fix(todo-continuation): detect text-based questions in last assistant message (fixes #5548)

[EvangelosMoschou]

Problem

The todo-continuation enforcer (boulder) was injecting a forced-continuation prompt whenever the model stopped with unfinished todos, even when the model had intentionally paused to ask the user for input via plain text rather than the question tool. The existing pending-question-detection only catches the question tool, so models that emit questions as text (e.g. "Please provide the preferred date and time?") got auto-resumed and convinced themselves to make a default decision on behalf of the user.

Root Cause

hasUnansweredQuestion in packages/omo-opencode/src/hooks/todo-continuation-enforcer/pending-question-detection.ts only checks for the question / ask_user_question tool invocations on the last assistant message. The Sisyphus prompt explicitly tells the model to "Ask ONE clarifying question" when intent is ambiguous, and most models answer with plain text rather than calling a tool — so this code path was the common case, not the edge case.

Fix

Extend hasUnansweredQuestion to also detect when the last assistant message's last text part ends with ? (after trimEnd()). The check is wired into the same call site in idle-event.ts (line 97), so the existing Skipped: pending question awaiting user response log line is reused and the countdown never starts.

Trade-off

This is a heuristic. A model that ends a sentence with ? unrelated to user input would also skip continuation. This is consistent with the existing question-tool heuristic (also best-effort) and can be tightened if false positives surface in practice. If users hit false positives, the easy escape hatches are:

• disabled_hooks: ["todo-continuation-enforcer"] to disable the entire hook
• /stop-continuation slash command to halt mid-session

TDD Evidence

• RED: text ending in ? → expected true, got false (bug reproduced in 3 of 4 new tests).
• GREEN: same input now returns true.
• Text NOT ending in ? → returns false (regression guard).
• Mixed text + tool_use → returns true (text question takes precedence after the existing question-tool check returns false).
• All 14 existing question-detection tests still pass.
• All 125 todo-continuation-enforcer tests pass.
• 79 pre-existing failures in the broader omo-opencode suite (fixEmptyMessagesWithSDK, executeSyncTask, plugin init performance) are unrelated to this change and exist on upstream/dev.
• bun run typecheck:packages reports the same pre-existing cleanup.ts: skippedPaths error as upstream/dev.

Files

• packages/omo-opencode/src/hooks/todo-continuation-enforcer/pending-question-detection.ts (+27/-0) — add lastTextEndsWithQuestion helper, extend hasUnansweredQuestion
• packages/omo-opencode/src/hooks/todo-continuation-enforcer/pending-question-detection.test.ts (+65/-0) — 4 new test cases

Fixes #5548

---

Summary by cubic

Prevent forced continuation when the assistant asks a question in plain text by detecting a trailing '?' in the last assistant message. Skips the continuation countdown the same as the question tool (fixes #5548).

• Bug Fixes
  • Extend hasUnansweredQuestion to treat the last assistant text ending with ? (after trimEnd()) as a pending question.
  • Heuristic may skip on unrelated ?, consistent with the existing tool-based check.
  • Add 4 tests (text question, multi-line, non-question, mixed text+tool); existing tests still pass.
  • Update packages/shared-skills/scripts/frontend-refs-manifest.mjs to whitelist aside.md and motherduck.md in designOriginals, fixing the provenance gate CI failure.

^{Written for commit 25ed87a. Summary will update on new commits.}