0 Fork 0

mirror of https://github.com/garrytan/gstack.git synced 2026-11-08 09:48:20 +01:00

Files

T

Garry Tan 55f8151fbe v1.91.30.0 fix: honest eval failure reports, Bun floor with safe auto-update, make-pdf and /review on Codex (#3033 )

* docs: approved reliability follow-ups plan (Oct 2026)

* D1: score overlay wrong answers as comparison results (contract v4)

A completed answer that fails the fixture's output check is now a valid
measurement with taskCorrect false and answerError set, in either arm.
Only execution, scope and metric failures invalidate a trial. ON
correctness still gates the case, so an OFF wrong answer stays in the
comparison. OVERLAY_CONTRACT moves to v4; v3 verdicts are not
reinterpreted. Replays the 37198445662 sonnet trials (OFF-0 wrong).

* C4: score auq-matrix substance with a 3-sample judge panel

Each skill keeps one native capture in the same execution; the
recommendation-substance grader now runs as a JUDGE_PANEL_SAMPLES panel on
that capture and gates on the mean against the unchanged minimum of 4. A
failed sample fails the panel without resampling. Approved at the gate
2026-10-04 for censuses after this change.

* E1: one Bun version source; setup refuses below the 1.3.3 floor, updater checks before advancing

bin/gstack-bun-version.sh holds the security floor (1.3.3) and the tested
version (1.4.0, the CI pin and engines.bun). setup refuses Bun below the
floor before writing anything, warns between floor and tested, warns and
continues on an unparseable version, and prints a 1.4.0 install hint.
Prereleases sort below their release (semver).

gstack-session-update now fetches, reads the incoming release's floor and
checks the bun setup will run before fast-forwarding to that exact
revision; below it the update is held with reason bun-too-old and a
resume command, leaving the revision and installs untouched.

Fixture copies of setup (team-mode, claude-code migration) now include the
sourced file; the slow-git lock test sleeps on fetch, the updater's
network step.

* E3: record the owned Chromium's PID over CDP so the #2709 reap runs on Playwright 1.62

Playwright 1.62's Browser has no process(), so chromiumProcInfo was always
null and browse stop never reaped a surviving headless Chromium. The owned
headless launch now reads the browser PID from CDP SystemInfo.getProcessInfo
(bounded at 2s; a hung or failed call logs once and records nothing). The
reaper's start-time and cmdline identity checks are unchanged, so a recycled
PID or an externally owned browser is never signalled.

* A1: ship docs gate records post-child hashes with gstack-docs-candidate, saves each record once

Stored transcripts (37198445662 gate-census-2) show the parent's reads of
release files take ~1 s; the parent's own model time dominates (completion
268 s of 392 s, store ~270 s of 387 s), and 45-78 s of it is retyping hash
maps, child output and records into several artifacts after the child
returns. The child contract is unchanged.

* A2: /plan-ceo-review sends the mode handoff line before any other tool call

plan-ceo-mode-routing's HOLD SCOPE session spends its 240 s post-selection
window on model work: 37176837432 skipped the handoff helper and paraphrased
the mode, 37198445662 ran it, then wrote the plan, loaded sections and probed
Codex for 114 s before the posture chat. The handoff line is the posture.

* B3: accept wrapped unchanged rows in the cropped Edit preview

PR run 37176835584 timed out with the floor's owned Edit menu on screen:
unchanged diff rows soft-wrap under their gutter with a blank marker
column, and a trimmed empty unchanged row keeps only its number. The
parser accepted wrapped +/- rows only, so it never verified the owned
edit. Both forms are now parsed and still compared with the exact source
lines; the captured viewport is replayed with negative controls.

* A3: record per-phase headroom diagnosis for docsync, plan-mode-no-op, mode-routing and gate slice 11

* E4: /review Step 1 keeps going when the base fetch fails and reports stale base coverage

Under Codex's read-only .git, 'git fetch && merge-base && diff --stat'
stopped at the fetch, so the empty output read as 'Nothing to review'.
The fetch is now separate, prints BASE_REFRESH fresh or stale with the
local base revision, and the review states stale coverage.

* D2: reword the format-approach question prompt

Both plan-ceo-review-format-approach refusals in census 37195203538
(Fable 5.1 model_refusal_no_fallback, category reasoning_extraction:
"duplicating model outputs") were anchored on the session prompt, whose
capture instruction asked for the "verbatim" question and "the exact
prose you would have shown" in a "format-capture test". The approach
case now asks for the same question in the same file as the step's
output, without that wording. The other three plan-format cases keep
their unchanged prompt (no refusals in 72 trials).

* C2: /qa defines its outcome filename fields; QA report templates name the measured timing fields

Low judge samples in failing panels named both: qa workflow (37186854666)
'an undefined {user} in the outcome filename', and qa-only workflow
(37198445662) 'template asks for Duration / durationMs totals while
reporting.md mandates Probe budget and Guarded command time'. The qa
union stays under its 1.11 size ratio (93292 of 93294 bytes).

* 0.1: session ledger written by the claude -p, Agent SDK, PTY and Codex runners

Each armed model session appends one best-effort line to
$GSTACK_EVAL_DIR/session-ledger.jsonl: a stable session key, the timeout it
armed, monotonic elapsed time, its end reason (completed, session_timeout,
observer_timeout, api_error, refusal, aborted, error), one sanitized evidence
line and, for stream-JSON runners, a compact liveness summary (longest
in-request silence with nothing outstanding, last event kind, tool open at
end). Partial messages now stream for every claude -p session and for SDK
sessions without a custom query provider (runSharedInteractive opts in); their
rows feed only the summary, so transcripts keep their size unless
publicStreamDiagnostics is set. Structured refusal and API-error signals win
over phrases, and tool results are never read for them.

* 0.2: failure_cause beside failure_class, from runner facts and the session ledger

One pure failureCauseOf() beside failureClassOf() sets failure_cause
(contract > pre_turn_infra > api_error > refusal > provider_stall >
session_timeout > observer_timeout > assertion > unknown; no machine-set
'detector') with one bounded evidence line, and failureDetailOf() keeps a
Bun matcher's Expected/Received or a judge's failing dimensions in
failure_detail. Both record builders (classifyTrialShard, junitCensus) call
them and attach the trial's ledger sessions. failure_class, error and
TRIAL_ERROR_MAX are computed exactly as before; the new fields are optional,
bounded and validated only when present, and a frozen copy of the v1 reader
accepts every new record. A provider stall needs partial messages and
STALL_WINDOW_MS (120 s, eval-budgets.ts beside HEADROOM_WARN/FAIL) of
in-request silence with nothing outstanding; it stays a failure under
EVAL_POLICY v1.

* E4: fingerprint under a read-only .git via a private object dir owned by review-log

Codex's sandbox mounts .git read-only, so gstack-wtree's git add -A could not
write blobs and /review lost its start capture. gstack-review-log now creates
a private 0700 object directory under TMPDIR, passes it to gstack-wtree (which
writes new objects there with the common-dir object store, resolved before
exporting, as an alternate, so linked worktrees work), exposes it to the
bindReview/checkSharedLibsReuse consumers that read the tree and blobs, and
removes it on EXIT/INT/TERM after they finish. gstack-wtree prints one stderr
line and exits 1 on a failed write or unwritable TMPDIR instead of printing a
partial hash; review-log no longer swallows that line. Standalone gstack-wtree
keeps writing to the repository store so git diff between saved snapshots
(/ship) still works.

* 0.3: red lines name their values, evidence and an after-a-repair command; one published-text sanitizer

Each red line in the census report keeps its stored error and adds the
machine cause with its evidence, the failure_detail values (Expected/Received,
or each failing judge dimension with its mean, sample count and the sample
sentence naming it), the slice artifact and shard directory that hold the
transcript, a fetch command (eval:pass-rates --run <id> --case <id>), and
'after a repair:' with a command that selects only that case (E2E ids via
--case, judges via their selection and an exact test-name filter). A
session that completed and failed a check says so instead of 'assertion
(success)'. Report text that reaches a PR comment, issue or committed file
goes through scripts/lib/published-text.ts: control and bidi characters
stripped, credential-scan patterns redacted, mentions neutralized, HTML and
link brackets escaped inline, blocks in a fence longer than any backtick run,
length capped. Verdicts and collector verdict objects are unchanged.

