メニュー
How any AI collaborator (Codex, Conductor, Octopus, Compound Engineering, future-Claude) works alongside the QA Architect agent on this repo. Read before designing or shipping a non-trivial change.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | qa-architect |
| description | How any AI collaborator (Codex, Conductor, Octopus, Compound Engineering, future-Claude) works alongside the QA Architect agent on this repo. Read before designing or shipping a non-trivial change. |
| triggers | ["verdict","qa-architect","tribunal","blast radius","invariant","regression","quality drift"] |
| license | internal |
| compatibility | {"agent-framework":">=1.0.0-preview","microsoft-extensions-ai":">=10.3.0"} |
| metadata | {"triggers":["verdict","qa-architect","tribunal","blast radius","invariant","regression","quality drift"],"evidence-kinds":["contract_check","adversarial_replay","quality_snapshot"],"contract":{"path":"docs/contracts/2026-05-02-qa-verdict.contract.md","schema":"docs/contracts/qa-verdict.schema.json","schema-version":"0.1.0"}} |
| allowed-tools | ["qa_assess_blast_radius","qa_gap_analyze","qa_propose_tests","qa_verify_invariants","qa_replay_adversarial","qa_score_quality_drift","qa_lookup_defect_memory","qa_emit_verdict"] |
This is the rubric for collaborating with Guitar Alchemist's senior QA agent. Read it before opening a PR, proposing a refactor, or shipping a metric-moving change. It is a checklist, not a narrative — follow it in order.
Two surfaces of the same role:
QAArchitectAgent — Common/GA.Business.ML/Agents/QAArchitectAgent.cs. Synchronous, in-process, callable from any C# code. Returns a QaVerdict.qa-architect-cycle.ixql — Demerzel/pipelines/qa-architect-cycle.ixql. Long-running, governance-grade, runs on PR open/sync and on a daily 06:00 UTC sweep. Persists verdicts to state/quality/verdicts/.They both speak the same protocol: the QaVerdict contract.
docs/contracts/2026-05-02-qa-verdict.contract.md (see also references/qa-verdict-contract.md bundled with this skill).docs/contracts/qa-verdict.schema.json.Available from the GaQaMcp MCP server. Treat them as your hands — call them rather than re-implementing.
| Tool | Call when… | Returns |
|---|---|---|
qa_assess_blast_radius | Before designing a non-trivial change | blast_radius object: layers touched, one-way doors crossed, invariants at risk, score 0..1 |
qa_gap_analyze | After writing code, before opening a PR | Array of followup candidates surfacing untested critical paths |
qa_propose_tests | When adding new components or breaking out abstractions | Array of test stubs (path + body) grounded in repo patterns |
qa_verify_invariants | Before merging changes that touch Core / Domain / Analysis / AI/ML | evidence{kind: contract_check} items, one per invariant |
qa_replay_adversarial | Before merging changes to chord/voicing/embedding pipelines | evidence{kind: adversarial_replay} |
qa_score_quality_drift | When changes may move a metric in state/quality/*.json | evidence{kind: quality_snapshot} with delta + guardrail check |
qa_lookup_defect_memory | Before designing — surfaces "we got burned by this before" | Array of past followups matching the touched components |
qa_emit_verdict | At the end, to persist a verdict | Path of the persisted JSON + verdict_id |
Run this before the PR is opened. Skip steps only when you can name the reason.
qa_lookup_defect_memory over the components you'll touch. If a past followup matches, your change must address it explicitly or document why it doesn't apply.qa_assess_blast_radius on the diff. If estimated_blast_score >= 0.4, expand testing per WP-2 of the plan: add property tests, fuzz inputs, or contract tests across the boundary you're crossing.qa_verify_invariants for the layer you're changing. The 5-layer dependency rule and the OPTIC-K dim invariant are both checked here.qa_propose_tests for any new component. Use the proposals as starting scaffolds, not as final tests.qa_gap_analyze last. Address must_fix candidates before opening the PR.The verdict you'll see in PR comments and at state/quality/verdicts/. Read in this order:
risk_tier — what action is required.
