mirror of
https://github.com/garrytan/gstack.git
synced 2026-08-21 05:27:14 +02:00
410b4928e7caa229b92827fa3e8cc95a49bb757b
4
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Garry TanandClaude Fable 5
|
410b4928e7 |
v1.66.0.0 feat: test/evals/CI speedup — 90s truthful free suite, diff-billed evals, required Linux lane (#2593)
* ci: bump CI image Bun 1.3.10 -> 1.3.13
Matches the local toolchain and brings native `bun test --shard=M/N` /
--parallel to CI (needed by the free-test lane and shard runner work).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* ci: stop version bumps rebuilding the eval Docker image (cache key trio)
Three coupled fixes, atomic because any subset is worse than none:
1. Image tag keys on hashFiles(Dockerfile.ci, bun.lock) — package.json is
out: its version field changed on 60/60 recent commits, forcing a ~2min
image rebuild per PR for a dependency set only bun.lock determines.
2. ci-image.yml now pushes that same content-hash tag (previously only
:latest/:sha, so the weekly prebuild never warmed the tag the eval
matrix actually looks up) and both eval workflows get registry layer
cache (cache-to export gated to same-repo runs; fork tokens cannot
write GHCR).
3. Dockerfile bakes /opt/node_modules_cache/.bun.lock and the runtime
Restore-deps guard diffs bun.lock instead of package.json — otherwise
every version-only bump made all 14 matrix jobs fall back to a live
bun install, which is slower than today's behavior.
Worst-case failure mode is self-healing: a missing tag or cache falls
back to exactly the previous rebuild-and-install path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* ci: stop double-running lint + skill-docs on every PR commit
Both fired on unrestricted push AND pull_request, so each PR push ran
them twice (12 duplicate (headSha, workflow) pairs in the last 200 runs).
push is now main-only; pull_request covers PR branches.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* ci: run actionlint from the prebuilt image (16s -> ~2s)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* ci: right-size five single-core jobs to ubicloud-standard-2
actionlint, skill-docs, version-gate, pr-title-sync, and the evals report
job never exceed one core; standard-8 was ~4x the cost for zero wall-clock.
build-image and the eval matrix keep standard-8.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* ci: fix workflow_dispatch concurrency collisions (head_ref || run_id)
head_ref is empty on workflow_dispatch, so every manual dispatch of these
four workflows shared one empty-suffix group and cancelled each other.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* ci(windows): cache bun installs; run the curated suite, not a hand list
- actions/cache on ~/.bun/install/cache keyed on bun.lock (install was
35-45s of both 55-64s jobs, all network) and Bun pinned to 1.3.13 to
match the other lanes.
- windows-free-tests now runs `bun run test:windows` (the runner's
--windows-only curation) instead of a hand-listed 13-file subset that
had drifted from the registry it sampled. POSIX-bound tests get
excluded in ONE place (the curation patterns), not two.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* evals: retry 1, not 2, on every paid path
Measured on the llm-judge shard: --retry 2 amplified 25 tests into 46
executions (+84%), with retried runs at 138s vs a 10-12s baseline (429
backoff), and a permanently-failing test paying 3x. One retry still
absorbs one-off flakes; chronic flakes become visible fix-work instead
of silent wall-clock.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* evals: split skill-e2e-review into three per-file CI shards
Bun runs describe blocks as concurrency barriers, so the e2e-review CI
job executed its tests serially: 741s of an 860s PR critical path for
tests whose slowest member is 224s. The per-file matrix is the repo's
parallelism unit, so the split moves:
- Retro E2E + retro-base-branch -> test/skill-e2e-retro.test.ts
- review/ship base-branch + Review Dashboard Via Attribution
-> test/skill-e2e-review-attribution.test.ts
- sql-injection / enum-completeness / design-lite stay in
test/skill-e2e-review.test.ts
One 741s job becomes three ~180-250s jobs. Locally the worst paid shard
drops from 1705s (94.7% of the 1800s kill) to under 700s. Test names,
bodies, suite strings, and eval-store collectors are unchanged, so
baselines carry over. Matrix rows added to both eval workflows
(attribution is gate-only, so no periodic row); the report job's
hardcoded runner count is gone (drift-proof).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: gate security-bench on SECURITY_BENCH=1, not model-cache existence
The existsSync gate ran ~12s of ONNX inference (plus a HuggingFace
dataset fetch) on every free-suite run on any dev box that had ever
warmed the classifier, while CI (no cache) silently skipped it. Now
explicit opt-in: SECURITY_BENCH=1 bun test browse/test/security-bench.test.ts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: watchdog E2E in 1.5s instead of 22.7s (tunable poll interval)
server.ts gains BROWSE_WATCHDOG_INTERVAL_MS (floor 50ms, default 15s
unchanged). The #994 stay-alive test runs a 250ms tick and waits for the
stay-alive log line instead of blind-sleeping 2s + 20s past the
production interval.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: dedupe coverage gates; route both walks through skill-census
skill-coverage-floor duplicated two matrix assertions (registry
completeness, gate-tier floor) with a DIFFERENT hand-rolled directory
walk — matrix's skipped nothing, floor's skipped node_modules/docs/test.
