feat(external_websocket_sync): mirror ACP elicitations to Helix and accept answers

[lukemarsden]

Companion to helixml/helix#3009, which pins this branch via ZED_COMMIT in sandbox-versions.txt. Merge this one first.

crates/external_websocket_sync/ mirrors the Zed agent thread into Helix, but its entry mapping in thread_service.rs matched only UserMessage | AssistantMessage | ToolCall and _ => returned everything else. acp_thread::AgentThreadEntry::Elicitation was silently dropped, so when Claude Code asked the user a question through AskUserQuestion — which claude-agent-acp turns into an ACP form elicitation that blocks the turn — Helix saw nothing answerable and the turn eventually died with Tool use aborted.

What this adds

• SyncEvent::{ElicitationRequested, ElicitationResolved, ElicitationResync, ElicitationResponseAck} with their to_outgoing_message() arms, and one elicitation_status_str() mapping Zed's Canceled to the wire's cancelled so the two spellings cannot spread.
• Emission from AcpThreadEvent::ElicitationRequested / ElicitationResponded and the EntryUpdated arm, with the requested_schema serialized verbatim — Helix renders the options, so nothing may be lossily flattened.
• A heartbeat plus a full resync on reconnect / open_thread, so Helix can tell "the agent still holds this question" from "the process that held it is gone" without guessing from the reconnect itself.
• A respond_elicitation command: GLOBAL_ELICITATION_RESPONSE_CALLBACK + pending queue, dispatched to a GPUI task calling AcpThread::respond_to_elicitation, which already no-ops safely on an unknown id or a non-Pending status. The ack reports accepted / noop / not_found rather than failing silently.
• E2E Phase 18 covering egress and resolution, driven through the synthetic seam so a CI phase never depends on the model choosing to call a tool.

This deliberately does not touch agent_ui: the Zed agent panel's elicitation card renders with no option controls, and it stays that way — Helix is self-sufficient.

Testing

The full dockerized suite is green with Phase 18 (single zed-agent round):

[zed-agent] Phase 18: ✅ question recorded pending with a 363-byte schema carrying both options
[zed-agent] Phase 18a: ✅ respond_elicitation delivered to Zed
[zed-agent] Phase 18a: ✅ Zed's ack round-tripped: question is now cancelled (agent_no_longer_holds)
[zed-agent] Phase 18b: ✅ question is terminal (accepted) and no longer answerable
[zed-agent] Phase 18: ✅ a normal turn still completes after the question cycle
  E2E TEST PASSED

Adding a phase also pushed a zed-agent round to ~190s, past run_e2e.sh's DEFAULT_TIMEOUT=300 * AGENT_COUNT; that budget is raised to 450s/round, which matters for CI since it uses the default.

Not verified: the two-agent matrix (E2E_AGENTS="zed-agent,claude"). The zed-agent round passes; the claude round stalls at Phase 1 — before any elicitation code runs — inside the runtime-installed claude-agent-acp startup. Believed unrelated, but unverified. No live end-to-end run has been done yet either; that is in progress and both PRs should wait for it.

🤖 Generated with Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[nessie993]

Superseded by the merged transport in #95 and the merged Helix implementation in helixml/helix#3220. Closing to keep one maintained implementation path.