plat 479

← coding-agents / claude

PLAT-479 — Clarification choices as cards for every attended coding chat, with Claude's native AskUserQuestion (PR 270)

Field	Value
State	deployed
Priority	-
Product	coding-agents
Area	claude
Summary	fixed on main (PR 270 plus its review fixes), needs a restart: selectable cards for clarification questions in every attended coding chat; Claude's native question uses the same lifecycle; verified live.

Coordination	Value
State	fixed on main; needs a rebuild and restart
Date	2026-10-04
Owner	coding-agent-bridge
Related	PLAT-473 (native tool policy), PR 270

Source

Workflow chat showed a coding agent's clarification options as static text under an agent-specific label. PR 270 renders selectable cards in the workflow and chat transcripts, registers a shared request_clarification tool for attended, writable coding-agent chats, and routes Claude's native AskUserQuestion through the same answer lifecycle (answers validated and durably settled before delivery; cancelled or stale prompts close without assuming a choice). Muse's native delivery is unchanged.

Review (read-only, then fixed by the owner's session)

A reviewer found the lifecycle sound (answers validated, then durable, then delivered; two answers at once get a 409; a user cannot answer or read another user's prompt; scheduled, bot, child and read-only sessions do not get the tool) and these issues:

• Dependency pins pointed at unmerged branches (mcpagent a639bf5, provider b92093d). Fixed: the provider commit was rebased onto main (2ccfe00), the mcpagent commit merged with the fixes below (9b5da2b), and the builder pins both.

• Claude hook could fail the whole turn when python3 is missing or the bridge has no session credentials. Fixed: native questions are left off for that turn (AskUserQuestion not offered) and the turn runs.

• Hook scripts (which hold a session token) were never removed. Fixed: scripts older than 24 hours are swept.

• Option labels with surrounding spaces were denied (the server trims labels, the hook did not). Fixed: the hook trims them.

• A question lost in a restart stayed answerable in the clean conversation view (the transcript view already closed it). Fixed: one shared withClosedQuestions used by both views.

• CI: the new clarification prompt section was not declared in the three product.yaml files, which the product-surface contract tests check (AgentWorks and Crew failed). Declared in the AgentWorks, Crew and Code product files next to native-subagents; it still applies only to attended builder chats.

Done / verified

• Targeted Go tests (Clarification, ClaudeNativeQuestion, PromptSections) and 15 frontend tests pass; tsc -b clean.
• TestClaudeNativeQuestionLive against real Claude Code (warm session, then the hook answers a two-question AskUserQuestion; history survives): pass, 41 s.

Left

• After a page reload the closed set is empty, so a lost question shows pending once; one click closes it again. The server cannot settle a prompt it no longer holds.
• Not run: Claude in Full CLI as a per-user server account or under the Mac sandbox; Codex's tool timeout against the 30-minute wait; Muse and Cursor with request_clarification. Cursor's native AskQuestion is not blocked (no hook fires for it) and a headless Cursor can still stall on it.
• The header length is only in the tool schema, not enforced in parseCodingAgentClarification (low).

Register notes

PLAT-479, fixed on main (PR 270 plus its review fixes), needs a restart: selectable cards for clarification questions in every attended coding chat; Claude's native question uses the same lifecycle; verified live.