[malibio]

Summary

• Deletes the held-off-menu-type decisions scenario and the scoring only it used (heldToOfferedTypes, afterRefusal, createdRecord, two summary lines, their tests). Keeps the control's unheld/held tests.
• Why: the held menu is bug-report, checkbox, task, text (the built-in Implementing a Task skill links task), so the request never tested the hold. Rewording to person (absent from the menu in 3/3 reps) fails the other way: the model never names it, searches and updates the bug report instead, 0 calls refused, so the pass asserts nothing. Not reworded further to win a rep.
• Refusal of an off-menu call stays covered by held-off-menu-node and the off_menu_type unit tests.

Measured (gemma-4-e4b-q4km, isolated daemon, 3 reps, DB reset between)

• held-on-menu-type: 3/3 on this branch and on main.
• held-off-menu-node: 0/3 on this branch and 0/3 on unchanged main (an update_node writing the seeded spec's signed-off date runs with type_refused=false off_menu_ran=false). Not caused by this change, but the issue's criterion that it "still passes 3 of 3" does not hold; it needs its own look.

test:changed passes.

🤖 Generated with Claude Code

[malibio]

Code Review Summary

Verdict: APPROVE (comment-only, self-authored PR). Net positive: removes a scenario whose premise was measured false, along with the scoring only it used. The diff is a clean deletion (21 additions, 215 deletions across decisions.ts and decisions.test.ts).

Checked:

• No dead code left. Grep for held-off-menu-type, heldToOfferedTypes, afterRefusal, createdRecord finds nothing in scripts/, packages/ or the docs checkout. persisted, turnHeld, offMenuCallRan, notHeld and CLARIFICATION_OPENER all still have live callers (held-off-menu-node, the control, extra).
• Comment accuracy. The replacement comment in the fixture list matches the measurements in the PR and issue (menu bug-report, checkbox, task, text; the person rewording was never named). It records why the scenario is gone, which is what the issue asked for.
• Tests. Control tests (unfenced -> "nothing here was measured", held -> pass) are kept and simplified; the refusal-scoring tests went with the code they covered.
• Acceptance criteria.
  • Criterion 1 (request names a type absent in 3/3) is superseded by the issue's own fallback ("if no such request keeps the turn held and the type off the menu, delete"). Fair deviation: person was off-menu but never named, so it can't assert a refusal.
  • Criterion 2 is met via "removed with the reason recorded".
  • Criterion 3 is only half met (see below).
  • Criterion 4: PR states test:changed passes.

Findings

Critical Issues

None.

Suggested Improvements

• Acceptance criterion 3 (held-off-menu-node passes 3 of 3) is not met, and the PR is honest that it is 0/3 on unchanged main too. That is the right call for scope (not caused by this change), but this PR now removes the other scenario that exercised a refused off-menu call end to end, and the comment in decisions.ts says refusal is "covered end to end by held-off-menu-node". While that scenario fails 0/3, the claim is true only in the sense that it exists. Please make sure the pre-existing failure has its own tracked issue (the PR body says "needs its own look" but does not link one) so that the end-to-end refusal coverage this PR leans on is not silently red. Per the team's fold-in preference, it is also acceptable to say in the PR why it is not folded here (a different scenario, a different cause: update_node ran with type_refused=false off_menu_ran=false, which may be a seeding or menu issue rather than scoring).
• Do not close held-off-menu-type's type is on the held menu, so the scenario does not test the hold #3693 as fully satisfied without noting the unmet criterion 3 on the issue; closing with a link to the follow-up is enough.

Nitpicks

• Nit: the tombstone comment in decisions.ts is five-plus lines of history in a list of live fixtures. It is useful as a "don't re-add this" guard, but consider trimming to the conclusion (a create outside the linked set can't be tested with the current skill set) since the measurements live in the PR/issue.

[malibio]

Code Review Summary

Verdict: APPROVE (comment-only, self-authored PR). Net positive: removes a scenario whose premise was measured false, along with the scoring only it used. The diff is a clean deletion (21 additions, 215 deletions across decisions.ts and decisions.test.ts).

Checked:

• No dead code left. Grep for held-off-menu-type, heldToOfferedTypes, afterRefusal, createdRecord finds nothing in scripts/, packages/ or the docs checkout. persisted, turnHeld, offMenuCallRan, notHeld and CLARIFICATION_OPENER all still have live callers (held-off-menu-node, the control, extra).
• Comment accuracy. The replacement comment in the fixture list matches the measurements in the PR and issue (menu bug-report, checkbox, task, text; the person rewording was never named). It records why the scenario is gone, which is what the issue asked for.
• Tests. Control tests (unfenced -> "nothing here was measured", held -> pass) are kept and simplified; the refusal-scoring tests went with the code they covered.
• Acceptance criteria.
  • Criterion 1 (request names a type absent in 3/3) is superseded by the issue's own fallback ("if no such request keeps the turn held and the type off the menu, delete"). Fair deviation: person was off-menu but never named, so it can't assert a refusal.
  • Criterion 2 is met via "removed with the reason recorded".
  • Criterion 3 is only half met (see below).
  • Criterion 4: PR states test:changed passes.

Findings

Critical Issues

None.

Suggested Improvements

• Acceptance criterion 3 (held-off-menu-node passes 3 of 3) is not met, and the PR is honest that it is 0/3 on unchanged main too. That is the right call for scope (not caused by this change), but this PR now removes the other scenario that exercised a refused off-menu call end to end, and the comment in decisions.ts says refusal is "covered end to end by held-off-menu-node". While that scenario fails 0/3, the claim is true only in the sense that it exists. Please make sure the pre-existing failure has its own tracked issue (the PR body says "needs its own look" but does not link one) so that the end-to-end refusal coverage this PR leans on is not silently red. Per the team's fold-in preference, it is also acceptable to say in the PR why it is not folded here (a different scenario, a different cause: update_node ran with type_refused=false off_menu_ran=false, which may be a seeding or menu issue rather than scoring).
• Do not close #3693 as fully satisfied without noting the unmet criterion 3 on the issue; closing with a link to the follow-up is enough.

Nitpicks

• Nit: the tombstone comment in decisions.ts is five-plus lines of history in a list of live fixtures. It is useful as a "don't re-add this" guard, but consider trimming to the conclusion (a create outside the linked set can't be tested with the current skill set) since the measurements live in the PR/issue.

[malibio]

Addressed: shortened the tombstone comment (nit). Skipped: filing a follow-up for held-off-menu-node 0/3. It fails the same on unchanged main and its cause is not yet known (an update_node on the seeded spec ran with type_refused=false off_menu_ran=false), so it is recorded on #3693 for the owner to triage rather than as a guessed issue. No re-review needed: comment-only change.