Routing guard refusal quotes the words it matched

[EdbertChan]

Summary

A checker blocks local workers from publishing code when a managed workflow should run the work instead.

Its refusal showed only broad action labels. Authors could not see which words caused the refusal.

The checker discarded the matched text after it classified each action.

It now includes each match and nearby text in the refusal. Tests preserve the current decisions for past-work and non-code wording.

Review Claim

Approve refusals that show each matched word with nearby text without changing which requests the checker blocks.

Review Lane

behavior

Review Unit

engine-runtime

Safety Invariant

The checker uses the same matching rules and block-or-allow decision. Only the refusal text changes.

Assumptions: This safety invariant is unconfirmed. The final-only output requirement prevents a confirmation question.

Slice Rationale

This slice changes one refusal message and its data shape. A separate change can check the command that runs instead of request wording.

Non-goals

• No change to which requests are blocked or allowed.
• No new action words or changes to noun, negation, or hyphen handling.
• No change to the workflow advice in the refusal.

Architecture

Before

flowchart LR
    A["Worker request"] --> B["Classify publication actions"]
    B --> C["Action labels"]
    C --> D["Format refusal"]

Loading

After

flowchart LR
    A["Worker request"] --> B["Classify publication actions"]
    B --> C["Action labels, matched text, and context"]
    C --> D["Format refusal"]

Loading

Test Plan

Test Plan

• python3 engine/skills/make-pr/scripts/preflight.py --base origin/main — preflight passed for the engine runtime unit.
• python3 -m unittest discover -s engine/hooks/agent-routing-guard/tests -v && python3 scripts/check_hook_test_coverage.py engine/hooks/agent-routing-guard — 31 tests passed, and hook coverage passed.
• python3 scripts/check_no_new_comments.py --base origin/main — no new comments found.
• bash scripts/scrub-handoff-artifacts.sh — no handoff artifacts found.
• node scripts/validate-pr-body-local.mjs --body-file /tmp/routing-guard-refusal-pr.md --base main — PR body validation passed.

Revert Plan

Revert Plan

• Safe to revert? Yes.
• Revert command: git revert <merge-commit-sha>
• Post-revert steps: None.
• Data migration? No.

---

Note

Low Risk
Hook decision logic is untouched; only refusal formatting and the internal hit type change, with tests asserting unchanged classifications.

Overview
When the agent-routing-guard blocks an Agent spawn for publication wording, refusals now quote what matched, not only verb labels like commit, push.

publication_verbs returns PublicationVerbHit records (label, matched substring, ±40 char context). block_message and unchecked_message insert {hits} lines such as commit: "…" before the existing routing guidance. Block vs allow logic is unchanged—same regex and negation/noun rules; only stderr copy is richer.

Tests add publication_labels for label-only assertions, lock in a few edge-case verdicts, and assert each hit’s matched_text and window appear in the entrypoint refusal.

^{Reviewed by Cursor Bugbot for commit 58ebfaa. 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_5319b103-66a8-4009-8932-df2985c1429b)