self-improve

name	self-improve
description	Repo-aligned evolutionary improvement loop with tournament selection. Generates multiple candidate plans, implements and verifies them, then keeps the best result. Requires explicit approval gates for iteration.

Self-Improve — Tournament-Style Improvement (Repo-Aligned)

Generate multiple candidate improvement plans, implement them one at a time, verify against eval criteria, and keep the best result.

Hard Compatibility Rules

• Follow AGENTS.md lifecycle and Approval Gate. No silent long-running loops.
• No external state directories, no implicit git worktrees/branches. Keep changes in the current working tree.
• Iteration is bounded:
  • Max 2 compilation/test fix retries per repo rules
  • Max 3 candidate plans per iteration
  • Max 2 iterations unless the user explicitly approves more

When to Use

• User asks to “optimize”, “improve”, “iterate until good”, or “run self-improve”
• eval-harness has defined objective evidence (commands + pass/fail criteria)

Protocol

Step 0 — Preconditions

• Confirm objective, scope boundary, and acceptance criteria exist (prefer eval-harness output).
• Confirm the files allowed to change.

Step 1 — Produce Candidate Plans (N=2..3)

For each plan:

• One testable hypothesis
• Specific file targets
• Verification commands (reusing eval-harness)
• Expected risk level and rollback (what to revert if it fails)

Step 2 — Execute One Plan at a Time

For each candidate, in order:

• Implement minimal changes for the hypothesis
• Run verification commands
• Record outcome as PASS/FAIL with evidence

Step 3 — Select Winner

• Winner is the plan that best satisfies acceptance criteria with lowest risk.
• If all candidates fail, produce a concise root-cause analysis and propose a re-plan (do not keep stacking fixes).

Step 4 — Iteration Gate

If acceptance criteria not met:

• Propose exactly one next iteration (new hypothesis or re-plan)
• Ask the user to approve continuing (no infinite loops)

Output Format

• Baseline evidence (if available)
• Candidate list (hypothesis + files + verification)
• Result table (candidate → PASS/FAIL → evidence)
• Winner summary (what changed and why)