* E2: one make-pdf resolver for readiness and every $P block on env-var hosts

make-pdf exported $P from its readiness block, which Codex and the other
env-var hosts lose between blocks. binaryAssignment gains 'make-pdf': the
MAKE-PDF SETUP readiness check and the runtime prelude both use it, so every
block that runs $P derives the same make-pdf/dist/pdf binary, with
MAKE_PDF_BIN (dev builds) winning when it is executable. Claude output keeps
its repo-local-first probe; only the false 'subsequent blocks' comment
changed. Root plus P=... fits the 400-byte prelude budget; the fixed
MAKE_PDF_OVERRIDE segment is the one exemption, named in runtime-root.ts and
the conformance test.

* 0.1: eval:pass-rates --headroom, --reds and --run report views; 85% headroom alarm in --gate

A new pure module, scripts/lib/eval-history.ts, reports per case the slowest
armed session against the timeout it armed (censored when it timed out,
insufficient below 3 samples, unknown without a session ledger, with the
case wall shown separately as an upper bound), the census critical path from
the slice jobs' timings, the verdict red ledger per census by failure class
and machine cause with shrunk per-case rates, and the all-green probability
(an approximation: the product of 1 - p_i, with an interval). --run <id>
lists one census's reds with their values and history and downloads only the
slice artifacts those reds name, reporting any over 64 MB instead of skipping
them. eval-flake-rank.ts only routes flags: --help works offline, an unknown
flag or case id exits 2 before any fetch, --case narrows the JSON and the
alarms, every view prints its run/branch scope, and --gate also alarms on a
session above HEADROOM_FAIL (85%) of its budget.

* E5: GSTACK_DESIGN_IMAGE_MODEL override; the live model check always tests the defaults

The image_generation tool model gains an env override validated against a
conservative gpt-image-<version> pattern; a bad value fails before any
request and the error names the variable. Pairing and rejection hints name
the tool model actually requested. design/scripts/live-model-check.ts now
ignores GSTACK_DESIGN_MODEL and GSTACK_DESIGN_IMAGE_MODEL (and says so), so
an override cannot mask a retirement of a default, and its exit 2 (no key)
says nothing was checked.

* C3: /deslop-shared-libs names a concrete risk for each recommendation, never 'none'

Both shared-libs-opportunity-judgment reds (37193478719 t3, 37195203538 t1)
failed adoption_and_risk because the top recommendation said 'Main risk:
none found inside the repo' / 'none in behavior'. The skill asked to 'name
the main risk or uncertainty' without saying what counts.

* B4: --case selects loop-registered cases; an unknown id exits before any process

caseSelection() resolves the one paid file that registers a case and how a
trial selects it: by its literal Bun test name, by the id a registration loop
names its test at runtime (coverage audits, TPA, journeys), or by running the
whole file when it registers no other case (plan-mode-no-op, whose test names
are not the id, so the old name pattern ran nothing). A whole-file case runs
one file shard per trial and is judged by shard status, as CI judges it.
--list prints the selection. Every E2E_TIERS id now resolves except
carve-section-loading, which fifteen files register (CI runs each as its own
shard) and which still fails with the owning files named.

* 0.3: after-a-repair runs the shard file when --case cannot resolve a case

A case that several files register (carve-section-loading) cannot be selected
by --case; its red line now prints the bun test command for the shard file
that went red, so every printed command selects exactly one case.

* C1: pre-register workflow-judge control bundles for qa, qa-only, review and setup-deploy

Positive (production bundle), negative (headings only) and one targeted
execution-blocking defect per judge, one 3-sample panel each, built with the
eval's own reader, prompt builder, options and floors.

* F: cost_known: a trial whose harness captured no billing says so; totals never hide it

cost_usd stays a finite number. Trial and eval records gain an optional
cost_known boolean (absent means known, so old records still parse; the
validator accepts only a boolean). Session-ledger rows record whether the
session's terminal result carried total_cost_usd; PTY and Codex sessions
never do, and SDK captures without a terminal result do not. A trial is
cost-known only when it has eval records and every record and session
captured billing, so JUnit-only PTY cases (125 of 298 trials in run
37198445662 recorded 0) now read cost unknown. The census headline prints the
known sum plus the unknown count, never a bare total over unknown costs, and
no total when every cost is unknown; the history import keeps the flag.
PTY transcripts are not priced: test/helpers/pricing.ts estimates exclude
cache writes, so they cannot give the billed cost.

* B1: replay corpora for the five repeat-offender cases

shared-libs-plan-callers, plan-ceo-section-loading, auto-decide-preserved,
plan-design-review-plan-mode and ship-docsync-late-result each get an
inventory of every wave census in scope (all artifacts still available)
and minimal, redacted detector or actor inputs replayed through the real
detector in the free suite: census passes, every distinct census red, and
derived negative controls labelled as derived. Fixed false reds stay as
expected passes with their history; the two genuine auto-decide product
failures (37176837432, 37179171083) stay refused. A contract test bounds
each entry at 64 KB, scans it for HIGH credentials, requires a pass and a
refusal per case, and reports entries whose series identity is stale.

Building the auto-decide corpus exposed a latent native detector misread:
a bold 'Mode application and rationale.' heading was parsed as a
withdrawn Mode field (37186854666). A field status now ends at its
sentence.

* B2: read approved stale-fill decisions from more valid ledger records

The structured ledger reader is authoritative before the prose fallback,
but it missed two valid approved records in passing censuses
(37186854666: a camelCase writeProfile subject; 37195203538: a
slash-joined 'D1 / STALE-FILL' row id), which then rested on the prose
detector that caused both historical reds. Both forms are accepted; an
unresolved row, an unrelated decision, the copied fixture plan and
non-write camelCase words stay refused.

* C1: record workflow-judge control results (negatives separated; qa and review miss a removed step)

* C2: record every low workflow-judge sample in the wave censuses with its disposition

* 0.3: published lines cannot break a fixed fence

evals.yml wraps collector-outcomes failures in a fixed three-backtick fence
for the PR comment, so those lines (and eval:pass-rates output meant for the
tracking issue) go through sanitizeFixedFenceLine: cleaned like every
published line, with backticks turned into quotes. Inline published text
also neutralizes backticks.

* A1: fit the post-child record wording under ship's 1.428 parity size cap

* E4: inline the Step 1 base refresh to fit review's 1.199 parity size cap

* B2: inventory of gate and periodic phrase-only detector checks

docs/evals/detector-inventory.md lists every pass/fail decision in gate
and periodic E2E cases that rests on phrases in model free text, with
its structured source (if any), whether the contract is user-visible,
and a disposition: already structured-first, convert candidate, phrase
witness required, or no structured source. Convert candidates are
ordered by census red history.

* 0.2: failureCauseOf takes the stall window instead of importing it; the PTY ledger never throws

eval-store.ts imported STALL_WINDOW_MS from eval-budgets.ts, which free tests
replace with mock.module stubs (review-finalization-budget, office-hours,
qa-fix-loop, qa-only), so loading eval-store failed there. failureCauseOf is
now dependency-free: the record builders pass STALL_WINDOW_MS. A PTY session
whose handle cannot show its text (test doubles) records no panel instead of
throwing from runPtySession's cleanup. The 37186854666 replay also runs
through classifyTrialShard.

* A2/auto-decide: the CEO handoff chat repeats the helper's printed line verbatim

Claude Code 2.1.284 collapses the helper's Bash output ('Ran 1 shell
command'), so the 'Auto-decided review mode → …' line reaches the user only
through chat. In lane D's auto-decide-preserved corpus every passing capture's
chat starts with that line, and the reds paraphrased it ('HOLD SCOPE was
auto-selected…', 'Mode decision done…', "I'm defaulting to HOLD SCOPE…").
The template now says verbatim and why.

