Codex Desktop: long chats stream slowly and the one-line thinking/activity summary goes stale (all models) #53478

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

26.1007.30912, build 20156, prod release channel. The app's update check reported up to date.

What subscription do you have?

Pro, $200

What platform is your computer?

macOS 27.0.1 (build 26A434) — Darwin 27.0.0 arm64 arm

What issue are you seeing?

In a long-running local Codex desktop chat, responses feel very slow, text streaming appears slow, and the single-line thinking/activity summary remains old rather than giving a useful indication of current work. The UI offers a thinking transcript, but it is unclear whether the displayed activity reflects what is currently happening.

I have observed this across all models I have tried, including Sol and Astra, versions 5.6 and 6, and Sol 6.1. This is not limited to one model. This cross-model behavior is user-reported; controlled comparisons have not been performed.

This makes it difficult to distinguish continued work, model latency, tool waits, and a stalled or stale UI.

What steps can reproduce the bug?

Observed scenario, not a deterministic clean-start reproducer:

1. Continue a long-running local project chat that has used shell commands, image inspection, background processes, and subagents.
2. Send follow-up instructions while the task is working.
3. Observe long waits between visible assistant updates, slow perceived streaming, and an old single-line thinking summary.
4. Compare the visible activity with saved session timestamps.

What is the expected behavior?

The displayed activity should accurately reflect current execution, or clearly show that it is the last available summary. Response text should stream as it arrives. The UI should make prolonged waits understandable and distinguish current work from stale activity.

Additional information

Read-only investigation on October 13, 2026 (Asia/Kolkata):

• Across eight turns with a saved first assistant message, the elapsed time from task_started to that saved message ranged from 8.0 to 77.9 seconds; median 37.2 seconds, mean 38.2 seconds.
• These are persisted message timestamps, not measured time to first token, rendering latency, or tokens per second.
• In the latest interrupted turn, the last visible assistant progress message was saved at 20:21:47 UTC on October 12. Tool results and model activity continued through about 20:27:08 UTC. The turn was interrupted at 20:35:34 UTC. This is approximately 14 minutes without another assistant progress message; it does not prove the backend was inactive for the entire interval.
• A follow-up user message was saved at 20:33:16 UTC, before the interruption.
• The latest model request recorded approximately 265,000 input tokens, mostly cached. Large context is a possible latency contributor, not an established root cause.

Possibly related reports: #51919 and #45702 concern stale thinking views or missing completed answers; #25779 and #21134 concern long-thread performance in Codex Desktop. This report concerns the main Codex task view and perceived latency; a common root cause is unverified.