mirror of
https://github.com/garrytan/gstack.git
synced 2026-05-01 19:25:10 +02:00
main
5 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Garry Tan
|
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> |
||
|
Garry Tan
|
a81be53621 |
v1.10.0.0: fix AskUserQuestion cadence + Pros/Cons format upgrade (#1178)
* fix(preamble): reorder AskUserQuestion Format above model overlay + rewrite Opus 4.7 pacing directive
Root cause of plan-review regression (v1.6.4.0): model overlays rendered
ABOVE the pacing rule in every SKILL.md, so Opus 4.7 read "Batch your
questions" first and absorbed it as the ambient default. The overlay's
claimed subordination ("skill wins on pacing, always") didn't stick —
literal-interpretation mode reads physical order, not claimed hierarchy.
Part 1 of 4 (plan: ~/.claude/plans/system-instruction-you-are-working-polymorphic-twilight.md):
scripts/resolvers/preamble.ts
- Move generateAskUserFormat above generateModelOverlay in section array
- Comment explains why — prevents future refactors from silently reverting
model-overlays/opus-4-7.md
- Replace "Batch your questions" block with "Pace questions to the skill"
- New wording makes one-question-per-turn the default when the skill
contains STOP directives; batching becomes the explicit exception
Regenerated 30 SKILL.md files via bun run gen:skill-docs.
Verified:
- With --model opus-4-7: Format renders at line 359, Model-Specific
Patch at 373, "Pace questions" at 419 (Format comes first, overlay
second, pacing directive intact).
- bun test passes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(plan-reviews): tighten STOP/escape-hatch directives across 4 templates
Part 2 of 4 (plan: ~/.claude/plans/system-instruction-you-are-working-polymorphic-twilight.md).
Codex caught that v1.6.3.0's reasoning collapsed on Opus 4.7: the old
escape-hatch wording ("If no issues or fix is obvious, state what
you'll do and move on — don't waste a question") let the literal
interpreter classify every finding as having an "obvious fix" and skip
AskUserQuestion entirely. Reviews became reports.
Per-template hardening (16 sites total, verified by rg):
plan-ceo-review/SKILL.md.tmpl (13 sites):
- 12 inline STOP directives: replace the full escape-hatch clause with
"zero findings → say so and proceed; findings → MUST call AskUserQuestion
as a tool_use, including for obvious fixes."
- 1 Escape hatch bullet in CRITICAL RULE section: tightened.
plan-eng-review, plan-design-review, plan-devex-review (1 site each):
- Each template's Escape hatch bullet tightened to match the new CEO wording,
adapted for each review's domain (issue/gap, decision/design/DX alternatives).
After regeneration: rg "don't waste a question" returns 0 across all
*SKILL.md.tmpl and *SKILL.md files. "zero findings, state" wording
present 16 times (matches prior count of escape-hatch sites).
bun test passes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(preamble): upgrade AskUserQuestion format to Pros/Cons decision brief
Part 4 of 4 (plan: ~/.claude/plans/system-instruction-you-are-working-polymorphic-twilight.md).
Every AskUserQuestion now renders as a decision brief, not a bullet list:
D-numbered header, ELI10, Stakes-if-we-pick-wrong, Recommendation, Pros/Cons
with ✅/❌ markers per option, closing Net: tradeoff synthesis.
scripts/resolvers/preamble/generate-ask-user-format.ts
- Full rewrite. Preserves prior rules (Re-ground, ELI10, Recommend,
Completeness, Options) and adds:
- D-numbering per skill invocation (model-level, not runtime state)
- Stakes line (pain avoided / capability unlocked / consequence named)
- Pros/Cons block with min 2 ✅ + 1 ❌ per option, min 40 chars/bullet
- Hard-stop escape: "✅ No cons — this is a hard-stop choice" for
genuine one-sided choices (destructive-action confirmations)
- Neutral-posture handling (CT1-compliant): (recommended) label
STAYS on default option to preserve AUTO_DECIDE contract; neutrality
expressed as prose in Recommendation line only
- Net line closes the decision with a one-sentence tradeoff frame
- Rule 11: tool_use mandate (prose "Question:" blocks don't count)
- Self-check list before emitting
test/skill-validation.test.ts
- Update format assertions to check for new Pros/Cons tokens
(Pros / cons:, Recommendation: <choice>, Net:, ELI10, Stakes if we
pick wrong:, ✅, ❌) across all tier-2+ skills
- Old "RECOMMENDATION: Choose" expectation removed (the new format uses
mixed-case "Recommendation:" with no literal "Choose")
test/skill-e2e-plan-format.test.ts
- Add v1.7.0.0 format token regexes (PROS_CONS_HEADER_RE, PRO_BULLET_RE,
CON_BULLET_RE, NET_LINE_RE, D_NUMBER_RE, STAKES_RE)
- Existing RECOMMENDATION_RE loosened to accept mixed-case "Recommendation:"
(canonical v1.7.0.0 form) alongside all-caps (legacy). Tests are
additive — the strict new-format gate is the upcoming cadence eval.
Regenerated 30 SKILL.md files via bun run gen:skill-docs.
Verified:
- bun test: 319 pass (1 pre-existing security-bench fixture oversize
failure on main, unrelated — confirmed via git stash test on main HEAD)
- New format tokens render in all tier-2+ skills (plan-ceo-review,
plan-eng-review, ship, office-hours verified)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test: gate-tier units + periodic Pros/Cons evals for AskUserQuestion format
Part 3 of 4 (plan: ~/.claude/plans/system-instruction-you-are-working-polymorphic-twilight.md).
Gate-tier (E1, free, runs on every `bun test`):
test/preamble-compose.test.ts — pins the composition order
Asserts AskUserQuestion Format section renders BEFORE Model-Specific
Behavioral Patch in tier-≥2 preamble output. Covers claude default,
opus-4-7 overlay, tier 2/3, and codex host. Catches any future edit
to scripts/resolvers/preamble.ts that silently reverts the order.
test/resolver-ask-user-format.test.ts — pins the Pros/Cons contract
14 assertions against generateAskUserFormat output: D<N>, ELI10,
Stakes if we pick wrong:, Recommendation: <choice>, Pros / cons:,
✅/❌ markers, min 2 pros + 1 con rules, hard-stop escape exact
phrase, neutral-posture CT1 rule ((recommended) label preserved for
AUTO_DECIDE), Completeness coverage-vs-kind, tool_use mandate
(rule 11), self-check list, D-numbering model-level caveat.
test/model-overlay-opus-4-7.test.ts — pins the pacing directive
Asserts raw overlay file + resolved overlay output contain "Pace
questions to the skill" and NOT "Batch your questions". Verifies
INHERIT:claude chain still works (Todo-list, subordination wrapper),
Fan out / Effort-match / Literal interpretation nudges preserved.
Also asserts claude base overlay does NOT carry the Opus-specific
pacing directive (no cross-contamination).
