training-signoff-worker

name	training-signoff-worker
description	Align env/training surfaces with runtime truth and lock down parity signoff artifacts.

Training Signoff Worker

NOTE: Startup and cleanup are handled by worker-base. This skill defines the WORK PROCEDURE.

When to Use This Skill

Use for env observation/action/reward alignment, self-play wrapper behavior, compiled PufferLib/PuffeRL integration, evaluation outputs, universal parity fixture expansion, and final signoff/scorecard features.

Required Skills

None.

Work Procedure

1. Read the feature, its assertions, .factory/library/architecture.md, .factory/library/environment.md, and .factory/library/user-testing.md.
2. Inspect the current runtime/training boundary files before editing: yomi_hustle/env/, yomi_hustle/training/, compiled backend files, and the relevant tests.
3. Confirm which runtime systems the training/signoff feature depends on; do not paper over missing runtime behavior with wrapper-only hacks.
4. Add failing targeted tests first. For signoff features, this may mean fixture or artifact-generation tests in addition to unit/integration coverage. Ensure every assertion in the feature's fulfills list is directly tied to at least one test assertion, fixture expectation, or signoff artifact check.
5. Implement the minimal alignment or signoff changes needed, then run targeted tests.
6. Perform at least one deterministic manual sanity check and record it in interactiveChecks (for example, structured payload round-trip, mirrored wrapper reset, compiled vec smoke rollout, or scorecard artifact inspection).
7. Run manifest validators before handoff. For signoff features, the full test suite is mandatory.
8. In the handoff, spell out which runtime surface, training surface, and artifact surface are now aligned and which assertions were closed.

Example Handoff

{
  "salientSummary": "Aligned the richer V2 action interface with wrapper replay/rollback and extended the compiled training stack to use it during evaluation. Added action-interface, wrapper, and integration tests plus a machine-readable signoff artifact check.",
  "whatWasImplemented": "Updated the trainer-facing action interface, wrapper payload handling, and evaluation entrypoints so parity-critical structured actions survive encode/decode, rollback, replay, and compiled vec execution. Added/updated tests for masks, wrapper round-trips, training smoke coverage, and signoff artifact validation.",
  "whatWasLeftUndone": "",
  "verification": {
    "commandsRun": [
      {
        "command": "uv run pytest tests/test_training_action_interface.py tests/test_self_play_wrapper.py -v",
        "exitCode": 0,
        "observation": "Structured payload and wrapper legality tests passed."
      },
      {
        "command": "uv run pytest tests/test_pufferlib_integration.py tests/test_training_metrics.py -v",
        "exitCode": 0,
        "observation": 
      
      
         
         
         
      
    
     
      
         
         
      
    
  
   
     
      
         
         
          
             
             
          
        
      
      
         
         
          
             
             
          
        
      
    
  
   

When to Return to Orchestrator

• The feature depends on runtime parity that is not yet implemented in earlier milestones.
• Optional RL dependencies or compiled backend prerequisites are unavailable in a way you cannot repair locally.
• Artifact/signoff requirements need mission-level scope changes rather than a local feature fix.