Controlled Lockcheck Autofix

Project-local skill for unattended QQ frontend lockcheck automation.

Read references/workflow.md first. Read references/failure-taxonomy.md when classifying a failed round. Read references/guardrails.md before proposing or applying any fix.

Hard Boundaries

• Keep the current frontend design and visible interaction unchanged.
• Treat tree20260315 as the reference behavior for core invoice-grab rules.
• Do not expand invoice discovery rules just to chase a failing round.
• Airbnb and similar non-target documents must keep exclusion evidence, but must not be counted as missing truth invoices.
• If a candidate fix touches FROZEN or GUARDED ZONE, stop and emit a blocking report instead of editing.

Required Workflow

1. Prepare an isolated controlled run with prepare_qq_lockcheck_run.py.
2. Launch monitor and frontend, then rely on controlled autostart instead of manual clicking.
3. Wait for the round to finish or block.
4. Run postprocess_qq_lockcheck.py.
5. Classify the round from process_log.txt, monitor_status.json, strict_truth_audit.json, and qq_lockcheck_report.json.
6. Only continue unattended iteration when the failure is inside the bounded autofix scope.

Output Expectations

• Persist per-round classification.json, decision.json, and round_summary.md.
• Keep run-root evidence paths, top blockers, and packaging-gate metrics in every round summary.
• State explicitly whether the round passed packaging gate: P0=0 and user_should_review_count<=6.