ac-review

name	ac-review
description	Review pending convention candidates. Approve, reject, or edit each candidate before it reaches CLAUDE.md.
disable-model-invocation	true

Auto-Context: Convention Review

Review candidates that have been promoted to review_pending status. No convention reaches CLAUDE.md without explicit user approval through this skill.

Important rules:

• ALWAYS write each decision to disk immediately after the user responds. Never batch decisions.
• Use jq for all JSON operations (atomic writes with tmp+mv pattern).
• Use jq -nc for JSONL changelog entries.
• Present one candidate at a time so each decision is persisted before the next.

---

Step 1: Read Pending Candidates

Read .auto-context/candidates.json in the project root.

Filter for entries where stage == "review_pending":

jq '[.[] | select(.stage == "review_pending")]' .auto-context/candidates.json

If no review_pending candidates found, tell the user:

No candidates pending review. Candidates need 3+ observations across 2+ sessions to qualify. Run /ac-status to see the current pipeline.

And stop.

If candidates exist, display: "Found N candidates ready for review."

---

Step 2: Present Each Candidate

For each review_pending candidate, present in this format:

### Candidate N of M

**Convention:** {text}
**Confidence:** {confidence}
**Observations:** {observations} across {sessions_seen length} sessions
**First seen:** {created_at}
**Evidence:**
- `{file}:{line}` -- {snippet}
  (for each item in evidence array)

**Action:** approve / reject / edit

Wait for the user's response after each candidate. Do NOT present the next candidate until the current one is processed and written to disk.

---

Step 3: Process User Decision

On "approve":

1. Read current .auto-context/conventions.json
2. Read session_count from .auto-context/lifecycle.json:

   jq -r '.session_count // 0' .auto-context/lifecycle.json
   

   If lifecycle.json is missing or corrupt, use 0 as fallback.
3. Create a new convention entry:
   • text: candidate's text (unchanged)
   • confidence: 0.7 (promoted extraction confidence)
   • source: "extraction" (preserve original source)
   • stage: "active"
   • created_at: candidate's created_at (preserve original)
   • approved_at: current ISO 8601 timestamp
   • session_id: candidate's session_id (preserve original)
   • last_referenced_session: session_count from lifecycle.json
   • observations: candidate's observations count
   • sessions_seen: candidate's sessions_seen array
4. Append to conventions.json atomically (jq + tmp + mv)
5. Remove the candidate from candidates.json atomically (filter by text match)
6. Log to .auto-context/changelog.jsonl:

   jq -nc --arg ts "$(date -u +%Y-%m-%dT%H:%M:%SZ)" --arg text "{text}" \
     '{ts:$ts, action:"approved", text:$text, reason:"user approved via /ac-review", from_stage:"review_pending", to_stage:"active"}' \
     >> .auto-context/changelog.jsonl
   

7. Confirm: "Approved: {text}"

On "reject":

1. Remove the candidate from .auto-context/candidates.json atomically
2. Log to .auto-context/changelog.jsonl with action: "rejected", reason: "user rejected via /ac-review", from_stage: "review_pending", to_stage: "rejected"
3. Confirm: "Rejected: {text}"

On "edit":

1. Ask the user for the modified convention text
2. Follow the same approve flow but use the user's edited text
3. Log to changelog.jsonl with action: "approved_edited" and include both original and modified text in the reason field
4. Confirm: "Approved (edited): {new_text}"

---

Step 4: Post-Review Summary

After all candidates are reviewed, show a summary:

Review complete: N approved, M rejected, K edited

If any were approved:

Approved conventions will appear in CLAUDE.md at next session start. Run /ac-status to see current pipeline state.