RSPEED-2885: sanitize model and MCP output in all response paths

[Lifto]

RSPEED-2885: sanitize model and MCP output in all response paths

QE found that the non-streaming /v1/responses path still leaks server internals:

• model: exposes full provider routing path (google-vertex/publishers/google/models/gemini-2.5-flash)
• output: includes mcp_list_tools items with full server-side tool definitions

Root cause: _sanitize_response_dict() only handled instructions and tools. The streaming path had ad-hoc output filtering at a separate call site that the non-streaming path lacked. Neither path sanitized model.

Fix:

• Move output array filtering into _sanitize_response_dict() — both paths now strip mcp_list_tools, mcp_call, mcp_approval_request items from server-deployed MCP servers
• Add model sanitization — strips provider prefix, keeps model name (e.g. gemini-2.5-flash)
• Remove redundant ad-hoc filtering from the streaming generator

Tested:

• Unit tests: 63 pass (15 for _sanitize_response_dict including new coverage for output and model)
• Integration tests: new tests for both streaming and non-streaming paths verifying sanitization end-to-end
• Local stack: verified with curl against both streaming and non-streaming endpoints — model, instructions, and output all sanitized correctly

Summary by CodeRabbit

• Bug Fixes

  • Removed server-deployed MCP output items from responses while preserving user-relevant MCP items.
  • Cleaned up model names by stripping provider routing prefixes for clearer model values.

• Tests

  • Expanded unit and integration tests for response filtering, instruction substitution, and model-name handling in streaming and non-streaming flows.
  • Updated end-to-end scenarios and test helpers to use a short-model placeholder in expectations.

[coderabbitai]

Walkthrough

Response sanitization now removes server-deployed MCP output items of specific types and strips provider routing prefixes from model values; duplicate streaming-path filtering was removed. Tests and e2e fixtures were updated to expect the shortened model name (MODEL_SHORT) and to cover streaming and non-streaming sanitization behavior.

Changes

Cohort / File(s)	Summary
Response Sanitization Logic src/app/endpoints/responses.py	Extended _sanitize_response_dict() to: 1) remove server-deployed MCP output items with types mcp_list_tools, mcp_call, or mcp_approval_request when server_label matches configured_mcp_labels; 2) strip provider routing prefixes from response["model"] (remove substring before last /). Removed duplicate output-filtering from the streaming response_generator() path.
Unit tests (response sanitization & streaming) tests/unit/app/endpoints/test_responses.py	Added json import and updated moderation-refusal fixture shape. Expanded tests to cover: filtering server MCP items from output, handling missing output, stripping provider prefixes from model, combined sanitization, and integration-style non-streaming and streaming assertions (parsing SSE lines).
End-to-end test fixtures & helpers tests/e2e/features/responses.feature, tests/e2e/features/responses_streaming.feature, tests/e2e/features/environment.py, tests/e2e/utils/utils.py	Updated e2e expectations to use {MODEL_SHORT} for returned model. Added context.default_model_short in environment setup (last path segment of default model). replace_placeholders now substitutes {MODEL_SHORT} using context.default_model_short.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~22 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Docstring Coverage	✅ Passed	Docstring coverage is 88.89% which is sufficient. The required threshold is 80.00%.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly and specifically describes the main change: sanitizing model field and MCP output across all response paths, which directly aligns with the core fixes implemented throughout the changeset.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

✨ Simplify code

• Create PR with simplified code

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[tisnik]

LGTM

[radofuchs]

LGTM, just one update in tests needed

[asimurka]

Why the provider stripping from model cannot be done on client side? This introduces some implicit behavior that violates the OpenResponses standard.