Run released workflows with the current run, resume, revise and verify (#271) - #289

Refs #271. This is the Workflows family, part 2a. Part 1 (#287) froze v0.2.0 and v0.3.0 workflow stores written by each release's own CLI and read them with the current readers. This part runs the current run, resume, revise and apply --verify over those same stores. The issue stays open.

What

Tests only: no fixture bytes and no source under src/ change.

• How it runs. Execution goes through WorkflowCoordinator.run, the call the durable job's worker makes; the hermetic suite starts no job. Fake codex and kimi-code node agents record their prompts. The workflow run / resume CLI is tested up to a stubbed job launch.
• run of released planned, created and revised rows (a revision leaves its parent runnable):
  • each node gets the stored goal and its own stored prompt, never the other node's task; the final node also gets its dependency's output;
  • calls_used continues from the released count;
  • the result is a current final artifact, approved, whose patch and bundle digests are recomputed from the final node's stored bytes;
  • every released workflow file (planner output, spec) and event is kept as written.
• resume against the stored absolute deadline. The clock is fixed from the row's own deadline_at, because three of these rows were stored without a heartbeat.
  • After the deadline: every crash and failure ends workflow deadline exceeded before any dispatch, with the deadline unchanged.
  • Before it, crashes: a dead owner's run and a recovered one reuse the released draft bundle and refund the killed node's attempt (3 → 4 calls). The released events continue in sequence.
  • Before it, a published final artifact: it is reused and approved, and v0.2.0's is not backfilled with final_agent. It is refused, without approving, replacing or re-running anything, when its patch bytes, plan digests, source fingerprint, node, workflow or repository disagree.
  • Before it, exhausted nodes: nothing is dispatched, and the run ends no runnable nodes remain before final_node completed.
  • Refusals: from completed, cancelled and planned, and while the released owner's pid lives. A refused takeover appends a failed resume execution after the released history and changes nothing else.
  • Output: the completion is judged by this invocation's own execution, even when the row's state or a later execution says otherwise.
• workflow revise: a v0.2.0 parent gains a lineage, its only changed column; a v0.3.0 lineage continues past its last version.
• The launch: the team peek follows default_team for a v0.2.0 row and not for a v0.3.0 row that recorded none. A final state is refused before any launch.
• apply --verify of the released final artifact, also after a prune:
  • v0.2.0 gets the gate's own allowlist only, with the verify: notice;
  • v0.3.0 also gets the final agent's list;
  • an unlisted variable never arrives, and the common ones do.
• Control plans enabled refuse a released workflow planned without one, before any agent or Toolgraph call.
• docs/architecture/persisted-contracts.md: the Workflows row names these.

No current reader or executor misread a released workflow.

Part 2b (not here; still "not yet" in the row). Each needs new release-written fixtures:

• control-plan workflows written by a release;
• a node bundle saved just before a crash, reclaimed on resume (the released crash fixtures have no bundle for the killed node, so this part covers completed-dependency reuse only);
• recovery of an interrupted workflow-origin apply.

Verification

• suite-run CLEAN; ruff check . clean.
• 24 injected regressions in workflow.py and cli.py each fail an intended assertion. They include:
  • the stored deadline reset;
  • no refund, or a refund for a durable failure;
  • the live-lease check, execution history or approval dropped;
  • the control-plan refusal removed, or an advisory call before it;
  • the lineage, team and verify-agent rules broken;
  • invocation() returning nothing or the latest entry;
  • the plan digests dropped;
  • each artifact-reuse guard removed;
  • the planner output truncated;
  • either node's task leaking into the other's prompt.
• Codex reviewed the design in two rounds before any code. The implementation went three rounds:
  • Round 1: NEEDS-FIX with 4 Major, fixed in 3e40cda.
    • The Toolgraph seam raised an exception that run_preflight swallows.
    • Digests, invocation and released files were unchecked.
  • Round 2: NEEDS-FIX with 2 Major and 1 Minor, fixed in 0cf71ad.
    • The invocation check from round 1 equated this invocation with the last entry.
    • The artifact-reuse guards were unexercised.
  • Round 3: unprimed, given the diff only. SHIP with no findings.

🤖 Generated with Claude Code