mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-29 08:13:12 +00:00
f5c32dff985b7984020799c8ffd741e3ceaedb95
1
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
7be17261bc |
v0.32.3.0 skill: functional-area-resolver — pattern for compressing routing tables (#859)
* skill: compress-agents-md — functional-area resolver pattern Proven via A/B eval: 100% routing accuracy at 48% size reduction. Converts granular per-skill resolver rows into functional-area dispatchers with '(dispatcher for: ...)' sub-skill lists. Includes: - SKILL.md with full pattern docs, before/after examples, eval results - routing-eval.jsonl with 5 fixtures - Anti-patterns (resolver-of-resolvers pipe table = 15% accuracy) * skill: rename compress-agents-md → functional-area-resolver, cite prior art The contribution is a pattern (functional-area dispatcher with `(dispatcher for: ...)` clauses), not a file. Rename describes the contribution; triggers broaden to cover both AGENTS.md and RESOLVER.md phrasings. SKILL.md rewrite: - Three-model A/B table (Opus 4.7 / Sonnet 4.6 / Haiku 4.5) replaces the original Sonnet-only claim. Functional-areas beats baseline by +13 to +17pp training (lenient) across all three models at 48% the size. - Strict + lenient scoring documented side by side. Lenient (predicted shares dispatcher area with expected) matches production agent behavior. - Preconditions added: refuse to compress if file <12KB or working tree dirty. - Multi-file routing precedence section for the v0.31.7 RESOLVER.md/AGENTS.md merge case. - Mandatory verification step (≥95% via the harness). - Daily-doctor.mjs reference scrubbed (didn't exist in gbrain). - Three prior-art citations: AnyTool (arXiv:2402.04253), RAG-MCP (arXiv:2505.03275), Anthropic Agent Skills progressive disclosure. The pattern is the static-prompt analog of runtime hierarchical routing. routing-eval.jsonl: 8 positive (5 original + 3 broadened triggers) + 4 adversarial negatives targeting skillify, skill-creator, book-mirror, concept-synthesis to prove broadened triggers don't over-capture adjacent meta-skills. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * evals: A/B harness for functional-area-resolver (gateway-routed, strict + lenient scoring) evals/functional-area-resolver/ lives outside skills/ deliberately. The skillpack bundler walks skills/<skill>/ recursively, so an eval surface in there would copy harness + variants + fixtures + tests into every downstream install. The pattern (in SKILL.md) ships everywhere; the eval evidence stays in the gbrain repo. What ships: - Three variant resolvers in variants/ — baseline.md (verbose 25KB) and functional-areas.md (compressed 13KB) extracted from a real production AGENTS.md at git commits 93848ff3b^ and 93848ff3b (owner PII scrubbed). resolver-of-resolvers.md derived mechanically by stripping (dispatcher for: ...) clauses — the ablation case. - 20 hand-authored training fixtures + 5 held-out blind fixtures. - harness-runner.ts — TypeScript runner via gbrain gateway. Flags: --model {opus|sonnet|haiku|<full-id>}, --variants-dir, --variants for description-length sweeps, --parallel N (rate-lease bound), --limit N for smoke runs, --yes for non-TTY. - Every output row carries BOTH `correct` (strict) and `correct_lenient` (predicted shares dispatcher area with expected). Lenient matches production behavior. - Receipt header binds (model, prompt_template_hash, fixtures_hash, harness_sha, ts, cmd_args). Re-runs are auditable. - harness.mjs — thin Node shim that spawns the TS runner via bun. - rescore.mjs — zero-cost lenient re-score of an existing JSONL. - harness-runner.test.ts — 45 unit tests (no API key needed) covering every pure function plus the dispatcher-list parser. The prompt template is load-bearing: without the "drill into (dispatcher for: ...) list" instruction, every compression variant collapses to ~30-60%. Documented in SKILL.md and README.md. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * evals: baseline receipts (Opus 4.7 + Sonnet 4.6 + Haiku 4.5, 2026-05-11) Three canonical 225-row receipts (3 variants × 25 fixtures × 3 seeds per model). Each receipt header binds (model, prompt_template_hash, fixtures_hash, harness_sha, ts) so the published SKILL.md numbers are reproducible. Training corpus (n=20, lenient): baseline | Opus 81.7% | Sonnet 86.7% | Haiku 73.3% | 25KB functional-areas | Opus 98.3% | Sonnet 100% | Haiku 88.3% | 13KB resolver-of-resolvers | Opus 63.3% | Sonnet 41.7% | Haiku 65.0% | 10KB functional-areas beats baseline by +13 to +17pp across all three models at 48% the size. resolver-of-resolvers' Sonnet collapse (41.7%) is the SKILL.md "compression without dispatcher clause is broken" claim, observed. Held-out (n=5, lenient) saturates at 100% across most cells (Sonnet × resolver-of-resolvers is 73.3% — the same failure mode visible on a smaller sample). ~$3 API spend across all three runs. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * skill: wire functional-area-resolver into RESOLVER.md + manifests skills/RESOLVER.md gets a new row in Operational, adjacent to skillify. Triggers: "Compress my resolver", "AGENTS.md too large", "RESOLVER.md too big", "functional area dispatcher", "shrink routing table". skills/manifest.json adds the new entry and bumps manifest version 0.25.1 → 0.32.3.0 (loadOrDeriveManifest reads this for sync-guard). openclaw.plugin.json adds functional-area-resolver to the skills array and bumps version 0.25.1 → 0.32.3.0 so install receipts stop being stale (src/core/skillpack/installer.ts:307-311 uses manifest version on every install). Verified: - gbrain check-resolvable --json: 42/42 reachable, 0 errors. - gbrain routing-eval: 70/70 pass (100% structural). - bun test test/skillpack-sync-guard.test.ts: passes (manifest in sync). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * v0.32.3.0 skill: functional-area-resolver — pattern for compressing routing tables Headline: compress a 25KB AGENTS.md down to 13KB without losing routing accuracy. Pattern proven across Opus 4.7, Sonnet 4.6, and Haiku 4.5 — beats the verbose baseline by +13 to +17pp at 48% the size. Empirical (training, n=20, 3 seeds, lenient): baseline 25KB: Opus 81.7% | Sonnet 86.7% | Haiku 73.3% functional-areas 13KB: Opus 98.3% | Sonnet 100% | Haiku 88.3% resolver-of-resolvers 10KB: Opus 63.3% | Sonnet 41.7% | Haiku 65.0% The (dispatcher for: ...) clause is the load-bearing signal. Strip it (the resolver-of-resolvers variant) and Sonnet collapses to 41.7% — the failure case the pattern's authors predicted, now observed. Files in this release: - VERSION + package.json bumped to 0.32.3.0 (4-segment per CLAUDE.md). - CHANGELOG.md: full empirical story, cross-model table, three prior-art citations (AnyTool, RAG-MCP, Anthropic Agent Skills progressive disclosure). - TODOS.md: nine v0.33.x follow-ups (dogfood on gbrain's own RESOLVER.md, CLI promotion to gbrain routing-eval --ab-compare, held-out corpus growth, cross-vendor Gemini+GPT verification, per-row description length sweep, structural compression to ~10KB, hierarchical area-of-areas, embedding pre-router, adversarial fixtures, prompt-design ablation doc). - llms-full.txt regenerated. Bisect-friendly history on this branch: |