fix(anthropic): map model_context_window_exceeded stop reason to length (#48874)

* test(anthropic): cover model_context_window_exceeded finish reason on chat and responses

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(anthropic): map model_context_window_exceeded stop reason to length

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(rust): mirror model_context_window_exceeded finish reason mapping

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

---------

Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm-rust/crates/core-utils/src/core_helpers.rs

@@ -13,6 +13,7 @@ const FINISH_REASONS: &[(&str, &str)] = &[
     ("max_tokens", "length"),
     ("refusal", "content_filter"),
     ("compaction", "length"),
+    ("model_context_window_exceeded", "length"),
     ("guardrail_intervened", "content_filter"),
     ("content_filtered", "content_filter"),
     ("content_filter", "content_filter"),
@@ -74,6 +75,7 @@ mod tests {
     #[case::end_turn("end_turn", "stop")]
     #[case::stop_sequence("stop_sequence", "stop")]
     #[case::max_tokens("max_tokens", "length")]
+    #[case::model_context_window_exceeded("model_context_window_exceeded", "length")]
     #[case::refusal("refusal", "content_filter")]
     #[case::guardrail_intervened("guardrail_intervened", "content_filter")]
     #[case::content_filtered("content_filtered", "content_filter")]

litellm/litellm_core_utils/core_helpers.py

@@ -266,6 +266,7 @@ _FINISH_REASON_MAP: Final[dict[str, OpenAIChatCompletionFinishReason]] = {
     "tool_use": "tool_calls",
     "refusal": "content_filter",
     "compaction": "length",
+    "model_context_window_exceeded": "length",
     # Cohere
     "COMPLETE": "stop",
     "ERROR_TOXIC": "content_filter",

tests/integration/providers/test_anthropic_context_window_stop_wire.py

@@ -0,0 +1,122 @@
+import json
+import uuid
+from collections.abc import Mapping
+from typing import Final
+
+import pytest
+from integration._support.client import Gateway
+from integration._support.wire import Reply, Request, wire_server
+from pydantic import JsonValue
+
+MODEL: Final = "claude-haiku-4-5"
+
+
+def _sse(event: Mapping[str, JsonValue]) -> bytes:
+    return f"event: {event['type']}\ndata: {json.dumps(event)}\n\n".encode()
+
+
+def _message(identity: str, stop_reason: str | None) -> dict[str, JsonValue]:
+    return {
+        "id": "msg_" + identity,
+        "type": "message",
+        "role": "assistant",
+        "model": MODEL,
+        "content": [{"type": "text", "text": "alpha bravo"}],
+        "stop_reason": stop_reason,
+        "stop_sequence": None,
+        "usage": {"input_tokens": 199878, "output_tokens": 12799},
+    }
+
+
+def _stream(identity: str, stop_reason: str) -> tuple[bytes, ...]:
+    return (
+        _sse({"type": "message_start", "message": {**_message(identity, None), "content": []}}),
+        _sse({"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}}),
+        _sse({"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": "alpha bravo"}}),
+        _sse({"type": "content_block_stop", "index": 0}),
+        _sse(
+            {
+                "type": "message_delta",
+                "delta": {"stop_reason": stop_reason, "stop_sequence": None},
+                "usage": {"output_tokens": 12799},
+            }
+        ),
+        _sse({"type": "message_stop"}),
+    )
+
+
+def _anthropic(identity: str, stop_reason: str):
+    def respond(request: Request) -> Reply:
+        assert request.method == "POST" and request.target == "/v1/messages", request.target
+        if json.loads(request.body).get("stream"):
+            return Reply(content_type="text/event-stream", chunks=_stream(identity, stop_reason))
+        return Reply(body=json.dumps(_message(identity, stop_reason)).encode())
+
+    return respond
+
+
+def _events(gateway: Gateway, path: str, body: Mapping[str, JsonValue]) -> tuple[dict[str, JsonValue], ...]:
+    with gateway.client.stream(
+        "POST", path, json={**body, "stream": True}, headers={"Authorization": f"Bearer {gateway.key}"}
+    ) as response:
+        assert response.status_code == 200, response.read().decode()
+        return tuple(
+            json.loads(line.removeprefix("data: "))
+            for line in response.iter_lines()
+            if line.startswith("data: ") and line != "data: [DONE]"
+        )
+
+
+def _chat_finish_reason(gateway: Gateway, model: str, *, stream: bool) -> JsonValue:
+    body: Final = {"model": model, "max_tokens": 64000, "messages": [{"role": "user", "content": "copy the list"}]}
+    if stream:
+        reasons: Final = tuple(
+            choice["finish_reason"]
+            for chunk in _events(gateway, "/v1/chat/completions", body)
+            for choice in chunk["choices"]  # comprehension-ok: chunk -> choices flatten
+            if choice.get("finish_reason")
+        )
+        assert len(reasons) == 1, reasons
+        return reasons[0]
+    response: Final = gateway.request("POST", "/v1/chat/completions", body)
+    assert response.status_code == 200, response.text
+    return response.json()["choices"][0]["finish_reason"]
+
+
+def _responses_status(gateway: Gateway, model: str, *, stream: bool) -> JsonValue:
+    body: Final = {"model": model, "max_output_tokens": 64000, "input": "copy the list"}
+    if stream:
+        terminal: Final = tuple(
+            event["response"]
+            for event in _events(gateway, "/v1/responses", body)
+            if event["type"] in ("response.completed", "response.incomplete", "response.failed")
+        )
+        assert len(terminal) == 1, terminal
+        return terminal[0]["status"]
+    response: Final = gateway.request("POST", "/v1/responses", body)
+    assert response.status_code == 200, response.text
+    return response.json()["status"]
+
+
+@pytest.mark.parametrize("stream", [False, True], ids=["json", "stream"])
+@pytest.mark.parametrize(
+    ("stop_reason", "finish_reason"),
+    [("model_context_window_exceeded", "length"), ("max_tokens", "length"), ("end_turn", "stop")],
+)
+def test_chat_maps_anthropic_stop_reason(gateway: Gateway, stop_reason: str, finish_reason: str, stream: bool) -> None:
+    identity: Final = uuid.uuid4().hex
+    with wire_server(_anthropic(identity, stop_reason)) as wire, gateway.scenario() as scenario:
+        model: Final = scenario.model(model=f"anthropic/{MODEL}", api_base=wire.url, api_key="scripted-anthropic-key")
+        assert _chat_finish_reason(gateway, model, stream=stream) == finish_reason
+
+
+@pytest.mark.parametrize("stream", [False, True], ids=["json", "stream"])
+@pytest.mark.parametrize(
+    ("stop_reason", "status"),
+    [("model_context_window_exceeded", "incomplete"), ("max_tokens", "incomplete"), ("end_turn", "completed")],
+)
+def test_responses_maps_anthropic_stop_reason(gateway: Gateway, stop_reason: str, status: str, stream: bool) -> None:
+    identity: Final = uuid.uuid4().hex
+    with wire_server(_anthropic(identity, stop_reason)) as wire, gateway.scenario() as scenario:
+        model: Final = scenario.model(model=f"anthropic/{MODEL}", api_base=wire.url, api_key="scripted-anthropic-key")
+        assert _responses_status(gateway, model, stream=stream) == status

tests/unit/litellm_core_utils/test_core_helpers.py

@@ -256,6 +256,7 @@ class TestMapFinishReasonAnthropic:
             ("max_tokens", "length"),
             ("tool_use", "tool_calls"),
             ("compaction", "length"),
+            ("model_context_window_exceeded", "length"),
             ("content_filtered", "content_filter"),
         ],
     )