Sub-roadmap: DAG runtime phase 3 — worker execution, heartbeat, and evidence-gated completion

[devkade]

Sub-roadmap: DAG runtime phase 3 — worker execution, heartbeat, and evidence-gated completion

Parent roadmap: #167
Related: #168, #169, #188, #191
Runtime-first sequence: phase 3 of 4
MVP inclusion: yes — this is the MVP completion boundary

Summary

Define and implement the minimum runtime execution loop: provision workers, dispatch claimed tasks, monitor heartbeat, collect worker output/evidence, and require evidence before task completion.

This is the MVP boundary for the runtime-first roadmap. Once this phase works, Ilchul has a real parallel runtime foundation that can accelerate verification and learning work.

Goal

Ilchul should be able to run ready/claimed tasks through worker sessions and complete tasks only when evidence is present.

Scope

• WorkerState model.
• Worker provisioning through existing local substrate where available.
• Adapter/substrate boundary usage from Design: define portable agent adapter and execution substrate contracts #188.
• Readiness nonce or equivalent readiness proof.
• Dispatch of claimed tasks to workers.
• Heartbeat tracking.
• Worker status: ready, busy, unhealthy, completed-retained, safe-to-close, stale-registry, cleanup-released, closed.
• Structured worker report contract.
• EvidenceRef collection from worker report, logs, test output, diff refs, or artifact refs.
• Task transition to verifying/completed only with evidence refs.
• Operator-visible worker/task status.

Non-goals

• No automatic merge/integration.
• No full RewardRecord calibration loop.
• No policy simulator-driven worker selection unless already available as a dry-run hint.
• No destructive cleanup of retained worktrees or tmux sessions.

Acceptance criteria

• WorkerState is defined and persisted.
• Worker readiness is recorded.
• Claimed task can be dispatched to a ready worker.
• Worker heartbeat is tracked.
• Missing heartbeat marks worker unhealthy or stale according to policy.
• Structured worker report is captured.
• Evidence refs are required for task completion.
• Completed-retained and safe-to-close states are distinguishable.
• MVP smoke run can execute at least two independent tasks in parallel through isolated workers or a documented local substitute.

Verification

• Unit tests for worker status transitions.
• Unit tests for evidence-required completion.
• Integration test with fake adapter/substrate.
• Smoke test with local substrate if available.
• Fixture showing parallel task execution through completion.

[devkade]

Design decision

needs-design is resolved by #167's Development Order.

Decisions:

• Sub-roadmap: DAG runtime phase 3 — worker execution, heartbeat, and evidence-gated completion #196 is the runtime MVP completion boundary.
• Worker execution must use the Design: define portable agent adapter and execution substrate contracts #188 adapter/substrate boundary and should be testable first with a fake adapter/substrate.
• MVP requires worker registry/state persistence, dispatch of claimed ready tasks, heartbeat tracking, stale/unhealthy projection, structured worker report capture, EvidenceRef extraction, and evidence-required task completion.
• Completed-retained and safe-to-close states remain distinct.
• Worker reports and captured output are evidence sources, not completion authority by themselves.
• No automatic merge/integration, reward learning loop, or destructive cleanup is in scope.

Implementation can proceed as Step 5 of #167. Exit gate is a smoke run executing at least two independent tasks in parallel through fake workers or a documented local substitute and completing only with evidence refs.