Skip to content

Deepen completed Harness session interpretation - #458

Merged
JacobStephens2 merged 3 commits into
mainfrom
issue-457
Oct 6, 2026
Merged

JacobStephens2 merged 3 commits into
mainfrom
issue-457

Conversation

@JacobStephens2

@JacobStephens2 JacobStephens2 commented Oct 6, 2026 •

Copy link
Copy Markdown
Owner

Summary

Failed Harness turns now fail the Run even when the CLI exits zero. One consuming interpretation owns retained-data authority, completion reporting, and validation for sessions and Muse/OpenCode Model/Effort checks.

adapter.interpretation(worktree, prompt)
  condense(stream lines)
  finish(execution)
    recover optional Muse records / OpenCode export
    return { optional report, validated outcome }
Sessions
  print report once
  use successful final message and Resume evidence

Closes #457

Evidence

  • Before: a_zero_exit_failed_claude_turn_leaves_the_pr_draft_without_a_resume failed: the Run returned 0 and reported its PR ready after Claude emitted error_max_turns.
    After: the regression passes: the Run fails, leaves the PR draft, and starts no Resume.
  • The direct Model/Effort-check regression preserves exactly interrupted; check-export interruption also stops descendants before any work.
  • Interpretation and fake-CLI coverage verifies all six Harnesses, both retained-text policies, failure/diagnostic precedence, recovery reporting after BrokenPipe, Session log bytes, and interruption during OpenCode export.
  • Validation: cargo fmt --check, cargo check, and cargo clippy --all-targets -- -D warnings passed. cargo test passed on the final implementation.
  • thirdshift-code-review against main: addressed the Standards finding by sharing duplicated test setup and the Spec finding by preserving interruption wording in captured checks.

Test seams

  • Harness interpretation: construct through the adapter, condense recorded lines, finish after execution, and assert the completion report and outcome.
  • Run delivery and Resume: fake Harness CLIs verify failed turns cannot become ready pull requests or start a Resume.
  • Model/Effort checks and process ownership: fake Muse/OpenCode CLIs verify refusal diagnostics, retained acquisition, interruption, and descendant cleanup.

Unaddressed findings

Standards

None. The duplicated test helper finding was addressed.

Spec

None. The captured-check interruption wording finding was addressed.

Merge Danger

Door: two-way

Revert restores the previous interpretation implementation; no persisted-data migration is needed.

Blast Radius: Harnesses

Completion behavior covers all six Harnesses and Muse/OpenCode preflight checks.

Built with codex · gpt-6.1-sol · xhigh

@JacobStephens2
JacobStephens2 merged commit 8dd33a8 into main Oct 6, 2026
13 checks passed
@JacobStephens2
JacobStephens2 deleted the issue-457 branch October 6, 2026 22:18
This was referenced Oct 7, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Deepen Harness sessions behind one completed interpretation

1 participant