swarm-fix

You are in Swarm Suite navigator mode. The user typed /swarm-fix $ARGUMENTS. The $ARGUMENTS (if any) names a specific finding ID; if empty, work through all confirmed findings in priority order.

Do this:

1. Call kb_navigator_state(project_path=<cwd>).
2. Pipeline check: if not in fix stage, tell the user and ask whether to advance.
3. Find candidates:
   • If $ARGUMENTS is a finding ID: that's the target.
   • Else: kb_search_findings(severity=high, status=confirmed) + (medium, confirmed). Sort by severity DESC, then by confidence DESC.
4. Show the user the prioritised list (max 5 entries) and confirm: "Fix all of these, or pick a subset?"
5. For each finding to fix: a. Snapshot tests once at the start: snapshot_tests(session_id). b. Open or reuse a fix session: start_session(review_session=..., arch_session=...). c. Decide PGVE vs one-shot: if the finding is in security / threading-safety / error-handling (high-iteration-risk categories), drive PGVE -- kb_start_pgve(task_spec=<concise description>) then submit_candidate / evaluate_candidate cycle until accepted or budget exhausted. Otherwise just propose_fix once. d. Always show the proposed diff to the user before apply. Never silently apply_approved. e. After the user approves: apply_approved(...) for the consensus fix, or apply_single(...) for an individual one. f. After each fix iteration: kb_check_quality_gate(...). Surface the recommendation:
   • continue -- offer next finding.
   • stop_clean -- offer to advance to verify stage.
   • stop_circuit_breaker -- STOP, tell user the cycle is unstable and recommend manual review.
6. Once done, end with: "Fixes applied -- want to run verify next?"

Anti-patterns to avoid:

• Applying any fix without showing the diff first.
• Calling apply_approved if kb_check_quality_gate last returned stop_circuit_breaker -- escalate to user.
• Looping forever on the same finding -- PGVE has a hard budget (default 5 candidates); respect it.