feat(conversations): take queued messages into a running Claude or Codex turn

[blakeandrewwood]

Messages sent while a run is in flight still go to the queue. Now a run whose provider can take input mid-turn takes them in at a tool boundary, instead of only after the turn ends.

• Claude: chat turns take the prompt on stdin (--input-format stream-json --replay-user-messages), and stdin stays open for the whole turn. Queued messages are written as user lines, and the replayed echo confirms each one was taken in.
• Codex: chat turns run codex app-server over stdio instead of codex exec; agent runs keep exec. Setup is initialize, then thread/start or thread/resume. The prompt and each queued message go as turn/start, which joins the active turn when there is one. On a first turn the prompt waits for the thread id from thread/started. A userMessage item confirms each message was taken in. Existing exec thread ids resume fine.
• Engine: while a tool is running, the run pops the queue and writes each message into the turn. A message queued when no tool is running stays queued, and it goes as the next turn as before. When a message enters, the reply so far and the user message are saved in order, and a new ConvEvent.user tells clients to refetch. The run keeps reading until every message it wrote has been answered, then closes stdin.
• OpenCode: unchanged, since it queues until the run ends. Steering there needs a long-running opencode serve.

Checked against the real CLIs (Claude 2.1.281, Codex 0.154.0) by hand, and end to end in e2e tests with mock-claude standing in for both providers.

🤖 Generated with Claude Code