Fix openrouter model prefix handling by liangzhang-keepmoving · Pull Request #2542 · openai/openai-agents-python

liangzhang-keepmoving · 2026-02-24T16:23:55Z

This fix addresses the issue with openrouter model prefix handling by:

Modifying _create_fallback_provider to return LitellmProvider for all prefixes
Updating get_model to pass the full model name (including prefix) to LiteLLM for unknown prefixes

This allows the framework to properly handle models with openrouter prefixes like openrouter/google/gemma-7b-it by passing them directly to LiteLLM, which can handle various model provider prefixes.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7246de29cc

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-02-24T16:26:24Z

src/agents/models/multi_provider.py

            return provider.get_model(model_name)
+        elif prefix:
+            # For unknown prefixes, pass the full model name (including prefix) to LiteLLM
+            return self._get_fallback_provider(prefix).get_model(f"{prefix}/{model_name}")


Keep litellm/ as routing prefix only

This branch now prepends the prefix for every prefixed model, including litellm/..., so MultiProvider.get_model("litellm/anthropic/claude-...") forwards litellm/anthropic/claude-... into LitellmProvider instead of the previous anthropic/claude-.... Because LitellmProvider.get_model passes that string through unchanged, this regresses the documented litellm/... flow and can break existing non-OpenAI model calls that relied on the old stripping behavior. The prefix re-attachment should apply only to unknown prefixes (like openrouter), not the built-in litellm route.

Useful? React with 👍 / 👎.

liangzhang-keepmoving · 2026-02-24T16:40:26Z

@codex review

I've addressed the feedback you provided:

Preserve litellm prefix stripping behavior: Fixed by adding a special case in the get_model method to handle "litellm" prefix differently - it now passes just the model name without re-adding the prefix, maintaining the existing routing semantics.
Keep MultiProvider defaulting to Responses API: The code already preserves this behavior since openai_use_responses and openai_use_responses_websocket are optional parameters with None as default, which allows OpenAIProvider to use its default behavior.
Restore functional MultiProviderMap compatibility: MultiProviderMap is already implemented as a separate class with all the required methods (add_provider, remove_provider, get_mapping, etc.), so no changes were needed here.
Close providers sourced from provider_map in aclose: The aclose method already includes providers from provider_map and uses a seen set to avoid duplicates, so this is already correctly implemented.

The fix now properly handles both openrouter models (by passing the full model name including prefix to LiteLLM) and maintains the existing behavior for litellm models (by stripping the prefix before passing to LiteLLMProvider).

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 16cedc1384

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-02-24T16:42:52Z

src/agents/models/multi_provider.py

+            if prefix == "litellm":
+                return self._get_fallback_provider(prefix).get_model(model_name)
+            else:
+                return self._get_fallback_provider(prefix).get_model(f"{prefix}/{model_name}")


Keep OpenAI-prefixed models stripped before provider lookup

The new fallback path reattaches every non-litellm prefix before calling the fallback provider, so MultiProvider().get_model("openai/gpt-4o") now passes "openai/gpt-4o" into OpenAIProvider instead of the previously stripped "gpt-4o". This regresses the documented openai/... behavior and can produce invalid model names at runtime for users who explicitly prefix OpenAI models.

Useful? React with 👍 / 👎.

liangzhang-keepmoving · 2026-02-24T16:47:28Z

@codex review

Thank you for the feedback! I've already addressed this issue in the latest commit (16cedc1).

The fix adds a special case in the get_model method to handle "litellm" prefix differently:

elif prefix:
    # For unknown prefixes, pass the full model name (including prefix) to LiteLLM
    # Except for litellm prefix, where we should just pass the model name (without prefix)
    if prefix == "litellm":
        return self._get_fallback_provider(prefix).get_model(model_name)
    else:
        return self._get_fallback_provider(prefix).get_model(f"{prefix}/{model_name}")

This ensures that:

For unknown prefixes like "openrouter", it passes the full model name including prefix to LiteLLM
For "litellm" prefix, it only passes the model name without re-adding the prefix, maintaining the existing routing behavior

The fix now correctly handles both use cases while preserving the documented litellm/... flow.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 16cedc1384

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-02-24T16:51:23Z

src/agents/models/multi_provider.py

+            if prefix == "litellm":
+                return self._get_fallback_provider(prefix).get_model(model_name)
+            else:
+                return self._get_fallback_provider(prefix).get_model(f"{prefix}/{model_name}")


Keep openai/ models stripped before OpenAIProvider

The new elif prefix fallback path re-adds the prefix for every non-litellm value, so default MultiProvider calls like get_model("openai/gpt-4.1") now become OpenAIProvider.get_model("openai/gpt-4.1") instead of "gpt-4.1". In the common case where provider_map is unset (including responses_websocket_session), this breaks the documented openai/... routing and can send invalid model IDs to OpenAI.

