Feature request: return explicit state for every wait_threads target

What version of the Codex App are you using?

Current Codex Desktop task tools exposing wait_threads with per-target threadId, optional hostId, and optional afterCursor.

What subscription do you have?

Not subscription-specific.

What platform is your computer?

Not platform-specific.

What issue are you seeing?

wait_threads wakes when the first requested target completes or needs attention, but it does not provide a stable contract that returns the explicit current state of every requested target.

For multi-target coordination, the caller needs to know which targets are queued, running, waiting for input, completed, failed, or unavailable after every wait. An omitted target or a compact progress entry is not enough to distinguish unchanged state from unknown state, and it encourages extra reads or accidental duplicate dispatch.

The wait response should contain a complete per-target state vector on every return, including when one target wins the wake and when the call times out.

A backward-compatible shape could be:

{
  "wake": {
    "reason": "turnCompleted",
    "threadId": "thread-a",
    "turnId": "turn-a-3"
  },
  "targets": [
    {
      "threadId": "thread-a",
      "hostId": "host-1",
      "state": "completed",
      "cursor": "cursor-a-3",
      "triggeredWake": true,
      "turnId": "turn-a-3"
    },
    {
      "threadId": "thread-b",
      "hostId": "host-1",
      "state": "running",
      "cursor": "cursor-b-8",
      "triggeredWake": false
    }
  ]
}

The exact state names are flexible, but they should form a documented closed set such as queued, running, needs_attention, completed, failed, and unavailable. Each requested target should appear exactly once, in request order, with its current cursor or revision and whether it caused the wake.

Expected behavior

• Every requested target appears exactly once in every successful wait_threads response.
• Each target has an explicit current state, even when another target caused the wake.
• Timeout responses still carry the complete state vector.
• Each target carries its latest cursor or revision so the next wait can advance without reconstructing state.
• Per-target failures are represented on that target as typed state/error data rather than only in a detached aggregate error list.
• Terminal final text can remain cursor-suppressed after delivery; state must not disappear with it.
• Existing top-level wake behavior can remain unchanged.

Why this is useful

A complete state vector makes one bounded multi-target wait sufficient for orchestration. The parent can resume only running targets, surface only attention-required targets, and retain terminal knowledge without polling every task separately.

It also prevents a missing or compacted target entry from being interpreted as idle, failed, or never dispatched.

Related issues

• #40397 requests exact turn correlation between send_message_to_thread and wait_threads. This request is complementary: it asks for complete current state for every target in a multi-target wait, whether or not exact turn targeting is also added.
• #26822 reports missing terminal child state in the separate collaboration-agent wait surface. This request is specifically for the first-party wait_threads task API and its per-target response contract.
• #24951 concerns timeout enforcement in multi-agent waiting. This request concerns the data returned when a bounded wait does return.

Start at the wait_threads task API and trace its response schema, wake behavior, and timeout path. Review related issues #40397, #26822, and #24951 for surrounding constraints. Done means every requested target has one explicit state entry, including on timeout, with cursor or revision and per-target failure information.