Two 'same' gates disagreeing on the census is the bug class
test/helpers/skill-census.ts was written to kill. Registry assertions
now live in matrix only (with floor's better error message), both files
walk via skillCensus().authoredSkills, and floor keeps the per-skill
structural checks it owns.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* evals: EVALS_JOBS for shard processes; explicit within-shard concurrency
EVALS_CONCURRENCY was overloaded: the legacy bun-test path used it as
--max-concurrency (default 15) while the sharded runner read it as the
process count — exporting the legacy value gave 15 concurrent Bun
processes each spawning claude (the 429 storm). Now: EVALS_JOBS = shard
processes (default 4); EVALS_CONCURRENCY = bun --max-concurrency inside
a shard (default 4, explicit in shard args — omitting it made
within-shard parallelism silently differ from the legacy path). Stale
49/59 header math replaced with the live-count rule.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* evals: enforce detach-timeout floor from the live shard census
New free tripwire: eval:bg:gate / eval:bg:periodic --timeout must cover
ceil(shards/jobs) x shard-timeout x 1.05, recomputed from the actual paid
test census every run. Hand-derived numbers go stale every time a paid
file lands — the review split just proved it: periodic's 28800s dropped
BELOW its new 32130s worst case (raised to 32400s here). An undersized
watchdog kills healthy runs and the tail reports never-started.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* evals: preflight ping once in the sharded parent, not per shard
The Anthropic fail-fast ping ran at module load in every paid test file
importing e2e-helpers — ~30 paid claude -p calls (30s timeout each) per
full sharded run for one bit of information. The parent now pings once
before spawning shards and sets EVALS_PREFLIGHT_OK=1; the module-load
path honors the flag. Extracted to test/helpers/anthropic-preflight.ts
(injectable spawn seam) with regression pins in both directions: the
flag must skip, its absence must ping exactly once, dead API must throw.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* evals: split touchfiles into pure data + selection logic + facade
touchfiles.ts listed ITSELF in GLOBAL_TOUCHFILES, so adding one test's
dep entry forced the full ~$38 / 30-45min suite — measured on 21.9% of
recent commits (42/192). The self-reference existed because data and
logic shared a file: any edit COULD be a selection-logic change.
Now: touchfiles-data.ts (the four maps, literals only, zero imports —
the future map-diff target), test-selection.ts (matchGlob/detectBase
Branch/getChangedFiles/selectTests), and touchfiles.ts as a re-export
facade so all ~12 import sites are untouched. GLOBAL_TOUCHFILES drops
the self-ref, adds test-selection.ts (logic stays maximally
conservative), and TEMPORARILY adds touchfiles-data.ts until the
map-diff change lands. New free test pins the literal-only property
(comment-aware state-machine scan with a self-test) and facade export
parity (===), so neither can silently rot.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: free runner — strict output, parallel execution, stable shard indices
Three coupled changes to scripts/test-free-shards.ts:
1. STRICT OUTPUT: runFreeShard streams through the paid runner's
BunTestOutputClassifier — exit 0 without bun's 'Ran N tests across M
files' summary, with (fail) lines, or with a wrong file count is a
FAILURE (anti-truncation backstop at the runner layer), plus an
external wall-clock timeout that SIGKILLs the process group
(timed-out distinct from failed; exit 124 vs 1). Also fixes a latent
shard-bleed: file selectors now use exactTestFileSelectors (relative
paths were substring filters that matched sibling roots).
2. PARALLEL: full-suite mode is one 'bun test --parallel' invocation
(Bun 1.3.13). Measured semantics recorded in the header: per-file
worker isolation, standard summary, and mid-suite process.exit
surfaces as a crashed-worker FAIL with exit 1 — strictly safer than
serial, where the same exit truncates silently. No static weight
lists; --shards M --shard i keeps deterministic hash partitioning for
CI matrices (native --shard rejected: round-robin renumbers when
files land). Spawned shards get throwaway GSTACK_HOME/TMPDIR so
parallel shards can't contend on real state. Per-shard epilogue
prints files/seconds/status every run.
3. Stable indices: assignFilesToShards no longer drops empty shards, so
a shard's index depends only on the file hash and requested count —
an empty CI matrix slot is a fast no-op success, not a renumbering.
package.json 'test' now delegates to the runner (TEST_ROOTS becomes the
single source of truth for roots; slop:diff tail preserved; the runner
inherits the 30s per-test timeout the old glob passed inline).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* ci: Linux free-test lane — ~400 files get CI coverage for the first time
New required, secretless free-tests job: the canonical runner's single
'bun test --parallel' invocation with strict-output classification on
ubicloud-standard-8. The free suite previously ran on NO Linux CI — only
a curated Windows subset ran anywhere — so every 'tests pass' claim
about main rested on contributors running them locally.
Secretless by design (no API keys; fork PRs finally get real test
signal) and pinned by test/free-tests-workflow-wiring.test.ts: canonical
runner invoked, zero secrets.* references, pull_request never
pull_request_target, and matrix-count/--shards agreement if anyone
switches to the sharded fallback.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* evals: map-diff selection — a touchfiles-data edit runs only what changed
Editing the eval dep-list data no longer forces the full ~$38 /
30-45min suite (measured on 21.9% of recent commits). When
touchfiles-data.ts is in the diff, selection now evaluates the BASE
version (git show -> mkdtemp -> spawnSync bun child printing the four
maps as JSON — sync because e2e-helpers selects at module scope) and
JSON-diffs per key: added entries, edited dep lists, and tier flips are
selected; keys removed from all maps are reported, never silently
dropped; a GLOBAL_TOUCHFILES edit still runs everything.
FAIL-CLOSED with named causes: missing-base-ref, git-show-failed,
import-failed, shape-mismatch each degrade to run-all and print
'selection: global — touchfiles-data changed (<cause>)' (D9 — silently
expensive beats silently wrong, but never silently). eval:select prints
'selected N of M, reason: ...' + removed tests; --base scopes the
map-diff too.
