diu waits for the word-list answer and says it in the same turn

[EdbertChan]

Summary

When a reply uses wording the user has had to ask about, diu now says so before the turn ends, and the agent rewrites it.

The five word lists become one question for a background model. The Stop hook waits for the answer, sixteen seconds in a live run.

If no answer arrives in time, the turn ends as it does today.

Review Claim

diu-stop builds one judge question from every phrases/plain-words-*.json, waits for the answer, and blocks the turn with a message that names the list and quotes the reply.

Review Lane

behavior

Review Unit

engine-runtime

Safety Invariant

No question is asked unless the payload names a transcript file that exists, so a harness payload without one, and every test, calls no model: the full diu-stop suite leaves zero job files behind. A verdict whose quote is not in the reply is dropped. No answer within the wait ends the turn unblocked.

Slice Rationale

This PR reads the word lists added by the PR below it, so it stacks on #446.

Delivering the answer on the user's next message was the first shape, and it was wrong: Claude Code has no way to push a message into an idle session, so a verdict nobody reads is a verdict the user never sees. Waiting is the only way to show it in the same turn.

Claude Code only for now. Cursor and Codex keep the behavior they have.

Non-goals

• No change to llm-judge, its inbox, or the word lists themselves.
• No Cursor or Codex wiring.
• The word count and unproven-claim checks are untouched.

Architecture

Before

graph TD
    S["Stop hook"] --> W["word count"]
    S --> C["unproven claim"]
    W --> B["block the turn, or let it end"]
    C --> B

Loading

After

graph TD
    S["Stop hook"] --> W["word count"]
    S --> C["unproven claim"]
    S --> P["one question from every plain-words list"]
    P --> J["background judge: one model call"]
    J --> V["answer, or nothing in time"]
    V --> Q["drop an answer that does not quote the reply"]
    W --> B["block the turn, or let it end"]
    C --> B
    Q --> B

Loading

Test Plan

Test Plan

• python3 -m unittest discover -s engine/hooks/diu-stop/tests, on the branch rebuilt onto current main: Ran 97 tests ... OK, and the job folder held zero files before and after, with no model calls running
• Live end-to-end, one fake turn through claude_stop_check.py with its own state folder: reply "No hook decides differently. Preflight passes and the review unit is engine-runtime." gave hook exit=2 after 16s and the message diu: the last reply used wording the user has had to ask about (plain-words-internal-names): "Preflight passes and the review unit is engine-runtime."
• Same reply under the first, five-question shape: two of five answers quoted the lists instead of the reply. The one question now says the quote must come from the reply, and the code drops a quote that is not in it.
• uvx ruff check . --select E9,F: All checks passed!
• check_no_new_comments.py, check_no_dated_provenance.py (both --base the diu word lists: five kinds of wording to avoid in replies #446 branch), check_hook_test_coverage.py engine/hooks/diu-stop: all exit 0
• python3 engine/skills/make-pr/scripts/preflight.py --base <the #446 branch>: exit 0, declare Review Unit: engine-runtime
• Stop hook timeout raised from 10 to 60 seconds to cover the wait. Claude Code's hooks guide gives command hooks a ten-minute default and states no cap; no test pins this hook's value.

Revert Plan

Revert Plan

• Safe to revert? Yes.
• Revert command: git revert <merge-sha>
• Post-revert steps: None. diu goes back to the word count and unproven-claim checks, and the word lists sit unread.
• Data migration? No

🤖 Generated with Claude Code

https://claude.ai/code/session_01Gnua5cuAAF1ey1b7zRqJCC

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_1d565321-990e-4a74-bc65-55e9b7c11f3a)

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_31dd4040-8250-4334-b58e-0c92f06e9c1b)

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request