. Not available on Windows yet. Full guide. |
gstack-wtree |
Working-tree fingerprint — prints a content hash of what's actually on disk (temp index seeded from the stat cache, ~40x cheaper than a full re-hash; untracked source counts, gitignored scratch doesn't). Identical content fingerprints identically through commits, rebases, amends, and squashes — it's what binds reviews and test evidence to content instead of commit SHAs. |
gstack-review-log |
Review-pass receipts — --start <skill> captures the working-tree fingerprint before a diff review; '<JSON>' --finish <token> consumes that single-use, repository/branch/skill-scoped receipt. Binding requires matching start/end content and reviewer-reported completed:true and converged:true; it is not independent proof that a model read the source. --check-shared-libs <token> reads a current finding as JSON on stdin and checks prior Skip decisions against the actual branch, capture and eligible raw source blobs. It returns reusable:false when proof is missing or unsafe, including older records without logger-versioned coverage. Final review logging computes shared-code fingerprints and coverage rather than trusting supplied proof. |
gstack-review-read |
Review freshness — emits review records with computed review_freshness.status and reason: CURRENT, STALE, or UNVERIFIED for diff reviews. /ship and /land-and-deploy use the same grade; a matching commit alone never certifies a diff review. Dashboard rules. |
gstack-evidence |
Verification-evidence ledger — run --label <lane> -- <cmd> transparently wraps any test command (the child's exit code always passes through) and records what ran against which working-tree fingerprint; check grades each label FRESH/STALE/MISSING with --expect-cmd, --max-age, and --allow-paths binding. /ship and /land-and-deploy cite fresh evidence instead of re-running suites. Per-run logs are 0600, capped at 2MB, pruned after 30 days; the ledger and logs stay machine-local by design. |
gstack-issue-guard |
Tracker-text trust envelope — fetches GitHub issue/PR text (issue <n>, pr-body, pr-comments, or --stdin) and wraps it in a labeled envelope so agents treat it as data: injection-shaped lines get labeled even through fullwidth and invisible-character evasion, and forged envelope banners are defused. Every tracker-text ingress in gstack routes through it, enforced by a CI scanner. |
gstack-ios-qa-daemon |
iOS QA daemon — Mac-side broker between an agent and a connected iPhone over USB CoreDevice. Loopback by default; --tailnet opens a Tailscale-facing listener with identity-gated capability tiers. Single-instance via flock on ~/.gstack/ios-qa-daemon.pid. See docs/howto-ios-testing-with-gstack.md. |
gstack-ios-qa-mint |
iOS allowlist manager — owner-grant CLI for the tailnet allowlist. grant/revoke/list against ~/.gstack/ios-qa-allowlist.json (mode 0600). Remote agents never auto-allowlist; this is the explicit-intent path. |
gstack-ios-qa-regen |
iOS bridge regenerator — deterministically installs the canonical DebugBridge package, generates typed state accessors, and records the installed gstack version. Safe to rerun after source changes or upgrades. |
The private paid CSO evaluation producer is a packaging contract, not an ordinary bun run build artifact. On macOS or Linux, a release operator compiles cso-eval-producer with the documented hardened Bun flags in the same clean build session as bun run build:cso, then installs it beside gstack-cso-launcher, gstack-cso-core, gstack-cso-watchdog, and the hidden .gstack-cso-generation manifest as one root-owned, nonwritable five-artifact unit. The producer rejects root execution, writable/symlinked/incomplete installations, and unreviewed provider CLI versions; each receipt binds all five artifact hashes, and collection rejects receipts from different unit identities. See the clean producer procedure. Paid producer evaluation remains unavailable on Windows because the detached watchdog has no Windows build.
Each producer gets a curated one-cell source copy with directories sealed to 0555 and files to 0444, plus pre/post content, Git, and mode checks. Claude receives that exact copy as a restricted read-only add-directory so its constrained launcher command can reach it. Gemini receives no source working directory or include-directory. Codex technically receives read-only filesystem access to the exact curated source root because the trusted helper inherits the Codex permission profile; its private provider work directory and exact cso-home artifact directory are the only write roots, every other root path remains denied, and the producer prompt requires source access through the helper. This is evaluation containment for an immutable public fixture, not a claim that Codex cannot directly read that fixture.
