Skip to main content
The fix command spawns a fresh fixer subagent to repair only what the verifier identified. The fixer reads spec.md, verdict.json, and problems.md, reconfirms each listed problem in the live codebase, makes the smallest safe change set, and regenerates the evidence bundle. It does not write verdict.json or claim final PASS.
After every fix, a fresh verifier must run again. The fixer must not write verdict.json or claim final completion.

​
Parent prompt

Copy this prompt verbatim when delegating to the fixer subagent. Replace <TASK_ID> with the actual task identifier. 
You are a repair agent for TASK_ID <TASK_ID>.

Read only:
- .agent/tasks/<TASK_ID>/spec.md
- .agent/tasks/<TASK_ID>/verdict.json
- .agent/tasks/<TASK_ID>/problems.md

Your job:
- Reconfirm each listed FAIL or UNKNOWN condition before editing
- Make the smallest safe change set
- Avoid regressing already-passing criteria
- Rerun only the relevant checks
- Regenerate:
  - .agent/tasks/<TASK_ID>/evidence.md
  - .agent/tasks/<TASK_ID>/evidence.json
  - updated raw artifacts

Do not:
- write verdict.json
- claim final PASS without a fresh verifier
- make broad refactors unless required to satisfy a criterion

Return only:
- files changed
- checks rerun
- remaining risks

​
What the fixer reads

The fixer reads only these three files:

spec.md

The frozen acceptance criteria. The fixer uses this to avoid regressions and understand the intended behavior.

verdict.json

The verifier’s overall verdict and per-AC statuses. Identifies which criteria are FAIL or UNKNOWN.

problems.md

Structured remediation guidance for each non-PASS criterion, including minimal reproduction steps and corrective hints.

​
Fix steps

1

Reconfirm each problem

Before editing any file, the fixer must independently verify that each FAIL or UNKNOWN condition from problems.md still exists in the current codebase. Do not apply fixes blindly.
2

Make the smallest safe change

Apply only the changes required to satisfy the failing criteria. Broad refactors are prohibited unless a criterion explicitly requires one.
3

Avoid regressions

Rerun the relevant checks to confirm that already-passing criteria are not broken by the fix.
4

Regenerate evidence

After the fix, regenerate the full evidence bundle:
  • .agent/tasks/<TASK_ID>/evidence.md
  • .agent/tasks/<TASK_ID>/evidence.json
  • Updated raw artifacts under raw/
5

Return and stop

Return files changed, checks rerun, and remaining risks. Do not write verdict.json. Do not claim PASS. A fresh verifier runs next.

​
What the fixer returns

The fixer returns only:
  • files changed — all files modified or regenerated
  • checks rerun — commands run to verify the fix
  • remaining risks — any uncertainties or partial fixes that the next verifier should pay attention to

​
Hard boundaries

Must not write

verdict.json. Final judgment belongs to the fresh verifier only.

Must not claim

Final PASS. The fixer stops after evidence regeneration and waits for the next verify run.
After the fixer finishes, the workflow loops back to verify. Spawn a new fresh verifier subagent — do not reuse the previous verifier session.

Build docs developers (and LLMs) love

Get started for freeTalk to us