using-commitment-control-json

name	using-commitment-control-json
description	Use when executing a v2 cybernetic run whose only approved task authority is commitment.control.json.

Using Commitment Control JSON

Purpose

Execute an approved v2 commitment without turning the execution strategy into human-approved control data.

The authoritative run files are:

commitment.control.json   immutable after approval
progress.jsonl            append-only execution facts
evidence/                 content-addressed subjects and evidence
final-report.json         completion claim, never self-authorizing

Design notes, plans, work packages, file lists, command lists, subagent choices, and strategy snapshots are optional agent-owned working material. They cannot change the commitment or authorize a completion claim.

Read references/runtime-control-detailed-rules.md before recording a typed wait, receipt, evidence subject, final Gate, or successor transition.

Start

1. Run:

   python3 .agents/skills/using-commitment-control-json/scripts/commitment_runtime.py \
     validate <run-dir>
   

2. Read commitment.control.json completely.

3. Treat objective, obligations, non_goals, and authority as immutable.

4. Keep strategy agent-owned; record revisions and diagnose repeated unchanged failure fingerprints instead of stopping for the user.

Liveness

Continue automatically for implementation defects, failed approaches, stale working plans, artifact mutations, new counterexamples, and unexpected conditions that can be investigated with current authority.

Pause only in one of these typed states:

• WaitingForHumanDecision: commitment meaning genuinely needs a human choice.
• WaitingForAuthority: a declared live/external/destructive action needs a fresh matching grant.
• WaitingForExternalDependency: required external state is unavailable.
• ImpossibleUnderCommitment: evidence proves the commitment cannot be met.
• SupersededBySuccessor: a human semantic decision produced a separately approved successor commitment; this run is terminal and never resumes.
• AbortedByUser: the user explicitly stops the run.

There is no generic Blocked state.

commitment_runtime.py state returns a machine next_action, not a user handoff prompt. Continue agent-owned actions in the current turn; ask a user only when its action_owner is user.

Re-route ordinary direct work before a consequential claim or action when it discovers durable obligations, exact-subject evidence, non-goals, or authority.

Authority

For a declared action: enter the authority wait, obtain an exact fresh human grant, resolve that wait, then record the executed action with evidence. Actions are default-denied; never execute a forbidden action. Local reversible work is not consequential authority.

A runtime human semantic decision never resolves the old commitment. Form and approve a successor, then record run.superseded; execute only the successor.

Evidence

Create one role-aware commitment.evidence_subject for the exact candidate. Bind every obligation, declared path, and result place to evidence for that subject. A candidate/configuration/environment change creates another subject; never mix subjects.

Use the event appender so validation and file locking happen before write:

python3 .agents/skills/using-commitment-control-json/scripts/commitment_runtime.py \
  append <run-dir> --event-json '<json-object>'

Events are hash-linked. A final report binds the chain head, selected evidence event hashes, and exact final-Gate event hash.

Completion

Read ../_shared/counterexample-gate-protocol.md before executing the final Gate.

After all obligation evidence is bound:

1. execute the independent falsification Gate against the exact subject and completion claim;

2. store a valid commitment.final_counterexample_review artifact and append counterexample.review.completed after the evidence it reviews;

3. if a counterexample is established, continue agent-owned repair and rerun the Gate; otherwise write final-report.json using only evidence events for that subject;

4. run:

   python3 .agents/skills/using-commitment-control-json/scripts/commitment_runtime.py \
     verify <run-dir>
   

Claim success only when permits_goal_achieved is true. A failed Gate names invalidated evidence; a pass needs changed replacement evidence for the same obligation and exact final-report/Gate binding.