Copy each completed cell's receipt together with state/cso-home/security/cso/ to the trusted adjudication host. Receipt entries are sorted paths relative to that directory and bind every retained file's size and SHA-256 plus an aggregate inventory hash. Provider homes, settings, sessions, and credentials live under separate disposable directories and are removed after the cell; they are never part of the retained artifact tree. The adjudicator must re-hash the transported tree against the receipt before trusting reports or repair bundles.
Paid producer qualification uses this explicit host matrix:
| Producer host | Daily/static cells | Comprehensive target execution |
|---|---|---|
| Codex CLI 0.153.4 | Supported under the custom permission profile | Fails closed because Docker/socket access is not granted to the model command sandbox; report the setup gap as partial |
| Claude Code 2.1.263 | Supported under restricted safe mode | Fails closed when the restricted launcher child cannot reach Docker; report the setup gap as partial |
| Gemini CLI 0.59.0 | Supported with isolated home/settings | Private release qualification host: the exact run_shell_command(<launcher>) policy can run the trusted helper without exposing a general shell |
Every setup-blocked comprehensive cell remains a miss in release-gate denominators. The evaluator does not silently count a host or workflow as supported when its containment policy prevents required setup.
./setup also registers one default-on Stop hook in ~/.claude/settings.json:
gstack-timeline-stop (closes dangling session-timeline entries when a session
is interrupted; fail-open — 2s internal budget, always exits 0, can never block
a session). Opt out persistently with ./setup --no-timeline-stop-hook — the
choice lands in the timeline_stop_hook config key, survives upgrades, and an
explicit "no" removes a live registration. GSTACK_TIMELINE_STOP_HOOK=no and
gstack-config set timeline_stop_hook no work too (flag > env > config).
./setup --no-team skips it for that run, gstack-settings-hook remove-source --source gstack-timeline-stop removes it by hand, and gstack-uninstall
removes it too.
Hook registration is canonical-only: every hook command points at the stable
~/.claude/skills/gstack install, never the tree setup ran from, so deleting
a worktree or Conductor workspace can't leave dead hooks erroring in your
sessions. Every ./setup run also heals first: gstack-settings-hook prune-stale --repoint removes dead gstack hook entries, re-points stale ones
at the stable install, and collapses duplicates, printing one line (and
writing a backup beside the file) only when it changed something.
Domain skills + raw CDP escape hatch
Two browser primitives in gstack's own engine (the fallback path when Aside isn't there) compound the agent over time:
$B domain-skill save— agent saves a per-site note (e.g., "LinkedIn's Apply button lives in an iframe") that fires automatically next time it visits that hostname. Quarantined → active after 3 successful uses → optional cross-project promotion via$B domain-skill promote-to-global. Storage lives alongside/learn's per-project learnings file. Full reference: docs/domain-skills.md.$B cdp <Domain.method>— raw Chrome DevTools Protocol escape hatch for the rare case curated commands miss. Deny-default: methods must be explicitly added tobrowse/src/cdp-allowlist.tswith a one-line justification. Two-tier mutex serializes browser-scoped CDP calls against per-tab work. Output for data-exfil methods is wrapped in the UNTRUSTED envelope.
Want raw CDP with no rails, no allowlist, no daemon — just thin transport from agent to Chrome? browser-use/browser-harness-js is a different philosophy (agent-authored helpers vs gstack's curated commands) and a good fit if you don't want gstack's security stack. The two can coexist: gstack's
$B cdpand harness can both attach to the same Chrome via Playwright'snewCDPSession.
Deep dives with examples and philosophy for every skill →
Karpathy's four failure modes? Already covered.
