Repository navigation
A reviewer turn ends when its verdict exists, and reviewers are told where the worktree is - #10684
Conversation
…the worktree is
Two of the three harness gaps the first live review run exposed (2026-09-06, attempt eca90231 on shell-typed-invocation). (1) The reviewer's turn ended only on the provider's stop reason, so a reviewer that wrote its verdict at round 19 kept spending rounds until the budget refused and its unit reported failure over a verdict the belt accepted. `HarnessTurnConfig` now carries `completion: HarnessTurnCompletion` — `CompletesOnProviderStop` for the worker, `CompletesWhenEntryPresent { directory, name }` for the reviewer and auditor — observed after each tool round by a successful listing of the verdict directory (`extdeps.filesystem.filesystem_io` `filesystem_entry_presence`, never a failed read), and `TurnCompletedByArtifact` exits zero with its own label and `turn.completed` event; an indeterminate listing is a counted `completion.indeterminate` event and the turn continues. (2) Reviewer and auditor guidance said the tool runs "in the candidate's worktree" without naming it, and every slow reviewer spent rounds guessing directories; `ToolRunsInTheCandidateWorktreeAt { path }` now carries the path, as the worker's `WorktreeAt` already did. The belt passes `verdict_dir` and `verdict_name` to the reviewer and auditor CLIs, which derive the path, so the completion subject and the guidance share one authority.
Not in this change: an isolated verdict slot per reviewer (the how-to-work reviewer read a sibling's verdict); the shared review directory stays and is a separate change.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019iXPtAqzPaNBZbaaM5rxrY
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e3bd8cab5c
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| match harness_write_events(config: config, events: stepped) { | ||
| HarnessEventsUnwritable { detail: d } => harness_event_write_refusal(config: config, detail: d) | ||
| HarnessEventsWritten => | ||
| HarnessEventsWritten => match harness_completion_observe(completion: config.completion) { |
There was a problem hiding this comment.
When an artifact-gated reviewer or auditor emits a terminal response without a tool call—either immediately or after a failed verdict-write—the loop takes the list_length(items: uses) == 0 branch before reaching this observation and returns TurnProviderStopped, which harness_turn_exit maps to exit 0. Consequently, CompletesWhenEntryPresent is not actually required: the belt retains a supposedly successful unit with no verdict and refuses to respawn it. The provider-stop branch should observe the configured artifact and fail when it is absent.
Useful? React with 👍 / 👎.
1 participant
What
Two of the three reviewer-harness gaps the first live review run exposed (2026-09-06, attempt
eca90231onshell-typed-invocation: two verdicts of six, four reviewers spent 20 rounds guessing directories or continued after writing).HarnessTurnConfig.completion: HarnessTurnCompletion—CompletesOnProviderStop(worker, probe) orCompletesWhenEntryPresent { directory, name }(reviewer, auditor). After each tool round the loop observes presence via a successful directory listing (filesystem_entry_presence), returnsTurnCompletedByArtifact(exit 0, labelcompleted-by-artifact,turn.completedevent with the artifact path). Indeterminate listing → countedcompletion.indeterminateevent, turn continues. The recursion is factored intoharness_turn_continueso the observation has one continuation.ToolRunsInTheCandidateWorktreeAt { path }replaces the unnamed premise in both guidance modules;harness_reviewer_guidance_at/harness_auditor_guidance_attake the worktree.gunbc_harness_review_argvand the audit argv passverdict_dir+verdict_name; the CLIs deriveverdict_pathfor the guidance and the seat actor, so the completion subject and the prompt share one authority.Evidence
Fresh build,
claim_batch --hermetic:harness_review_guidance_test2/2 (now asserts the worktree path is in the prose),harness_goal_audit_guidance_test3/3, newharness_turn_completion_test2/2 (artifact completion exits 0 while budget exhaustion exits 1; provider-stop completion is never reached).roadmap_belt_tick_clibelt_review_head_cliandharness_cliharness_reviewer_clitypecheck and run through to the wire with the new arguments.Not in this PR
An isolated verdict slot per reviewer (the how-to-work reviewer read a sibling's verdict from the shared review directory).
🤖 Generated with Claude Code
https://claude.ai/code/session_019iXPtAqzPaNBZbaaM5rxrY