The temporary conservative GLOBAL entry for touchfiles-data.ts is gone —
its changes route through the map-diff. 23 new free tests: pure-core
fixtures, selectTests wiring incl. a poison-injection guard, and a temp
git repo exercising every fail-closed cause end-to-end.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* evals: selection sees uncommitted work; git errors fail closed
getChangedFiles is now the deduped union of committed (base...HEAD),
staged+unstaged (git diff HEAD), and untracked (git status --porcelain
--untracked-files=all) — an agent that edits files and runs evals
BEFORE committing no longer gets the full $38 suite every time because
the committed diff looked empty. Clean tree still returns [] (run-all
by design for main-branch/periodic runs).
Git failures now THROW with the failing command, stderr, and 'set
EVALS_ALL=1 to deliberately run the full suite' — the old return []
silently became run-all, which is silently expensive. 11 new free tests
cover every source, dedupe, quoted paths, and both failure shapes via
an injectable spawn seam.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: revert GSTACK_HOME injection in the free runner — shared mutable state
The first full run under the strict runner surfaced 12 failures with one
root cause: injecting a single throwaway GSTACK_HOME per invocation made
6,900 tests share a MUTABLE scratch home. gstack-config tests wrote keys
into it; relink and update-check tests then read them (e.g. relink saw
skill_prefix left behind by a config test and produced prefixed names).
All 12 pass when run directly.
TMPDIR isolation stays (mkdtemp inside it is still per-call unique).
Tests needing GSTACK_HOME isolation mkdtemp their own per test — the
repo convention — and hermetic-env covers E2E children. The env-dump pin
now asserts GSTACK_HOME passes through UNTOUCHED so the injection can't
come back.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: rebase parity baseline to v1.64.0.0; fix capture-vs-check drift
The parity ratchet had quietly failed for 7 skills — v1.58-v1.64 growth
landed past the v1.57.7.0 anchors and nothing caught it because this
test had no CI lane (verified pre-existing: SKILL.md content is
byte-identical to origin/main). Same rebase protocol as
v1.53->v1.57.7.0; old baseline retained for the audit trail.
Root-caused a second latent bug while rebasing: captureBaseline recorded
SKELETON-ONLY bytes while the checker compares UNION bytes (skeleton +
carved sections/*.md), so a fresh capture read carved skills at ~2x
ratio (ship: 82KB captured vs 183KB checked). captureBaseline now takes
sectionedSkills and records unions for carved skills — capture and check
measure the same thing, so the NEXT rebase can't hit this. Four
CARVE_GUARDS skeleton caps re-ratcheted to current +headroom
(plan-ceo 92K, plan-eng 70K, office-hours 100K, design-consultation
70K), annotated inline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: package.json version matches VERSION (1.64.0.0)
v1.64.0.0 shipped with VERSION bumped but package.json left at 1.63.0.0
— the 'package.json version matches VERSION file' test fails on
origin/main today. Nothing caught it because that test had no CI lane
until this branch's free-tests job.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: fix variants-retry-after HTTP-date flake (TODOS P2)
toUTCString() truncates to whole seconds, so a +3000ms Retry-After date
could mean an effective wait of ~2001ms — flaking against the 2500ms
assertion floor ~1-2 in 9 runs under suite load. +4000ms puts the
truncation floor at 3001ms with the assertion floor safely below it.
Pulled forward from U4 because the free-tests lane is now a required
check and this flake would randomly block PRs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: skill-fixture helper — extract SKILL.md sections, don't copy files
extractSkillSections (fence-aware H2 scanner, loud-throw on missing
sections with available-heading list), extractSkillBody (drops the
shared generated preamble), extractSkillHead (frontmatter + first 30
lines, for routing fixtures). Pinned section lists per consumer, and
free-tier real-skill pins so a gen-skill-docs heading rename fails the
FREE suite instead of a paid run. skill-fixture.ts joins
GLOBAL_TOUCHFILES (fail-safe polarity: over-select).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(evals): review E2E fixtures extract sections — 1871 -> 207 lines
CLAUDE.md's extract-don't-copy rule, applied: the three review fixtures
carry only the sections the sql-injection/enum/design-lite prompts and
judges exercise (89% cut). Full-file copies made claude -p read 1871
lines per test — the direct cause of the 1705s worst shard (94.7% of
the 1800s kill).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(evals): retro E2E fixtures extract sections — 1821 -> 757 lines
Keeps every section the retro flow exercises incl. base-branch detect;
drops preamble, Global Retrospective Mode, Compare Mode (58% cut).
retro-base-branch was the single slowest CI test at 224s.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(evals): review-army fixture extracts sections — 1871 -> 650 lines
CS1's set plus Step 1.5 (PLAN COMPLETION AUDIT machinery) and Step 4.5
(army dispatch, quality_score, findings schema) that the 7 army tests
assert on. Pin test guards the three load-bearing strings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(evals): skillify fixtures via extractSkillBody — 63-83% smaller
Tests follow all 11 skillify steps, so the whole body stays; only the
shared generated preamble drops (skillify 1239->453, scrape 958->167).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(evals): context-skills fixtures via extractSkillBody — 74-82% smaller
context-save 1037->267 lines, context-restore 952->168; the 8 tests
exercise full save/restore/list flows so the body stays, preamble drops.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(evals): opus-47 discovery fixtures via extractSkillHead — ~95% smaller
Routing/fanout tests only read frontmatter + opening lines of the 14
installed skills (review 1871->54, office-hours 1706->80).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(evals): codex runner gains sections option — review variant 88% smaller
runCodexSkill/installSkillToTempHome accept sections?: string[] routed
through extractSkillSections; codex-review-findings wired (1465->181
lines). codex-discover-skill deliberately keeps the FULL copy — its
stderr assertions validate that the real generated artifact loads.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(evals): routing fixture installs skill HEADS, not ~18 full SKILL.md
Routing reads frontmatter only; extractSkillHead per skill (root
611->48, ship 1435->54 lines). This was the single worst fixture bloat
site: one fixture dir holding ~18 full skills.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* evals: parent-side shard skipping — a one-test diff runs 3 of 44 shards
The sharded runner spawned every shard regardless of diff; only the
child self-skipped, so a typical single-skill change still paid 44 Bun
boots + container-equivalent setup for shards with zero selected tests.
