Use to verify a feature/story meets all Definition of Done criteria before marking complete.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
DoD checklist enforcement.
Before checking any item below as DONE, perform a completion audit against the actual current state, not against intent or memory:
Do not rely on any of the following as proof of completion:
These clauses address two documented Mycelium failures: Eval Overfitting (2026-04-30 — agent encoded eval answers into documentation to "pass," gaming the proxy instead of checking the criterion) and SessionStart relay leak (2026-05-02 — agent declared turns complete while skipping protocol steps). Borrowed from OpenAI Codex /goal continuation template's completion-audit specification (logged 2026-05-03).
docs/adr/ was scaffolded by /mycelium:delivery-bootstrap)Skip this entire section unless diamonds/active.yml declares product_type: ai_tool. An AI feature is a behavioral contract, not a spec: it is judged on success thresholds, failure modes, and must-never constraints rather than line-item functionality. These criteria do NOT apply to software / content / service products — do not let them bleed into the product-agnostic checklist above.
canvas/ai-tool-metrics.yml#prompt_quality, not an engineering default — the PM owns the quality bar, not the engineer. A green eval run below the declared threshold is NOT DONE.behavioral_constraints entry verified non-violated. Each must-never rule in ai-tool-metrics.yml#behavioral_constraints has a populated verification and a recent last_verified. An unverified constraint is NOT DONE (treat-uncertainty-as-not-done, per the anti-bias clauses above).failure_modes derived from real outputs, with binary acceptance. ai-tool-metrics.yml#failure_modes.outputs_reviewed is set (target >= 20 real outputs reviewed) and every mode carries a BINARY acceptance_criterion naming who holds final judgment. Imagined failure modes do not count.regulatory.transparency_disclosure, EU AI Act Art. 50) — cross-check via /mycelium:regulatory-review.The anti-bias clause above says "a test that passes but doesn't exercise the new code is not coverage." That sentence has been in this skill as prose, and prose does not fail a build. Three shipped guards now decide it mechanically. Run them from the project root and paste the output as the evidence for the "tests" and "mechanisms wired" items — an unrun guard is the same false green it was written to catch.
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check_wiring.py" --root .
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check_negative_control.py" --root .
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check_test_authenticity.py" --root .
Each answers a different question, and passing one says nothing about the others:
| Guard | Question | The failure it catches |
|---|---|---|
check_wiring | Does anything CALL this mechanism? | Shipped, documented "auto-updated", invoked by nothing |
check_negative_control | Could this guard FAIL? | A check that is green because it can only be green |
check_test_authenticity | Does production code RUN when the test runs? | A test that imports nothing, asserts a tautology, or mocks away the thing it names |
Non-zero exit is a NOT-DONE, not a warning. Each prints the specific file and line; fix the finding or state the exemption with a reason. "It is hard to check" is not a reason.
Why this belongs at the moment of claiming done rather than in CI alone. The failure class is "built, not wired": an artefact whose tests pass, whose coverage rises, and which nothing calls. Every local signal an agent uses to decide it is finished — tests green, file written, coverage up — is blind to wiring, because wiring is a property of a PAIR and every default tool verifies a thing against itself. So the completion signal fires before the work is connected. These guards move that discovery from "months later, if ever" to "before the box gets ticked".
まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Accessibility audit, scoped to the surfaces a product actually has. Detects web / rendered_markdown / terminal / native_app / video_audio / document / headless, then applies only the criteria that bind. WCAG 2.1 AA in full for web; not at all for headless.
日本語の概要は準備中です。原文の説明を表示しています。
Bring Mycelium into a project that already has code. Detects that the repo predates the framework, asks before touching anything, then reads the codebase to draft what it CAN establish (delivery, solution shape) and — the actual point — names what it cannot (purpose, strategy, real user evidence). The output is a discovery backlog with a head start, never a filled canvas.
日本語の概要は準備中です。原文の説明を表示しています。
Design the smallest viable test to validate or invalidate a critical assumption. Based on Torres's assumption testing framework, organized by Gilad's AFTER model (Assessment → Fact-Finding → Tests → Experiments → Release Results).
日本語の概要は準備中です。原文の説明を表示しています。
Use before any research activity or significant decision. Reviews cognitive biases relevant to the current stage.
日本語の概要は準備中です。原文の説明を表示しています。
Use to evaluate whether current work aligns with Better Value Sooner Safer Happier. Run at diamond completion and periodically.
日本語の概要は準備中です。原文の説明を表示しています。
Lint canvas files for staleness, missing fields, inconsistent evidence types, and orphaned references. Run periodically or before major transitions.
日本語の概要は準備中です。原文の説明を表示しています。