Periodic-tier (E2, Opus-dependent, ~$1-2/run):
test/skill-e2e-plan-prosons.test.ts — 4 cases extending v1.6.3.0 harness
1. Format positive — every token present when plan has real tradeoff
2. Hard-stop NEGATIVE — plan with genuine tradeoff must NOT dodge to
"No cons — hard-stop choice" escape
3. Neutral-posture NEGATIVE — plan where one option dominates must emit
(recommended) label + "because <reason>", must NOT dodge to
"taste call" / "no preference"
4. Hard-stop POSITIVE — destructive-action plan may legitimately use
the hard-stop escape
test/helpers/touchfiles.ts — entries for all new eval cases
Dependencies: overlay, preamble.ts, generate-ask-user-format.ts, and
the 4 plan-review templates. Diff-based selection triggers the evals
whenever those files change. Also added entries for 7 expanded-coverage
cases (ship, office-hours, investigate, qa, review, design-review,
document-release) — test cases will land in follow-up PRs per skill.
Follow-ups noted in test file header:
- True multi-turn cadence eval (3 findings → 3 distinct asks) — current
harness captures one $OUT_FILE per session; multi-turn capture needs
new harness support.
- Expanded-coverage test cases for the 7 non-plan-review skills.
Verified:
- bun test: 349 pass (30 new + 319 baseline), 1 pre-existing security-bench
oversize failure on main (unrelated, unchanged).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test: regenerate golden fixtures + update ELI10 phrase check for v1.7.0.0
Pros/Cons format rewrite (
|
||
|
Garry Tan
|
69733e2622 |
fix(plan-reviews): restore RECOMMENDATION + Completeness split + Codex ELI10 (v1.6.3.0) (#1149)
* test: add AskUserQuestion format regression eval for plan reviews Four-case periodic-tier eval that captures the verbatim AskUserQuestion text /plan-ceo-review and /plan-eng-review produce, then asserts the format rule is honored: RECOMMENDATION always, Completeness: N/10 only on coverage-differentiated options, and an explicit "options differ in kind" note on kind-differentiated options. Cases: - plan-ceo-review mode selection (kind-differentiated) - plan-ceo-review approach menu (coverage-differentiated) - plan-eng-review per-issue coverage decision - plan-eng-review per-issue architectural choice (kind-differentiated) Classified periodic because behavior depends on Opus non-determinism — gate-tier would flake and block merges. Test harness instructs the agent to write its would-be AskUserQuestion text to $OUT_FILE rather than invoke a real tool (MCP AskUserQuestion isn't wired in the test subprocess). Regex predicates then validate the captured content. Cost: ~$2 per full run. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(plan-reviews): restore RECOMMENDATION + split Completeness by question type Opus 4.7 users reported /plan-ceo-review and /plan-eng-review stopped emitting the RECOMMENDATION line and per-option Completeness: X/10 scores. E2E capture showed the real failure mode: on kind-differentiated questions (mode selection, architectural A-vs-B, cherry-pick), Opus 4.7 either fabricated filler scores (10/10 on every option — conveys nothing) or dropped the format entirely when the metric didn't fit. Fix is at two layers: 1. scripts/resolvers/preamble/generate-ask-user-format.ts splits the old run-on step 3 into: - Step 3 "Recommend (ALWAYS)": RECOMMENDATION is required on every question, coverage- or kind-differentiated. - Step 4 "Score completeness (when meaningful)": emit Completeness: N/10 only when options differ in coverage. When options differ in kind, skip the score and include a one-line explanatory note. Do not fabricate scores. 2. scripts/resolvers/preamble/generate-completeness-section.ts updates the Completeness Principle tail to match. Without this, the preamble contained two rules (one conditional, one unconditional) and the model hedged. Template anchors reinforce the distinction where agent judgment is most likely to drift: - plan-ceo-review Section 0C-bis (approach menu) gets the coverage-differentiated anchor. - plan-ceo-review Section 0F (mode selection) gets the kind-differentiated anchor. - plan-eng-review CRITICAL RULE section gets the coverage-vs-kind rule for every per-issue AskUserQuestion raised during the review. Regenerated SKILL.md for all T2 skills + golden fixtures refreshed. Every skill using the T2 preamble now has the same conditional scoring rule. Verified via new periodic-tier eval (test/skill-e2e-plan-format.test.ts): all 4 cases fail on prior behavior, all 4 pass with this fix. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: bump version and changelog (v1.6.2.0) Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * test: add Codex eval for AskUserQuestion format compliance Four-case periodic-tier eval mirrors test/skill-e2e-plan-format.test.ts but drives the plan review skills via codex exec instead of claude -p. Context: Codex under the gpt.md "No preamble / Prefer doing over listing" overlay tends to skip the Simplify/ELI10 paragraph and the RECOMMENDATION line on AskUserQuestion calls. Users have to manually re-prompt "ELI10 and don't forget to recommend" almost every time. This test pins the behavior so regressions surface. Cases: - plan-ceo-review mode selection (kind-differentiated) - plan-ceo-review approach menu (coverage-differentiated) - plan-eng-review per-issue coverage decision - plan-eng-review per-issue architectural choice (kind-differentiated) Assertions on captured AskUserQuestion text: - RECOMMENDATION: Choose present (all cases) - Completeness: N/10 present on coverage, absent on kind - "options differ in kind" note present on kind - ELI10 length floor (>400 chars) — catches bare options-only output Cost: ~\$2-4 per full run. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * fix(preamble): harden AskUserQuestion Format + Codex ELI10 carve-out Follow-up to v1.6.2.0. Codex (GPT-5.4) under the gpt.md overlay treated "No preamble / Prefer doing over listing" as license to skip the Simplify paragraph and the RECOMMENDATION line on AskUserQuestion calls. Users had to manually re-prompt "ELI10 and don't forget to recommend" almost every time. Two layers: 1. model-overlays/gpt.md — adds an explicit "AskUserQuestion is NOT preamble" carve-out. The "No preamble" rule applies to direct answers; AskUserQuestion content must emit the full format (Re-ground, Simplify/ELI10, Recommend, Options). Tells the model: if you find yourself about to skip any of these, back up and emit them — the user will ask anyway, so do it the first time. 2. scripts/resolvers/preamble/generate-ask-user-format.ts — step 2 renamed to "Simplify (ELI10, ALWAYS)" with explicit "not optional verbosity, not preamble" framing. Step 3 "Recommend (ALWAYS)" hardened: "Never omit, never collapse into the options list." All T2 skills regenerated across all hosts. Golden fixtures refreshed (claude-ship, codex-ship, factory-ship). Updated the ELI10 assertion in test/gen-skill-docs.test.ts to match the new wording. Codex compliance to be verified empirically via test/codex-e2e-plan-format.test.ts. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * test: fix Codex eval sandbox + collector API Two test infrastructure bugs in the initial Codex eval landed in the prior commit: 1. sandbox: 'read-only' (the default) blocked Codex from writing $OUT_FILE. Test reported "STATUS: BLOCKED" and exited 0 without a capture file. Fixed: sandbox: 'workspace-write' for all 4 cases, allowing writes inside the tempdir. 2. recordCodexResult called a non-existent evalCollector.record() API (I invented it). The real surface is addTest() with a different field schema. Aligned with test/codex-e2e.test.ts pattern. With both fixed, the eval now actually measures Codex AskUserQuestion format compliance. All 4 cases pass on v1.6.2.0 with the gpt.md carve-out: RECOMMENDATION always, Completeness: N/10 only on coverage, "options differ in kind" note on kind, ELI10 explanation present. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * chore: bump version and changelog (v1.6.3.0) Adds the Codex ELI10 + RECOMMENDATION carve-out scope landed after v1.6.2.0's Claude-verified fix. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
Garry Tan
|
656df0e37e |
feat(v1.5.2.0): Opus 4.7 migration — model overlay, voice, routing (#1117)
* feat(v1.5.2.0): Opus 4.7 migration — model overlay, voice, routing
Adapts GStack skill text for Claude Opus 4.7's behavioral changes per
Anthropic's migration guide and community findings.