* C2: record the judge panels rerun on the final ship and review wording

* integrate: touchfiles and module-size ratchet entries for the wave's new modules

GLOBAL_TOUCHFILES gains session-ledger.ts, published-text.ts and
auq-substance-panel.ts; bin/gstack-bun-version.sh joins every entry that
lists setup or the updater; auq-matrix stays a rule case with its C4
evidence beside E2E_KINDS; eval-history, published-text and session-ledger
are registered new modules.

* E1: BROWSE SETUP install hint pins the tested Bun (1.4.0)

* E2: quote "$P" in make-pdf commands so install paths with spaces work

* E1: /gstack-upgrade checks the incoming Bun floor before pulling

gstack_bun_incoming_hold (bin/gstack-bun-version.sh) is now the one
pre-advance check: the auto-updater calls it before its fast-forward and
/gstack-upgrade's git path calls it after fetching. Below the incoming
release's floor the upgrade stops with BUN_TOO_OLD and the checkout,
VERSION and installs stay untouched. Vendored installs were already safe:
setup refuses before writing and the swap restores the previous install.

* 0.3: only main posts or closes tracking issues; reports carry alarms, red ledger, headroom and all-green

A branch-dispatched periodic or marathon run writes its report to the step
summary and a report artifact; the upsert and close steps run only on
refs/heads/main. The periodic body now keeps every pass-rate alarm line (the
old grep kept only the header), adds the red ledger and session headroom read
with an explicit --branch, the all-green estimate and the triage guide link,
and routes report text through scripts/lib/published-text.ts. The marathon
lane gains the matching close-on-green step.

Migrated: evals-workflow-wiring pinned report-summary.md in the upsert step;
the body is now written by 'Write the census report'.

* C3: prosons neutral-posture detector ignores a negated "taste call"

Both plan-review-prosons-neutral-neg reds (37179171083 t3, 37193478719 t3)
recommended A and closed with "a coverage call, not a taste call"; the
old /taste call/i matched the negation. The detector moves to a helper so
the two census captures replay free as negatives, with positive controls.

* B2: credit the CEO mode handoff only when the chat shows its line

Claude Code 2.1.284 collapses the gstack-ceo-mode-handoff Bash result, so
the printed AUTO_DECIDE line is user-visible only when a later chat message
of the session begins with it. The native detector now requires that
witness; the helper's tool result locates it but no longer replaces it, and
a generic mode statement cannot stand in once a handoff ran. Census
37182865432 is relabelled a genuine miss (expected: refused): the chat said
"I'm defaulting to HOLD SCOPE…" and "Mode decided: HOLD SCOPE". Corpus
replay: 5 credited, 3 refused. ceo-mode-handoff tests migrated to the
witness rule (deliberate behavior change).

* test: baseline the wave's approved detector and harness replay tests

B1 corpora, B3, C3, C4 and D2 replay captures through paid-case detectors,
actors or fixtures (test code by design); each entry carries its reason.

* E6: live Codex multi-block periodic case in the host-run Codex job

test/codex-e2e-multiblock-live.test.ts runs ./setup --host codex once from the
job's checkout into a fresh HOME and CODEX_HOME (auth.json copied from the CI
login home), asserts codex debug prompt-input lists the gstack router as a
real file, seeds one learning, and runs /learn's Show recent block through
codex exec: that later block must root itself through its prelude, run the
installed gstack-learnings-search and print the seeded learning; no block may
report a missing install or an unresolved gstack helper path.

The file joins CODEX_CI_FILES so the planner gives it to the host-run Codex
job. paid-test-durations.json carries a 180 s packing estimate (refreshed by
--write-durations after the first census) so the plan keeps every slice
within max-parallel. Oracles live in test/helpers/codex-multiblock.ts with
free replays in test/codex-multiblock.test.ts. Registration (E2E_TOUCHFILES,
E2E_TIERS periodic, E2E_KINDS rule) is the integrator's.

* E5: design live-model check as a periodic case

test/skill-e2e-design-model-smoke.test.ts runs design/scripts/live-model-check.ts
(default models only) under a temp HOME. Without OPENAI_API_KEY the case is a
Bun skip that the census lists as SKIPPED with no coverage credit; with a key,
exit 0 passes and any other exit fails with the script output (exit 2 under a
present key fails as 'nothing was checked'). paid-test-durations.json carries
a 90 s packing estimate. Registration (E2E_TOUCHFILES, E2E_TIERS periodic,
E2E_KINDS rule) is the integrator's.

* A1: docsync parent points children at a saved interface file

The /ship docsync fixture saved its observation interface only in the
parent prompt and told the parent to copy it into every child prompt,
about 30-40 s of parent generation per session (lane S, 37198445662).
docsSessionOptions now writes it to <home>/fixture-interface.md (0600) and
the parent tells each child to Read that file first. Fault-actor prompts,
which prepare already writes to a file, are unchanged. Fixture-only.

* A3: a JUnit case's recorded wall is capped by its shard's wall

junitCensus summed a case's JUnit test times as duration_ms; for
plan-mode-no-op's concurrent PTY tests that is 346 s inside a 176 s shard
(the "432-613 s" history). The case wall headroom shows is now
min(sum, shard elapsed), a true upper bound. Session headroom already
used the ledger's per-session clocks.

* A3: refresh paid-test durations from census 37198445662

Recorded 2026-09-29 estimates under-packed gate slice 11 (17 files
estimated 1078 s ran 1353 s; plan-review-report 76 s -> 198 s,
plan-ceo-plan-mode 37 s -> 112 s, safety-codex-consult 83 s -> 150 s).
Refreshed both tiers with --report --write-durations from the latest
census's slice walls. Old plan re-costed with observed walls: worst slice
~677 s (slice 11); refreshed plan: every slice estimated <= 539 s.

* C2: Browser Quick gets a feasible budget; QA loop names its directory and stop path

Judge rationales on the qa-only panels (2 of the last 3) cited the 30 s
Browser Quick budget: homepage + 5 targets, each a checkpoint write and a
guarded command, left ~5 s per probe. Quick is now 180 s (method text: 3
minutes); the qa-only-no-fix paid check expects the same native deadline.
PROBE_DIR points at its source line, a stopped loop goes to the final
report steps (annotations, materialize, Markdown), and /review and /ship
say that after smoke expiry only plan checks rerun in revalidation.

* C2: name the flows judges found implicit in review resolvers

Structured outside review: MISSING COVERAGE asks no fix question and is
never reported clean (was "preserve the existing user decision flow").
/review's HIGH-impact plan gate: spawned or non-interactive runs report
REQUIREMENTS MISSING and continue, writing no TODOs. Value cards: the byte
clamp applies to card lines; JSON records you write keep full values.

* C1: a free test catches a removed workflow step, which the judges missed

C1 controls showed the qa workflow judge passing /qa without Phase 10 and
the review judge passing /review without Step 3. The new test pins every
step/phase heading of qa, qa-only, review, ship, plan-ceo-review,
plan-eng-review and setup-deploy (derived from the templates' step
headings) and names any that goes missing or out of order; new step
headings must be pinned. The judge rubric is unchanged (EVAL_POLICY).

* size: fit the wave's wording under ship, review and qa parity caps

Measured after all lane merges: ship -79, review -210, qa -68 bytes over.
Trimmed the added wording first (post-expiry rule, MISSING COVERAGE,
plan-gate default, value-card JSON, stop line) and folded /review's
redundant INFORMATIONAL sentence and findings bullet. ship 268,036 (room
8), review 130,118 (room 1). qa needed 11 more bytes: cap 1.11 -> 1.1102
with the measured 93,305 / 84,049 = 1.11013 and its reason beside it.

* docs: troubleshooting sections for the wave's new messages

Bun floor (setup, auto-update, /gstack-upgrade), /review stale base and
fingerprint temp-dir errors, the design image model override, the
unrecorded Chromium PID, and a contributor section for census report
lines (pass-rates flags, failure causes, headroom, cost unknown, overlay
v4, stale corpora, auq substance). A free test checks every
troubleshooting.md#anchor printed by setup or bin scripts exists.

* TODOS: reliability follow-ups deferrals; close the Bun floor, make-pdf $P and #2709 reap

* 0.4: 'A census went red' guide; docs for causes, session ledger, headroom and new messages

docs/evals/census-red.md walks one red line to a diagnosis: prerequisites, the
report fields, eval:pass-rates --run/--reds/--headroom with expected output,
what unknown/insufficient/unrecorded data means, free replay, and the single
'after a repair' paid run allowed by TESTING_INTERNALS. Linked from
CONTRIBUTING, TESTING_INTERNALS and both census reports.

TESTING_INTERNALS documents the failure_cause glossary (one row per
FAILURE_CAUSES value, diagnostic only under EVAL_POLICY v1), failure_detail,
sessions[], cost_known, the session ledger, the pass-rates views and the 85%
headroom alarm, main-only tracking issues, overlay contract v4, auq-matrix's
3-sample substance panel and the E5/E6 periodic cases. troubleshooting.md
gains #bun-too-old and an anchor per new user-facing message from lanes P, D
and F. README names GSTACK_DESIGN_IMAGE_MODEL. test/census-red-docs.test.ts
pins the glossary to FAILURE_CAUSES, parses every guide command and checks the
links.

* v1.91.20.0: CHANGELOG and version for the reliability follow-ups

* regen: SKILL.md and sections for all hosts, digest, context-budget fixture

bun run gen:skill-docs --host all; goldens and llms.txt re-rendered
unchanged; bun test/helpers/capture-context-budget.ts.

* integrate: register codex-multiblock-live and design-model-smoke (periodic, rule)

* 0.4: census report and eval:pass-rates views link the census-red guide

* test: cookie approval fixture reverses the Bun install-hint pin (E1)

The historical approved cookie-workflow prompt carried BUN_VERSION=1.3.10;
the regenerated install hint now pins 1.4.0, so the fixture's reversal
list gains that edit like the earlier waves' entries.

* test: release floor counts the two new periodic cases (E5, E6)

* fix: /freeze, /guard, /unfreeze and /investigate work on Codex and other env-var hosts

Templates quoted "$HOME/.claude/skills/gstack/...", which the env-var hosts'
rewrites turned into "$HOME/.agents/skills/gstack/..." on Codex (a path a
global install never has; same class on the other hosts), and setup's
runtime roots had no freeze/bin. Quoted runtime paths now use
{{RUNTIME_ROOT}}: $GSTACK_ROOT on env-var hosts (the fence gets the
runtime prelude), the unchanged $HOME path on Claude. _link_runtime_dists
links freeze/bin and the careful/bin helper it sources into every
env-var host's root. Also fixes the same guess in /benchmark-models,
/retro and /diagram. test/env-host-runtime-assets.test.ts fails on the
previous code: no fence guesses a $HOME/<dir>/skills/gstack path, every
$GSTACK_ROOT path a fence uses is in the host's runtime root, and the
installed Codex /freeze and /unfreeze fences run from a fresh shell.
Template-reading tests now read the Claude render or expect the
placeholder (deliberate change).

* CHANGELOG/TODOS: lane W's workflow, cases and guide; the env-var host freeze fix

* E6: a block that reads a SKILL.md is not a missing install

The first live codex-multiblock-live run passed its real check (the /learn
block found the install and printed the seeded learning) but the oracle
flagged a sed of the learn SKILL.md, whose prelude source quotes the
'gstack: no install found (tried $_r)' text. Only an emitted line that
starts with the message and names a real root counts now; free replay
negative added.

* test: qa-only fixture keeps its caller-set 30 s deadline

The qa-only-no-fix fixture tells the model to start the guard with 30 s
(a caller limit), so its deadline assertion's default stays 30 000 ms; the
product's Browser Quick default (180 s) does not apply to it. Reverts the
helper and its test to their base.

* C2: keep PROBE_DIR 'owned' (qa-functional-prompt pins it); qa measured 93,311

* test: paid file count and periodic floor include the two new periodic cases

* test: ship docs freshness pin checks meaning after A1's wording trim

The snapshot still hashes selected release paths (--select narrows them),
generated outputs and docs; the pin now checks that instead of the
pre-trim sentence.

* E3: reap the recorded Chromium only after the daemon has exited

With the PID now recorded (E3), browse stop reaped Chromium while the
daemon was still shutting down; the daemon read that as a crash, exited 1
and left its state file (chromium-profile-isolation's headless stop test
failed 3/3 locally, passes on base). The CLI waits up to 5 s for the daemon
to exit before reaping a survivor.

* regen: qa and qa-only exploratory sections (PROBE_DIR wording)

* B1: batching counter reads a target declared mid-line; replay corpus

GSTA-23's census 37228573062 failed plan-eng-multi-finding-batching with
reviewCount=0 after 975 s although D1-D10 were asked one per finding,
each with its saved decision-ledger record. The report wrote 'Report
destination requested by the user. Review target (fixed): `PLAN.md` ...'
and the counter read target fields only at a line start, so no brief
inherited the plan identity. A target field may now open any sentence of
its line; a mid-sentence mention or another file still gives no identity.
Floor (3), budget and kind unchanged. Corpus: 9 earlier passing censuses,
the red (now credited at D5), and a derived batching refusal; 975 s noted
in the headroom diagnosis.

* B1: shared-libs corpus labels its git path and covers the supported shape

Every census entry was recorded in CI's eval image (git 2.43), where
gstack-safe-git falls back because --no-lazy-fetch is rejected; each entry
now says so (git_path: fallback). Two synthetic supported-path entries
(HEAD and a history citation in the brief, as git >= 2.44 produces) keep
the 37193478719 pass and the 37195203538 refusal, and a test asserts both
shapes are covered and the actor reads no git-path-specific text.

* v1.91.21.0: move off v1.91.20.0 (claimed by #3031)

* test: settle nav-guard commands before asserting; make two new tests Windows-safe

navigation-guard: Bun's expect(promise).rejects slows each Playwright round-trip
to ~1s while it waits (50 page.evaluate: 43ms plain vs 46.6s inside the matcher),
so the click's 5s timeout expired before the guard saw the link-local request
(CI: 'click: Timeout 5000ms exceeded ... waiting for scheduled navigations').
Settle the command first, then assert on the refusal; a source check keeps it so.

env-host-runtime-assets: PROBE_FILES now matches either path separator.
detector-corpus ship-docsync-late-result: JSON-escape the <ROOT> substitution so a
Windows temp path's backslashes do not break JSON.parse.

* fix: CI run/job ids are not phone numbers; context-recovery fixture gets the full runtime bin

pii.phone.e164 redacted the run id in the published eval report's fetch command
(--run 37235771700), breaking the command it prints. Digit-only spans in a run/job
flag, gh run argument, actions/runs/ or /job/ URL segment, or run_id/job_id key are
now exempt; the same digits elsewhere still report.

context-recovery-artifacts: the fixture's ./bin/ lacked gstack-skill-start and
gstack-context-recovery (the preamble's runtime scripts since v1.71), so the agent
spent its 10 turns hunting for them and hit error_max_turns with no final message
(run 37235771700, 11 turns). Link the whole runtime bin; diagnostic rerun passed
in 6 turns.

* test: replay run 37237194905's late-result capture through the docs authority detector

The only failure was call 20's write to a mistyped shard root (3GanyY for 3GanyU),
which the model disclosed; the detector is correct and the A1 post-snapshot
--out under the fixture home stays permitted. Named red recorded in TODOS.

* test: keep docsync-authority in the Windows profile

The replay test's literal helper path tripped the Windows curator's shebang-spawn
pattern, dropping the whole file from the Windows profile (test-free-shards
invariant). The snapshot --out case is already covered by docsync-command-grammar,
so it is removed here, and the Linux-path replay skips on win32.

* fix: gstack-docs-candidate --help works; docs gate red lines name the denying tool failure

PR run 37239016858 ship-docsync-completion: the parent ran
'gstack-docs-candidate --help' (the helper's usage error points at --help, which
failed). That undeclared command made docsToolFailures non-empty, which denied
the atomic-replacement proof of the child's legitimate native Edit, so the red
line blamed the Edit's temp file. A replay of the capture proves the atomic
replace once --help is declared.

