admin-bypass-sweep (4) an unanswered scope question resolves narrow

[EdbertChan]

Summary

One skill helps an agent merge many pull requests past the normal checks. First it must ask how much to merge, and get a fixed yes sentence.

The problem: in a real run, both questions went out in one message. The person typed only the yes sentence and said nothing about how much.

The agent read that silence as a yes to the bigger choice. It then rewrote five branches by force and merged two linked groups.

Both planned checks held. The gap was a third path: filling in the missing answer with the largest option.

The fix asks the two questions in separate turns. If the "how much" question gets no clear answer, the agent does only the smallest choice it offered.

Review Claim

When the person answers only the yes sentence, the merge skill does the smallest job it offered, never the biggest.

Review Lane

behavior

Review Unit

product-skill

Safety Invariant

A sweep of single-item groups reads the same as before; only a multi-item group needs the second answer, and an unclear offer resolves to the smallest one named.

Slice Rationale

This is one step of the sweep skill's intake: Step 2, where scope and consent are asked. Other intake steps ship as their own slices.

Non-goals

• Does not change the wording of the consent sentence.
• Does not add any other way to approve.
• Does not change Step 4's merge commands.
• Does not touch any other skill.

Test Plan

Test Plan

• python3 -m unittest discover -s product/skills/admin-bypass-sweep/tests -v

test_consent_only_reply_resolves_to_narrowest_scope (test_scope_intake.TestScopeIntake.test_consent_only_reply_resolves_to_narrowest_scope) ... ok
test_scope_answered_fixture_resolves_to_stated_scope (test_scope_intake.TestScopeIntake.test_scope_answered_fixture_resolves_to_stated_scope) ... ok
test_step_2_separates_scope_from_consent (test_scope_intake.TestScopeIntake.test_step_2_separates_scope_from_consent) ... ok

----------------------------------------------------------------------
Ran 3 tests in 0.000s

OK

• python3 engine/skills/make-pr/scripts/preflight.py --base main

unit    product-skill: 4 file(s)
declare Review Unit: product-skill
        ok      codify-has-code
        ok      no dated provenance
        ok      skills three-harness check
        ok      ecosystem boundaries
        ok      skill file refs
        ok      skill test coverage
        ok      skill trigger mechanism
        ok      skill trigger policy
        ok      subagent scope contract
        ok      all 17 scenario(s) behaved as declared
ok      preflight passed

• bash scripts/scrub-handoff-artifacts.sh

scrub-handoff-artifacts-ok

Fixtures added under product/skills/admin-bypass-sweep/tests/:

• scope_unanswered_consent_only_reply.md: the reply has only the consent sentence, so the fixture expects single-PR groups only.
• scope_answered_reply.md: scope is answered first ("Run the full grouped plan."), then consent comes in a separate turn, so the fixture expects the full plan.
• test_scope_intake.py: checks that Step 2 holds the separate-turn text and the narrowest-scope text, and that both fixtures state their expected result.

Revert Plan

Revert Plan

• Safe to revert? Yes
• Revert command: git revert <merge-sha>
• Post-revert steps: None. Reverting brings back the old Step 2 text, which does not say what to do with a reply that has only the consent sentence.
• Data migration? No

🤖 Generated with Claude Code

---

Note

Low Risk
Documentation and test harness for agent intake behavior only; no merge commands or runtime code paths change.

Overview
Step 2 intake for admin-bypass-sweep now splits scope from consent and defaults unclear scope to the smallest offered option.

The skill text no longer bundles scope confirmation with a vague reconfirm. The agent must ask only the scope question in one turn (narrowest choice first, then named stacks or the full plan), then request the STOP literal consent sentence in a later turn. If the human replies with only the consent sentence, that counts as authorization but not as choosing scope—execution stays at single-PR groups only; multi-PR stacks need an explicit scope answer naming them or a larger option.

Two markdown fixtures document the answered-scope vs consent-only paths, and test_scope_intake.py locks in the Step 2 wording and fixture expectations. Merge steps (Step 4) and the consent sentence itself are unchanged.

^{Reviewed by Cursor Bugbot for commit 9e2ef67. 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_9af8d706-d101-4092-9463-2ac70c282d88)