Reply checker: one block names every unproven sentence at once (part 3 on PR 416)

[EdbertChan]

Summary

The reply checker now lists flagged sentences from every paragraph in one block, so Claude can address them in one rewrite.

Previously, it stopped at the first claim without proof. Fixing that paragraph could reveal another block with the same message.

Each numbered item quotes a sentence, capped at 120 characters, and names its trigger. Shared guidance explains how to add proof or a valid tag.

Review Claim

One block reports a flagged sentence from every paragraph lacking proof; tagging both paragraphs in the two-claim repro clears the next check.

Review Lane

behavior

Review Unit

engine-runtime

Safety Invariant

Replies that passed still pass, and replies that were blocked remain blocked. Claim detection, evidence exemptions, marker validation, and word-limit behavior stay unchanged; only claim diagnostics expand.

Slice Rationale

Part 3 stacks on PR #416, targeting stack/EdbertChan/cat-unverified-tag/2-retry-checks-evidence. That PR checks rewrites; this slice changes what a block reports. The hook and its regression tests form one review unit.

Alternative considerations: the user requested a separate stacked PR; a count alone would not tell Claude which sentences need proof.

Assumptions: scope and safety invariant follow the supplied workflow. Detection remains paragraph-based, with one reported match per paragraph.

Non-goals

No changes to claim rules, tag format, evidence scope, word limits, rewrite-check activation, or the ship-claim hook. This does not enumerate multiple claims within a single paragraph.

Architecture

Before

Detection returned the first offending phrase and stopped.

After

find_unverified_claims() collects a trigger and sentence from each flagged paragraph, and main() formats the list into one block.

The existing find_unverified_claim() interface still returns the first phrase or None, preserving its callers.

Test Plan

Test Plan

• python3 engine/hooks/diu-stop/tests/test_hooks.py — 81 tests, OK, exit 0. Covers both sentences appearing together, both tags clearing a rewrite, only the remaining sentence being reported, and 120-character truncation.
• python3 engine/hooks/diu-stop/tests/test_fix_matrix.py — 5 tests, OK, exit 0.
• python3 engine/skills/make-pr/scripts/preflight.py --base origin/stack/EdbertChan/cat-unverified-tag/2-retry-checks-evidence — exit 0; output:

unit engine-runtime: 2 file(s)
declare Review Unit: engine-runtime
gate python3 scripts/check_hook_test_coverage.py engine/hooks/diu-stop
check_hook_test_coverage: OK (1 hook(s) checked)
ok preflight passed

• bash scripts/scrub-handoff-artifacts.sh — scrub-handoff-artifacts-ok, exit 0.

Revert Plan

Revert Plan

• Safe to revert? Yes; restores first-claim diagnostics while retaining PR [CAT-UNVERIFIED tag](2) A rewritten reply still gets checked for proof #416's rewrite checks.
• Revert command: git revert 3c6948d.
• Post-revert steps: rerun both Python test commands above.
• Data migration? No.

---

Note

Low Risk
Claim detection and pass/block outcomes are intended to stay the same; only block-message diagnostics change, aside from redundant duplicate imports that add no behavioral risk.

Overview
The Claude Stop hook’s unverified-claim feedback now reports every flagged paragraph in a single block instead of stopping at the first trigger phrase.

find_unverified_claims() walks the message in order, still using the same paragraph-level detection rules, and returns a (trigger, sentence) pair per hit. Each quoted sentence is extracted around the match and truncated to 120 characters so stderr stays readable. find_unverified_claim() remains a thin wrapper over the first hit for existing callers. When the hook blocks, stderr lists numbered sentences plus their triggers and asks for proof or a {{CAT-UNVERIFIED}} tag per paragraph, so one rewrite can address all outstanding claims.

Tests cover multi-claim blocks, partial fixes on retry (stop_hook_active), and sentence truncation.

Note: The diff also duplicates the sys.path setup and import markers block in claude_stop_check.py (lines 56–64); that looks accidental and worth cleaning up separately.

^{Reviewed by Cursor Bugbot for commit 60e9619. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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_0464ffb6-9477-4885-a9c7-3d184f27fabd)

[EdbertChan]

Mergify repair stopped: lowest open stack PR #433 is based on stack/EdbertChan/cat-unverified-tag/2-retry-checks-evidence, which still belongs to open PR(s) #416 outside this stack; leaving it alone to avoid dropping dependency changes

[EdbertChan]

Mergify repair stopped: required check failed: test. The retry cap was reached for current head 0a07aff.

[mergify]

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

• Queue this pull request

[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_0e324690-ddc2-481e-a83b-cd23fe6ac01a)