Finish the task. Account for what remains.
An unattended run can finish its code change and still leave an unknown remote effect, an unsettled charge or an unowned server. Use these local records to decide what can proceed and what needs reconciliation.
Free operations kit · No signup · v0.1 · October 1, 2026
Start with one task and the failure that costs you time today. Map the records onto your existing scheduler, tracer, gateway and supervisor.
Choose the decision you need to make
Decide whether the task completed
The worker says done. Your scheduler has only an exit code.
Record the authorized postcondition, independent evidence and effects separately. Accept legitimate no-ops; reconcile missing or contradictory evidence before replay. Cost and retained work remain separate obligations.
Own the server you start
A port is busy, or the task leaves a service running.
Bind the workload to an immutable identity and generation. Check which session and build the endpoint serves. Record retention approval before lease expiry, and verify cleanup without touching another session.
Bound the work still being billed
Concurrent calls, retries or missing usage undermine a cost cap.
Specify reservation, settlement and unsupported-route behavior for your existing gateway. Keep uncertain charges reserved. The executable ledger uses abstract fixture units; it does not certify a provider or gateway.
Review what a plugin can reach
A configuration or plugin update changes the authority in use.
Separate declared, inferred and verified access. Include host-side execution, receiver permissions and mutable definitions. A worked update shows why a valid signature does not approve a new write grant.
Hand over the evidence
The trace, file state and receiver acknowledgement disagree.
Correlate the tool claim with independent observations and receiver evidence. Preserve missing intervals and transient effects. Share an allowlisted packet with minimal payloads and an explicit next action.
Preserve the contract when adapters change
A model, provider, tool schema or fallback route changes.
Keep terminal denials, safe transient retries and unknown outcomes distinct. Hold unsupported schemas. A generic error message cannot replace an enforced boundary or a machine-readable recovery contract.
For structured handoffs, download the blank receipt envelope (.json) and the fictional worked envelope (.json). These are proposed local records, not a universal standard or the lab's input schema. Blank fields are unknown; worked examples are fictional and NOT RUN.
Run the reference lab
The lab executes fixed receipt, adapter, retry and budget contracts, plus a bounded Docker lifecycle experiment with two synthetic sessions. It checks endpoint identity, a busy port, a detached child and cleanup scoped to one session. Lease and lost-acknowledgement decisions are simulated and labeled separately.
- Download the executable lab (.py)
- Download the method and results (.md)
- Read the finding: Done is not a receipt
Python 3.10+; standard library only. Run with --contracts-only for the synthetic contract checks. The full path needs Docker and a pre-pulled python:3.12-slim image. It creates and removes its own bounded containers, uses no host port or credentials, and calls no model or paid provider. Inspect the script before running it.
Keep the evidence scope visible
The lab verifies its reference contracts and synthetic workloads on one configuration. It does not establish your client's behavior, distributed budget enforcement or adversarial-model resistance. The inventory and incident packet are review templates, not an automated scanner or redactor.
For a wider deployment review, use the harness boundary guide and its 24 separate NOT RUN scenario specifications. For an uncertain remote action, use the reconciliation walkthrough.