agent-plan-act-reflect

# agent-plan-act-reflect Run a single-agent correction loop under the public policy/checkpoint contract. This differs from agent-debate: plan-act-reflect revises one candidate against evidence; debate compares genuinely consequential alternatives. ## Use this skill for - A task with a runnable or otherwise deterministic acceptance contract. - A candidate likely to need more than one evidence-producing cycle. - A bounded optimization, refactor, or draft correction. Do not use it for open-ended ideation, an unbounded “until perfect” request, or semantic acceptance that belongs to a human. ## Preconditions Require: - one concrete goal - acceptance criteria - a readable policy_ref - a valid checkpoint_ref - an identified critique source The policy is the only source for cycle, retry, context, and child limits. This skill does not define fallback numeric limits. If agent-collab-harness is unavailable, perform at most the currently authorized single action and return to the human. Do not emulate an autonomous loop with copied limits. ## Cycle 1. Validate the policy and checkpoint. 2. Evaluate policy before any delegated-executor or reviewer spawn. 3. Plan the smallest action that could add acceptance evidence. 4. Act within the declared scope. 5. Run the critique source. 6. Add evidence references and observed metrics to the checkpoint. 7. Classify progress: - acceptance satisfied: stop with PASS. - same failure: increment same_failure_retries. - no new artifact...