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, test, source, decision, or blocker: increment no_evidence_cycles.
   • new evidence: reset the relevant no-progress counter.
8. Run agent-collab policy evaluate after the cycle.
9. Obey PolicyDecision:
   • continue: revise the plan using the new evidence.
   • checkpoint: save resumable state. For v2 scope=slice with auto_continue, use agent-collab checkpoint advance and continue the same authorized goal. No human override is needed for an ordinary eligible slice transition. For a v2 action checkpoint requiring context compaction, preserve evidence and authorization in a smaller linked packet, record measured active sizes, then re-evaluate before execution. Maintenance is not a human approval gate.
   • stop: obey its scope. An action stop prohibits repeating that action; the primary-agent may diagnose read-only or prepare an evidence-backed correction. A goal stop preserves the hard limit or human gate.
   • v1 decisions retain their original checkpoint/stop semantics until explicit migration; do not silently reinterpret an old record.

An infrastructure error is evidence of an error, not permission to retry. A retry requires the next policy evaluation to permit it.

State

Write .coord/par_.yml:

schema_version: 2
goal: "..."
policy_ref: "${AGENT_COLLAB_POLICY}"
checkpoint_ref: ".coord/task-checkpoint.json"
acceptance_criteria:
  - "..."
critique_source: "..."
cycles:
  - cycle: 1
    plan_summary: "..."
    artifact_refs: ["..."]
    evidence_refs: ["..."]
    verdict: "pass | fail | error | needs-human"
    next_action: "..."
final_status: "pass | checkpoint | stop | error | needs-human"

Write .coord/par__final.md with:

• final status
• last PolicyDecision
• acceptance evidence
• unsuccessful attempts
• unresolved risks
• human decision required, if any

Both files are scratch by default. Promote only explicit shipping or acceptance evidence.

Memory

Never write a lesson directly to canonical memory. Create a proposal under .coord/memory-proposals/ with evidence references. Applying it requires a recorded human approval and appends a new event; it never edits an older event.

Invariants

• Evaluate after every cycle and before every spawn.
• Preserve cumulative usage and failure history across slices, sessions, and executors. Unknown tokens/cost remain unknown, never zero. Explicit goal limits and native platform limits remain hard; absent limits are not invented.
• Use stable failure identities based on operation, target, relevant inputs, and error class. Renaming a task or switching executors is not a correction.
• A waiting external service is not a failed retry. Continue only with new evidence, a safe next step, and the required acceptance checks.
• Diagnose recoverable action failures and perform safe context maintenance before escalating. Do not ask the user to renew unchanged authorization. Count actual human intervention separately from automatic recovery or waiting; fewer pauses never justify bypassing a real gate or claiming unmeasured success.
• Agent self-critique is not independent acceptance.
• Human semantic gates cannot be replaced by an aggregate agent score.
• PASS requires cited acceptance evidence, not “looks good”.