mumei

Plays "what could go wrong" against a Wave's diff. Surfaces race conditions, edge cases, silent failures, and operability gaps that other reviewers miss. Triggered LAST in the review pipeline (after spec-compliance and security have completed) so it can avoid duplicating their findings.

Reviews a draft design.md against the approved requirements.md. Detects coverage gaps (ACs without a corresponding design element), missing architectural artifacts (no diagram, no Components, no Trade-offs), and Wave Plan defects (granularity unfit for tasks decomposition). Triggered automatically by /mumei:compose after each design draft. Returns PASS / NEEDS_IMPROVEMENT / MAJOR_ISSUES with structured findings.

Re-validates a single finding produced by another reviewer with fresh context. Returns valid / invalid / unsure. Triggered by /mumei:compose after the 3 reviewers complete (spec-compliance / security / adversarial) — invoked once per finding in parallel for severity=HIGH/CRITICAL findings. Filters false positives before they reach the user.

Independent evaluator that scores a candidate memory entry from a reviewer agent against a 7-axis rubric (generality, recurrence, longevity, coverage_gap, actionability, density, confidence) and decides ADD / UPDATE / SKIP. Triggered by /mumei:compose Phase 5 Stage 6 and /mumei:peruse after each review-pipeline candidate is emitted; runs once per candidate. Read-only — never writes to memory; the orchestrator persists ADD/UPDATE via hooks/_lib/memory.sh atomic helpers.

Writes a property-based test from an _Invariant declaration WITHOUT seeing the production implementation. Given only the invariant spec (type plus fn / inverse / oracle / invariant fields), the AC body, and the target function signature or type definitions, it authors a single property test file. Triggered by /mumei:compose during implement phase for each AC that carries an _Invariant line. The generated test is frozen as a golden file so the implement actor cannot later edit it to match a flawed implementation.

Render a detailed reliability view of a mumei feature — pass^3 over the most recent 10 trials plus a table of the last 10 trial rows from reliability-log.jsonl. Triggered only by explicit user invocation `/mumei:attest <feature>`. Reads .mumei/specs/<feature>/reliability-log.jsonl or .mumei/plans/<feature>/reliability-log.jsonl via hooks/_lib/reliability.sh. Exits non-zero with `feature not found: <feature>` to stderr when the feature directory is absent. The k=3 / window=10 parameters are fixed.

Render a one-line reliability summary of the active mumei feature (or a specified one). Prints `<feature> | pass^3: <value-or-N/A> (n=<n_trials>, window=10, k=3)` to stdout — exactly one line, no headers. Triggered only by explicit user invocation `/mumei:glance` (no args reads `.mumei/current`) or `/mumei:glance <feature>`. Outputs `no active feature` to stdout (not stderr) and exits 0 when `.mumei/current` is missing or points to a non-existent feature. Read-only via hooks/_lib/reliability.sh.

This skill should be used BEFORE any feature design. It runs structured gleaning with the user — asking 5 high-leverage questions per round, up to 3 rounds, to extract Goal / Scope / Constraints / Edges / Done. Output is saved to .mumei/scratch/<topic>.md and used as input for /mumei:compose. Triggers include "I want to add X", "we need a feature for Y", "let's brainstorm Z", or any vague feature request.

One-time setup for a project to use mumei. Detects existing CLAUDE.md / .claude/rules/, proposes additions about mumei's expectations (phase gates, Wave commits, review pipeline), and applies them with user approval. Triggers when the user says "set up mumei", "install mumei", or "initialize mumei in this project".

Generate a retrospective markdown for an archived (or about-to-be-archived) mumei feature. Reads requirements / design / tasks / spec-reviews / reviews / cost-log and emits muse.md with AC count, Wave count, review iter counts, fix-spiral detection, token cost, cache hit rate, and hook firing breakdown. Triggered by the user invoking /mumei:muse <feature> after /mumei:shelve (or before, if explicitly requested). Read-only with respect to feature content; writes only the muse.md output file.

24 hooks across 18 events