Key changes:
model-overlays/claude.md:
- Fan out explicitly (4.7 spawns fewer subagents by default)
- Effort-match the step (avoid overthinking simple tasks at max)
- Batch questions in one AskUserQuestion turn
- Literal interpretation awareness (deliver full scope)
hosts/claude.ts:
- coAuthorTrailer updated to Claude Opus 4.7
SKILL.md.tmpl:
- Expanded routing triggers with colloquial variants ("wtf",
"this doesn't work", "send it", "where was I") — 4.7 won't
generalize from sparse trigger patterns like 4.6 did
- Added missing routes: /context-save, /context-restore, /cso, /make-pdf
- Changed routing fallback from strict "do NOT answer directly" to
"when in doubt, invoke the skill" — false positives are cheaper
than false negatives on 4.7's literal interpreter
generate-voice-directive.ts:
- Added concrete good/bad voice example — 4.7 needs shown examples,
not just described tone. "auth.ts:47 returns undefined..." vs
"I've identified a potential issue..."
Regenerated all 38 SKILL.md files. All tests pass.
* refactor(opus-4.7): split overlay, align routing, fix trailer fallback
Follow-up to wintermute's initial Opus 4.7 migration commit (addresses
ship-quality review findings before v1.6.1.0 release).
Overlay split (model-overlays/):
- Move 4 Opus-4.7-specific nudges (Fan out, Effort-match, Batch your
questions, Literal interpretation) from claude.md into new
opus-4-7.md with {{INHERIT:claude}}
- claude.md now holds only model-agnostic nudges (Todo discipline,
Think before heavy, Dedicated tools over Bash)
- Prevents Opus-4.7-specific guidance leaking onto Sonnet/Haiku
- Uses existing {{INHERIT:claude}} mechanism at
scripts/resolvers/model-overlay.ts:28-43
scripts/models.ts:
- Add opus-4-7 to ALL_MODEL_NAMES
- resolveModel: claude-opus-4-7-* variants route to opus-4-7,
all other claude-* variants continue to route to claude
scripts/resolvers/utility.ts:
- Update coAuthor trailer fallback: Opus 4.6 -> Opus 4.7
(fallback was missed in the initial migration commit)
scripts/resolvers/preamble/generate-routing-injection.ts:
- Align policy with new SKILL.md.tmpl: soft "when in doubt, invoke"
instead of hard "ALWAYS invoke... Do NOT answer directly"
- Replace stale /checkpoint reference with /context-save +
/context-restore (skills were renamed in v1.0.1.0)
- Expand route coverage to match full skill inventory:
/plan-devex-review, /qa-only, /devex-review, /land-and-deploy,
/setup-deploy, /canary, /open-gstack-browser,
/setup-browser-cookies, /benchmark, /learn, /plan-tune, /health
scripts/resolvers/preamble/generate-voice-directive.ts:
- Voice example closing: "Want me to ship it?" -> "Want me to fix it?"
- Preserves directness while routing through review gates
SKILL.md.tmpl:
- Add routing triggers for skills that were missing from the list:
/plan-devex-review, /qa-only, /devex-review, /land-and-deploy,
/setup-deploy, /canary, /open-gstack-browser,
/setup-browser-cookies, /benchmark, /learn, /plan-tune, /health
- Within Opus 4.7 overlay, added scope boundary to
"Literal interpretation" nudge ("fix tests that this branch
introduced or is responsible for")
- Added pacing exception to "Batch your questions" nudge so skills
that require one-question-at-a-time pacing still win
Follow-up commit will regenerate SKILL.md files + update goldens.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore(opus-4.7): regenerate SKILL.md files + update golden fixtures
Mechanical consequence of the preceding source changes (overlay split,
routing alignment, voice example, routing expansion). No behavior change
beyond what that commit introduced.
- 36 SKILL.md files regenerated via bun run gen:skill-docs
- 3 golden fixtures updated (claude, codex, factory ship skill)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(routing): assert slash-prefixed skills + new policy + current names
Align gen-skill-docs.test.ts routing assertions with the remediated
routing-injection output:
- Expect '/office-hours' slash-prefixed form (matches SKILL.md.tmpl style)
- Add test asserting /context-save + /context-restore references
(guards against stale '/checkpoint' name regression)
- Add test asserting "When in doubt, invoke the skill" soft policy
(guards against "Do NOT answer directly" hard policy regression)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(binary-guard): replace xargs-per-file loops with fs.statSync + mode filter
The "no compiled binaries in git" describe block had two flaky tests:
- "git tracks no files larger than 2MB" timed out at 5s regularly because
it spawned one `sh -c` per tracked file via `xargs -I{}` (~571 shells
on every run, ~11s locally).
- "git tracks no Mach-O or ELF binaries" ran `file --mime-type` over every
tracked file (~3-10s, flaky near the timeout).
Both were pre-existing — not caused by any recent change — but showed up
as red in every local `bun test` run and masked legit failures in the
same suite.
Rewrites:
- 2MB test: `fs.statSync(f).size` in a filter. Millisecond-fast.
- Mach-O test: pre-filter to mode 100755 files via `git ls-files -s`,
then batch-invoke `file --mime-type` once across all executables.
With zero executables tracked, the `file` invocation is skipped.
Test suite: 320 pass, 0 fail, 907ms (was ~12.7s with 2 fails).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(team-mode): give setup -q / setup --local tests a 3-minute budget
./setup runs a full install, Bun binary build, and skill regeneration.
On a cold cache it takes 60-90s, comfortably above bun test's 5s default.
Both "setup -q produces no stdout" and "setup --local prints deprecation
warning" have been flaky-to-failing for a while with [5001.78ms] timeouts.
