Automatic rollback on test failure

[adamdaum]

Summary

When a feature implementation fails tests or breaks existing functionality, the loop continues forward with broken state. There's no mechanism to revert to the last known-good commit and retry or skip the feature.

Current Behavior

• Each iteration commits changes at Step 9
• If Step 6 (testing) fails, the agent may still commit broken code
• If regression testing (Step 3) detects a break, there's no automated recovery
• Manual git revert is required to restore working state

Proposal

• Before starting a feature, tag or record the current commit SHA as a checkpoint
• If testing fails after N retry attempts, automatically git reset --hard to the checkpoint
• Mark the feature with a "status": "blocked" or "attempts": N field in feature_list.json
• Skip blocked features in subsequent iterations and move to the next eligible feature
• Add a --max-retries flag (default: 2) to control retry behavior per feature

Workflow

1. Record checkpoint (git SHA)
2. Attempt feature implementation
3. Run tests
4. If tests pass → commit, continue
5. If tests fail → retry (up to max-retries)
6. If all retries fail → git reset to checkpoint, mark feature blocked, continue to next

Considerations

• Need to handle partially-staged changes cleanly during rollback
• Blocked features should be surfaced in the final summary
• Consider a --strict mode that halts the entire loop on any failure