- the helper prints its usage for --help/-h and exits 0
- the fixture interface and command grammar admit the helper's bare --help
- the atomic-proof denial names its check and the tool failures behind it
  (deny labels renumbered to their lines)
- replay with negative controls (undeclared command, temp left behind, renamed
  elsewhere); diagnostic rerun of ship-docsync-completion passed

* test: retro-base-branch fixture uses main and a shallow report path

PR run 37241493377: the agent wrote retro-output.md one directory above the
requested <tmp>/retro-base/retro-output.md (it dropped the retro-base segment),
so the report check found nothing. The fixture also ran plain git init, which
makes master on CI's git while the prompt says the repo is on main; the agent
logged that mismatch as a learning. Init on main and use the temp dir itself as
the repo. Diagnostic rerun passed (18 turns, 203s).

* test: qa-quick sets a caller probe limit and drains past its session budget

C2 raised Browser Quick's native budget to 180 s. qa-quick is the only gate case
on the native budget, and its wall went 194 s (main periodic, pre-C2) -> 254 s and
274 s (this PR's runs) -> 300 s session timeout (run 37242691913): the agent now
probes every navigation target with a checkpoint, then needs ~80 s to report.
Like qa-only-no-fix, the eval now sets a caller probe limit (30 s), scopes the
fixture, and gives bun's outer timeout the session drain margin instead of racing
the 300 s session. The product's 180 s budget and the session budget are unchanged.
Diagnostic rerun passed in 195 s.

* review/ship: use main's post-expiry smoke sentence; v1.91.24.0

GSTA-23's measurement on the review workflow judge: main's 'Post-expiry smoke
rechecks are not-run.' passed 10/10; the C2 wording passed 6/10 and the
'..., even in 4c.' variant failed in PR run 37248650879. qa-probe-gates pins
main's sentence before step 4c.

* test: record the plan-design corpus test in the test-of-test baseline

#3031's ratchet now counts tests that slice a paid test file's source; the B1
plan-design-review-plan-mode replay corpus reads test/skill-e2e-design.test.ts
like its sibling corpora already listed.

* fix: --case no longer deadlocks importing the running CLI

runCaseDiagnosis loaded runPaidShards with a dynamic import of
scripts/test-paid-shards.ts; when that file is the entry it is still in its
top-level await, so the import never resolved and `--case` hung after its
first line (on main since #3031, found by the per-case measurement runs). The
CLI now passes its runner in; library callers keep the lazy import.

* test: name the in-flight file and the native error when bun dies before its summary (#3037)

* Bun 1.4.2 everywhere CI and setup pin the tested version

Bun 1.4.0 on Windows aborts about 3% of free-suite shards with native
bad-handle errors (main too; #3037); the same head on 1.4.2 had none in 180
shard runs. Every current 1.4.0 pin moves together, as bun-version-drift
requires: setup-bun in every workflow (CSO orchestration included), the eval
image (sha256 from bun-v1.4.2's SHASUMS256.txt, re-hashed locally), GitLab,
the Ubicloud free-suite setup, gstack-bun-version.sh's tested version,
engines.bun and @types/bun, the install hint in generated skills, the Dia and
cookie native-qualification pins, and docs and tests that state the current
pin. Recorded 1.4.0 observations, fixtures, CHANGELOG history and the CSO
images' oven/bun:1.3.10 base stay.

* fix: --case runs a single-case file whole

A session option naming the case (testName: 'office-hours-section-loading')
made caseSelection pick test-name mode, whose filter matched no Bun test, so
every trial would run hollow. A file that registers only this case now runs
whole before any name match is tried.

* fix(office-hours): reviewer returns a hash-bound receipt instead of echoing its verdict

The spec reviewer now Writes its verdict once, seals it with
gstack-office-hours-review seal (validates the saved JSON and prints
OFFICE_HOURS_VERDICT round=N sha256=<hash> path=<verdict>), and returns
only that line. check --receipt recomputes the hash and fails closed on a
missing, malformed, stale-round, wrong-path or mismatched receipt. The
saved file stays the sole findings inventory; validators bind the receipt
to the saved bytes and the judge reads those bytes.

* fix(office-hours): only blocking findings send a design back for another review round

Review findings carry severity blocking|minor (verdict schema v2). PASS when
no blocking findings remain; minor findings are recorded in Reviewer
Concerns and the Spec Review table, never a reason for another round.
CONVERGENCE needs a persisting blocking prior. A persisting or unverified
blocking prior cannot be relabeled minor. The 3-round cap is unchanged.

* fix(office-hours): rounds 2 and 3 are delta re-reviews of the captured design diff

prepare saves each reviewed design as round-N.design.md and gives later
reviewers the exact line diff since the previous round. A new later-round
blocking finding must cite (changed_text) an excerpt of a changed line;
seal, check and the validators verify it against the snapshots and fail
closed when a citation or snapshot is missing. Other new observations are
minor. Severity, anti-downgrade and CONVERGENCE rules are unchanged.

* test(office-hours): preservation judge accepts COMPLETED with recorded minor findings

* fix(office-hours): parent fixes only blocking findings; callers may set --max-rounds

Between rounds the parent edits for blocking findings only; minor findings
stay recorded. gstack-office-hours-review takes --max-rounds 1|2|3 (default
3) on every command; reaching it stops at MAX_ITERATIONS. The marathon
fixture sets 2 rounds so round 1 (full) and round 2 (delta) run live.

* docs(changelog): office-hours review loop changes in v1.91.28.0

* fix(cso): helper JSON keeps run URLs only in the validated workflow slot

The shared redactor now treats GitHub Actions run ids as CI metadata
(c245d1bc2), so main's #3041 invariant that the same run digits stay
redacted outside qualification.workflow held only through the phone
false positive. Make it explicit in the CSO sanitizer.

* test(plan-ceo): split-overflow progress detector accepts the skill's candidate wording

Collection never ended when the model wrote 'E1 Slack:' / 'E1-SLACK:' or
labels like 'Include in this quarter' / 'Defer with flip trigger', so the
session ran the whole review (930-1464 s) and failed on its extra findings
or the case deadline (6 of 21 runs). The candidate check now accepts a
platform-bound E-id and the action is the bucket's leading verb, as the
skill names them (Include/Defer/Cut/Hold). The semantic judge, call floor
and include/defer/cut validation are unchanged. Replays the captured CI
37374526437, t2, t9 and t15 menus with negative controls.

* test(plan-ceo): split-overflow candidate identity is the one E1-E5 token

The skill fixes no header or question wording, and shape-matching kept
missing new forms (header 'D1.1 Slack' with 'SCOPE-E1:', 'D1.1 E1 Slack',
'E1 Slack DM bot ...:', a misspelled 'Mattermst'). Identity is now exactly
one E1-E5 token across the header and the question's first line, with no
other candidate's platform named. Options rules, judge, call floor and
include/defer/cut validation are unchanged. Replays all 31 measured runs;
the quoted-row controls now quote a second ledger id.

* v1.91.30.0: move past #3044's claimed 1.91.29.0

* test: register the split-overflow wording replay as a free fixture

* fix(skills): resolve judge-cited ambiguities in /qa and /plan-eng-review

- tasks JSONL keeps the raw branch as GIT_BRANCH instead of overwriting
  gstack-slug's sanitized BRANCH used for artifact names
- the Codex outside-voice block runs for ready or unverified, matching
  its routing rule
- plan-eng startup names the preamble's own setup prompts; the
  cross-project question waits for Step 0
- /qa 8c: a fix that applies but leaves the finding reproducing gets one
  revised fix, else it is reverted; browser verification names its before
  screenshot; the exploratory loop drops its duplicate 3-step summary

* fix(plan-eng): keep the startup steps in their pinned order

* fix(plan-eng): explicit spawned sessions follow the spawned rule at the scope gate; one announcement rule

A missing tool still sends the prose menu whatever the session type, but a
session the caller or host explicitly marks spawned selects the menu's
Recommendation without prose or BLOCKED. The pre-load announcement
sentence now states Step 0's single check.

