Merge 80d5668646 into add7dea9aa

litellm/llms/custom_httpx/llm_http_handler.py

@@ -895,11 +895,18 @@ class BaseLLMHTTPHandler:
         )
 
         # get config from model, custom llm provider
+        #
+        # optional_params["extra_headers"] is a second source of request headers
+        # here: providers that need headers of their own (github_copilot and its
+        # editor-version, for one) put them there in main.py, and so do callers
+        # using the extra_headers= kwarg. The caller's own headers= win on a
+        # clash, same as everywhere else.
+        _config_headers: Final = optional_params.get("extra_headers")
         request_headers: Final = cast(  # cast-ok: validate_environment is declared as a bare dict
             "dict[str, object]",
             provider_config.validate_environment(
                 api_key=api_key,
-                headers=headers or {},
+                headers={**(_config_headers if isinstance(_config_headers, dict) else {}), **(headers or {})},
                 model=model,
                 messages=messages,
                 optional_params=optional_params,

litellm/llms/openai/openai.py

@@ -794,7 +794,15 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
                 )
 
             if headers:
-                inference_params["extra_headers"] = headers
+                # merge, don't replace: optional_params["extra_headers"] already
+                # holds the provider's own headers (github_copilot puts its
+                # editor-version set there) plus whatever the caller passed as
+                # extra_headers=. Letting headers= overwrite the dict drops all of
+                # that, and the caller's values win on a genuine clash.
+                inference_params["extra_headers"] = {
+                    **cast("dict[str, str]", inference_params.get("extra_headers") or {}),
+                    **headers,
+                }
             if model is None or messages is None:
                 raise OpenAIError(status_code=422, message="Missing model or messages")
 