The test logic was fine, the budget wasn't. Bumped both to 180s via the
third-arg timeout.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(opus-4.7): E2E eval for fanout rate + routing precision
Closes the measurement gap flagged by the ship-quality review: "zero
tests exercise Opus 4.7 behavior; every skill-e2e hardcodes 4.6."
Two cases, both pinned to claude-opus-4-7:
1. Fanout rate (A/B)
- Arm A: regen SKILL.md with --model opus-4-7 (overlay ON, includes
"Fan out explicitly" nudge).
- Arm B: regen SKILL.md with --model claude (overlay OFF, only
model-agnostic nudges).
- Prompt: "Read alpha.txt, beta.txt, gamma.txt. These are independent."
- Measure: parallel tool calls in first assistant turn.
- Assert: arm A >= arm B.
2. Routing precision (6-case mini-benchmark)
- 3 positive prompts that should route (wtf bug, send it, does it work)
- 3 negative prompts that match keywords but should NOT route
(syntax question, algorithm question, slack message)
- Assert: TP rate >= 66%, FP rate <= 33%.
Cost estimate: ~$3-5 per full run. Classified as periodic tier per
CLAUDE.md convention (Opus model, non-deterministic). Runs only with
EVALS=1 env var, touchfile-gated so unrelated diffs don't trigger it.
Test plan artifact at
~/.gstack/projects/garrytan-gstack/garrytan-feat-opus-4.7-migration-eng-review-test-plan-20260421-230611.md
tracks the full specification.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(opus-4.7): rewrite fanout nudge to show parallel tool_use pattern
The original fanout nudge told 4.7 to "spawn subagents in the same turn"
and "run independent checks concurrently" in prose. An E2E eval on
claude-opus-4-7 reading 3 independent files showed zero effect: both
overlay-ON and overlay-OFF arms emitted serial Reads across 3-4 turns.
Rewrite follows the same "show not tell" principle the PR introduced for
voice examples. The nudge now includes a concrete wrong/right contrast
showing the exact tool_use structure:
Wrong (3 turns):
Turn 1: Read(foo.ts), then wait
Turn 2: Read(bar.ts), then wait
Turn 3: Read(baz.ts)
Right (1 turn, 3 parallel tool_use blocks in one assistant message):
Turn 1: [Read(foo.ts), Read(bar.ts), Read(baz.ts)]
Applies to Read, Bash, Grep, Glob, WebFetch, Agent, and any tool where
sub-calls don't depend on each other's output.
Effect on test/skill-e2e-opus-47.test.ts fanout eval: unchanged (both
arms still 0 parallel in first turn via `claude -p`). May land better in
Claude Code's interactive harness, where the system prompt + tool
handlers differ. Tracked as P0 TODO for follow-up verification in the
correct harness.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(opus-4.7): tighten ambiguous /qa routing prompt
"does this feature work on mobile? can you check the deploy?" was too
vague — a reasonable agent asks "which feature?" via AskUserQuestion
instead of routing to /qa. That's not a routing miss, it's an under-
specified prompt.
Replaced with "I just pushed the login flow changes. Test the deployed
site and find any bugs." — concrete subject + clear QA verb.
Result: pos-does-it-work went from MISS to OK, routing TP rate 2/3 -> 3/3.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(opus-4.7): rewrite scratch-root helper + add afterAll cleanup
First run of the Opus 4.7 eval exposed two test-setup gaps that made
results misleading:
- Only the root gstack SKILL.md was installed. Claude Code does
auto-discovery per-directory under .claude/skills/{name}/SKILL.md, so
without individual skill dirs the Skill tool had nothing to route to.
Positive routing cases all failed.
- `claude -p` does not load SKILL.md content as system context the way
the Claude Code harness does. The overlay nudges in SKILL.md were
invisible to the model, so the fanout A/B could not actually differ.
New `mkEvalRoot(suffix, includeOverlay)` helper, modelled on the pattern
in skill-routing-e2e.test.ts:
- Installs per-skill SKILL.md under .claude/skills/ for ~14 key skills
so the Skill tool has discoverable targets.
- Writes an explicit routing block into project CLAUDE.md.
- When includeOverlay is true, inlines the content of
model-overlays/opus-4-7.md into CLAUDE.md too. This is what makes the
fanout A/B observable in `claude -p`: arm ON gets the overlay in
context, arm OFF does not.
Plus an afterAll that re-runs gen-skill-docs at the default model so
the working tree is not left with opus-4-7-generated SKILL.md files
after the eval finishes (would break golden-file tests in the next
`bun test` run otherwise).
With this setup in place: routing went from 3/3 FAIL to 3/3 PASS
(correct skill or clarification in every positive case, zero false
positives on negatives). Fanout A/B is now a fair comparison; still
shows 0 parallel in both arms under `claude -p` (tracked as a P0 TODO
for re-measurement inside Claude Code's harness, where fanout may land
differently).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(todos): verify Opus 4.7 fanout nudge in Claude Code harness (P0)
v1.6.1.0 shipped a rewritten "Fan out explicitly" nudge with a concrete
tool_use example. Under `claude -p` on claude-opus-4-7, the A/B eval
showed zero parallel tool calls in the first turn for both arms
(overlay ON and OFF). Routing verified 3/3 in the same harness, so the
gap is specific to fanout and likely to `claude -p`'s system prompt +
tool wiring.
This TODO closes the measurement loop the ship-quality review flagged:
re-run the fanout A/B inside Claude Code's real harness (or a faithful
replica) before landing another Opus migration claim.
P0 because it is a ship-quality commitment from the v1.6.1.0 release
notes, not a nice-to-have.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore(release): v1.6.1.0 — Opus 4.7 migration, reviewed
Bump VERSION + package.json from 1.6.0.0 to 1.6.1.0. New CHANGELOG
entry describing the ship-quality remediation of PR #1117:
- Overlay split (model-agnostic claude.md + opus-4-7.md with INHERIT)
- Routing-injection aligned with SKILL.md.tmpl ("when in doubt" policy,
current skill names, full skill inventory)
- utility.ts trailer fallback updated
- Voice example closes through review gate instead of ship-bypass
- Literal-interpretation nudge bounded to branch scope
- Batch-questions nudge has explicit pacing exception
- First Opus 4.7 eval: routing verified 3/3, fanout A/B unverified
under `claude -p` (tracked as P0 TODO for next rev)
- Pre-existing test failures fixed: fs.statSync binary guard, 180s
setup timeout, golden-file updates
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(opus-4.7): key touchfile entries by testName, not describe text
TOUCHFILES completeness scan in test/touchfiles.test.ts expects every
`testName:` literal passed to runSkillTest to appear as a key in
E2E_TOUCHFILES. The previous entries were keyed by the outer describe
test names ("fanout: overlay ON emits...") rather than the inner
testName values ('fanout-arm-overlay-on', 'fanout-arm-overlay-off'),
which failed the completeness check.
