brooks-harness

brooks-lint — Maintenance Harness (Orchestrator)

This skill orchestrates work on the brooks-lint repo itself. It runs a sequential subagent pipeline: each stage is a dedicated agent defined in .claude/agents/. Spawn each with the Agent tool, subagent_type set to the agent name, and always model: "opus". Stages depend on each other in order, so this is a pipeline, not a parallel team.

Pipeline

[orchestrator]
   Phase 0  context check
   Phase 1  classify request → select stages
   Phase 2  run selected stages in order, with a QA loop-back:
            skill-author → eval-curator → consistency-qa ─(FAIL)→ back to author
                                              │ PASS
                                              ▼
                            trigger-boundary-auditor   (only if a description changed)
                                              ▼
                                      release-manager  (only if release requested)
   Phase 3  report + collect feedback

Phase 0 — Context check

Determine the run mode before doing anything:

• _workspace/brooks-harness/ exists + maintainer asks to redo part of a prior run → partial re-run: invoke only the affected stage(s), reusing prior notes.
• _workspace/brooks-harness/ exists + a fresh request → new run: move the old folder to _workspace/brooks-harness_prev/, start clean.
• No _workspace/brooks-harness/ → initial run: create it.

Run notes and the QA report live under _workspace/brooks-harness/. The real artifacts are the repo files themselves — agents edit skills/, evals/, manifests directly; _workspace/ only holds the run's notes and the PASS/FAIL verdict for audit.

Phase 1 — Classify the request

Pick the minimal set of stages. The QA stage is never skipped — every change is gated.

Phase 2 — Run the pipeline

Spawn each selected stage as a subagent in order. Pass each agent (a) the task contract and (b) the previous stage's summary. Agents write their summaries to _workspace/brooks-harness/; read them between stages.

1. skill-author — creates/edits the content. For a brand-new skill it invokes the new-skill scaffold. Returns the list of files touched + convention-relevant choices (new risk codes, new Step numbers, changed description trigger phrases).
2. eval-curator — if skill-author reported new/changed risk codes or modes, adds the paired happy-path + false-positive scenarios and runs npm run evals.
3. consistency-qa (gate — never skipped) — runs npm run validate + npm test + npm run evals, then the cross-document sync checks (manifests, all version-bearing text files, CHANGELOG, AGENTS/GEMINI book count, eval count). Writes a PASS/FAIL verdict. On FAIL: loop back to the agent named in the verdict (author or eval-curator), fix, then re-run QA. Repeat once; if it still fails, stop and report to the maintainer.
4. trigger-boundary-auditor — run only if a description field changed. It read-only audits the six shipped skills' trigger surfaces for false-triggering and routing collisions. Surface its findings; if it flags a real collision, loop back to skill-author.
5. release-manager — run only if a release was requested, and only after QA PASS. Cuts the release via the release skill.

Phase 3 — Report & feedback

Report: stages run, files changed, QA verdict, trigger-audit findings (if any), and the release URL (if any). Then offer the maintainer a feedback opening: "Anything to adjust in the result, the agent roles, or the pipeline order?" Record accepted changes in the CLAUDE.md harness change-log table.

Conventions this harness enforces

• All Agent calls use model: "opus" — harness quality tracks agent reasoning.
• consistency-qa must be general-purpose (it runs npm scripts); the trigger-boundary-auditor is read-only.
• No slash commands are created — short forms are auto-installed by the session-start hook.
• Direct-to-main: changes push to main without a PR (per repo CLAUDE.md); the global simplify→review→commit gate still applies to non-doc edits, but skill/guide content is markdown and follows the validate gate instead.

Error handling

• A stage that fails once is retried once with its error as input; a second failure stops the pipeline and reports to the maintainer (no silent skip).
• QA FAIL never proceeds to release.
• Conflicting data is reported with provenance, not deleted.
• High-risk git ops (--no-verify, --force, history rewrites) require explicit maintainer authorization — release-manager stops and asks.

Test scenarios

Normal flow — "add a brooks-security skill": Phase 1 selects author+eval+QA+audit. skill-author runs new-skill brooks-security, creates SKILL.md (with a sibling-carving "Do NOT trigger for:" clause) + guide; eval-curator adds an S-code happy-path + a false-positive scenario; consistency-qa runs the gate → PASS; trigger-boundary-auditor confirms no collision with brooks-review/audit. Report lists files + PASS.

Error flow — QA FAIL on book-count drift: maintainer adds a thirteenth book but edits only source-coverage.md. consistency-qa's cross-doc check finds README still says "twelve" → FAIL, attributed to skill-author. Orchestrator loops back; skill-author updates README/AGENTS/GEMINI wording; QA re-runs → PASS. No release was requested, so the pipeline ends at Phase 3.