The parent now computes selection once (mirroring e2e-helpers exactly:
EVALS_ALL -> run-all, empty union -> run-all, git errors propagate the
fail-closed throw) and drops shards where no selected test name maps in.
Mapping = quoted E2E map keys in the file's source UNION keys whose dep
list registers the file (constructed-name families need the second
direction). FAIL-OPEN everywhere it matters: run-all, non-skill-e2e
files, unreadable source, zero mapped names all keep the shard — the
child filter stays authoritative, so a parent bug can only run extra.
New taxonomy status skipped-by-diff (never conflated with
never-started); selection banner prints once; --list is selection-aware.
C6 lands in the same commit: a HARD tier-alignment test — every paid
skill-e2e file must be parent-mappable or provably fail-open-safe.
Note: this change-set's 14 dep-list registrations in touchfiles-data.ts
rode along in
|
||
|
Garry TanandClaude Opus 4.8
|
c7ae63201a |
v1.58.1.0 feat: hermetic local E2E + Conductor prose AskUserQuestion (#2004)
* feat: add shared call-time isConductor() helper
Single source of truth for Conductor host detection in TS consumers
(CONDUCTOR_WORKSPACE_PATH / CONDUCTOR_PORT). Reads the passed env at
call time, not a module-load snapshot, so unit tests can pin the env
inline without Bun --preload (esm-hoist-breaks-env-pin-bootstrap).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: harden question-preference-hook harness against ambient Conductor env
runHook copied all of process.env into the hook subprocess, so running the
suite inside Conductor (CONDUCTOR_WORKSPACE_PATH/PORT set) would leak those
markers. Strip them so the existing cases deterministically characterize
NON-Conductor behavior before the Conductor branch lands. Baseline: 15 pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: PreToolUse hook denies AskUserQuestion in Conductor, redirects to prose
Conductor disables native AskUserQuestion and routes through a flaky MCP
variant that returns '[Tool result missing due to internal error]'. The
hook now denies any AUQ call in a Conductor session and instructs the model
to render a prose decision brief instead (transport avoidance, not preference
enforcement) — firing for one-way doors too, with a typed-confirmation
requirement for destructive paths.
Precedence: never-ask auto-decide still wins (user already settled those);
Conductor prose is the fallback for everything else; non-Conductor behavior
is byte-for-byte unchanged. Restructured the per-question loop to compute
eligibility without early-returning so the Conductor branch can run as the
fallback while preserving memoryContext on every exit.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: Conductor renders AskUserQuestion decisions as prose by default
In Conductor, native AskUserQuestion is disabled and the MCP variant is
flaky, so skills now render every decision as a plain-text prose brief the
user answers by typing a letter — proactively, not as a failure reaction.
- Preamble emits CONDUCTOR_SESSION, gated on != headless so eval/CI inside
Conductor still BLOCKs instead of rendering prose to nobody.
- AskUserQuestion Format gains a Conductor-default-prose rule (auto-decide
preferences still apply first; prose decisions log via gstack-question-log
since PostToolUse never fires), a one-way/destructive typed-confirmation
rule, and a typed-reply continuation protocol for split chains.
- Regenerated all SKILL.md + ship golden fixtures; bumped affected carve
skeleton caps to absorb the always-loaded additions.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: deploy the Conductor AskUserQuestion hook (setup + upgrade migration)
The PreToolUse hook only delivers its Conductor-prose guarantee if it's
installed, but setup skips hook registration in non-interactive (conductor/CI)
setups. Two fixes so layer 3 actually deploys:
- setup: treat a Conductor workspace as an implicit opt-in for the PreToolUse
hook on the silent fall-through (never overriding an explicit opt-out).
- migration v1.58.0.0: re-register the hook for existing Conductor installs on
/gstack-upgrade, idempotent and respecting plan_tune_hooks=no.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: E2E for Conductor prose + fix auto-decide-preserved GSTACK_HOME bug
- New skill-e2e-conductor-prose (periodic): Conductor env + plan-eng-review
surfaces a prose decision brief, not a silent skip. Header documents this is
end-to-end behavior coverage; the deterministic Conductor guard is the
question-preference-hook unit test (the PTY harness can't register the MCP
variant — Codex #10).
- Fix the pre-existing bug in auto-decide-preserved: it seeded the never-ask
preference under GSTACK_HOME=tmpHome but never passed GSTACK_HOME into the
PTY run, so the spawned claude read the real ~/.gstack and the preference
was inert (Codex #9). Now passes GSTACK_HOME + CONDUCTOR_WORKSPACE_PATH to
prove auto-decide still wins over the Conductor prose redirect.