Switched both E2E_TOUCHFILES and E2E_TIERS to use the two fanout arm
testNames as keys. The routing sub-tests use a template literal
(`routing-${c.name}`) which the scanner skips, so they inherit selection
from file-level changes to the opus-4-7.md / routing-injection.ts paths
already covered by the fanout entries.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: gstack <ship@gstack.dev>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry Tan
|
22a4451e0e |
feat(v1.3.0.0): open agents learnings + cross-model benchmark skill (#1040)
* chore: regenerate stale ship golden fixtures
Golden fixtures were missing the VENDORED_GSTACK preamble section that
landed on main. Regression tests failed on all three hosts (claude, codex,
factory). Regenerated from current preamble output.
No code changes, unblocks test suite.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat: anti-slop design constraints + delete duplicate constants
Tightens design-consultation and design-shotgun to push back on the
convergence traps every AI design tool falls into.
Changes:
- scripts/resolvers/constants.ts: add "system-ui as primary font" to
AI_SLOP_BLACKLIST. Document Space Grotesk as the new "safe alternative
to Inter" convergence trap alongside the existing overused fonts.
- scripts/gen-skill-docs.ts: delete duplicate AI slop constants block
(dead code — scripts/resolvers/constants.ts is the live source).
Prevents drift between the two definitions.
- design-consultation/SKILL.md.tmpl: add Space Grotesk + system-ui to
overused/slop lists. Add "anti-convergence directive" — vary across
generations in the same project. Add Phase 1 "memorable-thing forcing
question" (what's the one thing someone will remember?). Add Phase 5
"would a human designer be embarrassed by this?" self-gate before
presenting variants.
- design-shotgun/SKILL.md.tmpl: anti-convergence directive — each
variant must use a different font, palette, and layout. If two
variants look like siblings, one of them failed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat: context health soft directive in preamble (T2+)
Adds a "periodically self-summarize" nudge to long-running skills.
Soft directive only — no thresholds, no enforcement, no auto-commit.
Goal: self-awareness during /qa, /investigate, /cso etc. If you notice
yourself going in circles, STOP and reassess instead of thrashing.
Codex review caught that fake precision thresholds (15/30/45 tool calls)
were unimplementable — SKILL.md is a static prompt, not runtime code.
This ships the soft version only.
Changes:
- scripts/resolvers/preamble.ts: add generateContextHealth(), wire into
T2+ tier. Format: [PROGRESS] ... summary line. Explicit rule that
progress reporting must never mutate git state.
- All T2+ skill SKILL.md files regenerated to include the new section.
- Golden ship fixtures updated (T4 skill, picks up the change).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat: model overlays with explicit --model flag (no auto-detect)
Adds a per-model behavioral patch layer orthogonal to the host axis.
Different LLMs have different tendencies (GPT won't stop, Gemini
over-explains, o-series wants structured output). Overlays nudge each
model toward better defaults for gstack workflows.
Codex review caught three landmines the prior reviews missed:
1. Host != model — Claude Code can run any Claude model, Codex runs
GPT/o-series, Cursor fronts multiple providers. Auto-detecting from
host would lie. Dropped auto-detect. --model is explicit (default
claude). Missing overlay file → empty string (graceful).
2. Import cycle — putting Model in resolvers/types.ts would cycle
through hosts/index. Created neutral scripts/models.ts instead.
3. "Final say" is dangerous — overlay at the end of preamble could
override STOP points, AskUserQuestion gates, /ship review gates.
Placed overlay after spawned-session-check but before voice + tier
sections. Wrapper heading adds explicit subordination language on
every overlay: "subordinate to skill workflow, STOP points,
AskUserQuestion gates, plan-mode safety, and /ship review gates."
Changes:
- scripts/models.ts: new neutral module. ALL_MODEL_NAMES, Model type,
resolveModel() for family heuristics (gpt-5.4-mini → gpt-5.4, o3 →
o-series, claude-opus-4-7 → claude), validateModel() helper.
- scripts/resolvers/types.ts: import Model, add ctx.model field.
- scripts/resolvers/model-overlay.ts: new resolver. Reads
model-overlays/{model}.md. Supports {{INHERIT:base}} directive at
top of file for concat (gpt-5.4 inherits gpt). Cycle guard.
- scripts/resolvers/index.ts: register MODEL_OVERLAY resolver.
- scripts/resolvers/preamble.ts: wire generateModelOverlay into
composition before voice. Print MODEL_OVERLAY: {model} in preamble
bash so users can see which overlay is active. Filter empty sections.
- scripts/gen-skill-docs.ts: parse --model CLI flag. Default claude.
Unknown model → throw with list of valid options.
- model-overlays/{claude,gpt,gpt-5.4,gemini,o-series}.md: behavioral
patches per model family. gpt-5.4.md uses {{INHERIT:gpt}} to extend
gpt.md without duplication.
- test/gen-skill-docs.test.ts: fix qa-only guardrail regex scope.
Was matching Edit/Glob/Grep anywhere after `allowed-tools:` in the
whole file. Now scoped to frontmatter only. Body prose (Claude
overlay references Edit as a tool) correctly no longer breaks it.
Verification:
- bun run gen:skill-docs --host all --dry-run → all fresh
- bun run gen:skill-docs --model gpt-5.4 → concat works, gpt.md +
gpt-5.4.md content appears in order
- bun run gen:skill-docs --model unknown → errors with valid list
- All generated skills contain MODEL_OVERLAY: claude in preamble
- Golden ship fixtures regenerated
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat: continuous checkpoint mode with non-destructive WIP squash
Adds opt-in auto-commit during long sessions so work survives Claude
Code crashes, Conductor workspace handoffs, and context switches.
Local-only by default — pushing requires explicit opt-in.
Codex review caught multiple landmines that would have shipped:
1. checkpoint_push=true default would push WIP commits to shared
branches, trigger CI/deploys, expose secrets. Now default false.
2. Plan's original /ship squash (git reset --soft to merge base) was
destructive — uncommitted ALL branch commits, not just WIP, and
caused non-fast-forward pushes. Redesigned: rebase --autosquash
scoped to WIP commits only, with explicit fallback for WIP-only
branches and STOP-and-ask for conflicts.
3. gstack-config get returned empty for missing keys with exit 0,
ignoring the annotated defaults in the header comments. Fixed:
get now falls back to a lookup_default() table that is the
canonical source for defaults.
4. Telemetry default mismatched: header said 'anonymous' but runtime
treated empty as 'off'. Aligned: default is 'off' everywhere.
5. /checkpoint resume only read markdown checkpoint files, not the
WIP commit [gstack-context] bodies the plan referenced. Wired up
parsing of [gstack-context] blocks from WIP commits as a second
recovery trail alongside the markdown checkpoints.
Changes:
- bin/gstack-config: add checkpoint_mode (default explicit) and
checkpoint_push (default false) to CONFIG_HEADER. Add lookup_default()
as canonical default source. get() falls back to defaults when key
absent. list now shows value + source (set/default). New 'defaults'
subcommand to inspect the table.
