[JacobStephens2]

Summary

--trial request.json
  validate protocol v1 and an isolated output folder
  shared Spec review session → one Resume if background work was killed
  JSON outcome, timings, cause, session ID and artifact paths → stdout
  Session traces, Model evidence and review reports → output_dir

Uses the existing Session prompt and Factory skills without a Command record.
The protocol and caller preparation requirements are documented in docs/trial.md.
Code review remains the follow-up in #1016.

Closes #1001

Test seams

• The thirdshift --trial <request.json> CLI, with a prepared repository and fake Harness executables: version rejection, outcome and timings, session ID, Model/Effort, report preservation, Resume and interruption; filesystem observations verify factory-log isolation, output reuse protection and cleanup.

Evidence

• Before: the acceptance test failed with not a GitHub issue URL: --trial; focused regressions exposed missing failed-session IDs, inherited ownership and leftover Factory-skill temporary files.
  After: cargo test --test trial passes all 11 Trial tests (one fixture entry point ignored), including Claude and Codex, both reports, unknown versions, Void causes, interruption and exactly one Resume.
• Before: the Spec reviewer's exact regression returned session_id: null after an observed Harness ID and interruption.
  After: cargo test --test trial an_interrupted_trial_keeps_an_already_observed_session_id -- --exact passes and returns fake-session-1; the regression is retained.
• cargo test --test background_work: 4 passed, one fixture entry point ignored. cargo check: passed.
• scripts/affected-tests issue-977 passed all 43 browser checks. Its duplicate local full Rust run was stopped after successful CI covered the exact same head, as docs/testing.md permits; Rust test-binary compilation took 3m 41s. No local full-suite completion is claimed.
• CI attempt 2 passed the full Rust suite, formatting and Clippy on 772251c; site and Console checks also passed. The initial Linux attempt hit ExecutableFileBusy starting a copied binary in the unchanged Console-install fixture. Its exact test passed locally, and the failed job passed on retry without code changes.

Merge Danger

Door: two-way

Blast Radius: sessions

Shared session storage now distinguishes an ownerless Trial from a failed Command-log allocation. Failed Harness sessions retain their ID; normal Run launch and Resume policy use the same implementation.

Review coverage

Independent Standards and Spec reviews used issue-977 at d05aa971f1e050efbc88744404c75d1433d9d5c4. Standards S1 (typed protocol states) and Spec P1 (interrupted-session ID) were addressed in 772251c and confirmed in follow-up reviews.

Both axes read all 11 changed files. Unread changed files: none under Standards and none under Spec. Generated-output exclusions: none.

Unaddressed findings

• Standards: none.
• Spec: none.

Built with codex · gpt-6.1-sol · xhigh