fix: report maxTokens and stopSequence in MCP sampling results

src/fast_agent/mcp/sampling.py

@@ -36,6 +36,9 @@
 
 logger = get_logger(__name__)
 
+# Stop reasons defined by the MCP sampling spec for a final (non-tool) result.
+_SAMPLING_STOP_REASONS = (LlmStopReason.STOP_SEQUENCE, LlmStopReason.MAX_TOKENS)
+
 
 @dataclass(frozen=True, slots=True)
 class _SamplingModelSelection:
@@ -180,11 +183,14 @@ def _sampling_response(
             stop_reason="toolUse",
         )
 
+    stop_reason = llm_response.stop_reason
+    if stop_reason not in _SAMPLING_STOP_REASONS:
+        stop_reason = LlmStopReason.END_TURN
     return CreateMessageResult(
         role=llm_response.role,
         content=TextContent(type="text", text=llm_response.first_text()),
         model=model,
-        stop_reason=LlmStopReason.END_TURN.value,
+        stop_reason=stop_reason.value,
     )
 
 

tests/unit/fast_agent/mcp/test_sampling.py

@@ -1,6 +1,11 @@
-from mcp_types import CreateMessageRequestParams, SamplingMessage, TextContent
+import pytest
+from mcp_types import CreateMessageRequestParams, CreateMessageResult, SamplingMessage, TextContent
 
+from fast_agent.config import MCPSamplingSettings, MCPServerSettings
+from fast_agent.mcp import sampling
+from fast_agent.mcp.prompt import Prompt
 from fast_agent.mcp.sampling import sampling_agent_config
+from fast_agent.types.llm_stop_reason import LlmStopReason
 
 
 def test_build_sampling_agent_config_with_system_prompt():
@@ -46,3 +51,38 @@ def test_build_sampling_agent_config_empty_system_prompt():
 
     # Verify instruction is the empty string as received in params.systemPrompt
     assert config.instruction == ""
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("llm_stop_reason", "expected"),
+    [
+        (LlmStopReason.MAX_TOKENS, "maxTokens"),
+        (LlmStopReason.STOP_SEQUENCE, "stopSequence"),
+        (LlmStopReason.END_TURN, "endTurn"),
+    ],
+)
+async def test_sample_reports_why_generation_stopped(
+    monkeypatch: pytest.MonkeyPatch, llm_stop_reason: LlmStopReason, expected: str
+) -> None:
+    class _StoppedLLM:
+        async def generate(self, *_args: object, **_kwargs: object):
+            return Prompt.assistant("partial", stop_reason=llm_stop_reason)
+
+    monkeypatch.setattr(sampling, "create_sampling_llm", lambda *_args: _StoppedLLM())
+    params = CreateMessageRequestParams(
+        max_tokens=5,
+        messages=[SamplingMessage(role="user", content=TextContent(type="text", text="Hello"))],
+    )
+
+    result = await sampling.sample(
+        params,
+        server_name="test",
+        server_config=MCPServerSettings(sampling=MCPSamplingSettings(model="passthrough")),
+        agent_model=None,
+        api_key=None,
+        app_context=None,
+    )
+
+    assert isinstance(result, CreateMessageResult)
+    assert result.stop_reason == expected