[JacobStephens2]

Summary

Keep a Session's completion-bearing reply when a later background-task notification produces a reply without a completion line. Claude and Grok share this selection, so every existing completion-line reader receives the same preserved reply. Later completion lines still win.

root assistant/result reply
  remember the whole reply when its last line is a completion line
completed turn
  return the remembered reply, falling back to the latest result

Closes #1405.

Evidence

• Before: At main (06f08b4c), cargo test --bin thirdshift harness::interpretation_tests::completion_line_survives_background_notification -- --exact failed with exit 101: both Grok and Claude returned the background-notification reply instead of PR ready.\nBug reproduced: fixed.
  After: The same command passed (1 test). The retained test also passes in the final interpretation suite.
• Before: The Spec review's exact command, cargo test --bin thirdshift harness::interpretation_tests::assistant_completion_line_survives_a_late_result, failed with exit 101 on the initial fix: the earlier completion line appeared only in an assistant message, while the late reply supplied the single result.
  After: The same command passed (1 test), and its regression test remains in the change.
• Confirmed cause: Result events overwrote final_message; earlier assistant text was available to trace recording but was ignored by reply selection. The shared interpreter now remembers completion-bearing root assistant and result replies. Payload validation, terminal outcomes, and child-message filtering remain with their existing owners.
• cargo test --bin thirdshift harness::interpretation: 89 passed, 1 ignored.
• cargo test --test bug_run: 14 passed, 1 ignored. Both Claude and Grok replay the single-result notification pattern for reproduced/not-reproduced endings; missing completion lines retain the exact existing error.
• cargo fmt --check and git diff --check: passed.
• scripts/affected-tests main: passed, including all 50 site tests and the full Rust suite. Nextest was unavailable, so the selector used cargo test; /usr/bin/time -p measured the Rust run at 1,373.54 seconds wall time, including compilation.

Test seams

• Harness Interpretation: raw Claude/Grok records through condense/decode and finish to Ended.final_message and require_final_message.
• Completion-line readers: the Ticket run command through Delivery diagnosis, and interpreted replies through Security review parsing.

Review coverage

Standards: 0 findings. Spec: 1 finding, reproduced and fixed; its regression test was kept. Both reviewers inspected every changed file, including the follow-up fix. Unread changed files: none. Generated-output exclusions: none.

Unaddressed findings

Standards

None.

Spec

None.

Security

None.

Security review outcome

No unaddressed introduced findings.

Merge Danger

Door: two-way

Revert the code change to restore prior reply selection.

Blast Radius: Sessions

Changes which reply completion-line readers receive from Claude and Grok when background notifications cause extra replies.

Built with codex codex-cli 0.164.0 · gpt-6.1-sol · xhigh