tests/unit/llms/github_copilot/test_chat_completion_headers.py

@@ -0,0 +1,172 @@
+"""Regression tests for #49062: caller headers must not wipe the provider's own
+request headers on the chat completions path.
+
+Two ways the headers got lost, both covered here:
+
+* the copilot editor-version set that main.py folds into
+  optional_params["extra_headers"] never reached the wire when the caller also
+  sent headers=, because the openai SDK handler assigned over the merged dict
+  instead of merging. That takes the copilot headers out as request headers and
+  as the provider's auth set, so github answers 400 "missing Editor-Version"
+* the generic http handler ignored optional_params["extra_headers"] entirely and
+  only looked at headers=, so every provider routed through it lost the
+  provider-specific headers main.py had collected
+
+No sockets are opened; respx intercepts the outbound request.
+"""
+
+from unittest.mock import MagicMock, patch
+
+import httpx
+import pytest
+import respx
+
+import litellm
+
+API_BASE = "https://api.githubcopilot.com"
+
+CHAT_COMPLETION_BODY = {
+    "id": "chatcmpl-headers",
+    "object": "chat.completion",
+    "created": 0,
+    "model": "claude-sonnet-4",
+    "choices": [
+        {
+            "index": 0,
+            "message": {"role": "assistant", "content": "hi"},
+            "finish_reason": "stop",
+        }
+    ],
+    "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+}
+
+COPILOT_HEADERS = {"editor-version": "vscode/1.95.0", "copilot-integration-id": "vscode-chat"}
+
+
+@pytest.fixture
+def copilot_endpoint():
+    with respx.mock(assert_all_called=False) as mock:
+        yield mock.post(url__regex=rf"{API_BASE}/chat/completions").mock(
+            return_value=httpx.Response(200, json=CHAT_COMPLETION_BODY)
+        )
+
+
+@pytest.fixture
+def no_copilot_login(monkeypatch):
+    """Skip the on-disk copilot login; the headers are what is under test."""
+    from litellm.llms.github_copilot.authenticator import Authenticator
+
+    monkeypatch.setattr(Authenticator, "get_api_key", lambda self: "ghu_fake_token")
+
+
+@respx.mock
+def test_extra_headers_survive_a_forwarded_client_header(copilot_endpoint, no_copilot_login):
+    """headers= is what the proxy fills in when it forwards client headers."""
+    litellm.completion(
+        model="github_copilot/claude-sonnet-4",
+        messages=[{"role": "user", "content": "hi"}],
+        max_tokens=8,
+        api_key="ghu_fake_token",
+        api_base=API_BASE,
+        headers={"x-litellm-session-id": "sess_123"},
+    )
+
+    sent = copilot_endpoint.calls.last.request.headers
+    assert sent["editor-version"] == "vscode/1.95.0"
+    assert sent["copilot-integration-id"] == "vscode-chat"
+    assert sent["x-litellm-session-id"] == "sess_123"
+
+
+@respx.mock
+def test_extra_headers_survive_a_second_extra_header(copilot_endpoint, no_copilot_login):
+    """A caller adding one header of its own keeps the provider's set intact."""
+    litellm.completion(
+        model="github_copilot/claude-sonnet-4",
+        messages=[{"role": "user", "content": "hi"}],
+        max_tokens=8,
+        api_key="ghu_fake_token",
+        api_base=API_BASE,
+        extra_headers={"x-corp-llm-call-id": "abc-123"},
+    )
+
+    sent = copilot_endpoint.calls.last.request.headers
+    assert sent["editor-version"] == "vscode/1.95.0"
+    assert sent["x-corp-llm-call-id"] == "abc-123"
+
+
+@respx.mock
+def test_caller_can_still_override_editor_version(copilot_endpoint, no_copilot_login):
+    litellm.completion(
+        model="github_copilot/claude-sonnet-4",
+        messages=[{"role": "user", "content": "hi"}],
+        max_tokens=8,
+        api_key="ghu_fake_token",
+        api_base=API_BASE,
+        extra_headers={"Editor-Version": "vscode/9.9.9"},
+    )
+
+    assert copilot_endpoint.calls.last.request.headers["editor-version"] == "vscode/9.9.9"
+
+
+@patch("litellm.llms.openai.openai.OpenAIChatCompletion.make_sync_openai_chat_completion_request")
+@patch("litellm.llms.openai.openai.OpenAIChatCompletion._get_openai_client")
+def test_sdk_path_merges_instead_of_replacing(mock_get_client, mock_request):
+    """Same bug at the unit level: headers= must not drop extra_headers=."""
+    from litellm.llms.openai.openai import OpenAIChatCompletion
+
+    parsed_response = MagicMock()
+    parsed_response.model_dump.return_value = CHAT_COMPLETION_BODY
+    mock_get_client.return_value = MagicMock()
+    mock_request.return_value = ({}, parsed_response)
+
+    OpenAIChatCompletion().completion(
+        model="claude-sonnet-4",
+        messages=[{"role": "user", "content": "hi"}],
+        model_response=litellm.ModelResponse(),
+        timeout=60.0,
+        optional_params={"extra_headers": dict(COPILOT_HEADERS)},
+        litellm_params={},
+        logging_obj=MagicMock(),
+        custom_llm_provider="github_copilot",
+        api_key="ghu_fake_token",
+        headers={"x-forwarded-by-client": "1"},
+        acompletion=False,
+    )
+
+    sent_data = mock_request.call_args.kwargs["data"]
+    assert sent_data["extra_headers"] == {**COPILOT_HEADERS, "x-forwarded-by-client": "1"}
+
+
+def test_http_handler_applies_extra_headers_from_optional_params():
+    """The generic handler serves most non-OpenAI providers; it used to drop these."""
+    from litellm.llms.base_llm.chat.transformation import BaseConfig
+    from litellm.llms.custom_httpx.llm_http_handler import BaseLLMHTTPHandler
+
+    config = MagicMock(spec=BaseConfig)
+    config.validate_environment.return_value = {"Authorization": "Bearer k"}
+    config.get_complete_url.return_value = "https://example.com/v1/chat/completions"
+    config.should_fake_stream.return_value = False
+    config.max_retry_on_unprocessable_entity_error = 1
+
+    response = httpx.Response(200, json=CHAT_COMPLETION_BODY)
+    with patch.object(BaseLLMHTTPHandler, "_make_common_sync_call", return_value=response):
+        BaseLLMHTTPHandler().completion(
+            model="m",
+            messages=[{"role": "user", "content": "hi"}],
+            api_base="https://example.com/v1",
+            custom_llm_provider="openai_like",
+            model_response=litellm.ModelResponse(),
+            encoding=None,
+            logging_obj=MagicMock(),
+            optional_params={"extra_headers": {"x-provider-header": "1"}},
+            timeout=5.0,
+            litellm_params={},
+            acompletion=False,
+            headers={"x-caller-header": "2"},
+            provider_config=config,
+        )
+
+    assert config.validate_environment.call_args.kwargs["headers"] == {
+        "x-provider-header": "1",
+        "x-caller-header": "2",
+    }