[Bug]: Orchestration message pointer is typed + submitted over the user's unsent prompt draft #22651

Activity

1. AmethystLiang commented on Sep 24, 2026

   

   AmethystLiang

   on Sep 24, 2026 – with Linear Code

   Contributor

   More actions

   Very valid. Looking into improving this for composer that we can detect draft
2. soonJ817 commented on Sep 30, 2026

   

   soonJ817

   on Sep 30, 2026

   More actions

   Still reproducible on 1.4.217 (macOS, Claude Code coordinator).

   The pointer is written to the PTY and a separate \r follows ~500 ms later (enterTimer in deliverPendingMessages), so an unsent draft is submitted together with the pointer. In our runs the trigger is almost always worker heartbeats.

   Workaround that avoids it today: keep a waiter whose --types includes heartbeat (a loop running check --wait --types worker_done,escalation,question,heartbeat that acks heartbeat-only deliveries). Pointer delivery skips types that have a live waiter, so nothing is typed into the prompt. Heartbeats that arrived before the loop got delivered_at set (pointer typed); after it they were consumed silently.

   Gap: between waits (right before/after processing a delivery) the pointer can still land on a draft. A draft check before writing (as proposed above), or keeping heartbeats out of pointer delivery (#14910 / #15173), is still needed.
3. ixiion001 commented on Oct 6, 2026

   

   ixiion001

   on Oct 6, 2026

   More actions

   The same happens through the CLI, not only with the orchestration nudge (1.4.220, macOS, Claude Code coordinator):

   1. The user types ZEBRA half line into the agent's prompt and does not press Enter.
   2. orca terminal read --screen --json shows it: .result.terminal.draft = "ZEBRA half line".
   3. orca terminal wait --for tui-idle returns satisfied: true at once.
   4. orca terminal send --text "…" --enter is accepted, and the agent receives one prompt: the user's half line and the
      sent text joined. Twice in two tries.

   A suggestion: wait --for tui-idle not satisfied while draft is non-empty (or a --for empty-draft condition), and
   terminal send refusing, or holding, while a draft is present unless --force. Until then we read draft before
   every send.
4. Race428 commented on Oct 7, 2026

   

   Race428

   on Oct 7, 2026

   More actions

   Reproduced on 1.4.221 with a Claude Code coordinator. The pointer was submitted together with the user's draft, so the agent received a single turn reading:

   You have 1 orchestration message. Run `orca orchestration check --run run_…`.
   ok
   

   where ok was the user's unsent reply to an earlier question. In this Run most pointers were for heartbeat or status ("Setup succeeded for worker …") deliveries, i.e. nothing the coordinator needed to act on.