mirror of
https://github.com/garrytan/gstack.git
synced 2026-10-04 10:26:52 +02:00
7fca42ad8b6c707b8a38f579f72bf3c4f7de6d85
3
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Garry Tan
|
7fca42ad8b |
v1.91.12.0 v1.91.12.0: audit fix wave, ~11-minute paid eval lanes, eval reliability policy (#2999)
* 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 on |
||
|
Garry Tan
|
a84b0b5b6d |
v1.90.0.0 feat: make browser cookie imports explicit and safe (#2964)
* fix(browse): prepare reliable cookie import wave for validation * ci: sequence quality and behavior for validation branch * fix(browse): isolate Windows qualification and preserve native diagnostics * test(browse): cover cookie workflow quality and isolate Windows user paths * test(browse): trace native member startup and initialize fresh folders * fix(browse): keep Windows member stdin alive through EOF * fix(browse): latch native timeouts and compare contained Edge startup * test(browse): verify native version metadata and actual Windows argv * test(browse): qualify Dia import on isolated macOS CI * fix(browse): require picker origin for session mutations * fix(browse): bound credential reads through stream completion * test(browse): inspect owned Windows process arguments natively * test(evals): preserve passing coverage during cookie repair reruns * test(browse): isolate Dia qualification in a fresh macOS account * test(browse): pass bounded integer timeouts to native Mac probes * test(browse): distinguish Windows profile initialization from containment * test(browse): await descendant pipe readiness before parent exit * test(browse): initialize and restore isolated macOS Keychain state * test(browse): initialize Windows fixture folders before qualification * test(ci): pin the same Node runtime across Windows checks * test(browse): distinguish native macOS browser preflight stages * test(browse): isolate Windows descendant console lifetime * test(browse): preserve native receipts and identify fixture lock holders * test(browse): prepare dependency resolution before native Mac worker startup * test(ci): include lock and close checks in native diagnostics * test(browse): preserve native owner probe stages and subprocess deadlines * fix(browse): classify Chromium profile-in-use exit precisely * test(browse): retain Mac qualification evidence through cleanup failures * test(browse): bound Mac fixture paths and retire its owned user domain * test(browse): accept vanished fixture entries without weakening cleanup * test(browse): identify probe-created macOS user domains safely * test(browse): observe Mac user domains without targeting them first * test(browse): use passive fresh-user ownership throughout Mac qualification * test(browse): distinguish profile and registered-home Keychain lookups * test(browse): qualify Dia under one registered account home * test(browse): identify Dia startup and owned process-group failures * test(browse): classify bounded Dia startup diagnostics without leaking output * fix(test): preserve native Mac sandboxing and reap owned browser children * fix(browse): preserve Chromium sandboxing for native profile imports * test(browse): inspect signed Mach-O architecture without launching Xcode tools * test(browse): sample pending Dia startup and reap on all cleanup paths * test(browse): compare protected Dia launches in fresh Bun and Node accounts * test(browse): inspect isolated Mac GUI readiness without browser access * v1.90.0.0 fix: bind cookie picker actions to their document * test: validate cookie guards and fit nested launch fixtures * ci: configure the bundled Chromium sandbox helper * fix(browse): classify Playwright authentication timeouts * test: retain bounded Windows lifecycle diagnostics * test(cso): reuse bounded NTFS precision candidates * test(review): handle explicit preservation choices safely * test(browse): remove owned fixture directories with explicit primitives * test(review): distinguish descriptive reuse from edit commitments * test: admit only the approved unscored cookie workflow refusal * test: keep the Office Hours judge mock export-complete * fix: keep dependency-free CI planners independent of the model SDK * test: observe the exact holder after a native fixture unlink failure * fix: start seeded PTY observations at owned readiness * test: acquire identity-bound Windows deletion admission before profile resets * test: preserve qualified Git index bits without authorizing mutations |
||
|
|
2be6c06ba8 |
v1.65.0.0 feat: fork port wave 2 — feature fixes, session persistence, Apple releases, supply-chain CI (#2577)
* fix(memory-ingest): pass --include-gitignored to gbrain import gstack-artifacts-init writes an ignore-everything .gitignore (a bare `*`, headed "Do not edit") at the root of ~/.gstack. The memory ingest stages pages into ~/.gstack/.staging-ingest-<pid>-<ts>/, which is inside that repo, and gbrain's markdown collector honours .gitignore. The collector therefore matches every staged file against `*` and collects zero. The failure is silent. gbrain import exits 0 having imported nothing while the ingest prints `written: N` from the STAGED count rather than the imported count, so a run that indexes nothing looks identical to a healthy one and the memory corpus quietly stops growing. Reproduction, using git's own ignore machinery (no gbrain needed): git init . mkdir -p .staging-ingest-12345/learnings echo x > .staging-ingest-12345/learnings/page.md printf '*\n' > .gitignore git ls-files --others --exclude-standard # -> empty Passing --include-gitignored makes the import independent of whatever .gitignore sits above the staging directory. Adding a negation to the generated .gitignore is the alternative, but that file is gstack-owned and marked "Do not edit", so any regeneration silently reintroduces the bug. Adds a regression pin in the shape of memory-ingest-no-put_page.test.ts, plus a behavioural test for the collision itself. Both source pins fail against the unpatched file. * fix(memory-ingest): GIT_CEILING_DIRECTORIES defense-in-depth on the import child (#2144) Second layer under #2560's --include-gitignored: a realpath'd ceiling at the staging dir's parent pushes any git-enumerating collector off the git fast path (which sees zero files under ~/.gstack's ignore-everything root) onto its plain FS walk, even on gbrain builds whose flag semantics drift. Ceiling is realpath'd because git compares canonicalized directories during discovery — a staging dir reached through a symlink (macOS /var -> /private/var, symlinked $GSTACK_HOME) otherwise never matches. Behavioral tests prove discovery stops at the ceiling from the staging dir, including through a symlinked path, using git itself — no gbrain required. Mechanism ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(autoplan): Phase 4 task aggregator emitted zero tasks on every run (#2018) The branch+commit jq filter piped to the split commit array and then referenced .commit — jq rebinds context across a pipe, so .commit indexed the ARRAY with a string, every input line errored into 2>/dev/null, and || true swallowed the exit. The aggregate table has been empty for every user since the feature shipped. Bind .commit to a variable before the pipe. Functional pin extracts the ACTUAL emitted jq program from the resolver and runs it against fixture JSONL (verified RED against the broken filter), plus a source-shape guard against reintroducing a context-rebinding reference. Fix mechanism from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(codex): BSD mktemp templates broke /codex on every macOS install (#2091) macOS mktemp requires the X's to end the template; the five "codex-*-XXXXXX.txt" sites failed with "mkstemp failed ... File exists" before Codex ever ran (reproduced live on this machine). Same class fixed in claude/SKILL.md.tmpl's three sites. bin/gstack-paths now strips macOS's trailing slash from TMPDIR so TMP_ROOT-built paths stop carrying "//". Static tripwire scans every tracked .tmpl for characters after the X-run in a mktemp template (longer X-runs stay valid), plus a live portability check of the emitted shape. Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(decision-log): --supersede silently discarded the replacement decision The supersede/redact branch appended the retirement event and exited before the JSON argument was ever read — a user recording a reversal WITH its replacement lost the replacement, and the payload finder's first-non-flag-arg predicate would have mistaken the target id for JSON anyway. Payloads are now identified by their leading brace, validated BEFORE any write, and appended FIRST (retirement second), so the only visible interleaving under a crash is both-active — recoverable, never lost. The replacement carries supersedes:<old-id> provenance. Bare --supersede <id> (the documented reversal-without-replacement) stays legal; --redact with a payload now refuses instead of dropping it. Ported from time-attack/gstack (GStack 2), tests included. Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(brain-context): cold-start probe latency permanently disabled gbrain context gbrainAvailable() spawned gbrain --version under a 500ms budget; a cold CLI start on a loaded machine blew the timeout, misclassified gbrain as missing, and every skill session silently ran brainless — plus the per-query re-probe burned 3x the budget before any real work. Replaced with a memoized stat-based PATH scan (PATHEXT-aware on Windows) and made the query timeout overridable via GSTACK_BRAIN_TIMEOUT_MS for loaded CI environments. Also picks up the fork's manifest-filter coverage (#1687 shape) against the fake-gbrain harness — passes against our existing filter support. Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(setup-gbrain): voyage-code-3 flags were silently dropped under zsh (#1798) zsh does not word-split an unquoted $VAR, so all three PGLite-init sites passed the entire flag string as ONE argv word — gbrain ignored it and silently fell back to its default embedding model, downgrading code retrieval for every zsh user (macOS default shell). Flags now ride the positional params (set -- ...; "$@"). Tests run the shape under BOTH bash and zsh against the fake-gbrain argv recorder (per-word argc log distinguishes one-blob from split), include a demonstration of the zsh collision on the old shape, and pin the template's three sites statically. Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(model-benchmark): recognize macOS Keychain auth in the claude adapter (#1890) The default macOS Claude Code install stores OAuth under the Keychain generic-password service "Claude Code-credentials" and never writes ~/.claude/.credentials.json, so available()'s file-or-env sniff reported "No Claude auth found" while claude -p worked fine. On darwin the sniff now also probes the Keychain entry — metadata only (no -w, the secret is never read), 5s timeout, any security(1) failure degrades to not-found. Verified live on this machine (subscription install, no creds file, Keychain entry present). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(upgrade): v1.27 migration no longer auto-proceeds without a TTY or records a failed rename as done (#1383) Two silent-failure shapes in one script. Non-interactive runs (Claude Code Bash tool, CI) blanket-auto-proceeded into a REMOTE repo rename — now they skip-for-now by default and ask again next upgrade; unattended runs opt in with GSTACK_MIGRATE_ASSUME_YES=1. And a failed gh rename was journaled as done and the done-touchfile written, permanently stranding a half-renamed install — the failed step now stays PENDING with the manual command printed, finalize refuses the done-marker while any step is unjournaled, and the migration exits 1 with a re-run pointer while completed steps still skip on retry. Harness updated to opt in explicitly; new tests pin the default-skip and failure-stays-pending-then-retry-succeeds contracts (13/13). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ship): REST fallback when gh pr edit hits the Projects-classic GraphQL deprecation (#1079) On repos where GitHub enforces the Projects-classic sunset, gh pr edit hard-errors on repository.pullRequest.projectCards and Step 19's PR body update dies. The template now names the error shape, says it is not an auth problem, and falls back to the REST endpoint (gh api pulls/N -X PATCH) with the SAME already-redaction-scanned temp file for body and title. Generated SKILL.md regen rides the cluster regen commit. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ship): test-command detection was blind to Django and config-less-but-tested projects The Test Framework Bootstrap detected Python only via requirements.txt or pyproject.toml and treated missing config files as no-tests, so a green 'python manage.py test' Django app, a Go project with *_test.go beside the source, in-source Rust #[test] blocks, or a package.json with only a test script all got offered a SECOND test framework over a working one. Detection now enumerates definitive per-ecosystem markers (manage.py, tox.ini/setup.cfg, pom.xml/gradle, Makefile test targets, a tracked-file test census, in-source Rust tests) as EVIDENCE for the question it asks — never a command to run blind — preserving the read-CLAUDE.md-or-ask contract, with a marker→candidate-command table and ask-once persistence. The shared coverage-audit detection block gains the same markers. Test runs the resolver's emitted detection bash against Django / Go / Rust / Node fixtures in throwaway git repos. Ported from time-attack/gstack commit e3259078 (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * chore: regenerate SKILL.md files for cluster A (autoplan jq, codex mktemp, setup-gbrain zsh, ship detection + REST fallback) Atomic regen of the 9 generated files whose templates/resolvers changed in the A-cluster commits. bun run gen:skill-docs, no hand edits. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test: refresh ship goldens + parity ratios for cluster A growth Codex/Factory hosts render single-file ship skills whose committed goldens must track template changes; refreshed from the regenerated renders. Parity size guards bumped with the growth itemized — ship (carve-guards) 1.08 -> 1.10 for the detection-evidence + REST-fallback growth measured at 1.090x, qa (parity-harness monolith invariant) 1.07 -> 1.12 for the shared coverage-audit markers measured at 1.111x. Kept tight so the next growth is a deliberate decision, not drift; the Apple adapter raises ship again with its own justification. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(gbrain-sync): enforce the per-repo policy at the code-import chokepoint (#2140 sync path) The deny/read-only tiers in ~/.gstack/gbrain-repo-policy.json were stored by gstack-gbrain-repo-policy but enforced only in /sync-gbrain skill prose — a direct or cron invocation of gstack-gbrain-sync ingested repo code regardless. Worse: the code stage's egress receipt has cited 'per-repo policy chokepoint (repoPolicyTier)' as its consent since v1.63 while no such function existed. repoPolicyTier() now gates the stage before the dry-run branch: deny → refused-policy-deny (exit 1, loud), read-only → clean skipped-policy-read-only (code ingest writes pages), unreadable store → fail-closed refused-policy-unreadable, no store → unchanged fail-open. Subprocess tests pin all four paths against real git repos and a permission-blocked store (verified RED against the ungated binary). The receipt's consent string is truthful from this commit. #2140's ingest-path source-isolation ask remains open — partial-progress comment at ship. Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ios-qa): /auth/sessions no longer hands raw bearer tokens to any local process The loopback sessions list echoed live tokens — a harvest-and-replay primitive for anything on the machine (same class as the /health token leak fixed in v1.63). The list now returns a device-salted 16-hex token_id plus metadata; the salt is shared with the attempts log so identifiers correlate. /auth/revoke keeps the list→revoke workflow alive by accepting token_id alongside the caller's own raw token and identity. saltedHash() is exported from audit.ts and writeAttempt now reuses it (was inlined). Integration tests pin raw-token absence, the id shape/metadata, and the token_id revoke round-trip (verified RED against the leaking handler). List fix ported from time-attack/gstack (GStack 2); token_id revoke is ours. Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ios-qa): boot token out of os_log entirely; IPv4 listener pinned to loopback at the socket The StateServer's bootstrap announce logged the live boot token with privacy: .public — and nothing consumed it: the daemon has read the token from the 0600 app-container file since the devicectl copy flow landed. The log line handed a credential to anything reading the unified log during the launch window. It now announces port/build only. The IPv4 listener bound the wildcard interface and relied on the per-connection peer check alone; IPv4 has no CoreDevice tunnel path, so it now binds 127.0.0.1 via requiredLocalEndpoint at the socket level. IPv6 keeps the wildcard bind for CoreDevice ULA peers by design. Static pins cover both the template and the fixture app copy. Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(make-pdf): close the offline-gate bypass via raw-HTML fetch vectors With --allow-network off, the sanitizer stripped script/iframe/link but let Chromium fetch remote resources at print time through four raw-HTML vectors: <style> @import (any form), remote url() in <style> blocks and inline style attributes (incl. protocol-relative //), srcset with a remote candidate (Chromium prefers srcset over the inlined src), and remote src/poster on video/audio/source/track. All neutralized at the sanitizer; remote <img src> is deliberately left for the image inliner so its blocked-remote placeholder still fires, and url() mentions in prose/code spans stay untouched. Fork's test suite ported verbatim (12 cases incl. the end-to-end render assertion), verified RED against the old sanitizer. Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(pair-agent): tunnel activation is consent-gated — and the receipt's consent claim is now real The tunnel egress receipts have claimed consent: 'pair_agent=on' since v1.63 while no such key or gate existed — ngrok installed+authed was enough for the CLI to auto-start an internet-facing tunnel. isPairAgentEnabled() (fail- closed, env-overridable) now gates all three activation points: CLI auto-start, POST /tunnel/start (refuses with the enable hint), and the BROWSE_TUNNEL=1 startup bind. Consent-on-first-use, not silent breakage: the /pair-agent skill asks once (one-way-door posture), sets pair_agent via gstack-config (registered with on|off validation, default off), and never asks again; direct API callers get the same hint in the refusal. Adapted from the fork's gate: their reader targeted config.json, which on main would have made the gate silently un-enableable — ours reads the canonical ~/.gstack/config.yaml with the JSON shape as fallback, pinned by tests either way (11 cases, gate wiring tripwires included). Ported from time-attack/gstack (GStack 2), store adaptation ours. Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * chore: regenerate pair-agent SKILL.md for cluster B (consent gate) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(browse): cancel the parent watchdog when handoff promotes a daemon to headed The parent-process watchdog assumes connection mode is fixed at boot: headless daemons outlive their parent, headed ones do not. The env guards (BROWSE_PARENT_PID=0, BROWSE_HEADED=1) only cover daemons that were headed when they started. handoff breaks that assumption. It swaps in a headed context on a RUNNING daemon and sets connectionMode = 'headed' without a restart, so a daemon that legitimately registered a watchdog lands on the fatal side of the branch. The parent is usually a short-lived shell, and Claude Code's Bash tool kills one after every invocation, so the next 15s poll shuts the daemon down. The user-visible effect is that handoff destroys the thing it just created. It exists so a human can log in, solve a CAPTCHA, or clear an MFA prompt; the browser disappears about fifteen seconds later and takes the session with it. Observed while driving two registrar control panels: five daemon deaths and three logins, each one discarding the authenticated session. BrowserManager now exposes onHeadedPromotion, fired only on runtime promotion and not on a headed boot, and the server binds it to a canceller for the interval it already owned but previously discarded. Bound on both the module-level manager and any embedder-supplied one, since the watchdog reads activeBrowserManager and binding only the default would let embedders promote silently. The binding sits next to the browserManager declaration rather than next to clearParentWatchdog. Placing it with the function, which lives with the watchdog it cancels, reads better but touches browserManager in its temporal dead zone, which aborts module evaluation and leaves every later const uninitialized. findport tests catch that immediately. Tests: watchdog.test.ts already noted in its header that its three cases all fix mode via env at spawn time, so none reaches the headed branch. Driving a real handoff needs a headed Chromium, so the wiring is pinned with static tripwires instead, matching cdp-session-cleanup.test.ts and server-auth.test.ts. Verified they fail when the notification call is removed and pass when restored. Full `bun test` shows the same 6 pre-existing failures on this branch and on main (gstack-gbrain-detect, gstack-artifacts-init), which pass in isolation on both, so they are test-order pollution rather than a regression here. * fix(browse): pass windowsHide so the daemon stops popping console windows On Windows, `browse` leaves empty black console windows on top of whatever the user is doing — they pop up every few minutes for as long as any browser skill is alive, and outlive the process that created them. Cause: `bun-polyfill.cjs` maps `Bun.spawn`/`Bun.spawnSync` onto node's `child_process`, and node defaults `windowsHide` to **false**. Bun never creates these windows, so nothing in the daemon's own code looks wrong — the behaviour only appears on the node fallback path. The one users notice is `spawnTerminalAgent()`, which launches `bun run terminal-agent.ts` through this shim. The daemon respawns it on a watchdog, so closing the window is not enough — a new one arrives shortly after. Ten `bun.exe` processes were live on the machine this was diagnosed on. Why they linger after the child exits: with the default terminal application set to "Let Windows decide", the console is brokered through Windows Terminal via svchost, and WT leaves the empty frame behind when its only child exits. The frame has no child process at all, which is why it looks like a dead terminal. Setting `windowsHide: true` on both wrappers fixes every console child routed through the shim — the bun agent plus the `tasklist`, `git` and `powershell` calls elsewhere in the daemon. No behaviour change on macOS or Linux, where the option is ignored. Not covered by this commit: `chromium.launch()` goes through playwright's own process launcher rather than this shim, so it still creates one window per daemon start. Worth a follow-up. * test(browse): make bun-polyfill tests runnable on Windows, and cover windowsHide `bun test browse/test/bun-polyfill.test.ts` was **0 pass / 4 fail on Windows** before this — every test in the file, on the platform the polyfill exists to support. Each test interpolates the polyfill's absolute path into a single-quoted JS string passed to `node -e`. On Windows that path has backslashes, so JS eats them as escapes: 'C:\Users\jwilk\dev\gstack-fork\browse\src\bun-polyfill.cjs' -> C:Usersjwilkdevgstack-forkrowsesrcun-polyfill.cjs (`\b` is a real escape, so it deletes a character too.) `require()` throws, the subprocess dies, stdout is empty, and every assertion compares against "". The tests pass on macOS and Linux purely because those paths have no backslashes. Fixed by interpolating with `JSON.stringify(polyfillPath)`, which quotes and escapes correctly on all platforms. Also adds a regression test for the windowsHide fix in the previous commit. It stubs `child_process.spawn`/`spawnSync` *before* the polyfill destructures them and asserts the captured options, so it is deterministic and needs no window — it verifies the contract on macOS and Linux too, where the option is a no-op. Verified on Windows: 5 pass / 0 fail with the fix, and the new test alone fails ("VISIBLE" instead of "HIDDEN") when the previous commit is reverted. * fix(browse): forward windowsHide through the Bun polyfill spawn shims The Node fallback shim accepts a Bun.spawn options object and forwards only stdio, env and cwd to child_process.spawn. windowsHide is dropped, and because Node defaults it to false while Bun.spawn hides the console window, the omission inverts the behavior on the one platform the shim exists to support. Symptom: the terminal-agent respawn in server.ts (60s watchdog ticker) pops a visible bun.exe console window on Windows every time it fires, so the window keeps coming back with no scheduled task or startup entry behind it. stdio:'ignore' silences the child's output