- scripts/resolvers/preamble.ts: preamble bash reads _CHECKPOINT_MODE
and _CHECKPOINT_PUSH, prints CHECKPOINT_MODE: and CHECKPOINT_PUSH: so
the mode is visible. New generateContinuousCheckpoint() section in
T2+ tier describes WIP commit format with [gstack-context] body and
the rules (never git add -A, never commit broken tests, push only
if opted in). Example deliberately shows a clean-state context so
it doesn't contradict the rules.
- ship/SKILL.md.tmpl: new Step 5.75 WIP Commit Squash. Detects WIP
count, exports [gstack-context] blocks before squash (as backup),
uses rebase --autosquash for mixed branches and soft-reset only when
VERIFIED WIP-only. Explicit anti-footgun rules against blind soft-
reset. Aborts with BLOCKED status on conflict instead of destroying
non-WIP commits.
- checkpoint/SKILL.md.tmpl: new Step 1.5 to parse [gstack-context]
blocks from WIP commits via git log --grep="^WIP:". Merges with
markdown checkpoint for fuller session recovery.
- Golden ship fixtures regenerated (ship is T4, preamble change shows up).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat: feature discovery flow gated by per-feature markers
Extends generateUpgradeCheck() to surface new features once per user
after a just-upgraded session. No more silent features.
Codex review caught: spawned sessions (OpenClaw, etc.) must skip the
discovery prompt entirely — they can't interactively answer. Feature
discovery now checks SPAWNED_SESSION first and is silent in those.
Discovery is per-feature, not per-upgrade. Each feature has its own
marker file at ~/.claude/skills/gstack/.feature-prompted-{name}. Once
the user has been shown a feature (accepted, shown docs, or skipped),
the marker is touched and the prompt never fires again for that
feature. Future features get their own markers.
V1 features surfaced:
- continuous-checkpoint: offer to enable checkpoint_mode=continuous
- model-overlay: inform-only note about --model flag and MODEL_OVERLAY
line in preamble output
Max one prompt per session to avoid nagging. Fires only on JUST_UPGRADED
(not every session), plus spawned-session skip.
Changes:
- scripts/resolvers/preamble.ts: extend generateUpgradeCheck() with
feature discovery rules, per-marker-file semantics, spawned-session
exclusion, and max-one-per-session cap.
- All skill SKILL.md files regenerated to include the new section.
- Golden ship fixtures regenerated.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat: design taste engine with persistent schema
Adds a cross-session taste profile that learns from design-shotgun
approval/rejection decisions. Biases future design-consultation and
design-shotgun proposals toward the user's demonstrated preferences.
Codex review caught that the plan had "taste engine" as a vague goal
without schema, decay, migration, or placeholder insertion points. This
commit ships the full spec.
Schema v1 at ~/.gstack/projects/$SLUG/taste-profile.json:
- version, updated_at
- dimensions: fonts, colors, layouts, aesthetics — each with approved[]
and rejected[] preference lists
- sessions: last 50 (FIFO truncation), each with ts/action/variant/reason
- Preference: { value, confidence, approved_count, rejected_count, last_seen }
- Confidence: Laplace-smoothed approved/(total+1)
- Decay: 5% per week of inactivity, computed at read time (not write)
Changes:
- bin/gstack-taste-update: new CLI. Subcommands approved/rejected/show/
migrate. Parses reason string for dimension signals (e.g.,
"fonts: Geist; colors: slate; aesthetics: minimal"). Emits taste-drift
NOTE when a new signal contradicts a strong opposing signal. Legacy
approved.json aggregates migrate to v1 on next write.
- scripts/resolvers/design.ts: new generateTasteProfile() resolver.
Produces the prose that skills see: how to read the profile, how to
factor into proposals, conflict handling, schema migration.
- scripts/resolvers/index.ts: register TASTE_PROFILE and a BIN_DIR
resolver (returns ctx.paths.binDir, used by templates that shell out
to gstack-* binaries).
- design-consultation/SKILL.md.tmpl: insert {{TASTE_PROFILE}} placeholder
in Phase 1 right after the memorable-thing forcing question so the
Phase 3 proposal can factor in learned preferences.
- design-shotgun/SKILL.md.tmpl: taste memory section now reads
taste-profile.json via {{TASTE_PROFILE}}, falls back to per-session
approved.json (legacy). Approval flow documented to call
gstack-taste-update after user picks/rejects a variant.
Known gap: v1 extracts dimension signals from a reason string passed
by the caller ("fonts: X; colors: Y"). Future v2 can read EXIF or an
accompanying manifest written by design-shotgun alongside each variant
for automatic dimension extraction without needing the reason argument.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat: multi-provider model benchmark (boil the ocean)
Adds the full spec Codex asked for: real provider adapters with auth
detection, normalized RunResult, pricing tables, tool compatibility
maps, parallel execution with error isolation, and table/JSON/markdown
output. Judge stays on Anthropic SDK as the single stable source of
quality scoring, gated behind --judge.
Codex flagged the original plan as massively under-scoped — the
existing runner is Claude-only and the judge is Anthropic-only. You
can't benchmark GPT or Gemini without real provider infrastructure.
This commit ships it.
New architecture:
test/helpers/providers/types.ts ProviderAdapter interface
test/helpers/providers/claude.ts wraps `claude -p --output-format json`
test/helpers/providers/gpt.ts wraps `codex exec --json`
test/helpers/providers/gemini.ts wraps `gemini -p --output-format stream-json --yolo`
test/helpers/pricing.ts per-model USD cost tables (quarterly)
test/helpers/tool-map.ts which tools each CLI exposes
test/helpers/benchmark-runner.ts orchestrator (Promise.allSettled)
test/helpers/benchmark-judge.ts Anthropic SDK quality scorer
bin/gstack-model-benchmark CLI entry
test/benchmark-runner.test.ts 9 unit tests (cost math, formatters, tool-map)
Per-provider error isolation:
- auth → record reason, don't abort batch
- timeout → record reason, don't abort batch
- rate_limit → record reason, don't abort batch
- binary_missing → record in available() check, skip if --skip-unavailable
Pricing correction: cached input tokens are disjoint from uncached
input tokens (Anthropic/OpenAI report them separately). Original
math subtracted them, producing negative costs. Now adds cached at
the 10% discount alongside the full uncached input cost.
CLI:
gstack-model-benchmark --prompt "..." --models claude,gpt,gemini
gstack-model-benchmark ./prompt.txt --output json --judge
gstack-model-benchmark ./prompt.txt --models claude --timeout-ms 60000
Output formats: table (default), json, markdown. Each shows model,
latency, in→out tokens, cost, quality (when --judge used), tool calls,
and any errors.
Known limitations for v1:
- Claude adapter approximates toolCalls as num_turns (stream-json
would give exact counts; v2 can upgrade).
- Live E2E tests (test/providers.e2e.test.ts) not included — they
require CI secrets for all three providers. Unit tests cover the
shape and math.
