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.
- 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.
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.
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.
- 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.
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.
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.
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.
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.
- 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'.
- 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.
- 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.
- 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.
- /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.
- 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.
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.
- 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.
Main shipped v1.91.9.0, so this wave becomes v1.91.10.0. Conflicts kept this
branch's planner-derived assertions. Main's new test-value eval gets rule
kinds for its three gate cases and its measured 332 s duration from #2998's
CI; census counts and the PR fallback floor follow the new file. The packing
test now weighs each tier's recorded durations the way the planner does.
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.
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.
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.
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.
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.
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.
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.
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.
- 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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
garrytan