sharp-review

name	sharp-review
description	Post-feature sharp review (锐评) —parallel reviewers, merge findings, sync task list

Sharp Review (锐评)

Post-feature review: 2 of 3 reviewers, each JSON-Schema-constrained, cross-checked and merged into a single memory entry .claude/memory/YYYY/MM/DD/sharp-review.md with rem frontmatter. Normally invoked by the Stop hook once enough change accumulates (Wave Gate); /sharp-review runs it manually. Trigger thresholds, profile-weighting math, mode internals, and config keys → reference/profiles-and-modes.md.

Execution mode (read first)

Review runs on git state, not the conversation — so offload it: main loop dispatches one sharp-review:sharp-review subagent to run Steps 1–6 and return only the Sharp review: <summary> line (pass it the hook's firedSources). Zero leakage into the main session; the worker fans out reviewers directly (Step 3) and must not re-dispatch (recursion).

Execution

Step 0 — Resolve plugin root

If $env:CLAUDE_PLUGIN_ROOT is empty (not inherited on some machines), resolve it via the PowerShell fallback in reference/windows-troubleshooting.md before running any sharp-review script; if still empty, report and stop as described there.

Step 1 — Pick profiles

Each run picks 2 profiles via weighted random draw without replacement, filtered to those whose source fired this round. The Stop hook records which trigger source(s) fired in reviewGate.firedSources; pass them so the pick is constrained to eligible profiles. Capture the JSON array:

node "$env:CLAUDE_PLUGIN_ROOT/scripts/pick-profile.js" --sources diff,docs

[
  { "key": "adversarial", "label": "adversarial review (对抗性审查)", "mode": null, "promptKind": "diff", "framing": "…", "reviewScope": "…" },
  { "key": "diff", "label": "diff review", "mode": null, "promptKind": "diff", "framing": null, "reviewScope": null }
]

If fewer than 2 profiles are eligible (e.g. --sources codebase → only architecture), the array has a single element. Reviewer-to-profile assignment is shuffled — no profile is predictably bound to a specific reviewer model.

Omit --sources for the full default rotation (manual run). Manual override: --profile architecture (returns single-element array). Selection is stateless (weights only). Profile table + weighting math + per-project profileWeights → reference/profiles-and-modes.md.

Step 2 — Gather context

Diff-sourced profiles (diff/security/adversarial) need the diff payload; agent-mode profiles (architecture/docs/deps) explore the repo but still need stats/range/seed — so always run diff-manifest.js (the ONLY allowed diff source; never run raw git diff or paste diff text):

node "$env:CLAUDE_PLUGIN_ROOT/scripts/diff-manifest.js"                       # uncommitted vs HEAD (default)
node "$env:CLAUDE_PLUGIN_ROOT/scripts/diff-manifest.js" --range "main...HEAD" --path "src/components"

--range/--path scope the review. Capture the JSON output — it carries mode, range, seed, stats, and (mode-dependent) diff | manifestText + excludedSummary. Full payload schema → reference/profiles-and-modes.md.

Step 3 — Run reviewers (fan-out)

Empty-diff gate: If ALL profiles honor the diff manifest (all have mode === null) AND mode === "empty": report in chat Sharp review skipped: no reviewable changes after filtering (<excludedSummary>) and stop. If at least one profile is agent-mode (architecture/docs/deps), proceed — agent-mode profiles explore the repo, not a diff.

Otherwise fan out via mcp__plugin_fabric_fabric__call (primary — direct API calls, no safety-classifier dependency) for each active reviewer; fall back to the Agent tool (Claude Code) / spawn_agent (Codex) if takeover is unavailable. Prerequisite: the dispatched sharp-review:sharp-review worker must list mcp__plugin_fabric_fabric__call in its tools: allowlist — an explicit allowlist excludes everything unnamed, so omitting it silently forces every reviewer onto the flaky Agent fallback (the cause of past all-reviewers- FAILED runs). Collect each reviewer's raw { findings: [...] } and feed the raw results to post-review.js --raw. deepseek/claude return JSON directly; codex review-mode returns prose — normalize per reference/direct-fanout.md § Codex prose normalization (which also defines the single []-vs-null rule). Do NOT merge or assign SR- ids yourself — the shared mergeFindings/renderReviewMarkdown in lib.mjs (invoked by post-review.js) owns that, so every host produces byte-identical output. Full procedure, seed-mod-3 rotation, raw.json schema, and positional alignment → reference/direct-fanout.md.

Step 4 — Write memory entry & sync

IMPORTANT on Windows: Do NOT use Bash redirection (>) with Windows paths — Bash treats backslashes as escape characters and creates stray files in the wrong location. Instead, write the raw.json from Step 3 using the Write tool, then call post-review.js via PowerShell:

node "$env:CLAUDE_PLUGIN_ROOT/scripts/post-review.js" --date <YYYY-MM-DD> --raw "$env:TEMP/claude-sharp-review/raw.json"

This writes .claude/memory/YYYY/MM/DD/sharp-review.md with rem frontmatter, then stamps its entry into .claude/rules/MEMORY.md in-process (no rem plugin needed).

Step 5 — Resolve findings

todo mark <SR-ID> fixed

This flips **Status:** OPEN → FIXED in sharp-review.md and re-derives the frontmatter — equivalent to hand-editing + post-review.js --rescan. (todo is the rem-owned CLI; sharp-review never calls task-engine.js directly.)

For the full file-ownership table (where findings, archives, and manual tasks live) → reference/task-system.md.

Step 6 — Report

Output in chat ONLY: Sharp review: <summary>. Do NOT dump findings in chat — this report step is the attention gate. By default findings go to the backlog (AI consumers like evolve read them there; humans triage async via todo). Only when a human asks to triage now do you route OPEN findings through the shared attention gate (shared/attention.mjs). Consumer-aware routing detail → reference/profiles-and-modes.md § Attention boundary.

Phase 2 — Task Audit

After the review:

1. Run todo — review open findings against code changed in this session. Items the review touched show ⚠ likely-resolved.
2. For each finding you fixed AND verified (tests pass / behavior confirmed) in this session, run todo mark <SR-ID> fixed immediately — do not leave it for the next review to rediscover. Do NOT mark a finding fixed if you only changed the file without confirming the issue is resolved.
3. Flag stale items (> 90d untouched) — todo report shows ⚠ stale markers.
4. Check in-flight Codex tasks via TaskGet — do not mark feature complete until verified.

Usage

Run /sharp-review after finishing a feature. No arguments needed.