Andrej Karpathy's AI coding rules (17K stars) nail four failure modes: wrong assumptions, overcomplexity, orthogonal edits, imperative over declarative. gstack's workflow skills enforce all four. /office-hours forces assumptions into the open before code is written. The Confusion Protocol stops Claude from guessing on architectural decisions. /review catches unnecessary complexity and drive-by edits. /ship transforms tasks into verifiable goals with test-first execution. If you already use Karpathy-style CLAUDE.md rules, gstack is the workflow enforcement layer that makes them stick across entire sprints, not just single prompts.
Parallel sprints
gstack works well with one sprint. It gets interesting with ten running at once.
Design is at the heart. /design-consultation builds your design system from scratch, researches what's out there, proposes creative risks, and writes DESIGN.md. But the real magic is the shotgun-to-HTML pipeline.
/design-shotgun is how you explore. You describe what you want. It generates 4-6 AI mockup variants using GPT Image. Then it opens a comparison board in your browser with all variants side by side. You pick favorites, leave feedback ("more whitespace", "bolder headline", "lose the gradient"), and it generates a new round. Repeat until you love something. Taste memory kicks in after a few rounds so it starts biasing toward what you actually like. No more describing your vision in words and hoping the AI gets it. You see options, pick the good ones, and iterate visually.
Works with impeccable. If you use impeccable too, gstack does not fight it. gstack runs impeccable's deterministic engine as a pre-pass in /design-review, /review, /ship, and /design-html when you have it installed (gstack never runs impeccable's installer or launcher; the first time a design skill finds no engine it asks once whether to download the engine binary, checksum-pinned and logged, into ~/.impeccable, and remembers your answer), speaks the same 61 rule ids in its own voice, reads PRODUCT.md, writes DESIGN.md in the open DESIGN.md format both tools read, and hands deferred findings to /impeccable <command>. Say no and nothing changes: no nag, no missing step. gstack-config set design_detector off turns the pre-pass off. Attribution for the material gstack derived from impeccable and the DESIGN.md spec is in NOTICE.md.
/design-html makes it real. Take that approved mockup (from /design-shotgun, a CEO plan, a design review, or just a description) and turn it into production-quality HTML/CSS. Not the kind of AI HTML that looks fine at one viewport width and breaks everywhere else. This uses Pretext for computed text layout: text actually reflows on resize, heights adjust to content, layouts are dynamic. 30KB overhead, zero dependencies. It detects your framework (React, Svelte, Vue) and outputs the right format. Smart API routing picks different Pretext patterns depending on whether it's a landing page, dashboard, form, or card layout. The output is something you'd actually ship, not a demo.
/qa was a massive unlock. It let me go from 6 to 12 parallel workers. Claude Code saying "I SEE THE ISSUE" and then actually fixing it, generating a regression test, and verifying the fix — that changed how I work. The agent has eyes now.
Smart review routing. Just like at a well-run startup: CEO doesn't have to look at infra bug fixes, design review isn't needed for backend changes. gstack tracks what reviews are run, figures out what's appropriate, and just does the smart thing. The Review Readiness Dashboard tells you where you stand before you ship.
Test everything. /ship bootstraps test frameworks from scratch if your project doesn't have one. Every /ship run produces a coverage audit. /qa creates native regressions when infrastructure is available and explicitly reports missing test coverage; CSS-only fixes may use browser evidence instead. 100% test coverage is the goal — tests make vibe coding safe instead of yolo coding.
/document-release is the engineer you never had. It audits relevant authored docs against the release diff and corrects factual drift before final verification. Risky changes return for approval. /ship invokes it on every run and owns TODOS, release metadata, generation and publication; the child reports updated, current or blocked documentation.
Aside is the browser gstack drives first. For browser surfaces, on a Mac with the Aside AI browser open, /qa, /qa-only, /design-review, /canary, /benchmark, /scrape, and /browse all run there — your real browser, with your real logged-in sessions, in tabs the agent opens for itself and closes when it's done. No cookie import, no "open the browser" step, no CAPTCHA handoff dance: hit a sign-in wall, sign in inside Aside, say "done", and the agent continues. Anything a page returns is treated as untrusted content — the agent takes syntax from it, never instructions. /make-pdf, /diagram, and design previews print and screenshot through Aside too (served from your machine on loopback, one render per script), and the planning skills do their web research through Aside's own agent before reaching for a search tool.
