- C2: Added check_review_summary() function to check_replied.py that correctly checks for bot replies after a review submission (using REST API timestamps instead of mismatched GraphQL node IDs). Updated build_threads.py type_map to use the new review_summary type.
  - C1: Kept -p flag (it works correctly for CI — tools still execute), updated docs to accurately describe print mode with tool use.
  - C3: Removed all --max-turns 20 references from docs (flag doesn't exist in Claude CLI).

  Important fixes:
  - I1: Aligned SKILL.md exit code 2 handling with build_threads.py — both now treat it as "proceed with caution".
  - I2: Fixed null-unsafe comment["user"]["login"] at two locations in check_replied.py to use .get() chains.
  - I3: Replaced --force-with-lease fallback with safe failure — logs warning and exits instead of risking overwrite of concurrent commits.
  - I4: Fork-based PRs now use a separate fork remote for push, keeping origin pointed at upstream so origin/${BASE_BRANCH} diffs still work.

  Minor fixes:
  - M1: Updated stale docs (push batching, allowed tools list, post-commit guardrails).
  - M2: Changed missing python3 from exit 0 to exit 1 (it's a hard requirement).
  - M3: Moved local tid out of the while loop body to function scope.