Task Finish Contract
Prevent mid-task stalls. Every task must finish with explicit state and evidence.
Problem
Agents often:
- Stop mid-task without explanation
- Output plans without execution
- Lack clear completion criteria
- Missing evidence artifacts
Workflow
1. State Output (Each Substantial Step)
**Goal**: What finished looks like
**Progress**: What has been done
**Next**: One concrete action executing now
2. Completion Proof Format
For tasks with 2+ steps, include:
**DONE_CHECKLIST**:
- [ ] Item 1 completed
- [ ] Item 2 completed
**EVIDENCE**:
- Executed: command/action summary
- Artifact: path/URL/id
- Verified: check command result
**NEXT_AUTONOMOUS_STEP**:
- One follow-up that runs without user input
3. Anti-Stall Rule
- Planning-only replies: max 1
- Next reply MUST contain execution evidence
- Never end with "I will now..." without tool result
Executable Completion Criteria
| Criteria | Verification |
|---|---|
| Goal stated | Select-String "Goal" memory/{date}.md matches |
| Progress tracked | Select-String "Progress" memory/{date}.md matches |
| Next action defined | Select-String "Next" memory/{date}.md matches |
| Evidence present | Artifact path/URL exists |
| No unresolved markers | `Select-String "TODO |
Privacy/Safety
- No sensitive data in completion evidence
- Artifact paths use relative or workspace paths
- No credentials in task logs
Self-Use Trigger
Use when:
- Starting any multi-step task
- Resuming after interruption
- Handoff to another agent
Finish what you start. Prove it with evidence.