[Juliusolsson05]

PR-4 of the conditions-framework stack — integrates Claude's AskUserQuestion picker with the modular conditions system and fixes the interrupted-question ghost render. Targets integration/conditions-framework.

Pairs with claude-code-headless#26 (gitlink bumped 9939c1c → a0eeb74), which makes the picker a first-class, deduped condition in the unified snapshot.

1. Ghost-render fix — the reported bug

"a question being interrupted might still render even tho we are like 4 messages past"

An AskUserQuestion tool_use blocks the agent until a tool_result lands. If the user interrupts instead of answering (Esc / sends a new prompt), no result is ever emitted, so the block stayed resultAt == null forever and BlockRow kept routing it to the live, clickable picker — it ghosted in the feed indefinitely.

foldEvent.ts turn_stopped now stamps resultAt on pending AskUserQuestion blocks. This is the transcript-backed dismissal — not screen-derived, so it can't flicker the way a screen-snapshot gate would. Scoped strictly to toolName === 'AskUserQuestion': every other pending tool can still resolve with a late result across a stop (stamping them would drop real output); AskUserQuestion is unique in that it cannot resolve after an interrupt.

2. Shared-union entry

claude.ask-user-question (+ ClaudeAskUserQuestionState / …Option) — the app-side mirror of the headless parser's shape, on this side of the submodule import wall.

3. Minimal renderer consumer

A 'QUESTION' dispatch attention badge when the live condition is present (selectors.ts). Purely informational — it cannot touch the inline picker's render gate (stays semantic, !resultAt) or its answering path, so it can't regress the working single-select flow.

Deliberately deferred to PR-5

The sendThenReparse answering driver (multi-select / free-text / multi-question) and the answerability gate that closes the documented "answered-via-terminal → stray digit" race. Those are coupled to the driver, not to this PR; building them here would half-wire the driver or risk regressing the working path.

Why no screen-snapshot shortcut

Presence/dismissal of the inline row stays semantic (the tool block's resultAt); the screen-derived condition is a secondary signal (badge now, answering affordances in PR-5). That split is the entire point of the framework over the old snap.askUserQuestion direction.

Verification

tsc -b (web + node): no new errors — 15 pre-existing failures in unrelated subsystems (dictation, pidusage, ghostJournal, …), all three touched files clean. Headless tsc --noEmit clean. Vendor-sync of conditions-core: zero drift.

⚠️ Merge order when authorized: claude-code-headless#26 first (with --merge, to keep the gitlink commit reachable), then re-point the gitlink if its merge SHA differs, then this PR.

🤖 Generated with Claude Code