fix: guard hook re-reads transcript for a lane stated in the same message (0.6.1)

[sofus-common]

Summary

• Fix: when the lane statement and the tool call are in one assistant message, the statement can reach the transcript after the PreToolUse hook reads it, so the first write is denied. When no lane is found, hooks/edit_guard.py now re-reads the transcript every 0.1 s for up to 2 s before denying.
• Dates 0.6.1 (with fix: guard hook recognizes lane stated via SendUserMessage or "Lane is X" (0.6.1) #8) and records the verification.

Verification

• Headless Claude Code 2.1.300 with claude-sonnet-5-5, typo fixture, agent states "The lane is Fast." then edits:
  • 0.6.0 hook: 4 denials, agent gave up (reproduces the report).
  • fix: guard hook recognizes lane stated via SendUserMessage or "Lane is X" (0.6.1) #8 hook without re-read: 3 of 4 sessions denied once; instrumented hook saw no assistant text at the first call.
  • With re-read: 6 of 6 sessions edited with no denial; the statement arrived within 0.7 s.
• New test_edit_guard.py case (statement appended 0.3 s after the hook starts) fails on the fix: guard hook recognizes lane stated via SendUserMessage or "Lane is X" (0.6.1) #8 hook and passes now.
• check_version.py (+ self-test), test_behavior_check.py, test_edit_guard.py pass.
• Not run: the 12-case behavior suite (hook-only release; docs/behavior-checks.md requires it for skill-instruction changes). Brief mode is unavailable headless, so the SendUserMessage path is unit-tested only.

Known limits

• A write with no stated lane is now denied after a 2 s wait instead of at once.

🤖 Generated with Claude Code