When Aside isn't there, gstack's own browser takes over — automatically. Linux, Windows, or a Mac with Aside closed: the same skills use the bundled headless Chromium that ./setup builds, produce the same evidence, and light up the features below that only make sense when the browser is gstack's rather than yours.
Real browser mode. /open-gstack-browser launches GStack Browser, an AI-controlled Chromium with anti-bot stealth, custom branding, and the sidebar extension baked in. Sites like Google and NYTimes work without captchas. The menu bar says "GStack Browser" instead of "Chrome for Testing." Your regular Chrome stays untouched. All existing browse commands work unchanged. $B disconnect returns to headless. The browser stays alive as long as the window is open... no idle timeout killing it while you're working.
Sidebar agent — your AI browser assistant. Type natural language in the Chrome side panel and a child Claude instance executes it. "Navigate to the settings page and screenshot it." "Fill out this form with test data." "Go through every item in this list and extract the prices." The sidebar auto-routes to the right model: Sonnet for fast actions (click, navigate, screenshot) and Opus for reading and analysis. Each task gets up to 5 minutes. The sidebar agent runs in an isolated session, so it won't interfere with your main Claude Code window. One-click cookie import right from the sidebar footer.
Personal automation. The sidebar agent isn't just for dev workflows. Example: "Browse my kid's school parent portal and add all the other parents' names, phone numbers, and photos to my Google Contacts." Two ways to get authenticated: (1) log in once in the headed browser, your session persists, or (2) click the "cookies" button in the sidebar footer to import cookies from your real Chrome. Once authenticated, Claude navigates the directory, extracts the data, and creates the contacts.
Prompt injection defense. Hostile web pages try to hijack your sidebar agent. gstack ships a layered defense: content filters (datamarking, hidden-element stripping, ARIA scrubbing, URL blocklist) on every page read, plus a 22MB ML classifier running locally in a sidecar subprocess that scans page-derived content before the agent sees it, with a verdict combiner that requires classifier agreement before blocking (prevents single-model false positives on Stack Overflow-style instruction pages). Everything runs on your machine, no network calls. Emergency kill switch: GSTACK_SECURITY_OFF=1. See ARCHITECTURE.md for the full stack.
Browser handoff when the AI gets stuck. Hit a CAPTCHA, auth wall, or MFA prompt? $B handoff opens a visible Chrome at the exact same page with all your cookies and tabs intact. Solve the problem, tell Claude you're done, $B resume picks up right where it left off. The agent even suggests it automatically after 3 consecutive failures.
/pair-agent is cross-agent coordination. You're in Claude Code. You also have OpenClaw running. Or Hermes. Or Codex. You want them both looking at the same website. Type /pair-agent, pick your agent, and a GStack Browser window opens so you can watch. The skill prints a block of instructions. Paste that block into the other agent's chat. It exchanges a one-time setup key for a session token, creates its own tab, and starts browsing. You see both agents working in the same browser, each in their own tab, neither able to interfere with the other. If ngrok is installed, the tunnel starts automatically so the other agent can be on a completely different machine. Same-machine agents get a zero-friction shortcut that writes credentials directly. This is the first time AI agents from different vendors can coordinate through a shared browser with real security: scoped tokens, tab isolation, rate limiting, domain restrictions, and activity attribution.
Multi-AI second opinion. In Codex, gstack sends outside reviews to Claude Code through /claude-code. In Claude Code, /codex sends them to OpenAI Codex. Other harnesses expose both skills and use Codex where automatic outside reviews are supported. Each skill supports code review, adversarial challenge, and consultation with session continuity. Routing follows the harness, so changing your configured model does not change the outside reviewer. Reports identify the provider that actually completed each review; unavailable outside coverage remains visible.
Safety guardrails on demand. Say "be careful" and /careful warns before any destructive command — rm -rf, DROP TABLE, force-push, git reset --hard. /freeze locks edits to one directory while debugging so Claude can't accidentally "fix" unrelated code. /guard activates both. /investigate auto-freezes to the module being investigated.