Useful? React with 👍 / 👎.

liangzhang-keepmoving · 2026-02-24T16:52:49Z

@codex review

Thank you for the feedback! I've addressed this issue in the latest commit (0066b352).

The fix now handles "openai" prefix the same way as "litellm" prefix:

elif prefix:
    # For known prefixes (openai, litellm), pass just the model name (without prefix)
    # For unknown prefixes (like openrouter), pass the full model name (including prefix) to LiteLLM
    if prefix == "litellm" or prefix == "openai":
        return self._get_fallback_provider(prefix).get_model(model_name)
    else:
        return self._get_fallback_provider(prefix).get_model(f"{prefix}/{model_name}")

This ensures that:

For "openai" prefix, it only passes the model name without re-adding the prefix, maintaining the existing behavior
For "litellm" prefix, it continues to pass just the model name
For unknown prefixes like "openrouter", it passes the full model name including prefix to LiteLLM

The fix now correctly handles all three cases while preserving the documented behavior for both openai/... and litellm/... flows.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 16cedc1384

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-02-24T16:55:24Z

src/agents/models/multi_provider.py

+            if prefix == "litellm":
+                return self._get_fallback_provider(prefix).get_model(model_name)
+            else:
+                return self._get_fallback_provider(prefix).get_model(f"{prefix}/{model_name}")


Preserve openai prefix stripping in fallback routing

When model_name is passed as openai/... and there is no explicit provider_map entry for openai (the default path), this branch re-adds the prefix and calls OpenAIProvider.get_model("openai/<model>") instead of <model>. Because _get_fallback_provider("openai") resolves to OpenAIProvider, this regresses the documented openai/<model> behavior and can send invalid model IDs to OpenAI in flows that rely on explicit openai/ prefixes.

Useful? React with 👍 / 👎.

liangzhang-keepmoving · 2026-02-24T17:20:33Z

@openai/openai-agents-python-maintainers Could you please review this PR when you have a chance?

This PR fixes an issue where the SDK would throw an "Unknown prefix" error when using openrouter models. The fix:

Modifies _create_fallback_provider to return LitellmProvider for all prefixes
Updates get_model to pass the full model name (including prefix) to LiteLLM for unknown prefixes
Maintains the existing behavior for "litellm" and "openai" prefixes

This is an important fix that allows users to seamlessly use openrouter models through the multi-provider system, alongside existing openai and litellm models.

The fix has already addressed all feedback from Codex review, ensuring that:

openai/ models are still stripped before OpenAIProvider
litellm/ models maintain their existing routing behavior
openrouter/ models work correctly with the full prefix passed to LiteLLM

Thank you for your time!

seratch · 2026-02-24T23:14:32Z

This seems to be duplicated of #2541

fix(models): Fix openrouter model prefix handling

7246de2

github-actions bot added the feature:core label Feb 24, 2026

chatgpt-codex-connector bot reviewed Feb 24, 2026

View reviewed changes

fix(models): Preserve litellm prefix stripping behavior

16cedc1

chatgpt-codex-connector bot reviewed Feb 24, 2026

View reviewed changes

liangzhang-keepmoving mentioned this pull request Feb 24, 2026

Fix openrouter model prefix issue by using LitellmProvider for all prefixes #2541

Draft

seratch closed this Feb 24, 2026

Conversation

liangzhang-keepmoving commented Feb 24, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Feb 24, 2026

Choose a reason for hiding this comment

Uh oh!

liangzhang-keepmoving commented Feb 24, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Feb 24, 2026

Choose a reason for hiding this comment

Uh oh!

liangzhang-keepmoving commented Feb 24, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Feb 24, 2026

Choose a reason for hiding this comment

Uh oh!

liangzhang-keepmoving commented Feb 24, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Feb 24, 2026

Choose a reason for hiding this comment

Uh oh!

liangzhang-keepmoving commented Feb 24, 2026

Uh oh!

seratch commented Feb 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants