fix: catch vLLM InternalServerError for overlong prompts

[mikasenghaas]

Summary

• vLLM returns overlong-prompt errors as HTTP 500 InternalServerError instead of 400 BadRequestError
• Extended handle_openai_overlong_prompt decorator to also catch InternalServerError and check for context-length phrases in the error text
• Added tests for both matching (converted to OverlongPromptError) and non-matching (passes through) 500 errors

Test plan

• Verified fix converts vLLM-style 500 with context length message to OverlongPromptError
• Verified non-context-length 500 errors still propagate as InternalServerError
• All existing error handling behavior unchanged (decorator still catches BadRequestError, still re-raises auth errors)

🤖 Generated with Claude Code

---

Note

Medium Risk
Expands exception handling to reinterpret some InternalServerError responses as OverlongPromptError, which could mask genuine 500s if the message matches the context-length phrases; tests reduce this risk by asserting non-matching 500s still propagate.

Overview
Extends OpenAI chat-completions overlong-prompt detection to also handle vLLM-style HTTP 500s by catching InternalServerError in handle_openai_overlong_prompt and mapping context-length messages to OverlongPromptError.

Adds regression tests covering both conversion of vLLM 500 context-length errors and pass-through behavior for unrelated 500s (e.g., "CUDA out of memory").

^{Written by Cursor Bugbot for commit f93ca5f. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

Test expects wrong exception type for non-matching 500

High Severity

The test test_vllm_non_overlong_internal_server_error_not_converted expects OpenAIInternalServerError to propagate, but the base get_response method in client.py wraps all non-auth, non-Error exceptions in ModelError. When the decorator re-raises InternalServerError, it's caught by except Exception as e: raise ModelError from e. The existing analogous test test_anthropic_non_overlong_bad_request_not_converted correctly expects ModelError instead.