fix(mcp): semantic tool filter now transforms tools to Chat Completions format

[unfitcoder101]

Fixes the schema mismatch reported in #32281 (persisted specifically in the MCP Semantic Tool Filter path, confirmed by @brian-sbc on v1.93.0-dev.1 after #32285/#32282 landed).

Root cause: _process_mcp_tools_to_openai_format calls _transform_mcp_tools_to_openai without passing target_format, which defaults to "responses" (flat Responses API shape). The semantic filter hook always needs the Chat Completions shape ({"type": "function", "function": {...}}) since it's used on /v1/chat/completions, so tools passed through the filter came out flat and failed strict schema validation on providers like hosted_vllm.

Fix: add a target_format param (default preserved for existing callers) and have the semantic filter hook explicitly pass target_format="chat".

Verified locally: reproduced the exact MCP tool → semantic filter → hosted_vllm path with a mocked tool list; confirmed tools now come out correctly wrapped in {"type": "function", "function": {...}}.

[greptile-apps]

Greptile Summary

This PR fixes a schema mismatch in the MCP semantic tool filter path: tools were being transformed to the flat Responses API shape ({"name": ..., "parameters": ...}) instead of the Chat Completions wrapper shape ({"type": "function", "function": {...}}) required by /v1/chat/completions providers such as hosted_vllm.

• Adds a target_format: Literal["responses", "chat"] parameter (defaulting to "responses") to _process_mcp_tools_to_openai_format, preserving backwards compatibility for all existing callers.
• The semantic filter hook now explicitly passes target_format="chat", ensuring tools exit the pipeline in the correct shape for strict-schema-validating providers.

Confidence Score: 4/5

The change is narrow, backward-compatible, and addresses a real schema mismatch on the semantic filter path. Existing callers are unaffected by the defaulted parameter.

The fix is correct and well-scoped: _transform_mcp_tools_to_openai already supported target_format and the test suite covers both branches. No test is added specifically verifying that the semantic filter hook now produces the wrapped shape — a targeted test for hook.py would increase confidence that this exact code path stays correct under future refactors.

litellm/proxy/hooks/mcp_semantic_filter/hook.py — the fix is applied here but no new unit test guards the target_format="chat" call in this file.

Important Files Changed

Filename	Overview
litellm/responses/mcp/litellm_proxy_mcp_handler.py	Adds target_format parameter (default "responses") to _process_mcp_tools_to_openai_format, threading it through to the existing _transform_mcp_tools_to_openai call. Default preserved; no existing callers affected.
litellm/proxy/hooks/mcp_semantic_filter/hook.py	Passes target_format="chat" to _process_mcp_tools_to_openai_format so the semantic filter hook receives tools in the Chat Completions shape required by /v1/chat/completions providers.

_{Reviews (1): Last reviewed commit: "fix(mcp): semantic tool filter now trans..." | Re-trigger Greptile}

[codecov]

Codecov Report

❌ Patch coverage is 25.00000% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/proxy/hooks/mcp_semantic_filter/hook.py	0.00%	2 Missing ⚠️
litellm/responses/mcp/litellm_proxy_mcp_handler.py	50.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[unfitcoder101]

Added a targeted unit test (test_expand_mcp_tools_uses_chat_format) covering _expand_mcp_tools directly with a mocked tool list — asserts the output uses the {type, function} wrapper shape rather than the flat Responses API shape. Passes locally.

[unfitcoder101]

Added a coverage-tracked unit test in tests/test_litellm/responses/mcp/test_litellm_proxy_mcp_handler.py (test_process_mcp_tools_to_openai_format_forwards_target_format) covering the exact line Codecov flagged. All 20/21 tests in that file pass — the one remaining failure (test_completion_with_function_tools_works_without_fastapi_installed) is a pre-existing local-venv-only failure unrelated to this change, confirmed by reproducing it identically with these commits stashed out.

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing unfitcoder101:fix/mcp-semantic-filter-clean (24076a2) with litellm_internal_staging (3bba363)}

[unfitcoder101]

Heads up on the Codecov patch-coverage flag — the uncovered lines in hook.py (HTTPException import, SemanticToolFilterContextWindowError) came in from upstream commits merged via litellm_internal_staging (53e5b22, 5421fdf, 1e8c2f7, etc.) after this PR was opened, not from this PR's actual change. My fix is isolated to _process_mcp_tools_to_openai_format's target_format forwarding, which already has dedicated regression tests (test_process_mcp_tools_to_openai_format_forwards_target_format, test_expand_mcp_tools_uses_chat_format) — both passing. Happy to discuss if maintainers want coverage added for the newer unrelated code too, but wanted to flag the diff isn't mine.

[unfitcoder101]

@ishaan-jaff Hey, this has been open a couple weeks with no review yet — happy to address anything if it needs changes, just let me know.