- Provider CLIs sometimes return non-JSON error text to stdout; the
parsers fall back to treating raw output as plain text in that case.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat: standalone methodology skill publishing via gstack-publish
Ships the marketplace-distribution half of Item 5 (reframed): publish
the existing standalone OpenClaw methodology skills to multiple
marketplaces with one command.
Codex review caught that the original plan assumed raw generated
multi-host skills could be published directly. They can't — those
depend on gstack binaries, generated host paths, tool names, and
telemetry. The correct artifact class is hand-crafted standalone
skills in openclaw/skills/gstack-openclaw-* (already exist and work
without gstack runtime). This commit adds the wrapper that publishes
them to ClawHub + SkillsMP + Vercel Skills.sh with per-marketplace
error isolation and dry-run validation.
Changes:
- skills.json: root manifest with 4 skills (office-hours, ceo-review,
investigate, retro) each pointing at its openclaw/skills source.
Each skill declares per-marketplace targets with a slug, a publish
flag, and a compatible-hosts list. Marketplace configs include CLI
name, login command, publish command template (with placeholder
substitution), docs URL, and auth_check command.
- bin/gstack-publish: new CLI. Subcommands:
gstack-publish Publish all skills
gstack-publish <slug> Publish one skill
gstack-publish --dry-run Validate + auth-check without publishing
gstack-publish --list List skills + marketplace targets
Features:
* Manifest validation (missing source files, missing slugs, empty
marketplace list all reported).
* Per-marketplace auth check before any publish attempt.
* Per-skill / per-marketplace error isolation: one failure doesn't
abort the batch.
* Idempotent — re-running with the same version is safe; markets
that reject duplicate versions report it as a failure for that
single target without affecting others.
* --dry-run walks the full pipeline but skips execSync; useful in
CI to validate manifest before bumping version.
Tested locally: clawhub auth detected, skillsmp/vercel CLIs not
installed (marked NOT READY and skipped cleanly in dry-run).
Follow-up work (tracked in TODOS.md later):
- Version-bump helper that reads openclaw/skills/*/SKILL.md frontmatter
and updates skills.json in lockstep.
- CI workflow that runs gstack-publish --dry-run on every PR and
gstack-publish on tags.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor: split preamble.ts into submodules (byte-identical output)
Splits scripts/resolvers/preamble.ts (841 lines, 18 generator functions +
composition root) into one file per generator under
scripts/resolvers/preamble/. Root preamble.ts becomes a thin composition
layer (~80 lines of imports + generatePreamble).
Before:
scripts/resolvers/preamble.ts 841 lines
After:
scripts/resolvers/preamble.ts 83 lines
scripts/resolvers/preamble/generate-preamble-bash.ts 97 lines
scripts/resolvers/preamble/generate-upgrade-check.ts 48 lines
scripts/resolvers/preamble/generate-lake-intro.ts 16 lines
scripts/resolvers/preamble/generate-telemetry-prompt.ts 37 lines
scripts/resolvers/preamble/generate-proactive-prompt.ts 25 lines
scripts/resolvers/preamble/generate-routing-injection.ts 49 lines
scripts/resolvers/preamble/generate-vendoring-deprecation.ts 36 lines
scripts/resolvers/preamble/generate-spawned-session-check.ts 11 lines
scripts/resolvers/preamble/generate-ask-user-format.ts 16 lines
scripts/resolvers/preamble/generate-completeness-section.ts 19 lines
scripts/resolvers/preamble/generate-repo-mode-section.ts 12 lines
scripts/resolvers/preamble/generate-test-failure-triage.ts 108 lines
scripts/resolvers/preamble/generate-search-before-building.ts 14 lines
scripts/resolvers/preamble/generate-completion-status.ts 161 lines
scripts/resolvers/preamble/generate-voice-directive.ts 60 lines
scripts/resolvers/preamble/generate-context-recovery.ts 51 lines
scripts/resolvers/preamble/generate-continuous-checkpoint.ts 48 lines
scripts/resolvers/preamble/generate-context-health.ts 31 lines
Byte-identity verification (the real gate per Codex correction):
- Before refactor: snapshotted 135 generated SKILL.md files via
`find -name SKILL.md -type f | grep -v /gstack/` across all hosts.
- After refactor: regenerated with `bun run gen:skill-docs --host all`
and re-snapshotted.
- `diff -r baseline after` returned zero differences and exit 0.
The `--host all --dry-run` gate passes too. No template or host behavior
changes — purely a code-organization refactor.
Test fix: audit-compliance.test.ts's telemetry check previously grepped
preamble.ts directly for `_TEL != "off"`. After the refactor that logic
lives in preamble/generate-preamble-bash.ts. Test now concatenates all
preamble submodule sources before asserting — tracks the semantic contract,
not the file layout. Doing the minimum rewrite preserves the test's intent
(conditional telemetry) without coupling it to file boundaries.
Why now: we were in-session with full context. Codex had downgraded this
from mandatory to optional, but the preamble had grown to 841 lines and
was getting harder to navigate. User asked "why not?" given the context
was hot. Shipping it as a clean bisectable commit while all the prior
preamble.ts changes are fresh reduces rebase pain later.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore: bump version and changelog (v0.19.0.0)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore: trim verbose preamble + coverage audit prose
Compress without removing behavior or voice. Three targeted cuts:
1. scripts/resolvers/testing.ts coverage diagram example: 40 lines → 14
lines. Two-column ASCII layout instead of stacked sections.
Preserves all required regression-guard phrases (processPayment,
refundPayment, billing.test.ts, checkout.e2e.ts, COVERAGE, QUALITY,
GAPS, Code paths, User flows, ASCII coverage diagram).
2. scripts/resolvers/preamble/generate-completion-status.ts Plan Status
Footer: was 35 lines with embedded markdown table example, now 7
lines that describe the table inline. The footer fires only at
ExitPlanMode time — Claude can construct the placeholder table from
the inline description without copying a literal example.
3. Same file's Plan Mode Safe Operations + Skill Invocation During Plan
Mode sections compressed from ~25 lines combined to ~12. Preserves
all required test phrases (precedence over generic plan mode behavior,
Do not continue the workflow, cancel the skill or leave plan mode,
PLAN MODE EXCEPTION).
NOT touched:
- Voice directive (Garry's voice — protected per CLAUDE.md)
- Office-hours Phase 6 Handoff (Garry's voice + YC pitch)
- Test bootstrap, review army, plan completion (carefully tuned behavior)
Token savings (per skill, system-wide):
ship/SKILL.md 35474 → 34992 tokens (-482)
plan-ceo-review 29436 → 28940 (-496)
office-hours 26700 → 26204 (-496)
Still over the 25K ceiling. Bigger reduction requires restructure
(move large resolvers to externally-referenced docs, split /ship into
ship-quick + ship-full, or refactor the coverage audit + review army
into shorter prose). That's a follow-up — added to TODOS.
