[openai_responses] Continuous compaction loop on startup

[vrascal]

What version of Kimi Code is running?

0.29.2

Which open platform/subscription were you using?

self-hosted

Which model were you using?

step-3.7-flash

What platform is your computer?

Linux 6.17.13.2-pve x86_64 x86_64

What issue are you seeing?

In Kimi 0.29.2 (Linux x64), when using the openai_responses API type with a self-hosted model, Kimi continuously triggers compaction immediately after startup.

The application repeatedly performs compaction without any user interaction:

● Compaction complete (29477 → 266 tokens) (Ctrl-O to show compaction summary)

● Compaction complete (270 → 832 tokens) (Ctrl-O to show compaction summary)

● Compaction complete (836 → 621 tokens) (Ctrl-O to show compaction summary)

This makes Kimi unable to enter the normal interactive state.

The same self-hosted model works correctly with Codex, so the model endpoint appears to be compatible. The issue only occurs when using Kimi 0.29.2 with the openai_responses configuration.

What steps can reproduce the bug?

Steps to reproduce

1. Configure Kimi 0.29.2 to use a self-hosted model.
2. Set the API type to openai_responses.
3. Start Kimi.
4. Observe that compaction starts repeatedly without any user interaction.

What is the expected behavior?

Kimi should start normally, connect to the self-hosted model, and wait for user input.

Compaction should only happen when the context length reaches the configured limit, not continuously during startup.

Additional information

No response