* eval policy v3: judge panels gate on the per-dimension median of 3 samples

A judge dimension now passes when at least 2 of its 3 samples meet the
unchanged threshold (judgePanelMedian); any other sample count or a
non-numeric sample fails closed. Thresholds, sample count, judge model and
prompts are unchanged; panel logs show every sample and the mean, and name
the median as the gate. Applies to the skill-llm-eval panels, the
workflow-judge cache, the auq-matrix substance panel and the calibration
replicas. EVAL_POLICY 2 -> 3 and HARNESS_VERSION 1 -> 2 (verdicts change),
so every case starts a new pass-rate series. Replays the captured panels
from jobs 112032744416 and 112049856586 as replayed evidence.

2026-10-06 18:07:50 +00:00

26 KiB

Raw Permalink Blame History

gstack reliability follow-ups (October 4, 2026): headroom, honest eval verdicts, last host gaps

Implementation plan

Context

Base: origin/main 2db0b3adc (v1.91.19.0, #3025). The severe fix wave left follow-ups in TODOS.md. Its eight periodic censuses on the wave branch (37174266054 … 37198445662) are the evidence base for this plan:

0 of 8 censuses were green; 21 verdict reds in total (2-3 per census).
By cause: 10 detectors, scripted actors or the CEO mode handoff (each repaired one at a time); 5 judge low samples; 3 provider-shaped events (one stream stall, two refusals); 2 fixture/product; 1 headroom.
27 of 28 failed trials are labelled assertion, including observer timeouts, so the reports themselves misdescribe what happened.
At the measured rate (about 1.2% red per verdict across ~193 verdicts) an all-green census is about 7% likely. The lever is the per-verdict red rate, and the biggest single source is brittle detectors and scripted actors, not provider events.
Slowest cases: plan-ceo-mode-routing up to 687 s and plan-mode-no-op 613 s; ship-docsync-completion at 88% and store at 79% of budget; census wall time 14-15 minutes against the ~10-minute target. review-army-perf (102-140 s of 300 s) and plan-design-review-plan-mode (142-240 s, no timeouts) are not headroom problems.
In census 37198445662, 20 of the 21 workflow judges that score actionability sat at exactly 4.00 and none scored 5: text edits can remove low samples but cannot lift a mean that the rubric caps at 4.

Release promises

P0. Census reds become rarer and honest. The PR reports, per census, verdict reds by machine class and the expected all-green probability, for the 8 pre-PR censuses and every census on this PR. Every red class that recurred has a structural fix (replay corpus, structured signal or speedup), not a one-off spelling patch.
P1. The slowest cases get headroom. Cases the headroom report flags above 75% of their armed budget are brought under it by doing less work; no budget is raised. The weekly eval:pass-rates --gate alarm (not a free check against a committed file, which could never fail on a code change) flags any case above 85%. The census wall target is 10 minutes, so the slowest gate slice must fall to about 450 s (today 711 s).
P2. Labels say what happened. Every trial record keeps failure_class (assertion | contract | timeout | infra, which verdicts read) and its 300-character error field exactly as today, and gains optional, bounded fields: failure_cause (contract, pre_turn_infra, api_error, refusal, provider_stall, session_timeout, observer_timeout, assertion, unknown), failure_cause_evidence, and failure_detail (assertion {expected, received} or judge {dimension, mean, samples, rationale}, each part capped). A machine cannot tell a detector miss from a product miss, so there is no detector cause; that is a diagnosis written by a person. Verdict outputs (panelVerdict, collector-outcomes.json, lane exit codes, INFRA re-dispatch eligibility) are identical under EVAL_POLICY v1, proven by running main's and the branch's verdict code over the 8 baseline censuses' artifacts and by running a frozen copy of the v1 reader over new records. This PR is judged under v1.
P2b. A red points to its evidence and the next step. Each red line in the census report and tracking issue shows the case and trial, the observed failure with its Expected/Received values (or, for a judge, the dimension, mean, samples and lowest-scoring rationale), the artifact that holds the transcript, and the next action: inspect and replay free first; a paid run only "after a repair", selecting just that case. Target: diagnosis from the report line alone in 5 minutes per red (today 15-20). Every model- or fixture-written string that reaches a public issue, PR comment or committed file passes one sanitizer: control characters stripped, mentions neutralized, a code fence longer than any backtick run, HTML escaped, length capped, and the repo's credential-scan patterns redacted (adversarial test: @mention, triple backtick, </details>, image link, ghp_/sk- token, 10 KB string).
P3. Every low judge sample has a cause. Each low workflow-judge sample in recent censuses is traced to a skill-text fix or recorded as a density-only complaint; judges are first checked on deliberately bad control bundles to learn whether they can discriminate at all.
P4. The wave's small product gaps close (Bun floor, make-pdf $P, the #2709 reap, /review's fingerprint under Codex's read-only .git, design model smoke, live Codex periodic case).

Workstream 0 (first, free): evidence

0.1 Session ledger, headroom report and red ledger. Runners (SDK, PTY, Codex) append one line per armed session to a small session ledger file in the trial's artifact dir (same pattern as contract-violations.jsonl): stable session key, the exact timeout value it armed, monotonic elapsed, end reason, and a compact liveness summary (longest in-request silence, last event kind, tool open at end). The record builders (including junitCensus for PTY and JUnit-only cases) attach it to trial records. A new module (not eval-flake-rank.ts) reports per case: elapsed vs armed budget per session (a case's headroom is its slowest session; timed-out samples are marked censored), the lane critical path from runner slice timings, and verdict reds by machine cause with per-case rates. Past censuses carry no session clocks, so history shows the case wall as a labelled upper bound and session headroom starts with the first census after this lands; below a minimum sample count a case is insufficient, never a pass. Reached as bun run eval:pass-rates --headroom | --reds | --run <id> with offline --help, unknown-flag and unknown-case rejection, --case applied to JSON and alarms, and the printed run/branch scope; --run downloads only the slices its red lines name. The all-green probability is Π(1-p_i) over per-case shrunk rates with an interval, labelled an approximation (independence is not proven).
0.2 Failure cause classifier (policy-neutral). One pure, dependency-free failureCauseOf() beside failureClassOf() (not in e2e-helpers.ts, which imports bun:test and selection machinery), called by both record builders from runner observations and the session ledger; precedence contract > pre_turn_infra > api_error > refusal > provider_stall > session_timeout > observer_timeout > assertion > unknown; it never changes failure_class, and conflicting evidence yields the higher-precedence cause with both recorded: a PTY idle turn showing an API Error: panel records api_error with the panel line (the PTY runner's observe() already stops on it; other runners match); refusals are detected from the structured stop reason first, phrases last; a stream stall is recorded only when the session streams partial messages and no event of any kind, thinking deltas included, arrives for 120 s (pre-registered; the longest in-message silence across 50 captured messages on Claude Code 2.1.284 was 9.6 s) while a model request is in flight and no tool call, permission prompt, hook or subagent is outstanding. Sessions without partial messages are never classified as stalls. Partial messages are enabled for this, but only the compact liveness summary is kept (full diagnostics would roughly quadruple transcript size). panelVerdict and lane verdicts are unchanged. Replays: 37195203538 (refusals), 37174266054 and 37176837432 (timeouts), and 37186854666 (it ran without partial messages, so it must come out session_timeout); the provider_stall positive is a synthetic fixture with an injected silence; controls: long-thinking, hung-tool, permission-pending, hook-running and subagent-running transcripts are not provider_stall; a tool result quoting refusal text is not refusal; one fixture per runner (SDK, PTY, Codex, JUnit).
0.3 Report and tracking issue. Expected/Received values (today sanitizeTrialError keeps only the first line, so 16 of 28 wave failures show nothing useful) and judge scores and rationale go into failure_detail, not error; every red names its artifact; rerun: becomes "after a repair:" and selects only that case; the weekly issue includes the full pass-rate alarm lines (today its grep keeps only the header), the red ledger and the all-green probability (a documented formula with a command anyone can rerun); branch-dispatched censuses no longer post to or close main's tracking issue (both steps gated on refs/heads/main, same in evals-marathon.yml), and the periodic workflow's concurrency group is scoped by branch so a branch census no longer cancels main's; branch and run identity are passed explicitly instead of git rev-parse --abbrev-ref HEAD.
0.4 "A census went red" guide. A short page linked from the report, the issue and CONTRIBUTING: prerequisites, the exact inspect and replay commands, expected output, what unknown data means, and when a paid run is allowed (only after a repair, per TESTING_INTERNALS). Durable evidence (the baseline red ledger, judge diagnoses) is committed under docs/evals/; scratch stays in .context/.

