Codex Desktop on macOS: image-input turns take 130s cold / 45s warm before a trivial response #33018

What version of the Codex App are you using?

• Codex Desktop release: 26.707.72221
• Bundled Codex CLI: 0.144.2

What platform is your computer?

• macOS 26.5.2
• Apple Silicon (arm64)

What issue are you seeing?

Image-input turns in Codex Desktop have very high time-to-first/final response, even when the prompt explicitly asks not to analyze the image and only return a fixed acknowledgement.

A controlled test used one PNG:

• Dimensions: 1448 × 1086
• File size: 1,200,858 bytes (~1.2 MB)
• Prompt: 图片链路测试，不分析图片内容，只回复“附件接收成功”。
  • English: “Image-pipeline test. Do not analyze the image content. Only reply ‘Attachment received successfully.’”

Results in the same thread:

1. First image turn immediately after fully restarting Codex:
   • Completed successfully
   • Duration: 130,737 ms
2. Second turn with the same image and same prompt:
   • Completed successfully
   • Duration: 45,373 ms

The ~85-second reduction suggests a significant cold-start component, but the warmed request still taking ~45 seconds for a fixed one-line response is unexpectedly slow.

Test thread ID: 019f5ff5-113b-7091-be41-ae835789ef74

Local timing evidence

For the warmed second turn:

• Desktop accepted turn/start in approximately 3 ms
• The turn completed without an error
• No tools were called
• No unknown conversation error was logged for this test
• The delay occurred after turn/start and before the final assistant response

This suggests the bottleneck is after local turn submission, likely in image upload/validation, multimodal preprocessing, upstream queueing, or time to first model output. The Desktop logs do not expose enough server-side phase timing to identify which stage dominates.

Text-only turns after the same restart responded normally.

What steps can reproduce the bug?

1. Fully quit and restart Codex Desktop on macOS.
2. Create a local task.
3. Attach one ordinary PNG (~1.2 MB, 1448 × 1086).
4. Use a trivial prompt that requests a fixed one-line acknowledgement and no image analysis.
5. Measure the turn duration.
6. Repeat the identical image/prompt in the same thread.

Observed: ~130.7 seconds on the cold request and ~45.4 seconds on the warmed request.

What is the expected behavior?

A small, single-image turn with a fixed acknowledgement should begin and complete promptly. If multimodal initialization or upstream queueing is slow, the UI should expose a phase-specific status instead of looking stuck.

Additional information

Possibly related to the broader pre-first-output observability problem described in #24260, but this report is specifically reproducible with image input on macOS and shows a large cold-versus-warm latency difference.

Useful diagnostics would include timestamps for:

• local image accepted
• upload started/completed
• image validation/preprocessing started/completed
• upstream request sent
• response headers / first byte received
• first model event received