Proactive skill suggestions. gstack notices what stage you're in — brainstorming, reviewing, debugging, testing — and suggests the right skill. Don't like it? Say "stop suggesting" and it remembers across sessions.
10-15 parallel sprints
gstack is powerful with one sprint. It is transformative with ten running at once.
Conductor runs multiple Claude Code sessions in parallel — each in its own isolated workspace. One session running /office-hours on a new idea, another doing /review on a PR, a third implementing a feature, a fourth running /qa on staging, and six more on other branches. All at the same time. I regularly run 10-15 parallel sprints — that's the practical max right now.
The sprint structure is what makes parallelism work. Without a process, ten agents is ten sources of chaos. With a process — think, plan, build, review, test, ship — each agent knows exactly what to do and when to stop. You manage them the way a CEO manages a team: check in on the decisions that matter, let the rest run.
Voice input (AquaVoice, Whisper, etc.)
gstack skills have voice-friendly trigger phrases. Say what you want naturally — "run a security check", "test the website", "do an engineering review" — and the right skill activates. You don't need to remember slash command names or acronyms.
Uninstall
Option 1: Run the uninstall script
If gstack is installed on your machine:
~/.claude/skills/gstack/bin/gstack-uninstall
This handles skills, symlinks, global state (~/.gstack/), project-local state, browse daemons, and temp files. Use --keep-state to preserve config and analytics. Use --force to skip confirmation.
Option 2: Manual removal (no local repo)
If you don't have the repo cloned (e.g. you installed via a Claude Code paste and later deleted the clone):
# 1. Stop browse daemons
pkill -f "gstack.*browse" 2>/dev/null || true
# 2. Remove per-skill directories whose SKILL.md points into gstack/
# (rm -rf, not rmdir — installed dirs also contain runtime-asset links)
find ~/.claude/skills -mindepth 1 -maxdepth 1 -type d ! -name gstack 2>/dev/null |
while IFS= read -r dir; do
link="$dir/SKILL.md"
[ -L "$link" ] || continue
target=$(readlink "$link" 2>/dev/null) || continue
case "$target" in
gstack/*|*/gstack/*)
rm -rf "$dir"
;;
esac
done
# Directories gstack created carry a .gstack-owned marker (the only signal on
# Windows, where installs are file copies with no symlink to read)
for marker in ~/.claude/skills/*/.gstack-owned; do
[ -f "$marker" ] && rm -rf "$(dirname "$marker")"
done
# Alias skills install as copies (no symlink to detect) — remove by name
rm -rf ~/.claude/skills/_gstack-command ~/.claude/skills/connect-chrome 2>/dev/null
# 3. Remove gstack
rm -rf ~/.claude/skills/gstack
# 4. Remove global state
rm -rf ~/.gstack
# 5. Remove integrations (skip any you never installed)
rm -rf "${CODEX_HOME:-$HOME/.codex}/skills/gstack"* 2>/dev/null
rm -rf ~/.factory/skills/gstack* 2>/dev/null
rm -rf ~/.kiro/skills/gstack* 2>/dev/null
rm -rf ~/.openclaw/skills/gstack* 2>/dev/null
rm -rf ~/.cursor/skills/gstack* 2>/dev/null
rm -rf ~/.config/opencode/skills/gstack* 2>/dev/null
# 6. Remove temp files
rm -f /tmp/gstack-* 2>/dev/null
# 7. Per-project cleanup (run from each project root)
rm -rf .gstack .gstack-worktrees .claude/skills/gstack 2>/dev/null
rm -rf .agents/skills/gstack* .factory/skills/gstack* 2>/dev/null
Manual removal leaves gstack's hook entries behind in ~/.claude/settings.json
(the uninstall script removes all of them for you, including entries whose
_gstack_source tag was stripped). Edit that file and delete every hook whose
command path points into .claude/skills/gstack/: the SessionStart auto-update
hook, the AskUserQuestion PreToolUse/PostToolUse hooks, and the Stop hooks
(session timeline, plus verify-gate if you opted in). Left in place, they error
on every matching event once the install directory is gone.