Workstream A: headroom on the cases the report flags (P1)

A1. ship-docsync completion (88%) and store (79%). Diagnose from the stored transcripts first. If the parent's reading of release-file contents dominates, the parent stops reading them and recomputes hashes with bin/gstack-docs-candidate (cheap, no model reading), comparing them with the child's claimed hashes; a mismatch stays a hard failure. Any change to the child's result contract updates the parser, child, actor and parent together, defines the hash algorithm and file set, and keeps negatives for stale results, missing files and unreported changes.
A2. plan-mode-no-op (432-613 s; likely the 711 s critical-path gate slice) and plan-ceo-mode-routing (687 s): time-boxed to the stored transcripts, report a per-phase breakdown (preamble, Step 0, model latency, tool time) first, then cut work in the skill or fixture; if model latency dominates, it becomes a named item rather than a cut that weakens the workflow.
A3. Any other case the 0.1 report puts above 75%. review-army-perf and plan-design-review-plan-mode get no work unless the report flags them.
A4. Fresh critical journey. The marathon lane runs the existing fresh /plan-eng-review case (plan-eng-review-artifact-full) on the PR head; time to the first question and to the written test plan are reported in the PR (non-blocking lane). A fresh /ship marathon case would be new harness work and goes to TODOS.
Before/after numbers come from named CI censuses, never local runs.

Workstream B: robust detectors and actors (the largest red class)

B1. Replay corpora for repeat offenders. shared-libs-plan-callers, plan-ceo-section-loading, auto-decide-preserved, plan-design-review-plan-mode and ship-docsync-late-result each get the stored passing and failing captures still downloadable (inventory first; artifacts expire after 30-90 days, and expired ones are listed as unavailable) replayed through their detector or actor in the free suite (≥1 pass and ≥1 fail each). Entries are minimal, redacted excerpts (the detector's input slice, size-capped, credential-scanned in CI), pinned to the series identity they came from, with stale entries reported for refresh. Each corpus entry records its original census outcome, the expected detector/actor result and a one-line justification, so a fixed false red is an expected pass with its history kept, and genuine negatives stay refused.
B2. Structured signals before phrases. Where a skill already writes a structured record (decision ledgers, review logs, handoff helper output, VERDICT: lines), valid current structured evidence (right session, after the action, not copied fixture input) is authoritative; phrase detectors run only when no structured evidence exists. Where the contract is user-visible output (a question shown, text rendered), the behavioral witness stays required and the record only locates it. Audit every periodic and gate detector for phrase-only checks and convert the ones with a structured source.
B3. Floor runner TTY preview (TODOS P2): accept collapsed context rows and gutterless wraps in currentEditPreview, still compared against exact source lines; replay the screen from PR run 37176835584.
B4. --case selects loop-registered cases (coverage audits, TPA, plan-mode-no-op) and an unknown id exits non-zero before any process starts.
B5. Pre-push hook hang behind shard neighbors: reproduce with the failing plan under xvfb and diagnose first; fix if the cause is in scope, otherwise a named open item.

Workstream C: judges (P3)

C1. Calibration controls first, only for judges that produced low samples (qa, qa-only, review, plan-ceo modes, plan-eng sections, setup-deploy as applicable). Score compression at 4 is a hypothesis, not a fact (the schema allows 5). Pre-register in docs/evals/ the positive, negative and targeted-defect control bundles, the expected outcome for each, and one 3-sample panel per judge, reusing scripts/judge-calibration-configs.ts. If a judge cannot separate the negative controls, list its cases in the PR as giving no coverage and bring a rubric proposal to Garry as a pre-registered EVAL_POLICY change.
C2. Trace low samples. Collect every low workflow-judge sample's rationale in the last N censuses into .context/; fix skill text only where a rationale names a concrete, execution-relevant gap (and tie it to an observable failure where one exists); density-only complaints are recorded, not chased. Regenerate, run parity, prompt-size and removed-clause checks, then one 3-sample panel per changed judge.
C3. Behavior cases in drift. plan-review-prosons-neutral-neg (flagged ACTION REQUIRED by eval:pass-rates) and shared-libs-opportunity-judgment: diagnose from stored transcripts and fix the cause, or propose CASE_QUARANTINE entries only if the diagnosis is detector/harness/model-latency and the entry rule is met.
C4. auq-matrix. Its recommendation-substance check gates a rule case on one judge sample and was red in 6 of the last 24 censuses. Gate question below: keep capture and judging in one execution and score that one capture with a 3-sample panel (mean, unchanged threshold), recording the evidence in E2E_KINDS.

Workstream D: provider events and the overlay control arm

D1. Overlay comparison control arm (gate question below). Today fixture.verify()/measurementsValid turns a control-arm wrong answer into an invalid measurement that fails the comparison. Proposed: a completed wrong answer in either arm stays a measured correctness result in the comparison report, and only predefined execution or measurement failures invalidate a trial, applied symmetrically; the overlay contract moves to v4. Because this changes what a gate scores after the 37198445662 red was seen, it needs Garry's approval and applies only to censuses after it.
D2. plan-ceo-review-format-approach refusals. Check whether its capture prompt ("verbatim … exact prose you would have shown") triggers the refusal classifier; reword the fixture if so.

Workstream E: small product gaps (P4)

E1. Bun versions from one source. Supported minimum stays 1.4.0 (engines.bun, the tested CI version); 1.3.3 is the security floor (below it D0's no-autoload flags are silently ignored, verified by a compiled probe, not a version comparison). Extend test/bun-version-drift.test.ts so setup, engines, the install hint (now 1.4.0 instead of 1.3.10), README and CI agree. Setup refuses below 1.3.3 and warns between 1.3.3 and 1.4.0 ("untested, upgrade with bun upgrade"), with a troubleshooting anchor; prerelease strings are compared by a defined rule; --help and --status keep working without Bun. The auto-updater fetches, reads the incoming release's requirement and checks the bun that setup will use before advancing the live checkout, leaving the live revision and installs untouched on rejection with a stable bun-too-old reason and resume command. The first upgrade run by an older updater cannot be protected; the CHANGELOG says so.
E2. make-pdf $P joins $B/$D in the shared runtime prelude for env-var hosts through one make-pdf resolver used for both readiness and every block, preserving the MAKE_PDF_BIN override and the real make-pdf/dist/pdf name (paths with spaces tested).
E3. #2709 reap on Playwright 1.62: record the Chromium PID via CDP SystemInfo.getProcessInfo (or the launch server) only for launches gstack owns; the reaper keeps checking recorded start time and command identity before signaling; tests cover the pinned launcher, bounded CDP failure, a recycled PID and an externally owned browser that is never reaped.
E4. /review fingerprint under Codex's sandbox: bin/gstack-wtree writes new objects into a private object directory (mktemp -d, 0700, under $TMPDIR, absolute common-dir objects as alternates resolved before exporting, linked worktrees supported) whose lifetime is owned by the review-log operation, because sharedLibsSnapshotCoverage() later reads the fingerprint's tree and blobs; it is removed (EXIT/INT/TERM) only after that consumer finishes, and a failed write exits 1 instead of printing a partial hash; a failed git fetch reports stale-base coverage naming the base revision used; gstack-review-log keeps the fingerprint helper's one-line stderr diagnostic (stdout stays hash-only), and /review Step 1.3's git fetch && chain no longer stops at "Nothing to review" under a read-only .git; tested through --start → --finish → shared-libs reuse under a read-only .git, with dirty tracked files, untracked files and a linked worktree.
E5. Design model smoke (periodic). The defaults already pin gpt-image-2 (design/src/models.ts:40); register design/scripts/live-model-check.ts as a periodic case so a retirement or entitlement change shows up before users hit it. A missing key (exit 2) is a visible skip with no coverage credit. The image tool model gains an env override (GSTACK_DESIGN_IMAGE_MODEL, validated against a conservative pattern, named in errors) so users can switch without editing code; the periodic check always tests the default, so an override cannot mask a retirement.
E6. Live Codex multi-block check as a periodic case in the host-run Codex job, reusing the job's checkout and running setup once so it does not become the new critical path.

Gate questions for Garry (decided before Workstream 0 lands)

EVAL_POLICY v1 lets an all-INFRA census be re-dispatched once, and counts only errors before the first model turn as INFRA. Should provider_stall (with the streaming evidence in 0.2) join INFRA under a v2 for future censuses? This PR is judged under v1 either way. Refusals stay a failure class of their own in both options (a refusal is model output users also see). Recommendation from the CEO review: labels only now (v1 unchanged); revisit v2 with the pooled provider-event base rate from 0.1 attached.

Also for approval: D1 (overlay control-arm scoring, applies to later censuses only) and C4 (auq-matrix scores its single capture with a 3-sample judge panel).

Delivery

One PR, a commit per item. Lanes (each commits sources only; one integrator owns touchfiles-data.ts, the ratchet file, generated SKILL.md, VERSION, CHANGELOG): F (session ledger, classifier, report, issue/workflow), P (A1, A2 headroom fixes), S (B1-B5 detectors and harness), D (C1-C4 judges, D1-D2), W (E1-E6 product); F, P, S and D start together, W's E1 waits for F's workflow edits. Freeze harness edits before the first paid PR-lane push, because every push re-runs the full gate. Patch version from the queue.

Validation

Each fix has a free regression test that fails on main; detector and classifier changes replay captured transcripts with negative controls.
Paid runs are pre-registered at the gate as an inventory: one diagnostic run per concrete repair of a rule case (never on unchanged inputs), a 3-trial panel for a changed behavior case, one 3-sample panel per changed judge and per C1 calibration set, the new E5/E6 cases once each to prove wiring, then the PR lane once on the frozen head (full gate, ~106 cases and 24 judges, because engines.bun and new files force full fallback), one periodic census on the frozen head, and the marathon run. Diagnostic runs never form a verdict.
Free suite once on Ubicloud at the end; PR gate lane; one periodic census on the frozen head.
Each workstream names the docs it updates (TESTING_INTERNALS, troubleshooting, the new red-census guide, CONTRIBUTING, AGENTS.md, README, TODOS.md). The 75%/85% thresholds and the stall window are named constants in one module; per-case exemptions go through a list with reason and expiry, never a raised budget.
Done when PR CI is green and the census on the frozen head has every red either machine-classified INFRA under the approved policy or listed as a named red with diagnosis in the PR and TODOS.md. A repair made after that census needs a fresh census on the new head (TESTING_INTERNALS:407-411). "Diagnosis complete" is not "release accepted": under v1, an INFRA-only census may be re-dispatched once; any other red left unresolved needs Garry's explicit exception before merge. The PR body carries the red-class table, the all-green probability before/after, and a "diagnosable from the report line alone" column.

Not in scope

Comparing gstack end to end against the stock host agent on product outcomes (the overlay harness covers model-level deltas; a full product A/B is its own project, filed to TODOS). Severe-wave items not taken here: C8 router panel, #2670 bash half, G2 TS twin, Codex telemetry/notice follow-ups, gate-outcome telemetry, design health subcommand, cmd.exe autoplan hook, /gstack-upgrade redesign, marketplace, placeholder convention, #2750, ios-qa tokens, #931, #1713, Windows SAC Node lane, navigation pre-send blocking, gbrain batch lookup, memory follow-ups, #2971 non-privacy pieces, redaction allowlist #2598.

Review record

Autoplan run started 2026-10-04 (GSTA-21). Source: TODOS.md on main 2db0b3adc.

CEO phase, 2026-10-04

Voices: Codex gpt-6-astra (completed, 8 findings, recommends staged releases), Claude native blind (hash 3163a6b1 matched, 2 critical: stall rule ambiguity, PR judged under its own new policy), Claude primary (8-census data, 15 amendments). Consensus: confirmed on all six dimensions that the draft optimized stale premises (A2/A3 not slow; real slow cases missing), that B as drafted would mislabel thinking or tool hangs and let the PR judge itself under v2, that A5's data source cannot support per-case headroom, and that judge margin cannot come from text alone. Disagreement (taste): Codex wanted separate releases; Claude voices kept one PR (owner rule). User challenge candidates were reframed as the gate's policy question (labels-only vs stall-as-INFRA) because all voices agree the change needs Garry's approval regardless. Rewritten plan above adopts AM-1..AM-15 plus native F1/F3/F4/F7/F10/F12/F13 and Codex 2/3/4/5/6/8.

DX phase, 2026-10-04

Voices: Codex gpt-6-astra (completed, 10 findings), Claude native blind (hash a661b072 matched; 2 critical: failure-class schema clash, non-deterministic free 85% check), Claude primary (DX 4.3 → 7.1; time from red census to diagnosis 15-20 min per red → target 5). Consensus confirmed on all six dimensions: the report line must carry evidence and the next action (not rerun:), the new classes must not touch the policy field or verdicts, the headroom check needs a defined clock, statistic and committed snapshot, the CLI needs help/validation, a two-minute triage guide, and E1 needs an upgrade recovery path. Taste: T-DX-1 Bun floor 1.3.3 (security) with 1.4.0 as tested version, chosen over a hard 1.4.0 refuse (Claude primary; avoids stranding git installs at SETUP_FAILED; Codex wanted one minimum, which 1.3.3 from one source satisfies); T-DX-2 separate failure_cause field (all three voices); T-DX-3 85% alarm also in eval:pass-rates --gate (accepted).

Eng phase (last), 2026-10-04

Voices: Codex gpt-6-astra (completed, 16 findings, recommends splitting), Claude native blind (hash 43124ba1 matched; high: error-field compat, unsanitized public text, headroom data gap, D1 is a post-hoc gate change), Claude primary (feasibility verified: thinking deltas present on Claude Code 2.1.284 with 9.6 s max in-message silence; past censuses carry no session clocks; full-fallback paid selection; census critical path 711 s; 16 amendments). Consensus confirmed on all six dimensions. Accepted: session ledger + pure classifier beside failureClassOf; bounded new fields with error untouched and a frozen-v1-reader test; one sanitizer for published text; no detector cause; synthetic stall positive; E4 object lifetime owned by review-log; E1 checks before advancing the checkout with 1.4.0 kept as supported minimum; D1 and C4 to the gate; main-only issue steps and branch-scoped concurrency; structured-first detectors keep behavioral witnesses for user-visible contracts; pre-registered paid inventory. Trimmed for proportion: the committed-snapshot 85% free check (could never fail on code), the fresh /ship marathon case, and the separate auq-matrix judge case; C1 runs only for judges with low samples. Disagreement (taste): Codex recommended splitting into several PRs; kept one PR per Garry's standing rule, with dependency-ordered lanes. Test plan: ~/.gstack/projects/garrytan-gstack/garrytan-reliability-followups-oct4-test-plan-20261004.md.

Final approval gate, 2026-10-04

Garry approved the plan as one PR. Decisions: (1) provider stalls are labels only (provider_stall cause, still a failure; EVAL_POLICY v1 unchanged); (2) D1 approved: a completed wrong answer in the overlay-OFF arm counts as a comparison result, not a failure, for censuses after this change (overlay contract v4); (3) C4 approved: auq-matrix scores its one capture with a 3-sample judge panel, mean against the unchanged threshold, for censuses after this change. Status: APPROVED.

26 KiB Raw Permalink Blame History