feat(runner)!: submit() continues a stored instance, and a kept run is left up for its submitter

[AlexS-scale]

Second step towards task run and eval run always going through the configured [runner] (discussed on #236). Two CLI options only worked in-process because the Runner contract couldn't carry them.

Resume (--context-json + --start-step). Runner.submit() takes an optional instance_id. Given one, the run continues that instance from its stored context, at metadata["start_step"]. A caller resumes by naming the instance rather than by shipping a context, so contexts stay out of runner payloads (Temporal limits one to 2 MB).

• LocalRunner seeds the run from the stored context, then lays the submitted metadata (per-run overrides, keys) over it.
• With no task_version given, it runs the version the instance records, so a newer version never runs over an older instance's context.
• An instance that records another task or version is refused: the run fails without running.
• Task.run now drops the recorded failures of the steps it is about to run again. A resumed run whose failed step now passes therefore ends COMPLETED, under any runner.

--keep. A run whose metadata[KEEP_KEY] is set and that ends on its own, passed or failed, is left up. Its submitter tears it down, for example from the instance's stored context with teardown_run. A cancelled run is torn down either way: what a step in flight deployed is only in the runner's live context.

Breaking: Runner.submit gains a parameter, so an override with the old signature no longer matches. LocalRunner is the only implementation in this repo.

Tests: a kept run that completes is left up, and one that is cancelled is torn down. A run given an existing instance, with no version, gets that instance's stored context, its id, start_step and the stored version. One given an instance of another version fails without running. A run drops the recorded failures of the steps it runs again.

Confidence Score: 5/5

The PR appears safe to merge; the previous version-mismatch finding is fixed.

Summary

Adds stored-instance resume and lets submitters keep resources after a run finishes.

• The runner can continue a stored task instance from its saved context.
• Kept runs leave their deployments up when they finish on their own.
• Rerunning a step clears its earlier failure from the run result.

The latest change rejects mismatched tasks and versions. No new actionable issues were found. Tests were inspected, not run.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A["Runner.submit"] --> B{"Stored instance?"}
  B -->|Yes| C["Use recorded version unless supplied"]
  C --> D{"Task and version match?"}
  D -->|No| E["Fail before running"]
  D -->|Yes| F["Restore context and apply overrides"]
  B -->|No| G["Create context"]
  F --> H["Task.run"]
  G --> H
  H --> I{"Finished on its own with keep?"}
  I -->|Yes| J["Leave resources up"]
  I -->|No| K["Tear down resources"]

Loading

_{Reviews (3) · Last reviewed commit: "fix(runner): a resume refuses an instanc..." · Reviewed by Greptile}