Steering the development cycle¶
Choose a coherent slice: an outcome you can verify and resume independently. Do not create one task per file or per phase.
Stages and their outputs¶
| Stage | Action | Expected output |
|---|---|---|
| Research | Explore sources and test hypotheses | Question, facts, options and unknowns |
| Plan | Define scope, criteria, ownership and verification | Executable plan, dependencies and permissions |
| Implement | Make the smallest coherent change | Explained diff, preserved work and targeted tests |
| Test | Run checks on the identified candidate | Commands, results, environment and limits |
| Judge | Compare evidence, requirements and standards | Verdict, findings, applicable score and independence |
| Iterate | Correct a defect or resolve a prerequisite | Bounded correction and evidence to rerun |
The coordinator chooses the next step and finalizes traceability. Reuse research or a plan that is still valid. Do not restart the whole cycle on every resumption, but do not skip a necessary check.
Develop with rapid feedback¶
For executable behavior: 1. Write or reuse a test at a public interface. 2. Observe a failure that actually reproduces the defect (red). 3. Make the minimal correction and observe success (green). 4. Refactor without changing behavior, then rerun tests. 5. Move to the next behavior and relevant regressions.
Expected results come from the requirement or an independent oracle, not the tested code. A documentation-format check does not prove agent behavior.
Read each stage's report¶
The final report always uses these seven fields, in order: 1. step: the phase actually performed. 2. modified files: changed paths, or N/A when none changed. 3. next step: the recommended phase, loop again, blocked or done. 4. current goal: the short objective addressed. 5. next goal: the next objective or missing prerequisite. 6. closed tasks: confirmed closed task IDs, or N/A. 7. explanation: what was done and why, evidence, limits and next gate.
Labels remain fixed; values are written in the user's language. The explanation retains relevant Goals/Non-goals, criteria, checks, verdict and session details. A blocked result or unchanged phase still uses all seven fields.
A proposed commit has no hash. A recommended session has not been launched. The report distinguishes modified files from files actually committed.
Judge and correct without changing the rules¶
Freeze the candidate and criteria before judgment. Tests are evidence, not acceptance. Required but unavailable independence remains a limit. Choose the scoring mode before implementation; a requested numerical score does not disappear because the task is small.
Correct the most precise defect, then rerun its checks and regressions. Keep evidence that remains valid; invalidate evidence whose inputs have changed. Do not lower a threshold or remove a test to obtain success. The default budget is three authorized attempts; waiting and diagnosis do not consume an attempt. When blocked or out of attempts, stop and prepare the handoff. Close only the scope actually verified and authorized.