Clean up CLAUDE.md
The uninstall script does not edit CLAUDE.md. In each project where gstack was added, remove the ## gstack and ## Skill routing sections.
Playwright
~/Library/Caches/ms-playwright/ (macOS) is left in place because other tools may share it. Remove it if nothing else needs it.
Aside
gstack never installed Aside, so it never uninstalls it. Keep it or remove it like any other app.
Free, MIT licensed, open source. No premium tier, no waitlist.
I open sourced how I build software. You can fork it and make it your own.
We're hiring. Want to ship real products at AI-coding speed and help harden gstack? Come work at YC — ycombinator.com/software Extremely competitive salary and equity. San Francisco, Dogpatch District.
GBrain — persistent knowledge for your coding agent
GBrain is a persistent knowledge base for AI agents — think of it as the memory your agent actually keeps between sessions. GStack gives you a one-command path from zero to "it's running, my agent can call it."
/setup-gbrain
Four paths, pick one:
- Supabase, existing URL — your cloud agent already provisioned a brain; paste the Session Pooler URL, now this laptop uses the same data.
- Supabase, auto-provision — paste a Supabase Personal Access Token; the skill creates a new project, polls to healthy, fetches the pooler URL, hands it to
gbrain init. ~90 seconds end-to-end. - PGLite local — zero accounts, zero network, ~30 seconds. Isolated brain on this Mac only. Great for try-first; migrate to Supabase later with
/setup-gbrain --switch. - Remote gbrain MCP — your brain runs on another machine (Tailscale, ngrok, internal LAN) or a teammate's server; paste an MCP URL and bearer token. Optionally pair with a local PGLite for symbol-aware code search in split-engine mode. Best for cross-machine memory without standing up a local DB.
After init, the skill offers to register gbrain as an MCP server for Claude Code (claude mcp add gbrain -- gbrain serve) so gbrain search, gbrain put, etc. show up as first-class typed tools — not bash shell-outs.
Keeping the brain current. Run /sync-gbrain from any repo to re-index its code into gbrain (incremental by default, --full for a full reindex, --dry-run to preview). The skill registers the cwd as a federated source via gbrain sources add, runs gbrain sync --strategy code, and writes a ## GBrain Search Guidance block to your project's CLAUDE.md so the agent prefers gbrain search/code-def/code-refs over Grep. The block is removed automatically if the capability check fails — no stale guidance pointing at tools that aren't installed.
Per-remote trust policy. Each repo on your machine gets one of three tiers:
read-write— agent can search the brain AND write new pages back from this reporead-only— agent can search but never writes (best for multi-client consultants: search the shared brain, don't contaminate it with Client A's work while in Client B's repo)deny— no gbrain interaction at all
The skill asks once per repo. The decision is sticky across worktrees and branches of the same remote.
GStack memory sync (different feature, same private-repo infra). Optionally pushes your gstack state (learnings, CEO plans, design docs, retros, developer profile) to a private git repo so your memory follows you across machines, with a one-time privacy prompt (everything allowlisted / artifacts only / off) and a defense-in-depth secret scanner that blocks AWS keys, tokens, PEM blocks, and JWTs before they leave your machine.
gstack-artifacts-init
Running gstack in Conductor? Conductor explicitly strips ANTHROPIC_API_KEY and OPENAI_API_KEY from every workspace's process env, so paid evals and gbrain embeddings won't work out of the box. Set GSTACK_ANTHROPIC_API_KEY and GSTACK_OPENAI_API_KEY in Conductor's workspace env config instead — gstack's TS entry points promote them to canonical names at runtime. Full details and the contributor checklist for adding the import to new entry points: Conductor + GSTACK_* env vars.
