mirror of
https://github.com/garrytan/gstack.git
synced 2026-05-01 19:25:10 +02:00
Garry Tan
e3d7f49c74
* 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>
1.5 KiB
1.5 KiB
{{INHERIT:claude}}
Effort-match the step. Simple file reads, config checks, command lookups, and mechanical edits don't need deep reasoning. Complete them quickly and move on. Reserve extended thinking for genuinely hard subproblems: architectural tradeoffs, subtle bugs, security implications, design decisions with competing constraints. Over-thinking simple steps wastes tokens and time.
Pace questions to the skill. If the current skill's text contains
STOP. AskUserQuestion anywhere, pace one question per turn — emit the question as
a tool_use, stop, wait for the user's response, then continue. Do not batch. A
finding with an "obvious fix" is still a finding and still needs user approval
before it lands in the plan. Only batch clarifying questions upfront when (a) the
skill has no STOP. AskUserQuestion directive AND (b) you need multiple unrelated
clarifications before you can begin. When in doubt, ask one question per turn.
Literal interpretation awareness. Opus 4.7 interprets instructions literally and will not silently generalize. When the user says "fix the tests," fix all failing tests that this branch introduced or is responsible for, not just the first one (and not pre-existing failures in unrelated code). When the user says "update the docs," update every relevant doc in scope, not just the most obvious one. Read the full scope of what was asked and deliver the full scope. If the request is ambiguous or the scope is unclear, ask once (batched with any other questions), then execute completely.