Review

<!-- Role: standalone (detached) review entry point. Shares hooks/_lib/review.sh + detectors with /mumei:peruse and /mumei:compose Phase 5, but runs without any feature dir and writes nothing under .mumei. Principle: fail-open, metadata-quarantined, evidence-gated. Same pipeline math as the vehicle reviews (mumei_review_apply_advisory_downgrade / ground_truth_high_count / aggregate_verdict / detached_report). -->

Review — standalone diff review

Runs mumei's review engine against an arbitrary diff and reports findings + verdict in the conversation. No .mumei footprint, no commits, no memory writes.

When to use

• The user wants a strong review of the current changes but is NOT driving the feature through /mumei:compose (spec) or /mumei:peruse (plan).
• The user keeps their own spec/SDD and wants mumei's detector + reviewer + gate pipeline on top of it (pass the spec file as the second argument).
• A change too small to plan, where the user still wants the strongest review.

When NOT to use

• An active mumei feature is mid-flight and the user wants its lifecycle review — use /mumei:compose (spec) or /mumei:peruse (plan) instead.
• The user wants findings persisted / phase advanced / archived — this skill is intentionally read-only.

Method

All steps run from the project root in a git repo. The skill never mutates .mumei, never commits, and never writes the ledger or agent memory.

Step 1 — Resolve base ref and diff

source "${CLAUDE_PLUGIN_ROOT}/hooks/_lib/detectors.sh"
source "${CLAUDE_PLUGIN_ROOT}/hooks/_lib/detectors-ext.sh"
source "${CLAUDE_PLUGIN_ROOT}/hooks/_lib/review.sh"
source "${CLAUDE_PLUGIN_ROOT}/hooks/_lib/residual.sh"

base_arg="$1"   # optional
spec_arg="$2"   # optional spec file for spec-compliance scope_source

if ! git rev-parse --git-dir >/dev/null 2>&1; then
  echo "not a git repository; /mumei:review needs a git repo." >&2
  exit 0
fi

# Base resolution: explicit arg → origin/HEAD → main → master.
if [[ -n "$base_arg" ]]; then
  base="$base_arg"
else
  base="$(git symbolic-ref --quiet --short refs/remotes/origin/HEAD 2>/dev/null \
    | sed 's#^origin/##')"
  [[ -z "$base" ]] && { git show-ref --verify --quiet refs/heads/main && base="main"; }
  [[ -z "$base" ]] && { git show-ref --verify --quiet refs/heads/master && base="master"; }
fi
if [[ -z "$base" ]]; then
  echo "could not resolve a base ref (no origin/HEAD, main, or master). Pass one explicitly: /mumei:review <base-ref>" >&2
  exit 0
fi

merge_base="$(git merge-base "$base" HEAD 2>/dev/null)"
if [[ -z "$merge_base" ]]; then
  echo "could not compute merge-base against '${base}'." >&2
  exit 0
fi

# Includes pushed commits since the branch point AND uncommitted working-tree
# changes (REQ-27.4).
diff="$(git diff "$merge_base")"
diff_lines="$(printf '%s\n' "$diff" | wc -l | tr -d ' ')"
if [[ -z "$diff" ]]; then
  echo "no changes against ${base} (merge-base ${merge_base}); nothing to review."
  exit 0
fi

Step 2 — Stage 0 detectors (detached, no feature dir)

Run the pluggable detector registry into a temp report. Absent tools are warn-skipped (REQ-27.5); Tier2 opt-in via MUMEI_DETECTOR_TIER2=1.

work_dir="$(mktemp -d -t mumei-review.XXXXXX)"
report="${work_dir}/detectors.json"
trap 'rm -rf "$work_dir"' EXIT
mumei_detector_run_all "$work_dir" "$report" "adhoc-review"

# Split detector findings by precision_class (REQ-27.8):
#   ground_truth (osv/secret/type/test) → surfaced directly (block-eligible)
#   candidate    (semgrep/codeql/linters) → fed to the Stage 4 gate
gt_detector="$(jq -c '[(.findings.HIGH + .findings.MEDIUM + .findings.LOW)[] | select(.precision_class == "ground_truth")]' "$report" 2>/dev/null || echo '[]')"
cand_detector="$(jq -c '[(.findings.HIGH + .findings.MEDIUM + .findings.LOW)[] | select((.precision_class // "candidate") == "candidate")]' "$report" 2>/dev/null || echo '[]')"

If MUMEI_BYPASS=1, skip detectors entirely (treat both arrays as []).

Step 3 — Launch reviewers (metadata-quarantined, cold)

Build each reviewer prompt with mumei_reviewer_prompt so the metadata-quarantine prefix (REQ-27.12) applies. Launch in parallel:

• security-reviewer — diff + ground_truth detector findings as <detector_findings ground_truth="true">. Cold otherwise (no PR description).
• adversarial-reviewer — diff + prior findings only (no spec, no metadata).
• spec-compliance-reviewer — ONLY when spec_arg is set; pass scope_source=<spec_arg>. Skip entirely when no spec file is provided (REQ-27.3).

Task(subagent_type: "mumei:security-reviewer", prompt: "<mumei_reviewer_prompt output> ...")
Task(subagent_type: "mumei:adversarial-reviewer", prompt: "...")
# only if spec_arg:
Task(subagent_type: "mumei:spec-compliance-reviewer", prompt: "... scope_source=${spec_arg}")

Step 4 — Per-finding adjudication gate

Collect reviewer findings + cand_detector into a candidate pool. For each candidate finding, use mumei_review_finding_needs_gate to decide:

• ground_truth findings (gt_detector) skip the gate (deterministic).
• candidate HIGH/CRITICAL → mandatory issue-validator (single adjudicator).
• candidate MEDIUM/LOW with reviewer confidence HIGH → may skip (same sampling rule as Phase 5 Stage 4).

Task(subagent_type: "mumei:issue-validator", prompt: "<finding> ...")

Keep decision == "valid" / valid_by_assertion; drop invalid; surface unsure with a marker. Build surfaced_json = gt_detector + validated candidate findings (each carrying precision_class, severity, and the validator object so the downgrade can read axes.reproducible).

Step 5 — Assemble the detached report + present

report_json="$(mumei_review_detached_report "$surfaced_json" "$reviewer_verdicts_json" "$diff_lines")"

mumei_review_detached_report applies the advisory downgrade (fail-open), counts ground_truth HIGH, caps surfaced by diff size with overflow → residual, and aggregates the verdict. Present to the user, in their language:

1. Verdict (PASS / NEEDS_IMPROVEMENT / MAJOR_ISSUES) with a one-line tally.
2. Surfaced findings — file:line, severity, severity_action (block / report_only), and the evidence the validator confirmed. Lead with blocking findings.
3. Residual — what was NOT covered (overflow + the AI-blindspot ceiling), so the user knows the review's limits.
4. State plainly that nothing was persisted and no commit was made.

Output

Conversational only. The skill writes NO files under .mumei, makes NO commits, and touches NEITHER the cross-feature ledger NOR agent memory. --save persistence is intentionally out of scope.

Don'ts

• Don't create .mumei/current, a feature dir, or any state.json.
• Don't run pre-review-detector.sh (it is feature-bound); call mumei_detector_run_all directly into a temp report.
• Don't treat candidate detector findings (semgrep / CodeQL / linters) as ground truth — they pass through the gate (fail-open). Only osv / secret / type-check / test-check block directly.
• Don't auto-suppress HIGH/CRITICAL findings; ungrounded ones surface as advisory.
• Don't commit, push, or advance any phase.