Full monty — every scenario, every flag, every bin helper, every troubleshooting step: USING_GBRAIN_WITH_GSTACK.md
Other references: docs/gbrain-sync.md (sync-specific guide) • docs/gbrain-sync-errors.md (error index)
Docs
| Doc | What it covers |
|---|---|
| Skill Deep Dives | Philosophy, examples, and workflow for every skill (includes Greptile integration) |
| Diagrams & Document Formats | Mermaid/excalidraw fences in PDFs, image sizing and safety defaults, --to html|docx, /diagram triplets |
| Builder Ethos | Builder philosophy: Boil the Ocean, Search Before Building, three layers of knowledge |
| Using GBrain with GStack | Every path, flag, bin helper, and troubleshooting step for /setup-gbrain |
| GBrain Sync | Cross-machine memory setup, privacy modes, troubleshooting |
| Architecture | Design decisions and system internals |
| Browser | How gstack drives Aside first (the contract, the cookbook, rendering, research), when the fallback engine kicks in, and the fallback's full $B command reference |
| Contributing | Dev setup, testing, contributor mode, and dev mode |
| Memorable recall bridge | Opt-in third-party workflow memory through gstack: two consents, what gstack hands over and can attest, removal, troubleshooting |
| Changelog | What's new in every version |
Privacy & Telemetry
gstack includes opt-in usage telemetry to help improve the project. Here's exactly what happens:
- Default is off. Nothing is sent anywhere unless you explicitly say yes.
- On first run, gstack asks if you want to share anonymous usage data. You can say no.
- What's sent (if you opt in): skill name, duration, success/fail, gstack version, OS. That's it.
- What's never sent: code, file paths, repo names, branch names, prompts, or any user-generated content.
- Change anytime:
gstack-config set telemetry offdisables everything instantly. - Every off-machine send is receipted. Any gstack-initiated network send — telemetry included — writes a hash-chained, tamper-evident receipt to
~/.gstack/security/egress.jsonlbefore the send; sensitive sinks refuse to send at all if the receipt can't be written. Audit withgstack-egress list, verify the chain withgstack-egress verify(exit 3 on tamper), see the standing consent settings withgstack-egress grants. The ledger records attempted sends so accidents are auditable — it's an audit trail, not a network firewall. - Optional third-party bridges are off by default and receipted too. The one that exists today, the Memorable recall bridge, hands your prompt to a locally installed vendor binary only after you run
gstack-memorable enable; every hand-off writes a receipt first, the consent shows up ingstack-egress grants, and what the vendor then sends is documented as the vendor's claim.
Data is stored in Supabase (open source Firebase alternative). The schema is in supabase/migrations/ — you can verify exactly what's collected. The Supabase publishable key in the repo is a public key (like a Firebase API key) — row-level security policies deny all direct access. Telemetry flows through validated edge functions that enforce schema checks, event type allowlists, and field length limits.
Local analytics are always available. Run gstack-analytics to see your personal usage dashboard from the local JSONL file — no remote data needed.
Troubleshooting
Skill not showing up? cd ~/.claude/skills/gstack && ./setup
/browse (or /qa, /design-review) says NEEDS_ASIDE or ASIDE_NOT_RUNNING? That's the probe telling you it's about to use the fallback browser. Want Aside? Open the app and sign in — aside --version should print a version and aside repl 'console.log("ok")' should print ok — then re-run. gstack never installs it for you. Want the fallback on purpose while Aside is open? GSTACK_SKIP_ASIDE=1 makes every skill, the renderer, and ./setup treat Aside as absent.
/browse fails on the fallback browser? cd ~/.claude/skills/gstack && bun install && bun run build
/make-pdf or /diagram can't render? Same two paths: with Aside open they print through Aside (bun run ~/.claude/skills/gstack/bin/gstack-render.ts some.html --screenshot /tmp/out.png tests it directly, and its first line, ENGINE=aside or ENGINE=browse, names the browser that actually rendered); without it they use the bundled browser, so bun run build is the fix.
Stale install? Run /gstack-upgrade — or set auto_upgrade: true in ~/.gstack/config.yaml
State in the wrong place, or a setting that won't stick? ~/.claude/skills/gstack/bin/gstack-paths --explain shows which directory gstack uses for its state and why. See docs/state-root.md.
Want shorter commands? cd ~/.claude/skills/gstack && ./setup --no-prefix — switches from /gstack-qa to /qa. Your choice is remembered for future upgrades.
