[App regression] “Add to chat” disappears for selected response text while the response is streaming

Open 💬 1 comment Opened Aug 8, 2026 by Kl-11

What version of the Codex App are you using (From “About Codex” dialog)?

Codex App 26.803.41515 (build 6321)

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Darwin 25.6.0 arm64 arm

What issue are you seeing?

When text is selected from an assistant response while that response is still streaming, the contextual selection toolbar omits Add to chat.

During generation, the toolbar still displays:

  • More details
  • Ask in side chat

As soon as the response finishes, selecting the same response text displays:

  • Add to chat
  • More details
  • Ask in side chat

This prevents the established response-review workflow at exactly the time it is useful: reading an answer while it is being produced and collecting several connected annotations for one follow-up.

Previously, each selected passage could be added to the existing composer draft without sending anything. The user could collect multiple response annotations, add ordinary text, review the complete follow-up, and press Send once. If the current turn was still active, that one composed follow-up entered the normal queue without interrupting the active turn.

The current behavior is not an intermittent text-selection failure. Selection works and the other contextual actions remain available; only Add to chat is suppressed based on response completion.

What steps can reproduce the bug?

  1. Start a Codex task that produces a response long enough to remain visibly streaming.
  2. While the response is still streaming, select text that has already appeared in the current assistant response.
  3. Observe that the selection toolbar shows More details and Ask in side chat, but not Add to chat.
  4. Wait for that response to finish.
  5. Select the same text again.
  6. Observe that Add to chat is now present.

What is the expected behavior?

Add to chat should remain available while an assistant response is streaming.

The complete intended workflow is:

  1. Select a passage from the streaming response.
  2. Choose Add to chat to add a structured response annotation to the current unsent composer draft.
  3. Continue reading and collect additional annotations in that same draft.
  4. Add or edit ordinary follow-up text if desired.
  5. Press Send once.
  6. If the active turn has not finished, queue that single combined follow-up normally without interrupting it.

Selecting or confirming an annotation must not automatically send it. Multiple annotations should remain collected together until the user deliberately presses Send.

Additional information

Inspection of the installed production client shows that the response target already exists during streaming:

  • The in-progress agentMessage is given a stable responseAnnotationTargetId.
  • The response renderer nevertheless uses that target only when item.completed is true.
  • While streaming, it therefore omits the response-annotation DOM attributes and the Add to chat callback.
  • Once completed becomes true, the same target is exposed and Add to chat appears.

This indicates that the missing action is caused by a client-side completion guard, not by the response lacking an annotation target. Removing that guard would restore the previous composer-and-queue workflow.

The exact first affected version has not been isolated. The same completion guard is present in the immediately preceding inspected build, 26.730.61639, so this report does not claim that 26.803.41515 introduced it.

Related, but not duplicate:

  • #25641 — the entire selection toolbar intermittently fails for some selections.
  • #33763 — existing response annotations are omitted from Copy as Markdown.
  • #37554 — PDF and document annotations submit immediately instead of batching.

In this issue, the selection toolbar is present and functional; Add to chat alone is deterministically hidden until the selected response completes.

View original on GitHub ↗

1 Comment

Allmight97 · 18 days ago

I've had this same issue Feedback ID: 019fe6ee-6dd7-72f3-be2f-b28b2f5001aa

Feature is really handy to have for streaming thinking summaries when Im' actively monitoring.