P0 → stop. Don't merge. Fix every must_fix followup, re-request review.P1 → must-fix is required before next release. Open a tracked issue if you cannot fix in this PR.P2 → should-fix. Address inline or annotate why deferred.P3 → informational only. No action required, but read it.verdict — what gating is enforced.
block → branch protection blocks merge. Pair with P0.merge_with_followups → mergeable; followups are tracked.pass → no must-fix items.informational → scheduled sweep or exploratory; not gating.reviewer_chain — who objected and why. If a single specialist role gave a low score, that's where the risk concentrates.followups[].rationale — the why of each item. If the rationale references a past incident or contract ID, take it literally.blast_radius.one_way_doors_crossed — non-empty here forces P0. These crossings require human sign-off, not a test fix.narrative — the human-readable summary, max 500 chars. Skim if you trust the structured fields; read closely if you don't.You do not have authority to:
schema_version bump proposal.risk_tier rollup rule (e.g. demote a P0 to P1).docs/plans/.Surface these to the human in the PR description with the heading Requires QA sign-off: and a one-paragraph justification.
docs/plans/2026-05-02-arch-qa-architect-tribunal-plan.mddocs/contracts/2026-05-02-qa-verdict.contract.mddocs/contracts/qa-verdict.schema.jsonCommon/GA.Business.ML/Agents/QAArchitectAgent.csDemerzel/pipelines/qa-architect-cycle.ixql (sibling repo)Apps/GaQaMcp/Post-merge intent-vs-delivery evaluator. Resolves the most recent squash-merged PR, captures its stated intent (title + Summary), diffs it against what actually landed, runs /octo:review for specialist commentary, scans Codex bot comments (P0/P1 degrade the grade), computes a high/medium/low alignment score, and writes a structured grade to state/quality/pr-grades/<merge-sha>.json. Closes the "agent declared done, nobody checked the goal" loop.
Run one bounded supervised autonomous cycle on the Guitar Alchemist repo. Picks the smallest unchecked roadmap slice inside allow_edit, implements it, runs the oracle, emits cycle evidence, and stops. Refuses to run unless dev-process-overseer reports loop-eligible and the preflight passes. Use when asked to "run one supervised loop", "advance the roadmap autonomously", or "do one cycle and stop".
Rebuild the OPTIC-K voicing index (state/voicings/optick.index). Invoke when schema hash bumps, tag enrichment changes, new chord qualities are added, or the leak-test / invariant-coverage tools flag regressions. Mandatory precondition: stop any process that has the index mmap-locked — GaApi and GaMcpServer both hold the file open at runtime, and the write step WILL fail with IOException otherwise.
Generalized Cherny-style autonomous improvement loop. Reads a domain-scoped baseline.json + oracle script, runs the oracle, picks the worst-scoring item, proposes a fix, validates rollback safety, commits if metric improved AND roundtrip passes. Use for chatbot-qa, embeddings, voicing-analysis, or any other domain that has a baseline + oracle. Refuses to run if killswitch present, governance halt active, or scope_boundary violated.
End-to-end autonomous (AFK) harness that develops AND QAs the GA chatbot. Orchestrates the existing Level-3 /auto-optimize dev loop + the semantic /ga-chatbot-qa-panel judge workflow, instruments every step to a dashboard the human can watch, and enforces branch-only / never-merge safety. Use to run unattended chatbot-quality improvement with full human visibility. Refuses to run if the backend is down, a killswitch/halt marker is present, or scope_boundary is violated.
Turns raw Sentrux structural-quality metrics (quality_signal, cycle count, coverage) into a ranked list of actionable refactor recommendations with starter sketches. Pulls live data from sentrux MCP tools, optionally cross-references AI annotations (@ai:business-value, @ai:smell) and the latest /test-plan output, then writes 5-10 prescriptive bullets to state/quality/sentrux-next-steps/<date-Z>.md so the Sentrux dashboard tab can render them. PROPOSES — never auto-refactors; the human picks which recommendation to take.