- Register both in touchfiles (periodic tier).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* v1.58.0.0 feat: Conductor renders AskUserQuestion decisions as prose
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: strip ambient Conductor env in memory-cache-injection hook harness
Same dev-in-Conductor leak fixed for question-preference-hook: this suite's
runHook copies process.env, so running it inside Conductor flipped the
defer-path memoryContext assertions into the [conductor] prose deny. Strip
CONDUCTOR_* so the cases characterize non-Conductor behavior. (CI is headless,
so this only bit local Conductor runs.)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: gstack-detach — run agent eval/bench jobs in their own session
Long agent-run jobs (30-60 min evals, benchmarks) die when the harness sends
SIGTERM to a background task's process group on turn boundaries / monitor
stops / interruptions (observed: 'script test:gate terminated by signal
SIGTERM'). gstack-detach runs the command in a fresh session (python3
os.setsid, or setsid on Linux, nohup fallback) so a group SIGTERM can't reach
it, and wraps it in caffeinate -i on macOS so idle-sleep can't kill it either.
Returns immediately; caller polls the logfile. Secrets stay in env, never argv.
The guard test pins the contract: the command runs in a different process
group than the caller and outlives the launching shell.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: eval:bg* scripts — detached eval runs for agents
Agent-facing convenience scripts that launch the eval suites through
gstack-detach so a harness SIGTERM can't kill a long run. eval:bg (diff-based),
eval:bg:all, eval:bg:gate, eval:bg:periodic — each returns immediately and
streams to /tmp/gstack-evals.log for polling. The plain test:evals / test:e2e
scripts stay foreground for humans.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs: CLAUDE.md — agents must run long evals via gstack-detach
Codifies the detached-execution default: agent-launched eval/benchmark runs go
through bin/gstack-detach (or the eval:bg* scripts) so a harness SIGTERM or
macOS idle-sleep can't kill a 30-60 min run, then poll the log with a
death-aware watcher. Humans keep foreground scripts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: harden gstack-detach against all four eval-infra killers
The basic bash detach fixed SIGTERM but a real run on a shared dev box hit
three more killers: cross-worktree API saturation (15-way concurrency x a
sibling worktree mass-timed-out the suite), a silent hang (periodic bun died
with no exit marker), and shared-/tmp log contamination (a concurrent
worktree's agent output bled into the log). Rewrite as a portable python3 tool
that bakes in all four fixes:
- fork + setsid: SIGTERM-proof (own session, survives harness polite-quit)
- caffeinate -i on macOS: no idle-sleep death
- --lock NAME (fcntl, machine-wide): concurrent worktrees SERIALIZE instead of
saturating the shared model API
- run-scoped default log (~/.gstack-dev/eval-runs/<label>-<slug>-<branch>-<ts>-<pid>):
no cross-worktree collision/contamination
- --timeout watchdog + a guaranteed '### gstack-detach EXIT=<code> ###' sentinel
on every terminal path: no silent hang, finished-vs-died always detectable
Guard test pins all four: detached pgid differs + outlives launcher, run-scoped
log path, watchdog EXIT=timeout, and lock serialization (second run WAITS).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: eval:bg* use run-scoped logs + machine lock + watchdog
Drop the shared /tmp/gstack-evals.log path (the cross-worktree collision that
contaminated a live run) for gstack-detach's run-scoped default, and add the
machine-wide gstack-evals lock (concurrent worktrees serialize, no API
saturation) plus per-tier watchdog timeouts (60/90/120 min). Each eval:bg*
prints its run-scoped log path to poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs: wire detached-eval guidance into /ship + correct CLAUDE.md flags
- /ship eval step (sections/tests.md): long eval suites launch via gstack-detach
(own session, machine lock, EXIT sentinel) so a turn boundary can't kill a
30+ min run mid-ship — the exact failure observed during this branch's ship.
- CLAUDE.md: correct the now-stale /tmp reference; document the --lock (serialize
worktrees, no API saturation), --timeout watchdog, run-scoped log, and the
guaranteed EXIT sentinel the poller breaks on.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* refactor: extract pure promotedEnv() from conductor-env-shim
Single source of truth for GSTACK_* key promotion semantics. The ambient
promoteConductorEnv() becomes a wrapper; behavior-preserving. Needed by the
hermetic env builder which must not mutate process.env.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: hermetic child-env builder for E2E runners
Allowlist scrub (basics/network/named-auth kept; CONDUCTOR_*, CLAUDE_*,
GSTACK_*, MCP_*, GBRAIN_*, operator credentials dropped), per-runner
extraAllow, overrides merge last, EVALS_HERMETIC=0 byte-identical escape
hatch read at call time (ESM-hoist safe). Sync memoized singleton temp dirs
(<runRoot>/.claude keeps the extractPlanFilePath contract), seeded
.claude.json for non-interactive first run, pid-aware GC of crashed runs.
19 free unit tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: session-runner spawns hermetic children + isolation canaries
claude -p children now get the allowlist-scrubbed env and a gated
--strict-mcp-config (EVALS_HERMETIC=0 restores operator env AND args).
Two gate-tier canaries make the clean room falsifiable: hermetic-canary
asserts env redirect + scrub + zero MCP servers + nonzero API-key cost
from the Bash tool_result (never model prose); hermetic-sentinel plants a
poisoned operator config (user CLAUDE.md + MCP server) and proves the
child cannot see it. Empirically verified on claude 2.1.175: print mode
needs no seed config (the seed serves the PTY path); the child CLI sets
CLAUDECODE for its own tools, so that scrub is pinned in unit tests, not
E2E. hermetic-env.ts joins GLOBAL_TOUCHFILES.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: PTY runner spawns hermetic claude sessions
launchClaudePty children get the allowlist-scrubbed env, a gated
--strict-mcp-config, and the session exposes hermeticConfigDir for
forensics (hermetic plan files live under <dir>/plans/ and still match
extractPlanFilePath via the /.claude dir-name contract). Seeded trust
state covers repo-cwd sessions; the 15s trust-watcher stays as fallback.