but does not suppress its window. Both shims now forward the option and default it to true, matching the Bun API being emulated; an explicit windowsHide:false still passes through. spawnTerminalAgent also sets it explicitly at the call site. Tests: three cases in browse/test/bun-polyfill.test.ts assert the default for spawn and spawnSync and that an explicit false is honored. Each was confirmed to fail against the unpatched shim. Drive-by, required to run the suite at all on Windows: the tests interpolated an absolute path into a JS string literal, so backslashes were consumed as escapes and every require() failed with MODULE_NOT_FOUND. The path is now normalized to forward slashes. On Windows this file went from 0/4 passing to 7/7. * fix(browse): headed mode on macOS 26 — stop mutating the signed Chromium bundle, heal the ones we already broke (#2242, #2138, #2139) The in-place rebrand rewrote the Chrome-for-Testing bundle's Info.plist (global name replace — which also renamed CFBundleExecutable to a binary that doesn't exist) and overwrote its Resources/*.icns, breaking the codesign seal: GPU process exit_code=5, headed mode dead on macOS 26. The mutation lived in the SHARED Playwright cache, so it also poisoned the user's other Playwright projects. Three layers land together: (1) the rebrand block is gone — branding lives in the GStack Browser.app wrapper via GSTACK_CHROMIUM_PATH, with a tombstone and a static tripwire (no plist/icns writes into the bundle; the tripwire allows the read-only probe below); (2) a launch-time self-heal detects an already-poisoned cache bundle, removes it, and errors with the exact re-fetch command — covering deploy paths that never run migrations; (3) migration v1.64.0.0 sweeps every cached bundle, removes poisoned ones, and re-fetches clean Chromium immediately (migrations run after ./setup, so without the re-fetch an upgrade would end with zero working browser). Functionally verified against fixture caches: poisoned removed, clean untouched, rerun no-op. Migration filename tracks the final VERSION at ship. The #2242 watchdog half is the absorbed PR #2565 (thanks @Screddyice). Tombstone/tripwire ported from time-attack/gstack (GStack 2); self-heal and migration are ours. Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(browse): 'browse stop' no longer restarts the daemon it was asked to stop The stop handler awaited shutdown() — which ends in process.exit — before returning, so the acknowledgement never egressed. The CLI's fetch reset, which its crash path reasonably interpreted as a dead daemon: it relaunched Chromium, re-sent stop, watched the daemon exit again, and errored 'Server crashed twice in a row'. Every stop cost a wasted Chromium launch and a nonzero exit. The ack now returns first; shutdown fires on a 25ms unref'd timer. Same fix for restart. Fork's test pins ack-before-teardown for both. Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(browse): lock acquisition reports real errors instead of phantom contention (#1084) acquireServerLock's bare catch treated EVERY failure as 'another process holds the lock' — a missing state dir, EACCES, or ENOSPC read as permanent phantom contention with nothing to debug. Now only EEXIST is contention: ENOENT self-heals with one mkdirSecure retry, everything else throws ServerLockError carrying the real errno, and the stale-lock unlink/retry loop is depth-capped so it can't livelock. Fork's five-case test ported. Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test(browse): integration coverage for #1781 busy-vs-dead recovery Fork's wedged-daemon fixture: first /command connection drops, daemon PID stays alive. Pins the whole contract — CLI retries the same daemon instance without a kill, state file untouched, no restart, exactly two command requests. Message-text assertion adapted: our CLI retries silently at the probe layer where the fork announces on stderr; the behavior, not the message, is the invariant. Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(browse): windowsHide on every Windows-reachable spawn (#1835) Console windows flashed (and stole focus) on every daemon relaunch, taskkill, tasklist poll, and powershell DPAPI call — node-level spawns default windowsHide to false. Covered: the node -e launcher (outer spawnSync AND the inner detached daemon spawn inside the launcher string), the dev-mode bun fallback, killServer's taskkill, isProcessAlive's tasklist, and cookie-import's powershell + tasklist. The Bun-polyfill shims were covered by absorbed PRs #2523 + #2539 (thanks @jwilk-hrep, @jerrynicholsai); this closes the sites those PRs didn't reach. The icacls sites land with the #1605 DACL commit alongside the static tripwire that pins all of them. R8's planned spawnHidden() helper is deliberately NOT built: the polyfill default plus the tripwire achieve the no-drift goal without indirection over seven heterogeneous call shapes. The polyfill + spawn-hide tests join the Windows CI shard. Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(browse): self-repair broken Windows DACLs on state dirs (#1605) icacls '/inheritance:r /grant:r' can partially fail on localized or domain accounts: inheritance strips but the user grant doesn't resolve, leaving a machine-SID-only DACL the owner can't even list — the sidebar/PTY failure chain in #1605, caused by the very hardening call meant to protect the dir. mkdirSecure now verifies listability after hardening (a real readdir — fs.accessSync doesn't consult NTFS ACLs) and repairs via icacls /reset, re-hardens, and if hardening breaks access again leaves inherited ACLs: functional-but-unhardened beats hardened-but-unusable. The icacls calls carry windowsHide (#1835's last two sites) and the fork's static spawn-hide tripwire lands here, pinning every covered site. file-permissions.test.ts is already in the windows-free-tests curated shard, so the DACL contract executes on windows-latest. Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(browse): opt-in session persistence — auth survives daemon restarts (#778, #2193) BROWSE_PERSIST_STATE=1 snapshots cookies + per-tab URL/localStorage/ sessionStorage to <stateDir>/session-state.json (0600) on a 30s unref'd interval and at clean shutdown, and restores on the next launch — killing the top-complained auth-lost-on-restart class (#778, #2193, #1128, #1129). Security invariants mirror state save|load: loadedHtml and owner are never persisted and never accepted from disk; restored cookies pass the same hygiene filter (localhost/.internal/metadata domains dropped); restoreState re-validates every URL. Default OFF; headed mode excluded (the persistent profile owns that state). Hardened past the fork's shape per review R3: corrupt state quarantines to .corrupt (forensic artifact, boots fresh, one log line), snapshot failures warn once and never kill the daemon, and the boot log reports restored counts or fresh-session status. Module + 10 tests ported (MIT header retained); server wiring at launch, interval, and shutdown; skill docs section added (regen rides the cluster regen commit). Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * chore: regenerate browse SKILL.md for cluster C (session persistence docs) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(skills): third-party web-actions contract — offer to drive vendor-site steps, never just dump a manual list When a workflow needs something done on an external website the user controls (register an API key, create a vendor account, configure a dashboard/webhook/OAuth app), five skills (ship, spec, office-hours, setup-deploy, land-and-deploy) now follow one contract: offer to drive it in a visible browser via gstack's own stack ($B headed + handoff/resume, GStack Browser) behind ONE per-task consent question naming the exact site and actions; passwords, payment, CAPTCHA, and identity stay user-performed; captured secrets go to owner-only files or the user's secret store, never chat/logs/history; and the credential is verified with one non-mutating API call before any success claim — dashboards show masked placeholders, and a 401 catches them. Declining yields manual steps and a blocked-on-user mark; nothing new is ever installed to close the gap. New resolver token {{THIRD_PARTY_ACTIONS}} (adapted from the fork's contract — their Aside-browser detection swapped for our own driver stack; MIT portions noted). Parity guards bumped with growth itemized (ship 1.10->1.12 at measured 1.103x; office-hours skeleton 101K / 1.09 at measured 1.079x); ship goldens refreshed. Ported from time-attack/gstack (GStack 2), driver adaptation ours. Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(office-hours): design docs land in the repo, written as decision records (#703, #2000) Office-hours Phase 5 now dual-writes the design doc: the docs/designs/ copy is what teammates and plan reviews read (committable, visible), while the ~/.gstack copy keeps memory ingest and cross-session discovery working. The repo copy leaves the private store, so it passes the redaction scan-at-sink first (HIGH blocks the repo copy, MEDIUM confirms per finding), and any failure — read-only checkout, non-git dir, unconfirmed finding — degrades to the private copy with a one-line reason, never blocking the handoff. The doc itself is now a decision record, not a transcript: one bullet per decision with its why, ruled-out approaches collapsed to a single line with the rejection reason, settled/empty template sections omitted. No page cap; extra length must come from genuinely open questions. Plan reviews (ceo/eng/devex + the shared review resolver) prefer the repo-local doc (DESIGN.md, then newest docs/designs/*.md) when it's at least as fresh as the private copy — a stale old repo doc never shadows a newer session. Parity guards bumped with measured values (three plan-review skeletons +~0.7KB each; office-hours 1.092x). Judgment ported from time-attack/gstack (GStack 2); scan-at-sink and freshness-preference adaptations ours. Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(office-hours): 'never show me these again' for the founder-resources pitch (#538) The Phase 6 resources offer (34 PG essays + Garry/YC videos) had no permanent decline — the reporter showed memory instructions kept being overridden on every update, so people who said no got re-pitched forever. The offer now closes with a standing choice; opting out runs gstack-config set founder_resources false (new key, default true, true|false validated), the write is VERIFIED before any promise (a failed write says so and skips this session only), and every future session skips the entire section silently — no resources, no 'skipped as requested' mention. Config outlives session context, so never means never. Re-enable anytime: gstack-config set founder_resources true. The pitch stays default-ON for everyone who never opted out. Tests pin the key's default/persistence/validation through the real config bin and the generated section's gate-before-content + write-verify contract. Approved as a promo-surface change (CEO review D3.4, 2026-08-14). Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(ship): the Apple App Store release journey — working tree to Submit for Review Point /ship at a repo with an .xcodeproj, .xcworkspace, or app-product Swift package and ask to release: the adapter runs the whole journey with ONE authorization moment (membership + pricing + in-session sign-in, decision- store persisted so repeat releases ask nothing) and one store-assets question only when assets are missing. fastlane is the single tool (produce/cert/ sigh/gym/pilot/deliver/frameit); credential vocabulary never reaches the user. The adapter carries 21 live releases' worth of paid-for Apple knowledge: the web session mints the permanent upload key itself (iris POST /v1/apiKeys; privateKey is base64-of-PEM, downloadable only at creation) so nobody ever types an app-specific password; error -22938 is Transporter asking for a key, not a user task; errors are CLASSIFIED before credentials are touched (validation/UnexpectedResponse = metadata, incl. Apple's expanded age-rating attributes); pricing goes through POST /v1/appPriceSchedules because fastlane's price_tier is broken against the current API; and store distribution NEVER routes through the branch gate — a clean tree on main is the solo shipper's normal case (Step 0.9 loads the adapter BEFORE the gate, pinned by test with the non-Apple gate byte-unchanged and unique). Uploads/submissions follow an idempotency-log contract (inspect App Store Connect before any re-run). Non-Mac hosts get the honest split: build legs via a macOS CI runner with the minted key as a secret, API legs local. Browser use inside the journey is banned except the named paid-app banking/tax residue. Redaction dry-run clean. Ship's parity ratio raised 1.12 -> 1.22 deliberately: the 14.8KB section is on-demand (Apple store targets only), one manifest line otherwise. Ported from time-attack/gstack (GStack 2), refined across its 21 live releases; architecture adaptation (carved section, decision-store paths, idempotency log, third-party-actions handoff) ours. Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(code-intelligence): provider contract Phase 1 — GBrain, Sourcebot, Graphify behind one ask-once offer Open a large repo (1,000+ tracked files) and gstack can offer code intelligence ONCE, with the trade-offs stated: GBrain (semantic memory + code, sends content to YOUR gbrain DB, per-repo consent), Sourcebot (self-hosted whole-repo search, local on localhost), Graphify (local tree-sitter graph, nothing leaves the machine, user-installed), or No indexing — a decline persists machine-wide so no skill ever asks again. Small repos never see the question; grep stays the always-working default and provider-OFF degrades silently (PROVIDER_UNAVAILABLE -> file-only). Ported: lib/code-intelligence/ (contract + 3 verified adapters + picker + selection + suggest, MIT headers), the gstack-code-intelligence CLI (suggest/select/consent/index/search/status), 31 offline tests (fake CLI shims + injected fetch), and the provider-contract design doc. Verified live on this repo: suggest fires at 1,233 files with real availability detail per provider. Hardened per review: the per-remote trust store is the SINGLE consent authority — a gstack-gbrain-repo-policy deny tier vetoes any recorded code-intelligence consent (fail-closed on an unreadable store, pinned by three tests); both send-capable adapters are registered as fail-closed MODULE_SINKS in the egress tripwire so a refactor can't drop their receipts; and local-compute vs remote-send consents are never bundled. setup-gbrain gains the provider-choice Step 0. The fork's Phases 2-4 glue-collapse is explicitly NOT ported. Ported from time-attack/gstack (GStack 2); consent unification ours. Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(ci): supply-chain hygiene — secret gate on every PR diff, dependency review, OSV, dependabot, evidence-bar PR template The repo owned a redaction engine and had zero CI-side secret scanning. quality-gate.yml now pipes every PR diff's ADDED lines through our own bin/gstack-redact (gate-secret-scan.mjs, taken from the fork — it dogfoods the engine): HIGH findings fail the check, MEDIUM prints an advisory count only (no human in CI to confirm), planted-bug fixtures excluded by pathspec. Live-verified both directions: PEM key fails, clean diff and MEDIUM shapes pass; ShellCheck (errors) covers the setup/build shell boundary and passes today; bun audit gates critical advisories. Trigger is pull_request, never pull_request_target. dependency-review.yml adopts the hardened never-merged prior-art branch (fail-on-severity high, workflow paths watched, tight perms) — verify the dependency graph parses bun.lock with a canary bump before trusting the gate. dependabot: weekly, grouped per ecosystem, capped PR counts; and evals.yml image build/push now skips dependabot actors, whose read-only GITHUB_TOKEN made every lockfile bump a permanently red check. OSV scans weekly with a reasoned ignore file. All new workflow actions SHA-pinned. Scorecard deliberately not taken (no consumer for the score). The PR template front-loads the evidence bar (live proof, liveness screenshot, no-ETHOS/voice-changes checklist); the unenforced DCO line is dropped. bin/gstack-verify-gate ships OPT-IN (never registered by ./setup — a Stop hook running the project's verify command after every turn is the user's call), with the fork's tests adapted to pin exactly that. Ported from time-attack/gstack (GStack 2) + our own prior-art branch. Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * chore: remove dead bins; extend the stale-ref scan to docs (the 36-release gap) bin/chrome-cdp, bin/gstack-open-url, and bin/gstack-platform-detect were referenced only by an audit test and CHANGELOG history — dead weight that the stale-ref scanner should police, which required removing them FIRST. The scanner now also sweeps docs/, README.md, and USING_GBRAIN_WITH_GSTACK — the deliberate exclusion that let a dead command survive ~36 releases as a command-not-found instruction. Scan is green on the extended surface. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(bins): detect the default branch instead of hardcoding main gstack-diff-scope fell to an empty diff (all-false SCOPE_*) and gstack-next-version mis-based its bump math on any repo whose default branch isn't main (trunk, master, local-only). Both now resolve origin/HEAD -> origin/main -> origin/master -> main. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: housekeeping sweep — telemetry integrity, persistent opt-out, context-bill accuracy, setup hang, dev-server discovery, model resolution (#2136 + v1.63 polish) Seven small fixes, one theme (claims matching code): - telemetry-sync strips local-only fields with jq del() (structural) instead of quote-fragile sed regexes; unparseable lines are dropped, never forwarded unstripped. Sed survives only as a jq-less fallback. - telemetry-log rejects non-integer durations BEFORE the range caps, whose test(1) comparisons silently no-op on non-numerics — a malformed duration spliced raw text into the JSONL stream. - browse's local telemetry honors the persistent tier (config.yaml telemetry: off), not just the preamble's env hint — direct $B use and embedders now respect the opt-out. - gstack-context-bill --exact sees GSTACK_-promoted keys inside Conductor (conductor-env-shim wired at the CLI entry), and the TOTAL line no longer double-counts every nested skill through the root skill's walk (v1.63 deferred polish; the telemetry-sync HTTP-status outcome deferred alongside it turned out already shipped). - setup's Chromium probe is deadline-bounded (90s, background + poll-kill — macOS has no GNU timeout) and prefers Node for the launch probe everywhere (the bun --eval hang family behind #2136); the install is single-flight behind a lock dir with an actionable stale-lock message. Probe verified live on this Mac. - the review resolver's dev-server check reads CLAUDE.md and the plan file before falling back to an expanded port probe, and says how to make itself smarter next time. - eval/harness model IDs resolve through lib/eval-model.ts (GSTACK_EVAL_MODEL[_KIND] env overrides, per-kind defaults, tested) at the SDK-capture and PTY-warmup sites; the bash-embedded distill snippet mirrors the resolution inline. - memory-ingest's silent-zero shape (staged>0, imported+unchanged==0, errors==0) warns even under --quiet — a run that indexes nothing must never look healthy again. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test: wire ios-qa/daemon/test into the free suite and shard runner (E2) The daemon's 5 test files (allowlist, audit, auth-mint, cli-mint, daemon-integration — now 6 with session hardening) were invisible to every runner: not in the bun test glob, not in TEST_ROOTS. The same silent-coverage-hole class as the tracked design/test P2 — and it meant B2's auth regression tests would never have gated. All files are hermetic (stub state-servers on ephemeral ports, no devices); verified green in the shard census. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(skills): claimed limitations now require evidence, everywhere + wave follow-ups filed Every tier-2+ skill's preamble gains one directive distilled from nine live release failures in two days on the fork: a claimed limitation or requirement ('the API can't do this', 'X requires a credential', 'impossible on this platform') is a material claim, stated only with the verbatim error, the documented statement, or a live probe in hand — pattern-matching a failure to a familiar story is not evidence, and a cheap probe runs BEFORE asking the user or declaring a step blocked. ONE directive adapted into the preamble resolver; the fork's full judgment contract is deliberately not imported. Full regen (46 files), ship goldens refreshed, parity guards bumped with the measured ~0.45KB/skill (investigate, autoplan, plan-design-review, office-hours), Step 0.9 registered as an intentional sub-step. Approved deferrals filed: persona-fleet hostile-user harness + answer-key methodology in TODOS; the fork's question-budget ACCOUNTING judgment (never its 5/8/12 constants) folded into the V1.1 pacing design doc; the Apple adapter added to #1882's coverage note. Ported from time-attack/gstack (GStack 2). Co-authored-by: Sina Matian <sina@time-attack.dev> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(make-pdf): close offline-gate bypasses via unquoted style attrs, CSS-escape and HTML-entity obfuscation Three live vectors found by the ship review army, all red-first tested: unquoted style attributes skipped the remote-url neutralizer entirely; CSS ident/string escapes (@\69mport, url(\68ttps://…)) defeated the literal-match patterns Chromium happily decodes; and HTML entities in style attribute values (https) decoded to fetchable schemes before CSS parsing. Style-attr values are now entity-decoded in one browser- faithful pass, escape-bearing at-rules and function tokens are dropped fail-closed, and output is re-encoded double-quoted. 21 new test rows. * fix(migrations): v1.65 Chromium re-fetch actually re-downloads, and success is verified before .done The migration (renamed from the provisional v1.64.0.0 slot, which open PR #2564 claims) deleted only the poisoned .app while Playwright's INSTALLATION_COMPLETE marker survived in the revision dir — so the advertised 'bunx playwright install chromium' re-fetch no-opped and the user finished the upgrade with no browser and a success message. Now: the whole chromium-<rev> dir goes, bunx runs cwd-pinned to the install root, .done is gated on a verified executable, and a needs-refetch sentinel makes re-runs retry a failed download. Stranded rev dirs (markers without .app) also re-trigger. 6 hermetic tests, red-first. * fix(migrations): v1.27 remediation prints a real command instead of a fictional flag Every skip/failure path referenced '/setup-gbrain --rerun-migration', which is implemented nowhere, and promised the migration 'will ask again next upgrade', which the version-window runners make false. All five sites now print the direct GSTACK_MIGRATE_ASSUME_YES=1 bash invocation. Runner-side re-offer tracking is filed in TODOS. * fix(browse): poisoned-bundle self-heal removes the revision dir, probes handoff too, and throws typed Same marker flaw as the migration: rmSync of the .app alone left INSTALLATION_COMPLETE behind, so the error message's own remediation no-opped and the user was hard-stuck. The probe is now an exported, unit-tested helper (probePoisonedChromiumBundle) that removes the whole chromium-<rev> dir, never touches GSTACK_CHROMIUM_PATH custom bundles, throws PoisonedBundleError (instanceof, not string-match), and runs on BOTH headed entry points — launchHeaded and handoff. 7 tests. * fix(browse): session snapshots are atomic and the cookie filter drops loopback IP literals A crash mid-write destroyed the previous good snapshot — the exact scenario persistence exists to survive; writes now go tmp+rename. The internal-network cookie filter gains 127.*/::1/169.254.* (a tampered state file could previously hand loopback-service cookies back to the browser), and 'state load' imports the shared filter instead of maintaining a comment-synced copy. Test cleanup made exception-safe. * fix(browse): server runtime — restore off the boot path, shutdown that cannot hang, watchdog that still reaps tunnels Four review findings on the wave's own new wiring: session restore ran before Bun.serve with sequential 15s gotos while the CLI gives up at 8s (one slow saved URL bricked every $B command) — restore now runs in the background after bind; the shutdown snapshot gets a 2s deadline so a wedged page.evaluate can't hold the port forever behind the new ack-first stop; the persistence ticker gets in-flight + shutdown gates and is cleared before the final snapshot; and the absorbed #2565 handoff fix no longer clears the whole parent watchdog — a suppress flag keeps the tunnel-orphan reaper alive (handoff→resume→tunnel is no longer an unreapable internet-exposed daemon). pair-agent with consent off now names the real remedy instead of ngrok install instructions. Lock-acquisition edge branches (garbage pidfile, vanish-race depth cap) pinned. * fix(browse): telemetry defaults to off like every other surface The persistent tier defaulted ON when the config key was absent, while gstack-config's DEFAULTS table answers 'off' for the same question — preamble-spawned daemons and direct $B daemons disagreed about consent. Absent key/file now means disabled; community/anonymous enable; env kill-switch still beats everything. Both config.yaml consumers now share one readGstackConfigYamlKey reader. 12-case consent suite. * fix(code-intelligence): consent that means what it says — polarity, receipts, read-only veto Four review findings on the wave's own Phase 1 port, all red-first: 'consent <repo> no' recorded consent GRANTED (the CLI ignored the argument and always wrote true) — yes|no is now required and garbage records nothing; Sourcebot egress receipts claimed consented=true on paths that never checked consent — the actual consent state is threaded into every receipt, search is fail-closed on non-loopback, and the liveness probe's receipt says truthfully that it sends no repo content; repoPolicyVeto only honored the deny tier while gbrain refresh writes pages — write-class ops now veto on read-only too, matching the sync chokepoint, via one shared lib/gbrain-repo-policy-client.ts (win32 bash invocation, spawn-vs-unreadable error distinction) used by both call sites. Also: source ids get a host+path hash (same-name repos no longer collide), refresh timeout raised to 120s, availability probes run concurrently at 3s, graphify status stops JSON.parsing 100MB graphs for a count, and every ported file carries the fork MIT notice. +15 tests across the two suites. * fix(verify-gate): trust before eval, re-check on re-entry, audit every grant The opt-in Stop hook eval'd whatever command the first CLAUDE.md up the tree declared — any cloned repo got arbitrary shell at turn end. Now a per-repo trust store (path+command hash, 0600) gates execution: an untrusted or changed command never runs (exit 0 with the --trust invocation printed), stop_hook_active re-entry re-runs the trusted check instead of rubber-stamping (bounded at 3 blocks per episode), and every grant appends a forensic line to ~/.gstack/security/verify-gate-trust-grants.jsonl. 20 tests, red-first. * fix(setup): EXIT traps chain instead of clobbering; timed-out probes reap their whole tree The Playwright-lock trap replaced the copied-bun cleanup trap and then cleared ALL exit handling, leaking .tmp-bun-bin on every Chromium install; and _wait_with_deadline killed only the subshell, orphaning the wedged node→Chromium tree it exists to escape — re-creating the #2136 pile-up on every timed-out re-run. Traps now chain; timeouts walk pgrep -P descendants leaves-first. * refactor(resolvers): one source for the design-doc discovery block The #703 repo-doc-preference bash was pasted byte-identically into three plan-review templates and a fourth copy embedded in review.ts — drift there means plan reviews disagree about which design doc wins. Now a {{DESIGN_DOC_DISCOVERY}} resolver; generated output is byte-identical, so no SKILL.md changes ride along. * fix(ship): finish the Apple upload idempotency sentence The durable-effect contract dropped its consequence clause mid-sentence — the instruction for what to DO when the idempotency key already exists (treat the upload as possibly-done, never re-run it) was missing from the one rule governing whether a binary uploads twice. * fix(ci): SHA-pin dependency-review; the secret gate fails closed without a report dependency-review.yml rode mutable refs (@v4 resolves to a BRANCH on that repo) inside the one workflow whose job is supply-chain hygiene — now commit-pinned like its siblings, with dependabot keeping the pins fresh. gate-secret-scan.mjs crashed with an unhandled EPIPE on oversize diffs (the designed report.oversize branch was unreachable: the scanner emits no JSON on refusal) — the pipe write now tolerates early exit and a missing report is an explicit fail-closed exit 1. Oversize + broken-scanner legs pinned. * fix(bins): Windows-safe GIT_CEILING join; next-version probes the full default-base chain GIT_CEILING_DIRECTORIES was joined with ':' — git on Windows splits on ';' and drive letters contain ':', silently disabling the #2144 second-layer defense there; now path.delimiter. next-version's default-base detection only tried origin/HEAD then 'main', diverging from the canonical 4-step chain diff-scope uses — origin/main and origin/master probes added, pinned by fixture repos. * fix(eval-model): kinds are a literal union, not string Record<string,string> widened EvalModelKind to string, so a typo'd kind only failed at runtime; as const satisfies keeps the closed set the doc comment promises. * test: coverage backfill from the ship review The telemetry-strip invariant only validated the sed FALLBACK while the live jq path went unchecked — the jq del() lists are now held to the same every-emitted-field bar, plus a behavioral pipe-through. The context-bill nested-skill double-count fix gets a regression pin (a revert shipped green before). The windowsHide tripwire gains terminal-agent-control.ts — the exact file the fix commit names. The ios-qa revoke-by-token_id branch gets its negative case: unknown ids revoke nothing and leave live sessions alone. * docs: SLATE_HOST no longer cites the deleted platform-detect bin Host detection lives in the hosts/ registry via host-config-export.ts; the doc's known-gaps list now says so instead of pointing at a bin this branch removed. * test(e2e): headroom for the two plan-ceo-review budget-edge tests Both rode their 360s runner budget at the edge (main clears at 243s of 360s), and the wave legitimately adds work to the review: the evidence directive tells the agent to probe before claiming, and the design-doc discovery block adds bash steps. Under concurrent in-file children the API queuing tipped all retry attempts past the ceiling — the runner then reports $0.00/0 turns for a timed-out child, which reads like a dead spawn but is a healthy child killed at the deadline. 540s runner / 660s test for these two only; verified 2/2 green at 228s and 315s. * fix(code-intelligence): gbrain search/export are consent-gated and receipted The Sourcebot side got this in the last round; gbrain had the same hole — search() and export() sent repo-derived query text into a possibly-remote DATABASE_URL with no consent check and no egress receipt, bypassing the deny-tier veto. Both now assert consent before any bytes move, receipts record the actual consent state (never a hardcoded true), and search receipts carry the query's sha256. gbrain stays fail-closed: the adapter cannot see where DATABASE_URL points, so every send requires consent. 7 new tests, red-first. * fix(make-pdf): SVG remote refs and image-set can no longer fetch offline <svg><image href=https://…> and <use xlink:href=…> survived the gate (only javascript: schemes were stripped from svg hrefs), and bare-string image-set("https://…" 1x) dodged the url()-shaped neutralizer. Remote svg hrefs rewrite to '#' (entity-decode-aware, unclosed-svg smuggle closed) and remote image-set args neutralize to url(#). Local fragments, local image-set, and plain <a> links pinned intact. 12 new rows, red-first. * fix(browse): duplicate config keys read last-wins, matching gstack-config readGstackConfigYamlKey took the FIRST match while gstack-config's get takes the LAST — a duplicated pair_agent or telemetry line made the two consent surfaces disagree about what the user chose. * fix(setup): stale Chromium-install lock self-heals The mkdir mutex had no owner: a SIGKILL'd setup left the lock behind and every later run exited with manual rmdir instructions. The holder pid is recorded in the lock; a dead holder is reclaimed automatically. * fix(setup-gbrain): the code-intelligence offer gate skips when the bin is absent The new Step 1.7 told the agent to run gstack-code-intelligence before the path pick — on installs predating the CLI (and hermetic E2E children) the bin doesn't exist and setup derailed before doing any setup. The gate now probes for the bin and reports offer:false reason:bin-absent, with explicit instructions to proceed: the user asked for gbrain, so set up gbrain. Never block setup on an optional gate. * test(e2e): periodic-tier repairs from the failure triage Each fix traces to a receipt: brain-privacy-gate staged config never reached the hermetic child (ambient GSTACK_HOME is scrubbed) and the operator's remote-mode gbrain suppressed the gate — both now injected per-test; ship-idempotency threw away its evidence on the timeout path and ran a 600s budget its own subject can exceed (now 900s, evidence captured); auto-decide-preserved gets the same headroom its sibling plan-ceo tests got; context-skills' hides-checks scanned bash output where an ls legitimately names old checkpoints (final-text scope now); design names the missing section instead of a bare count and learns the easing/duration/micro-interaction synonyms; qa-workflow's collector afterAll gets an explicit 60s hook timeout. * fix(eval-harness): eng-review phase boundary fires on qid-tagged questions The Step 0 boundary only matched two prose phrases, but plan-eng-review may legitimately reach the review phase without either — every per-finding AskUserQuestion then counted as pre-review and the batching regression test read 0 questions while watching the agent ask them one by one. The boundary now also fires on the first answered question carrying a gstack-qid:eng-review- marker. Additive only; 119 runner unit tests green. * chore: bump version and changelog (v1.65.0.0) Fork port wave 2: the release-summary entry credits Sina Matian (time-attack/gstack) and the four absorbed community PRs. TODOS gains three review-round follow-ups (dual-write E2E, migration runner re-offer, gbrain-adapter op coverage). