fix(azure_ai): strip tool-level extra fields on 400 and retry

[Sameerlite]

Relevant issues

Fixes LIT-3507
GitHub Copilot sends tools with a copilot_mcp_server_name field on each tool object. Azure AI Foundry (e.g. FW-Kimi-K2.6) rejects these with a 400:

Extra inputs are not permitted, field: 'tools[0].copilot_mcp_server_name'
Extra inputs are not permitted, field: 'tools[1].copilot_mcp_server_name'

LiteLLM was not stripping the field, so the 400 bubbled up to the caller.

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Changes

litellm/llms/azure_ai/chat/transformation.py

should_retry_llm_api_inside_llm_translation_on_http_error now returns True for "Extra inputs are not permitted" errors that reference tools[N].field paths, regardless of drop_params. The offending field is always safe to strip since the provider explicitly rejects it.

transform_request_on_unprocessable_entity_error calls a new _drop_tool_level_extra_fields helper that parses field names out of the error text (e.g. tools[0].copilot_mcp_server_name) and removes them from every tool object in request_data["tools"]. The fix is generic; any unknown tool-level field reported by Azure AI is stripped, not just copilot_mcp_server_name.

Regression test added in tests/test_litellm/llms/azure_ai/chat/test_azure_ai_transformation.py.

Screenshots / Proof of Fix

Start the proxy and run the repro script:

python litellm/proxy/proxy_cli.py --config litellm/proxy/dev_config.yaml --detailed_debug --reload --use_v2_migration_resolver 2>&1 | tee litellm.log

export LITELLM_API_KEY=sk-1234
export PROXY_URL=http://127.0.0.1:4000
export MODEL=fw-kimi-k2.6   # must be configured with azure_ai provider + Foundry api_base
./repro_fw_kimi_copilot_mcp_tools.sh

Before this fix: 400 BadRequest — Extra inputs are not permitted, field: 'tools[0].copilot_mcp_server_name'.

After this fix: the proxy strips copilot_mcp_server_name from every tool object on the first 400, retries transparently, and returns a successful completion.

Type

Bug Fix

---

Note

Low Risk
Scoped to Azure AI 422/400 retry and request mutation for provider-rejected tool keys; covered by a targeted regression test.

Overview
Azure AI chat translation now automatically retries after a 400 when the error mentions unknown fields on tool objects (e.g. tools[0].copilot_mcp_server_name from Copilot MCP), even if drop_params is off.

On retry, _drop_tool_level_extra_fields parses those field names from the error text and removes them from every entry in request_data["tools"], then the existing unprocessable-entity handling still runs. should_retry_llm_api_inside_llm_translation_on_http_error treats “Extra inputs are not permitted” as retriable when the message matches tools[N].….

A regression test covers stripping copilot_mcp_server_name while leaving tool function payloads intact.

^{Reviewed by Cursor Bugbot for commit ff6ceaf. Bugbot is set up for automated code reviews on this repo. Configure here.}

[greptile-apps]

Greptile Summary

This PR fixes a 400 error caused by Azure AI Foundry rejecting tool objects that contain provider-specific extra fields (e.g., GitHub Copilot's copilot_mcp_server_name). When a 400 response indicates tool-level extra-field rejections via the tools[N].field pattern, LiteLLM now automatically strips those fields and retries, without requiring the drop_params flag.

• A new _error_has_tool_level_extra_fields helper detects tool-level extra-field errors via regex, and a new _drop_tool_level_extra_fields method parses field names from the error text and removes them from all tool objects before the retry.
• should_retry_llm_api_inside_llm_translation_on_http_error is updated to trigger a retry for "Extra inputs are not permitted" errors whenever they reference tools[N].field paths, independent of the drop_params setting.
• A regression test covering detection, retry decision, and request transformation is added under the mock-only test directory.

Confidence Score: 5/5

Safe to merge — the change is narrowly scoped to Azure AI tool-level extra-field stripping and leaves all other retry/transform paths untouched.

The two new helpers are small and deterministic: _error_has_tool_level_extra_fields only pattern-matches a well-defined regex, and _drop_tool_level_extra_fields only removes fields explicitly named by the provider's error response. The retry gate change is intentional and correct — if Azure AI is already rejecting the field, stripping it is always safe. The re import is now at module level, the field-name regex uses [\w-]+ to handle hyphenated names, and the regression test covers the full detection-to-transform flow using mocks only.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/azure_ai/chat/transformation.py	Adds _error_has_tool_level_extra_fields and _drop_tool_level_extra_fields helpers; updates should_retry_llm_api_inside_llm_translation_on_http_error to retry on tool-level extra-field 400s without requiring drop_params; moves import re to module level. Logic is well-structured and the regex correctly uses [\w-]+ to capture hyphenated field names.
tests/test_litellm/llms/azure_ai/chat/test_azure_ai_transformation.py	Adds regression test for the copilot_mcp_server_name stripping fix; uses mocks throughout (no real network calls); covers detection, retry decision, and request transformation all in one test function.

_{Reviews (2): Last reviewed commit: "fix(azure_ai): move re import to top-lev..." | Re-trigger Greptile}

[codecov]

Codecov Report

❌ Patch coverage is 90.00000% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/llms/azure_ai/chat/transformation.py	90.00%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[Sameerlite]

@greptileai

[Sameerlite]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit ff6ceaf. Configure here.}

[mateo-berri]

LGTM; thanks!