Verified foreground via the plan-mode-no-op gate test.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: codex/gemini runners spawn hermetic children
Same allowlist scrub as the claude runners, with each provider's auth
surface re-admitted via extraAllow (codex: OPENAI_API_KEY/CODEX_* plus
its tempHome .codex copy; gemini: GEMINI_*/GOOGLE_* with real HOME for
~/.gemini auth). The gemini spawn previously inherited the full operator
env with no env property at all.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat: agent-sdk-runner spawns hermetic children via complete Options.env
The historical 'env: breaks SDK auth' failure was partial-env replacement:
Options.env replaces the child's entire environment, so objects lacking
ANTHROPIC_API_KEY killed auth. Passing the complete hermetic env (key +
PATH + redirected CLAUDE_CONFIG_DIR/GSTACK_HOME) works — validated live
via query() with a Bash tool call (success, real cost, Conductor vars
scrubbed). Per-test opts.env merges last; ambient key mutation still
works because the builder reads process.env at call time.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: static tripwire pins hermetic wiring in all five runners
Free-tier invariants: every runner builds child env via hermeticChildEnv,
no raw ...process.env spread at any spawn site, --strict-mcp-config gated
on isHermeticEnabled in both claude runners, and no test callsite passes
the operator env into a runner's override parameter (scoped to runner
calls — unit tests spawning gstack bin scripts directly are exempt).
Mirrors the terminal-agent-pid-identity / server-embedder-terminal-port
tripwire idiom.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: refresh codex/factory ship goldens with detached-eval block
|
||
|
Garry TanandClaude Opus 4.7
|
9e244c0bed |
v1.11.1.0 fix: plan-mode handshake + canUseTool test harness (#1182)
* feat: plan-mode handshake for interactive review skills Add a preamble-level STOP-Ask handshake that fires when the user invokes any of the 4 interactive review skills (plan-ceo-review, plan-eng-review, plan-design-review, plan-devex-review) while their Claude Code session is in plan mode. Without this gate, plan mode's "this supercedes any other instructions" system-reminder outranked the skills' interactive STOP gates and the skills silently wrote plan files without any per-finding AskUserQuestion. The handshake offers 2 options (exit-and-rerun, cancel) — the original third "stay and batch" option was dropped after two independent reviewers flagged it as a silent bypass of the skills' anti-skip rule. Architecture decisions (CEO+Eng review): - Preamble-level resolver, not per-template injection (Codex finding #2) - Position 1 in preamble composition: after bash block (_SESSION_ID live), before onboarding AskUserQuestion gates (so fresh-install users see the handshake first, not drowned in telemetry/proactive/routing prompts) - Generator-only `interactive: true` frontmatter flag, following the `preamble-tier` precedent (no host-config frontmatter allowlist edits) - Host-scoped to Claude via `ctx.host === 'claude'` check inside the resolver (simpler than `suppressedResolvers` which only gates `{{}}` placeholders) - One-way-door classification in scripts/question-registry.ts for all 4 skills so question-tuning `never-ask` preferences can't suppress the gate - Synchronous telemetry write to ~/.gstack/analytics/skill-usage.jsonl on handshake fire (captures A-exit and C-cancel outcomes that terminate the skill before end-of-run telemetry runs) Also adds an explicit STOP block to plan-ceo-review Step 0C-bis so the approach-selection question can't silently skip to mode selection. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * feat: extend agent-sdk-runner with canUseTool for AskUserQuestion interception Test harness at test/helpers/agent-sdk-runner.ts gains an optional `canUseTool` callback parameter. When a test supplies it, the harness flips `permissionMode` from `bypassPermissions` (overlay-harness default) to `default` so the SDK actually invokes the callback on every tool use, and auto-adds `AskUserQuestion` to `allowedTools` so Claude can fire it at all. Exports a `passThroughNonAskUserQuestion` helper so tests that only want to intercept AskUserQuestion can auto-allow every other tool with one line: `return passThroughNonAskUserQuestion(toolName, input)`. This is the foundation for D14 — every future interactive-skill E2E test can now assert on AskUserQuestion shape and routing. Previous E2E tests at `test/skill-e2e.test.ts` explicitly instructed the model to skip AskUserQuestion ("non-interactive run") which meant no test could actually verify the question content or routing. 6 new unit tests in test/agent-sdk-runner.test.ts cover: - permissionMode flips to 'default' when canUseTool supplied - permissionMode stays 'bypassPermissions' when canUseTool absent - canUseTool callback reaches the SDK options - AskUserQuestion auto-added to allowedTools when canUseTool supplied - AskUserQuestion NOT added when canUseTool absent - passThroughNonAskUserQuestion helper returns allow+updatedInput Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * test: plan-mode handshake E2E coverage and unit assertions Adds 6 E2E test files and 8 new unit assertions to verify the plan-mode handshake works end-to-end and stays correct under regeneration. E2E tests (gate-tier, paid, EVALS=1 EVALS_TIER=gate): - test/skill-e2e-plan-ceo-plan-mode.test.ts — handshake fires before any Write/Edit when plan-mode distinctive phrase is present; 2-option shape (Exit/Cancel); option A routes