Want namespaced commands? cd ~/.claude/skills/gstack && ./setup --prefix — switches from /qa to /gstack-qa. Useful if you run other skill packs alongside gstack.
Codex says "Skipped loading skill(s) due to invalid SKILL.md"? Your Codex skill descriptions are stale. Fix: cd "${CODEX_HOME:-$HOME/.codex}/skills/gstack" && git pull && ./setup --host codex — or for repo-local installs: cd "$(readlink -f .agents/skills/gstack)" && git pull && ./setup --host codex
Windows users: gstack works on Windows 11 via Git Bash or WSL. Aside is macOS-only, so on Windows (and Linux) the browser skills, /make-pdf, and /diagram always use gstack's bundled browser. Node.js is required in addition to Bun — Bun has a known bug with Playwright's pipe transport on Windows (bun#4253). The browse server automatically falls back to Node.js. Make sure both bun and node are on your PATH. Native /cso additionally requires Windows PowerShell and Visual Studio 2022 Build Tools with the Desktop development with C++ workload; setup leaves that skill explicitly unavailable when they are absent.
On Windows without Developer Mode (MSYS2 / Git Bash), setup falls back to file copies instead of symlinks because ln -snf produces frozen copies that don't refresh on git pull. Re-run cd ~/.claude/skills/gstack && ./setup after every git pull so your skill files match the repo. setup prints a one-line note reminding you. Unix and WSL keep symlinks and don't need the re-run.
Chromium install failed or hung during ./setup? The bundled browser is
best-effort: setup records the reason, finishes registering every skill, and
prints which skills are affected (/qa, /qa-only, /design-review,
/browse, make-pdf, /diagram, /pair-agent). With Aside open, the browser
skills keep running in Aside and only the fallback engine is missing;
/pair-agent always needs the bundled browser. Fix the cause and re-run
./setup. Knobs:
GSTACK_PLAYWRIGHT_INSTALL_TIMEOUT=<seconds> raises the download bound
(default 600) on slow links; GSTACK_SKIP_PLAYWRIGHT=1 skips the Chromium
install entirely (CI, no-browser boxes); GSTACK_CHROMIUM_NO_SANDBOX=1 is the
fix when Chromium installs but cannot launch because the host blocks
unprivileged user namespaces (Ubuntu 24.04+ AppArmor default, #2157).
Setup ended with "Not registered (a skill you own already uses the name; left untouched)"?
gstack only deletes or links over a skill entry it can prove it created: a
symlink into gstack, a directory carrying the .gstack-owned marker ./setup
writes into every directory it creates, or a SKILL.md that is byte-identical to
gstack's or carries the generated <!-- AUTO-GENERATED from ... --> banner. A
qa/ or ship/ you wrote yourself is left untouched by ./setup,
gstack-relink, and both prefix-mode flips, and the linker names it in the
final summary. Rename or move yours, or switch modes (./setup --prefix /
--no-prefix) so the names stop colliding. If you started your own skill from
a generated gstack SKILL.md and then edited it, that file is moved to
~/.gstack/backups/skills/<timestamp>/<skill>/SKILL.md before gstack's is
linked in, never deleted.
Claude says it can't see the skills? Make sure your project's CLAUDE.md has a gstack section. Add this:
## gstack
Use /browse from gstack for all web browsing. Never use mcp__claude-in-chrome__* tools.
Available skills: /office-hours, /plan-ceo-review, /plan-eng-review, /plan-design-review,
/design-consultation, /design-shotgun, /design-html, /review, /deslop-shared-libs, /test-audit, /ship, /land-and-deploy,
/canary, /benchmark, /browse, /open-gstack-browser, /qa, /qa-only, /design-review, /scrape,
/setup-browser-cookies, /setup-deploy, /setup-gbrain, /sync-gbrain, /retro, /investigate,
/document-release, /document-generate, /codex, /cso, /autoplan, /pair-agent, /careful, /freeze,
/guard, /unfreeze, /gstack-upgrade, /learn.
License
MIT. Free forever. Go build something.