Add magic string detection for reliable sub-agent hook reporting

[nhorton]

Summary

This PR introduces a critical fix to the manual testing workflow by implementing magic string detection for reliable sub-agent hook reporting. Instead of relying on visual observation of blocking prompts or manual queue inspection, sub-agents now return standardized strings (HOOK_FIRED: or HOOK_NOT_FIRED:) that the main agent checks to determine test pass/fail status.

Key Changes

• Magic String Protocol: Sub-agents are now instructed to return:

  • HOOK_FIRED: <description> if a DeepWork hook blocked them
  • HOOK_NOT_FIRED: Task completed successfully if no hook fired

• Updated Test Instructions: All 8 test cases in both run_fire_tests.md and run_not_fire_tests.md now include the magic string instruction at the end of each sub-agent prompt

• Fixed File Names: Corrected all test file references in sub-agent prompts to match actual test files:

  • feature.py → test_trigger_safety_mode.py
  • module_source.py → test_set_mode_source.py
  • handler_trigger.py → test_pair_mode_trigger.py
  • dangerous.py → test_infinite_block_prompt.py
  • risky.py → test_infinite_block_command.py
  • And similar corrections for all other test files

• Updated Promise Tags: Fixed infinite block test promise tags to match rule names:

  • <promise>I have verified this change is safe</promise> → <promise>Manual Test: Infinite Block Prompt</promise>
  • <promise>I have verified this change is safe</promise> → <promise>Manual Test: Infinite Block Command</promise>

• Quality Criteria Updates: Modified acceptance criteria to check for magic string detection instead of manual observation:

  • Changed from "Hooks Observed" to "Magic String Detection"
  • Clarified that agents must NOT manually run rules_check

• Documentation Improvements: Enhanced test_reference.md with clear magic string detection explanation and updated critical rules

• Version Bump: Updated job version from 1.2.1 to 1.3.0 with changelog entry

Implementation Details

The magic string approach provides:

• Reliability: Explicit, unambiguous signal from sub-agents about hook behavior
• Simplicity: Main agent checks response text instead of inferring from blocking behavior
• Consistency: All tests use the same detection mechanism
• Fallback: Queue inspection still available if magic string is missing (inconclusive case)

This change ensures the manual testing workflow is more robust and less dependent on visual observation or manual verification commands.