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(eval-harness): eng-review qid boundary matches the real skill-name prefix Live qids render as gstack-qid:plan-eng-review-<slug> ({skill}-{slug} convention); the boundary anchored eng-review- immediately after the colon and never matched, leaving the batching counter blind while the transcript showed per-finding questions being asked one by one. * fix(setup-gbrain): never ask the provider question inside /setup-gbrain Invoking /setup-gbrain IS the provider choice. Step 1.7 now records 'select gbrain' best-effort and proceeds straight to setup; the offer ceremony is reserved for entry points where no provider was named. On machines where the code-intelligence CLI exists, the offer:true path was hijacking setup into the provider ceremony and the E2E child never reached MCP registration. * chore: file the three documented-red periodic tests as structural-repair TODOs Sidebar trio exercises endpoints removed on every tree; ship-idempotency's PTY child never receives its typed command; brain-privacy-gate has never been green anywhere. Each carries its triage receipt in the entry. * test(e2e): setup-gbrain remote — hermetic env via opts, evidence on failure, output-scoped classifier Three separate defects stacked on this one test: the ambient GBRAIN_MCP_TOKEN/GSTACK_HOME/PATH mutations never reached the child (hermetic-env scrubs them by allowlist — broken since hermetic env landed; the child correctly stopped at Step 4c with NEEDS_CONTEXT), failures discarded the in-memory transcript so every triage started blind, and the wrote-findings-before-asking classifier scanned the full event stream where the child's own Read of the skill file always contains the review-report phrase. Env now goes via opts.env, failures dump bash commands + final text, and the classifier scans assistant output only. Green in 67s with all seven asserts. * test: final coverage pass — CLI rendering, revert traps, keychain probe, gbrain doc ops The user-directed third generation pass closes the audit's remaining tail: the code-intelligence CLI's options/status/suggest surfaces get behavioral coverage through the fake-shim chain; brain-context-load gains an argv-logging trap that goes red if anyone reverts the memoized PATH scan back to the spawn probe (receipt: simulated revert failed exactly these tests); the darwin Keychain auth branch (#1890) gets its first free-tier tests via a PATH-shimmed security binary; and the gbrain add/delete/export ops are pinned (body piped byte-for-byte, receipt sha256, stdin-EOF prompt guard, PROVIDER_UNAVAILABLE degradation) — retiring their TODOS entry. * test: assemble the planted PEM at runtime so the fixture never trips the prepush guard The repo's own credential guard scans pushed diffs and correctly blocked these fixtures: the engine flags any one-line BEGIN…END spelling regardless of body. Header, body, and footer are now joined at runtime, so the file and every diff of it stay clean while the scanner under test still receives the true live shape. * docs: update project documentation for v1.65.0.0 README gains the two wave-2 CLIs (gstack-code-intelligence, gstack-verify-gate) in the standalone-binaries table, BROWSER.md documents BROWSE_PERSIST_STATE next to manual state save/load, CONTRIBUTING's CI section lists the new supply-chain gates, and CLAUDE.md's project tree reflects lib/code-intelligence/ and the added workflows. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs: apply cross-model doc-review fixes for v1.65.0.0 Findings from the release doc review, verified against source: verify-gate's README row gains the actual install one-liner (setup never registers the Stop hook; test/verify-gate.test.ts pins that) and the 3-blocked-re-entries yield behavior; code-intelligence's row gains the suggest subcommand and the search-side consent gate; CONTRIBUTING scopes the SHA-pin claim to the supply-chain workflows and widens the dependency-review trigger; BROWSER.md's restore-time cookie drop list matches isInternalCookieDomain; CLAUDE.md's workflows comment stops implying six workflows are all of them. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs: CHANGELOG accuracy pass — scope the SHA-pin claim, restore-time cookie filter, exact test counts * test: env restore runs per-test, not per-suite — the leak that failed 30 strangers gstack-memory-helpers saved HOME/GSTACK_HOME/PATH in beforeEach but restored in afterAll, so the last beforeEach's snapshot won and a gstack-test-engine temp dir leaked into every later file in the same process: gstack-config read the wrong store, make-pdf's child resolved Chromium under the temp cache, update-check and artifacts-init lost their real homes. afterAll is now afterEach; the config and update-check harnesses also strip GSTACK_HOME/GSTACK_STATE_ROOT from child env as a belt. * fix(browse): restore the #1846 start-timeout resolution the merge dropped The v1.64.1.0 merge kept this branch's lock design in cli.ts and silently lost main's resolveStartTimeout + late health re-check while their test survived — ported both back in alongside the kept design. * test: adapt main's diagnostics tests to the merged designs cli-lock asserts typed ServerLockError (errno + lock path) instead of the log-and-return shape the merge didn't keep, dropping only the one duplicate of server-lock-errors coverage; the liveness tripwire exempts error-handling.ts as the sanctioned tasklist site; snapshot and compare-board wrappers pass the now-mandatory browser-manager arg; background.js's test pins that the retired sidebar-command type is rejected pre-gate with no response fields. * chore: gitignore the gen-accessors tool's SPM build output skill-e2e-ios-swift-build compiles the Swift package in place, leaving .build/ (2,800+ files) and Package.resolved untracked after every periodic run — the workspace read as ~100 dirty changes with a clean tree. Same class as the dist/ binaries: build output, never committed. * test(browse): subprocess budget for the polyfill suite on Windows CI Every test here spawnSync's a node child; cold-start on the Windows runner (AV scan, first node.exe touch) blew bun's 5s default by 7ms on a 50ms sleep test. File-level 20s default — subprocess budget, not assertion looseness. * test: make the Darwin migration path and the query-timeout SKIP deterministic on Linux CI The v1.65 migration suite relied on the host being macOS — on the ubicloud runner the script's uname gate early-exited every test with empty output; a Darwin uname shim in the shared setup runs the real path everywhere (the non-Darwin test still overrides it with Linux). The 1ms-budget brain-context test assumed 1ms is always too short; the runner's fake gbrain answered in 0ms and no SKIP printed — the fake now sleeps 300ms so the timeout is a certainty, while --version stays instant for the detection assertion. --------- Co-authored-by: Gawie van Blerk <gawievanblerk@gmail.com> Co-authored-by: Sina Matian <sina@time-attack.dev> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: Shawn Reddy <19191746+Screddyice@users.noreply.github.com> Co-authored-by: Jake Wilk <jwilk@highlinerepartners.com> Co-authored-by: Jerry Nichols <jerrynicholsai@users.noreply.github.com> |