fix(anthropic-adapter): exclude output_config from extra_kwargs in Anthropic adapter

[invoicepulse]

Relevant issues

Related to #22963

Pre-Submission checklist

• I have Added testing in the tests/test_litellm/ directory
• My PR passes all unit tests
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai

Type

🐛 Bug Fix

Changes

I ran into this bug while setting up Claude Code CLI with LiteLLM proxy routing to an Azure-hosted Fireworks model (GLM-5). Requests kept failing with:

400: Extra inputs are not permitted, field: 'output_config'

PR #22990 fixes this in the Azure transformation layer (gpt_transformation.py), but I found that requests coming through the Anthropic adapter path (POST /v1/messages → adapters/handler.py) bypass that fix entirely.

In _prepare_completion_kwargs(), the excluded_keys set only contains "anthropic_messages". All other Anthropic-specific params (like output_config) pass through extra_kwargs to non-Anthropic providers, which reject them.

Fix: Added "output_config" to excluded_keys in adapters/handler.py line 171.

Files changed:

• litellm/llms/anthropic/experimental_pass_through/adapters/handler.py — added output_config to excluded_keys
• tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_output_config_excluded.py — 3 unit tests

Test plan

• 3 unit tests pass — verifying output_config exclusion, existing anthropic_messages exclusion preserved, and valid kwargs still pass through

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 16, 2026 2:00am

[invoicepulse]

@greptileai

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing invoicepulse:fix/anthropic-adapter-output-config-excluded (a60c7a5) with main (548e7eb)}

[greptile-apps]

Greptile Summary

This PR fixes a bug where Anthropic-specific output_config parameter was leaking through the adapter path (POST /v1/messages → adapters/handler.py) to non-Anthropic providers (e.g., Azure OpenAI, Fireworks), causing 400: Extra inputs are not permitted errors. The fix adds "output_config" to the excluded_keys set in _prepare_completion_kwargs, mirroring the fix previously applied to the Azure gpt_transformation.py path in PR #22990.

Key changes and observations:

• handler.py: One-line fix adding "output_config" to the excluded_keys set alongside the existing "anthropic_messages" exclusion
• test_output_config_excluded.py: Three new unit tests that are correctly isolated (no network calls), covering the exclusion, backward-compatibility of anthropic_messages, and valid extra kwargs passthrough
• The fix prevents the 400 error, but output_config (which encodes reasoning effort) is silently dropped rather than being translated to an equivalent parameter for the target provider — the behavior is consistent with how anthropic_messages is treated, but worth noting that reasoning effort intent is lost
• The excluded_keys set is growing as a point-fix pattern; future Anthropic-specific parameters introduced in extra_kwargs will need to be added manually one by one

Confidence Score: 4/5

• Safe to merge — the change is minimal, well-tested, and stops a real 400 error without affecting the Anthropic-native path.
• The one-line fix is correct and well-scoped. The tests are purely in-process unit tests with no network calls, consistent with the repository's testing policy. The only concern is that output_config is silently dropped rather than translated, which may silently degrade reasoning quality for users who set it — but this is consistent with the existing treatment of anthropic_messages and is an acceptable tradeoff for a bug-fix PR. No backwards-incompatible changes are introduced.
• No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/experimental_pass_through/adapters/handler.py	Single-line fix adding "output_config" to excluded_keys set; prevents Anthropic-specific parameter from being forwarded to non-Anthropic providers. The fix stops the 400 error but silently drops reasoning effort semantics without any translation to target-provider equivalents.
tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_output_config_excluded.py	Three unit tests validating output_config exclusion, anthropic_messages exclusion, and valid extra_kwargs passthrough. Tests are pure unit tests with no network calls, consistent with repo conventions. Uses fragile relative sys.path.insert anchored to CWD rather than file.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["POST /v1/messages\n(Anthropic Messages API)"] --> B{Is target\nAnthropic?}
    B -- Yes --> C["AnthropicConfig.transform_request()\ngpt_transformation.py\noutput_config handled natively"]
    B -- No --> D["LiteLLMMessagesToCompletionTransformationHandler\n_prepare_completion_kwargs()"]
    D --> E["Iterate extra_kwargs\n(includes output_config from Claude Code)"]
    E --> F{key in\nexcluded_keys?}
    F -- Yes\n'output_config'\n'anthropic_messages' --> G["❌ Skip / Drop parameter\n(was causing 400 errors\nbefore this PR)"]
    F -- No --> H["✅ Add to completion_kwargs"]
    H --> I["litellm.acompletion(**completion_kwargs)\nForwarded to Azure / Fireworks / etc."]
    G --> I

Loading

_{Last reviewed commit: a60c7a5}

[greptile-apps]

Hardcoded exclusion may not scale as Anthropic adds new parameters

output_config is now hardcoded as a single entry in excluded_keys. The broader problem is that any future Anthropic-specific parameter that ends up in extra_kwargs will silently be forwarded to non-Anthropic providers until another one-off exclusion is added here.

More critically, output_config carries semantically meaningful information about reasoning effort. When the adapter silently drops it, the downstream model receives no instruction about reasoning effort — which may result in incorrect or degraded output quality without any error being raised.

A more robust approach would be to also translate known Anthropic-specific parameters (like output_config) into their target-provider equivalents (e.g., reasoning_effort) rather than unconditionally dropping them. However, if the intent here is purely to prevent 400 errors rather than preserve semantics, this fix is consistent with what was done in gpt_transformation.py for PR #22990.

At minimum, consider whether output_config should be handled as a named parameter in _prepare_completion_kwargs (like thinking, tool_choice, etc.) and translated to a provider-agnostic form, rather than being silently discarded through the exclusion list.

[greptile-apps]

Fragile relative path in sys.path.insert

os.path.abspath("../../../../..") resolves relative to the current working directory at the time the test is run, not relative to the file itself. This is consistent with the existing test file in this directory (test_anthropic_experimental_pass_through_adapters_transformation.py), but the pattern is fragile — it silently resolves to the wrong path if the test is invoked from a directory other than the file's own directory.

The safer approach is to anchor the path to the file's location:

Suggested change

	sys.path.insert(0, os.path.abspath("../../../../.."))
	sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../../..")))

[Sameerlite]

FIx for #22963 was merged. Closing this