closedloop-technologies Parallel Coding Agent Loop
Use this skill to coordinate independent Attempts for the same Run objective.
Inputs
- Run goal
- Active policy
- Attempt count
- Base branch and commit
- Required evaluation commands
- Budget limits
Procedure
- Create a ClimbHill run with the shared goal.
- Register one attempt per planned attempt with a distinct branch name.
- Give each attempt the same goal, policy snapshot, and required checks.
- Keep attempts independent until after evaluation.
- For each attempt, attach patch path, branch, head commit, cost, and evaluation records.
- Compare attempts only after policy and evaluation results are recorded.
- Recommend promote, reject, or recombine based on evidence.
Outputs
- Run ID
- Attempt IDs
- Branch plan
- Evaluation matrix
- Recommendation
Rubric
- Attempts are independent enough to compare.
- Every attempt has recorded policy and evaluation status.
- Costs are recorded or explicitly marked unknown.
- No attempt is promoted without human review.