Tests: 420/420 pass on gen-skill-docs.test.ts + host-config.test.ts.
Goldens regenerated for claude/codex/factory ship.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ci): install Node.js from official tarball instead of NodeSource apt setup
The CI Dockerfile's Node install was failing on ubicloud runners. NodeSource's
setup_22.x script runs two internal apt operations that both depend on
archive.ubuntu.com + security.ubuntu.com being reachable:
1. apt-get update (to refresh package lists)
2. apt-get install gnupg (as a prerequisite for its gpg keyring)
Ubicloud's CI runners frequently can't reach those mirrors — last build hit
~2min of connection timeouts to every security.ubuntu.com IP (185.125.190.82,
91.189.91.83, 91.189.92.24, etc.) plus archive.ubuntu.com mirrors. Compounding
this: on Ubuntu 24.04 (noble) "gnupg" was renamed to "gpg" and "gpgconf".
NodeSource's setup script still looks for "gnupg", so even when apt works,
it fails with "Package 'gnupg' has no installation candidate." The subsequent
apt-get install nodejs then fails because the NodeSource repo was never added.
Fix: drop NodeSource entirely. Download Node.js v22.20.0 from nodejs.org as a
tarball, extract to /usr/local. One host, no apt, no script, no keyring.
Before:
RUN curl -fsSL https://deb.nodesource.com/setup_22.x | bash - \
&& apt-get install -y --no-install-recommends nodejs ...
After:
ENV NODE_VERSION=22.20.0
RUN curl -fsSL "https://nodejs.org/dist/v${NODE_VERSION}/node-v${NODE_VERSION}-linux-x64.tar.xz" -o /tmp/node.tar.xz \
&& tar -xJ -C /usr/local --strip-components=1 --no-same-owner -f /tmp/node.tar.xz \
&& rm -f /tmp/node.tar.xz \
&& node --version && npm --version
Same installed path (/usr/local/bin/node and npm). Pinned version for
reproducibility. Version is bump-visible in the Dockerfile now.
Does not address the separate apt flakiness that affects the GitHub CLI
install (line 17) or `npx playwright install-deps chromium` (line 33) —
those use apt too. If those fail on a future build we can address then.
Failing job: build-image (71777913820)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore: raise skill token ceiling warning from 25K to 40K
The 25K ceiling predated flagship models with 200K-1M windows and assumed
every skill prompt dominates context cost. Modern reality: prompt caching
amortizes the skill load across invocations, and three carefully-tuned
skills (ship, plan-ceo-review, office-hours) legitimately pack 25-35K
tokens of behavior that can't be cut without degrading quality or removing
protected content (Garry's voice, YC pitch, specialist review instructions).
We made the safe prose cuts earlier (coverage diagram, plan status footer,
plan mode operations). The remaining gap is structural — real compression
would require splitting /ship into ship-quick vs ship-full, externalizing
large resolvers to reference docs, or removing detailed skill behavior.
Each is 1-2 days of work. The cost of the warning firing is zero (it's
a warning, not an error). The cost of hitting it is ~15¢ per invocation
at worst, amortized further by prompt caching.
Raising to 40K catches what it's supposed to catch — a runaway 10K+ token
growth in a single release — without crying wolf on legitimately big
skills. Reference doc in CLAUDE.md updated to reflect the new philosophy:
when you hit 40K, ask WHAT grew, don't blindly compress tuned prose.
scripts/gen-skill-docs.ts: TOKEN_CEILING_BYTES 100_000 → 160_000.
CLAUDE.md: document the "watch for feature bloat, not force compression"
intent of the ceiling.
Verification: `bun run gen:skill-docs --host all` shows zero TOKEN
CEILING warnings under the new 40K threshold.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ci): install xz-utils so Node tarball extraction works
The direct-tarball Node install (switched from NodeSource apt in the last
CI fix) failed with "xz: Cannot exec: No such file or directory" because
Ubuntu 24.04 base doesn't include xz-utils. Node ships .tar.xz by default,
and `tar -xJ` shells out to xz, which was missing.
Add xz-utils to the base apt install alongside git/curl/unzip/etc.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(benchmark): pass --skip-git-repo-check to codex adapter
The gpt provider adapter spawns `codex exec -C <workdir>` with arbitrary
working directories (benchmark temp dirs, non-git paths). Without
`--skip-git-repo-check`, codex refuses to run and returns "Not inside a
trusted directory" — surfaced as a generic error.code='unknown' that
looks like an API failure.
Benchmarks don't care about codex's git-repo trust model; we just want
the prompt executed. Surfaced by the new provider live E2E test on a
temp workdir.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(benchmark): add --dry-run flag to gstack-model-benchmark
Matches gstack-publish --dry-run semantics. Validates the provider list,
resolves per-adapter auth, echoes the resolved flag values, and exits
without invoking any provider CLI. Zero-cost pre-flight for CI pipelines
and for catching auth drift before starting a paid benchmark run.
Output shape:
== gstack-model-benchmark --dry-run ==
prompt: <truncated>
providers: claude, gpt, gemini
workdir: /tmp/...
timeout_ms: 300000
output: table
judge: off
Adapter availability:
claude: OK
gpt: NOT READY — <reason>
gemini: NOT READY — <reason>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test: lite E2E coverage for benchmark, taste engine, publish
Fills real coverage gaps in v0.19.0.0 primitives. 44 new deterministic
tests (gate tier, ~3s) + 8 live-API tests (periodic tier).
New gate-tier test files (free, <3s total):
- test/taste-engine.test.ts — 24 tests against gstack-taste-update:
schema shape, Laplace-smoothed confidence, 5%/week decay clamped at 0,
multi-dimension extraction, case-insensitive matching, session cap,
legacy profile migration with session truncation, taste-drift conflict
warning, malformed-JSON recovery, missing-variant exit code.
- test/publish-dry-run.test.ts — 13 tests against gstack-publish --dry-run:
manifest parsing, missing/malformed JSON, per-skill validation errors
(missing source file / slug / version / marketplaces), slug filter,
unknown-skill exit, per-marketplace auth isolation (fake marketplaces
with always-pass / always-fail / missing-binary CLIs), and a sanity
check against the real repo manifest.
- test/benchmark-cli.test.ts — 11 tests against gstack-model-benchmark
--dry-run: provider default, unknown-provider WARN, empty list
fallback, flag passthrough (timeout/workdir/judge/output), long-prompt
truncation, prompt resolution (inline vs file vs positional), missing
prompt exit.
New periodic-tier test file (paid, gated EVALS=1):
- test/skill-e2e-benchmark-providers.test.ts — 8 tests hitting real
claude, codex, gemini CLIs with a trivial prompt (~$0.001/provider).
Verifies output parsing, token accounting, cost estimation, timeout
error.code semantics, Promise.allSettled parallel isolation.
Per-provider availability gate — unauthed providers skip cleanly.
This suite already caught one real bug (codex adapter missing
--skip-git-repo-check, fixed in
|