to ExitPlanMode cleanly - test/skill-e2e-plan-eng-plan-mode.test.ts — same contract for plan-eng - test/skill-e2e-plan-design-plan-mode.test.ts — same contract for plan-design; exercises C-cancel branch instead of A-exit - test/skill-e2e-plan-devex-plan-mode.test.ts — same contract for plan-devex - test/skill-e2e-plan-mode-no-op.test.ts — negative regression: handshake must NOT fire when distinctive phrase is absent; skill proceeds normally through Step 0 (REGRESSION RULE guardrail against breaking existing interactive-review sessions) - test/e2e-harness-audit.test.ts — free unit test asserting every `interactive: true` skill has at least one canUseTool-using test file (prevents future drift where a skill opts in without coverage) Shared helper test/helpers/plan-mode-handshake-helpers.ts centralizes the canUseTool interceptor + distinctive-phrase injection so the 4 sibling E2E tests are thin wiring (~20 LOC each) and can't drift out of sync. Unit assertions added to test/gen-skill-docs.test.ts: - handshake section present in all 4 Claude-generated SKILL.md files - handshake section absent from non-interactive Claude skills (ship, review, qa, office-hours, codex, retro, cso) - handshake section absent from non-Claude host outputs (.agents, etc.) - 0C-bis STOP block present in plan-ceo-review/SKILL.md at correct position (between the "Present these approach options" line and "### 0D-prelude" header) - handshake resolver wired BEFORE generateUpgradeCheck in preamble composition order 6 new gate-tier entries added to test/helpers/touchfiles.ts so any change to the handshake resolver, preamble composition, skill templates, question registry, one-way-door classifier, or agent-sdk-runner fires the relevant E2E tests. test/touchfiles.test.ts updated for the new selection count (plan-ceo-review/** now triggers 15 tests, up from 8). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * chore(v1.11.1.0): VERSION bump + CHANGELOG entry + TODOS follow-ups Bumps from main's v1.11.0.0 to v1.11.1.0 (PATCH — bug-fix release, no new user-facing artifacts). CHANGELOG entry covers the plan-mode handshake, agent-sdk-runner canUseTool extension, and the 2 follow-up TODOs. CHANGELOG order: v1.11.1.0 (this) → v1.11.0.0 (workspace-aware ship, merged from main) → v1.10.1.0 (overlay efficacy harness). No duplicate headers. Syncs package.json version to match VERSION per the Step 12 idempotency invariant (both files must agree or /ship halts). TODOS.md: - Preserves the Testing/security-bench-haiku-responses P1 added on main - Adds P1 "Structural STOP-Ask forcing function" — broader class of the bug this release fixes - Adds P2 "Apply interactive: true to non-review skills (office-hours, codex, investigate, qa, retro, cso)" Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> |
||
|
Garry TanandClaude Opus 4.7
|
e3d7f49c74 |
feat(v1.10.1.0): overlay efficacy harness + Opus 4.7 fanout nudge removal (#1166)
* refactor: export readOverlay from model-overlay resolver Needed by the overlay-efficacy eval harness to resolve INHERIT directives without going through generateModelOverlay's full TemplateContext. * chore: add @anthropic-ai/claude-agent-sdk@0.2.117 dep Pinned exact for SDK event-shape stability. Used by the overlay-efficacy harness to drive the model through a closer-to-real Claude Code harness than `claude -p`. * feat(preflight): sanity check for agent-sdk + overlay resolver Verifies: SDK loads, claude-opus-4-7 is a live API model, SDKMessage event shape matches assumptions, readOverlay resolves INHERIT directives and includes expected content. Run with `bun run scripts/preflight-agent-sdk.ts`. PREFLIGHT OK on first run, $0.013 API spend. * feat(eval): parametric overlay-efficacy harness (runner + fixtures) `test/helpers/agent-sdk-runner.ts` wraps @anthropic-ai/claude-agent-sdk with explicit `AgentSdkResult` types, process-level API concurrency semaphore, and 3-shape 429 retry (thrown error, result-message error, mid-stream SDKRateLimitEvent). Pins the local claude binary via `pathToClaudeCodeExecutable`. `test/fixtures/overlay-nudges.ts` holds the typed registry. Two fixtures for the first measurement: `opus-4-7-fanout-toy` (3-file read) and `opus-4-7-fanout-realistic` (mixed-tool audit). Strict validator rejects duplicate ids, non-integer trials, unsafe overlay paths, non-safe id chars, and missing overlay files at module load. Adding a future overlay nudge eval = one fixture entry. * test(eval): unit tests for agent-sdk-runner (36 tests, free tier) Stub `queryProvider` feeds hand-crafted SDKMessage streams. Covers: happy-path shape, all 3 rate-limit shapes + retry, workspace reset on retry, persistent 429 -> `RateLimitExhaustedError`, non-429 propagation, process-level concurrency cap, options propagation, artifact path uniqueness, cost/turn mapping, and every validator rejection case. * test(eval): paid periodic overlay-efficacy harness `test/skill-e2e-overlay-harness.test.ts` iterates OVERLAY_FIXTURES, runs two arms per fixture (overlay-ON, overlay-OFF) at N=10 trials with bounded concurrency. Arms use SDK preset `claude_code` so both include the real Claude Code system prompt; overlay-ON appends the resolved overlay text. Saves per-trial raw event streams to `~/.gstack/projects/<slug>/transcripts/` for forensic recovery. Gated on `EVALS=1 && EVALS_TIER=periodic`. ~$3/run (40 trials). * test: register overlay harness in touchfiles (both maps) Entries for `overlay-harness-opus-4-7-fanout-toy` and `opus-4-7-fanout-realistic` in E2E_TOUCHFILES (deps: model-overlays/, fixtures file, runner, resolver) and E2E_TIERS (`periodic`). Passes `test/touchfiles.test.ts` completeness check. * fix(opus-4.7): remove "Fan out explicitly" overlay nudge Measured counterproductive under the new SDK harness. Baseline Opus 4.7 emits first-turn parallel tool_use blocks 70% of the time on a 3-file read prompt. With the custom nudge: 10%. With Anthropic's own canonical `<use_parallel_tool_calls>` block from their parallel-tool-use docs: 0%. Both overlays suppress fanout; neither improves it. On realistic multi-tool prompts (audit a project: read files + glob + summarize), Opus 4.7 never fans out in first turn regardless of overlay. Zero of 20 trials. Not a prompt problem. Keeping the other three nudges (effort-match, batch questions, literal interpretation) pending their own measurement. Harness is ready for follow-up fixtures — add one entry to `test/fixtures/overlay-nudges.ts` to measure any overlay bullet. Cost of investigation: ~$7 total across 3 eval runs. * chore: bump version and changelog (v1.6.5.0) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(eval): extend OverlayFixture with allowedTools, maxTurns, direction Per-fixture tool allowlist unblocks measuring nudges that need Edit/Write (e.g. literal-interpretation 'fix the failing tests' needs write access). Per-fixture maxTurns lets harder prompts run longer without changing the default. `direction` is cosmetic metadata for test output labeling. Also adds reusable predicates and metrics: - lowerIsBetter20Pct / higherIsBetter20Pct — 20% lift threshold vs baseline - bashToolCallCount — count of Bash tool_use across the session - turnsToCompletion — SDK-reported num_turns at result - uniqueFilesEdited — Edit/Write/MultiEdit file_path set size test/skill-e2e-overlay-harness.test.ts now threads fixture.allowedTools and fixture.maxTurns through runArm. * test(eval): 3 more overlay fixtures to measure remaining Claude nudges Measures three overlay bullets that haven't been tested yet: - claude-dedicated-tools-vs-bash — claude.md says 'prefer Read/Edit/Write/ Glob/Grep over cat/sed/find/grep'. Fixture prompts 'list every TypeScript file under src/ and tell me what each exports' and counts Bash tool_use across the session. Overlay-ON should drop it by >=20%. - opus-4-7-effort-match-trivial — opus-4-7.md says 'simple file reads don't need deep reasoning.' Fixture uses a trivial one-file prompt (config.json lookup) and measures turns_used. Overlay-ON should be <=80% of baseline turns. - opus-4-7-literal-interpretation — opus-4-7.md says 'fix ALL failing tests, not just the obvious one.' Fixture seeds three failing test files with deliberately distinct failure modes and counts unique files edited. Overlay-ON should touch >=20% more files. Adding a fourth fixture for any remaining overlay nudge is a single entry. The harness is now proven on: fanout (deleted after measurement), dedicated tools, effort-match, and literal-interpretation. * fix(eval): handle SDK max-turns throw gracefully Some @anthropic-ai/claude-agent-sdk versions throw from the query generator when maxTurns is reached, instead of emitting a result message with subtype='error_max_turns'. The runner treated that as a non-retryable error and killed the whole periodic run on the first fixture that exceeded its turn cap. Added isMaxTurnsError() detector and a catch branch that synthesizes an AgentSdkResult from events captured before the throw, with exitReason='error_max_turns' and costUsd=0 (unknown from the thrown path). The metric function still runs against whatever assistant turns were collected, so the trial produces a usable number. Hoisted events/assistantTurns/toolCalls/assistantTextParts and the timing counters out of the inner try so the catch branch can read them. No behavior change on the success path or on rate-limit retry paths. * test(eval): bump maxTurns to 15 for claude-dedicated-tools-vs-bash The prompt 'list every TypeScript file under src/ and tell me what each exports' needs 1 turn for Glob + ~5 for Reads + 1 for summary. Default maxTurns=5 was not enough; prior run threw from the SDK on this fixture and tanked the whole periodic eval. Bumping to 15 gives headroom. The runner now also handles max-turns gracefully even if a future fixture underestimates, so this is belt and suspenders. * test(eval): Sonnet 4.6 variants of the 5 Opus-4.7 fixtures Same overlays, same prompts, same metrics, `model: 'claude-sonnet-4-6'`. Tests whether the overlays behave differently on a weaker Claude model where baseline behavior is shakier. Sonnet trials cost ~3-4x less than Opus so these 5 add ~$4.50 to a full run. Measurement result from the first paired run (100 trials total, ~$14.55): - **Sonnet + effort-match shows real overlay benefit.** With the overlay on, Sonnet takes 2.5 turns on a trivial `What's the version in config.json?` prompt. Without, it takes exactly 3.0 turns in all 10 trials. ~17% reduction, below the 20% pass threshold but the signal is clean: overlay-ON distribution [2,2,2,2,2,3,3,3,3,3] vs overlay-OFF [3,3,3,3,3,3,3,3,3,3]. - All other Sonnet dimensions flat (fanout, dedicated-tools, literal interpretation). Same as Opus on those axes. - Opus effort-match remains flat (2.60 vs 2.50, +4% slower with overlay). Implication: model-stratified. The overlay stack helps Sonnet on some axes where it does nothing on Opus. Wholesale removal would hurt Sonnet. Per-nudge per-model measurement is the right move going forward. * chore: bump version to 1.10.1.0 Updates VERSION, package.json, CHANGELOG header, and TODOS completion marker from 1.6.5.0 to 1.10.1.0. --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |