Skill Refiner — intent-eval-lab coordination (specs, glossary, DRs, blueprints)

[jeremylongshore]

Skill Refiner — intent-eval-lab coordination

Per-repo coordination epic for the Skill Refiner buildout. Tracks every change this repo absorbs from plan 027-PP-PLAN-skill-refiner-snoopy-fluttering-comet-v4-2026-05-26.

Scope

• Blueprint B § 7 extension — skill-refiner-pass/v1 normative spec
• Canonical glossary (014-DR-GLOS) update — SkillVersion + Skill Refiner + 3-layer hooks (sinker/line/hook) + ScoreRecord + EditProposal + agentskills.io spec-pin
• New normative spec doc: 028-AT-SPEC-skill-refiner-pass-v1-predicate.md
• New architecture doc: 029-AT-ARCH-skill-refiner-3-layer-hooks.md (with D1/D3/D6 ASCII diagrams)
• New standard doc: 030-AT-STND-tri-linkage-discipline.md (D7 diagram)
• ISEDC Session 7 DR — ratifies URI namespace addition
• Evidence Report SPEC.md template (v1.0.0-draft)
• Hugo HTML renderer template

Architectural commitments honored

AC-2, AC-3, AC-8 (3-layer hook architecture), AC-11 (agentskills.io spec compliance), AC-12 (tri-linkage discipline)

Children (Beads)

• Parent coord epic: bd_000-projects-rqwk
• Leaf tasks: bd_000-projects-rqwk.1, bd_000-projects-rqwk.2, bd_000-projects-rqwk.3
• Diagram beads: bd_000-projects-rqwk.4, bd_000-projects-rqwk.5, bd_000-projects-rqwk.6, bd_000-projects-rqwk.7, bd_000-projects-pu35.6

Reference

Plan 027 lives in this repo at 000-docs/027-PP-PLAN-skill-refiner-snoopy-fluttering-comet-v4-2026-05-26.md.

Spec snapshots from Step 0 (2026-05-27): research/agentskills-spec-v1.0.0.md, research/claude-docs-spec-tree-2026-05-27.md, research/2605.23904-skillopt-digest.md.

---

Bead: bd_000-projects-rqwk
Doc: 000-docs/027-PP-PLAN-skill-refiner-snoopy-fluttering-comet-v4-2026-05-26.md

— Jeremy Longshore
intentsolutions.io

[jeremylongshore]

Linked bead: bd_000-projects-rqwk (source of truth). Use bd-sync to mirror updates.

[jeremylongshore]

bd-sync — cluster tracking (IEP reconcile 2026-06-10): the following 26 bead(s) are now cross-referenced to this umbrella issue (source of truth is bd; bd-sync mirrors state changes here):

• bd_000-projects-rqwk.1 — Author intent-eval-lab/000-docs/028-AT-SPEC-skill-refiner-pass-v1-predicate.md (normative …

• bd_000-projects-rqwk.3 — Update 014-DR-GLOS-canonical-glossary.md with Skill Refiner brand canon entries

• bd_000-projects-rqwk.2 — Author intent-eval-lab/000-docs/029-AT-ARCH-skill-refiner-3-layer-hooks.md (architecture +…

• bd_000-projects-rqwk.7 — Embed D6 — bead dependency graph in iel/029-AT-ARCH Appendix A

• bd_000-projects-rqwk.6 — Embed D3 — 3-layer hook architecture diagram in iel/029-AT-ARCH § 3

• bd_000-projects-rqwk.5 — Embed D2 — 5-repo IEP ecosystem diagram in iel/011-AT-ARCH § 12

• bd_000-projects-rqwk.4 — Embed D1 — 3-product agent-rig stack diagram in iel/029-AT-ARCH § 1

• bd_000-projects-0r8m.3 — Design skill-refiner-pass/v1 predicate schema (intent-eval-core/schemas/v1/skill-refiner-p…

• bd_000-projects-0r8m.2 — Implement SkillVersion Zod validator (src/validators/v1/skill-version.ts) via pnpm run cod…

• bd_000-projects-0r8m.1 — Design SkillVersion entity schema at intent-eval-core/schemas/v1/skill-version.schema.json…

• bd_000-projects-jsy3.6 — Pick Phase B target skill (/audit-tests vs /validate-skillmd — open question from plan § 7…

• bd_000-projects-jsy3.9 — Publish /j-rig plugin to claude-code-plugins marketplace (after MVP demo evidence publishe…

• bd_000-projects-jsy3.8 — Run /j-rig MVP demo end-to-end on chosen Phase B target skill + produce j-rig Evidence rep…

• bd_000-projects-jsy3.5 — Wire /j-rig plugin subcommands (refine bootstrap/propose/shadow/promote/status) — thin wra…

• bd_000-projects-jsy3.4 — Implement Hook hook (L3) — PostToolUse:Bash on git commit/push matcher runs agentic gate a…

• bd_000-projects-jsy3.3 — Implement Line hook (L2) — Stop hook captures end-of-turn rollouts; fires Refiner in backg…

• bd_000-projects-jsy3.2 — Implement Sinker hook (L1) — PostToolUse:Edit/Write on SKILL.md files runs validate-skillm…

• bd_000-projects-jsy3.1 — Scaffold /j-rig plugin in claude-code-plugins (plugin.json + .claude-plugin/hooks/hooks.js…

• bd_000-projects-3zol.8 — Release ceremony: ship @j-rig/refiner@0.1.0 to npm with sigstore provenance + AAR doc

• bd_000-projects-3zol.7 — Build Skill Refiner CLI (5 commands: bootstrap, score, propose, apply, status) in @j-rig/r…

• bd_000-projects-3zol.6 — Implement append-only event log (.j-rig/refiner/log.jsonl) + content-addressed store (.j-r…

• bd_000-projects-3zol.5 — Implement Skill Refiner apply() + accept() — pure SkillDoc→SkillDocV2 transformation and s…

• bd_000-projects-3zol.4 — Implement Skill Refiner propose() — wrap j-rig optimizer ChangeProposal with content-addre…

• bd_000-projects-3zol.3 — Implement Skill Refiner score() — delegate to J-Rig Binary Eval with tiered model routing …

• bd_000-projects-3zol.2 — Implement Skill Refiner bootstrap() — synthesize eval set from SKILL.md frontmatter + desc…

• bd_000-projects-3zol.1 — Scaffold @j-rig/refiner package + value types module (SkillDocHash, EvalSetHash, ScoreReco…

• Jeremy Longshore
  intentsolutions.io

[jeremylongshore]

bd-sync from bd_000-projects-0r8m.3: skill-refiner-pass/v1 predicate URI MINTED via ISEDC Class-1 ADR (DR intent-eval-lab/000-docs/082-AT-DECR-isedc-skill-refiner-pass-v1-predicate-uri-2026-06-17.md; PR #170). The SAK-charter-Q3 'no predicate URI without a separate Class-1 ADR' gate is now SATISFIED for this predicate. DECISIONS: (Q1) FLAT https://evals.intentsolutions.io/skill-refiner-pass/v1 (NO /authoring/ segment; chamber isolation via signing root + schema lane + file tree). (Q2) determinant-only signed in-toto body: verdict+reason[]+refiner_strategy_id+{skill_version_id,parent_version_id,source_snapshot_hash}+eval_set_ref+edit_proposal_hash+behavioral_delta+named_dim_deltas[]+alpha+test_statistic_kind; provenance bulk (transcripts/prose/dashboard) -> EvidenceBundle by hash. (Q3) STAGING-FIRST; production-signable ONLY when ALL 4 AND-gated triggers hold: SPEC.md normative section lands + DNSSEC/CAA green on evals. + SEPARATE authoring signing root live (DR-081 no-shared-root) + >=1 real SkillVersion clears the gate on a frozen signed eval-set. (Q4) mirror gate-result/v1 in-toto exactly. (Q5) v1 frozen; additive/loosening->Class-1 ISEDC; breaking->skill-refiner-pass/v2 (inherits DR-064). STILL GATED FOR BUILD: kernel schema + src/predicates/skill-refiner-pass-v1.ts is authoring-chamber kernel work (bandwidth-gated; SkillVersion entity is a DR-028 one-way-door); does NOT ship until SPEC.md lands; NO production signature until the 4 triggers.

[jeremylongshore]

bd-sync from bd_000-projects-0r8m.3: skill-refiner-pass/v1 kernel SCHEMA SHIPPED (staging, signing_mode=ln): PR jeremylongshore/intent-eval-core#56 squash-merged into main. Footprint: schemas/v1/skill-refiner-pass.schema.json + src/predicates/skill-refiner-pass-v1.ts + Zod validator + Python Pydantic model + PREDICATE_URIS registry + schemas/v1/index.json + 002-doc predicate-table row + 32 predicate tests; 1220 kernel tests green, all required CI passed. Reviewed against DR-082 line-by-line: FLAT URI (no /authoring/), determinant-only body, additionalProperties:false, in-toto body-only validation (digest=subject), NO SkillVersion entity (typed-id refs only). NOT production-signable — the 4 DR-082 Q3 triggers remain pending (SPEC.md normative section, DNSSEC/CAA, the SEPARATE authoring signing root, a real SkillVersion clearing the gate). One housekeeping follow-on: re-pin json-schema-to-zod (the _generated/ reference dir drifts on the local tool version; reference-only, not CI-verified).

[jeremylongshore]

bd-sync close on bd_000-projects-rqwk.1: skill-refiner-pass/v1 normative predicate spec authored + merged — intent-eval-lab PR #171 landed it as doc 083-AT-SPEC-skill-refiner-pass-v1-normative-spec (planned number 028 drifted to 083 between 2026-05-27 planning and 2026-06-17 authoring) + companion 084-AT-SPEC signing-trust-root design. DR-086-aligned (staging-first, freeze=first-production-Rekor), required lint lanes green, normative content verified byte-unchanged on the lint-fix pass. (GH issue left open — covers other beads)

[jeremylongshore]

bd-sync close on bd_000-projects-rqwk.11: Blueprint A Phase-D anti-goal amendment shipped — lab PR #192 merged: added § 3.8 'NOT a creator product for SKILL.md generation (Phase D anti-goal)' to 011-AT-ARCH (the constitution), executing ratified DR-028 T2. Records the binding-but-falsifiable anti-goal + the verbatim Karpathy+Gregg re-open trigger (self-gen lift > +8pp on kernel-pinned eval set on any frontier release OR internal acceptance-rate parity), grounded in SkillsBench evidence, DR-028 cited as ratifying authority. Conservative reading; one file; all required checks green. (GH issue left open — covers other beads)

[jeremylongshore]

bd-sync close on bd_000-projects-3zol.2: Skill Refiner bootstrap() DONE (audit-verified): packages/refiner-core/src/bootstrap.ts — deterministic EvalSet synthesis from SKILL.md frontmatter+description; UUIDv7 lineageId + semver evalSetVersion + refreshDueAt. (GH issue left open — covers other beads)

[jeremylongshore]

bd-sync close on bd_000-projects-3zol.3: Skill Refiner score() DONE (audit-verified): packages/refiner/src/score.ts — delegates to j-rig eval via injectable EvalRunner; pass_rate->ScoreRecord behavioral dim; AC-5 tiered routing (haiku|sonnet, opus unrepresentable at this entry point). (GH issue left open — covers other beads)

[jeremylongshore]

bd-sync close on bd_000-projects-3zol.4: Skill Refiner propose() DONE (audit-verified): packages/refiner/src/propose.ts — wraps refiner-core RefinerStrategy; AnthropicCompletionClient (SDK-free Messages API); opus-lock type+runtime guard; content-addressed EditProposal (refinerModel+refinerStrategyId). (GH issue left open — covers other beads)

[jeremylongshore]

bd-sync close on bd_000-projects-3zol.5: Skill Refiner apply()+accept() DONE (audit-verified): refiner-core/src/apply.ts (pure anchored SkillDoc->SkillDoc, parent-hash match, EditApplicationError) + accept.ts (DR-028 P0-RATIFY-1 Pareto strict-improvement gate, non-regression on named dims, α=0.05 z-test). (GH issue left open — covers other beads)

[jeremylongshore]

bd-sync close on bd_000-projects-3zol.6: event log + content store + best-pointer DONE (audit-verified): packages/refiner/src/store.ts RefinerStore — content-addressed .j-rig/refiner/store/, append-only log.jsonl (RefinerEvent union), mutable best-pointer; injected fs seam; idempotent. (GH issue left open — covers other beads)

[jeremylongshore]

bd-sync close on bd_000-projects-3zol.7: Skill Refiner CLI 5 commands DONE (audit-verified): packages/refiner/src/cli.ts registerRefineCommand() — bootstrap/score/propose/apply/status; offline commands work fully, key-gated commands fail loudly with guidance (no fabrication). (GH issue left open — covers other beads)

[jeremylongshore]

bd-sync close on bd_000-projects-rqwk.2: Skill Refiner 3-layer hook architecture doc shipped — lab PR #194 merged: doc 096-AT-ARCH documents sinker(L1=PostToolUse:Edit/Write, deterministic /validate-skillmd, never blocks) / line(L2=Stop, captures rollouts + fires refine loop, never blocks) / hook(L3=PreToolUse:Bash on git commit|push, Opus regression gate, the ONLY blocking layer, emits signed skill-refiner-pass/v1 on clean pass). Confirms L3=PreToolUse:Bash (the v4.1 mechanism fix) + flags stale PostToolUse label in plan 027 §6.5. 2 ASCII diagrams; all 9 required checks green. (GH issue left open — covers other beads)

[jeremylongshore]

bd-sync close on bd_000-projects-3zol.8: Skill Refiner release ceremony COMPLETE — @intentsolutions/refiner-core@0.1.0 + @intentsolutions/refiner@0.1.0 PUBLISHED to npm with SLSA provenance v1 (keyless sigstore via GitHub Actions OIDC), via .github/workflows/publish-refiner.yml run 27916699310 (tag refiner-v0.1.0; refiner-core then refiner, each pnpm publish --provenance; post-publish workspace-dep-rewrite guard passed). Scope @intentsolutions/* per CEO directive 2026-06-21 (existing NPM_TOKEN owns it, consistent with @intentsolutions/core); @j-rig stays the brand + CLI identity (j-rig refine). AAR finalized doc 027-AA-AACR-refiner-v0.1.0-release-2026-06-21 (PR #151). The Skill Refiner product is publicly launched. (GH issue left open — covers other beads)