harness-candidate-pr-loop

name	harness-candidate-pr-loop
description	Run the full candidate harness cycle: worktree isolation, invariant validation, benchmarking, comparison, artifact generation, and draft PR preparation.

harness-candidate-pr-loop Skill

Purpose

Execute the full slow-loop candidate cycle without merging anything:

create worktree → apply bounded plan → validate invariants → benchmark baseline/candidate → compare → record artifacts → prepare PR

Steps

1. Create an isolated candidate worktree with worktree.py.
2. Apply the reviewed MutationPlan in that worktree only.
3. Run invariant validation:

   python ${CLAUDE_PLUGIN_ROOT}/scripts/harness_core/validate_invariants.py \
     --harness-dir $CLAUDE_PROJECT_DIR \
     --mutation-plan ${CLAUDE_PLUGIN_ROOT}/state/mutation-plan.json
   

4. Run baseline and candidate benchmarks with benchmark_runner.py.
5. Compare them with compare_candidates.py.
6. Persist the decision in harness-run-ledger.sqlite.
7. Render HARNESS_EVOLUTION_STATE.md, OPEN_BREAKS.md, and the PR body template.
8. Optionally open a draft PR with gh pr create.

Hard Stops

• Never merge or auto-merge.
• Abort if held-out regressions appear.
• Abort if invariant validation fails.
• Abort if worktree isolation is not active.

Required Artifacts

• ${CLAUDE_PLUGIN_ROOT}/state/mutation-plan.json
• ${CLAUDE_PLUGIN_ROOT}/state/harness-report.json
• ${CLAUDE_PLUGIN_DATA}/harness-run-ledger.sqlite
• ${CLAUDE_PLUGIN_DATA}/run-artifacts/