ultracode在长上下文的时候不适配openai接口

[kexinoh]

如图....

[OnlyTerp]

Fixed in integration branch integrate/open-prs-and-issues (PR #16):

Tool-only assistant turns: OpenAI-compat backends that enforce strict JSON schema now get content: null (not empty string "") when an assistant message has only tool_calls and no text. This matches what strict upstreams expect.

Context-length 400s: When upstream returns errors mentioning context length / maximum context, the proxy now attaches a clearer hint in logs/responses pointing at session compaction, shorter prompts, or switching to a larger window route (including [1m] where configured).

If you're still hitting failures after updating, please paste the exact upstream error text (or a redacted screenshot) and which backend route/model you had selected — that'll help confirm whether it's the tool-only case, a true overflow, or something else.

[devin-ai-integration]

Already resolved on main (landed via PR #16): openai_compat tool-only assistant turns now send content: null for strict-schema backends, and upstream context-length errors get a clearer hint (compaction / shorter prompt / larger-window [1m] route). This behavior is verified by the offline 'openai_compat long-context hygiene: tool-only content=null + context hint' self-test in test_proxy.py, which passes in CI on #28. Closing as fixed; please reopen with the exact upstream error text + selected route if you still hit it.