The Loop Auto

Purpose

Provide the complete bundled fallback for one bounded automatic v0.1 mission. Auto is an attended-authority execution mode, not a supervisor, queue selector or endless monitor.

Use when

Use when the user has declared exactly one asset, outcome, done gate, frozen budget and authority scope and wants the lifecycle to continue without confirmation only across gates already covered by that authority.

Required inputs

• Validated run ID or run-creation authority, one declared asset and immutable track.
• One objective, exact done gate, frozen budgets and bounded mission lifetime.
• Visible current grant, owner, lease expectations and stop controls.
• Source or repository state, current evidence, issue ledger and retry limits.
• Capability catalog and bundled fallback coverage for every required stage.

Reject an open-ended objective, undeclared queue, missing kill switch, missing budget or authority whose scope cannot be shown.

Procedure

1. Create or load one bounded run; display asset, outcome, track, authority, expiry, budgets and stop controls.
2. Validate the event chain, reconcile an unmatched pending operation without replay, repair projections and acquire the required lease.
3. Before every stage and action, run the ordered safety preflight and halt on any failed gate.
4. Route to a qualified installed provider or the complete bundled fallback, persisting the decision first.
5. Execute Strategize, required Spec pack, one Build slice, Test and bounded Resolve cycles.
6. Use durable pre-callback intent for every side effect. On an unknown local outcome fail; on an unknown outward outcome wait for human verification. Never replay it.
7. Emit stage-boundary heartbeats and account for time, attempts, mutations, outward actions and cost.
8. Stop at green, a configured gate, exhausted budget, scope change, authority need, stop signal or retry limit, then execute truthful Close.
9. Auto must not select unrelated work, extend its own lifetime or invent a new mission.
10. Auto must not enter an unbounded monitor or work-selection loop. An empty or completed mission terminates with a handoff.

Output contract

At each boundary return run ID, mode auto, declared asset, track, stage, owner and lease; authority warning; budget usage; provider and fallback reason; artifacts, evidence and issues; last completed action; pending effect; current gate; and next stage or halt.

The final result is complete only on passed evidence and zero blockers; otherwise use the exact truthful state and reason.

Evidence gate

Every crossed gate must already be covered by visible authority and frozen scope. Persisted evidence, not self-report, advances stages. A no-third-party mission remains fully operable through bundled fallbacks. Unexecuted required checks block green.

Track requirements

For the code track, establish repository identity and isolation before unattended edits and keep remote Git or deployment actions gated. For the non-code track, freeze audience, format, sources, factuality and review method and keep sending or publication gated.

Safety and authority

Auto cannot elevate itself, add work, install a provider or weaken audit, evidence, ownership, leases, the kill switch or faithful failure. A present or indeterminate stop prevents callback entry. Removing a stop never resumes work without explicit recovery.

Optional delegation is capability-gated. If unavailable, execute the same bounded packet serially or inline. A watcher may report state but remains read-only and cannot become the supervisor or recovery actor.

Self-refutation

Before continuing, ask whether the next action still serves the same declared asset and done gate, whether current authority covers it and whether evidence or issue state contradicts progress.

Halt conditions

Halt on approval need, inaccessible external state, stop signal, lease conflict or expiry, namespace change, budget exhaustion, reopened issue, three red Test/Resolve passes, scope expansion, unknown effect or a human security, legal, privacy, money or destructive decision. Never turn the halt into monitoring or select replacement work.

Core execution does not require network access or the source checkout. The links below are optional public documentation.

References

• Stage contracts
• Skill routing
• Code and non-code tracks
• Autonomy policy
• Run state and leases
• Issue ledger
• Evidence contract
• Workflow dispatch
• Watcher contract
• Harness capability map