fix(llm): map max_completion_tokens to max_output_tokens for Responses API#438
Merged
pancacake merged 1 commit intoMay 3, 2026
Conversation
…s API When the Responses API path is selected for newer OpenAI models (gpt-5.x, o1, o3, o4) and an extra kwarg of `max_completion_tokens` flows in (typically from `get_token_limit_kwargs(model, n)`), the OpenAI SDK raises `TypeError` from `responses.create()` before any HTTP request leaves the client. `_should_fallback_from_responses_error()` only catches HTTP errors that carry a `status_code`, so the call is never retried via Chat Completions; the user just sees a stack trace. The bug exists on both OpenAICompatProvider and AzureOpenAIProvider, in the chat and chat_stream paths (4 sites total). Add `adapt_chat_kwargs_to_responses()` that translates the alias to `max_output_tokens` and is used at all four merge sites. The helper: - preserves the existing `None`-drop semantics of the previous dict-comprehension merge, - only applies the alias when `max_output_tokens` is not explicitly set by the caller, - does not mutate the input. Closes HKUDS#437.
Collaborator
|
Thanks for your contribution! |
2 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
When the Responses API path is selected for newer OpenAI models (gpt-5.x, o1, o3, o4) and an extra kwarg of `max_completion_tokens` flows in, the OpenAI SDK raises `TypeError` from `responses.create()` before any HTTP request leaves the client. Closes #437.
Why
`get_token_limit_kwargs(model, n)` returns `{"max_completion_tokens": n}` for newer models (the Chat Completions name). Both `OpenAICompatProvider` and `AzureOpenAIProvider` route those kwargs through `client.responses.create()` when `_should_use_responses_api()` matches, but the Responses API only accepts `max_output_tokens`. The SDK rejects the unknown kwarg with `TypeError` inside the client, and `_should_fallback_from_responses_error()` only catches HTTP errors with `status_code` — so the call is never retried via Chat Completions.
The reporter in #437 hit this with `LLM_BINDING = openai` and `LLM_MODEL = gpt-5.5` on v1.3.5, which routes through `OpenAICompatProvider`.
How
Add `adapt_chat_kwargs_to_responses()` to `provider_core/openai_responses/converters.py` and use it at all four merge sites: `chat` and `chat_stream` on both `OpenAICompatProvider` (lines 698, 781) and `AzureOpenAIProvider` (lines 126, 154).
The helper:
Tests
`tests/services/llm/test_openai_responses_converters.py` covers seven cases: passthrough, None-drop, the rename, None for the alias, explicit-name precedence, empty input, and input non-mutation. `pytest tests/services/llm/test_openai_responses_converters.py -v` is green; the rest of `tests/services/llm/` and `tests/services/config/test_llm_probe_config.py` pass with no new failures from this branch (75/76; the one unrelated failure is a missing `data/user/settings/agents.yaml` fixture).
Note on #390
There is an existing open PR (#390, "[codex] Normalize Azure OpenAI max token aliases") that addresses a related problem at the factory layer for Azure only by translating to `max_tokens`. The reproduction in #437 uses `LLM_BINDING = openai`, which goes through `OpenAICompatProvider` and is outside that PR's scope. Happy to defer if a different shape is preferred.