fix(anthropic): omit temperature instead of sending null when it is unset (#1794)

[Anai-Guo]

Fixes #1794.

Problem

mcp_server/src/config/schema.py defaults llm.temperature to None, and
mcp_server/src/services/factories.py passes it straight through to the core
LLMConfig — with the same comment repeated at all five provider branches:

# None is intentional for reasoning models; core LLMConfig stores it
# verbatim and downstream clients omit temperature when it is None.
temperature=config.temperature,  # type: ignore[arg-type]

OpenAIClient honors that contract (openai_client.py):

# Omit temperature entirely for reasoning models — don't even send None.
if not is_reasoning_model and temperature is not None:
    request_kwargs['temperature'] = temperature

AnthropicClient did not — it forwarded self.temperature to
messages.create() unconditionally. The Anthropic SDK serializes an explicit
None rather than dropping the key, so the request body carries
"temperature": null and the API rejects it:

400 - {'type': 'error', 'error': {'type': 'invalid_request_error',
       'message': 'temperature: Input should be a valid number'}}

Which I confirmed at the wire level against a mocked httpx transport
(anthropic 0.120.2):

call	key on the wire	value
messages.create(..., temperature=None)	present	null
messages.create(...) (omitted)	absent	—
messages.create(..., temperature=0.0)	present	0.0

Under the MCP server this is invisible to callers: the episode queue catches the
exception, logs it and marks the task done, so every add_memory still reports
success while nothing lands.

Fix

Build the messages.create() kwargs and add temperature only when it is not
None, mirroring OpenAIClient._create_completion. The guard is is not None,
not truthiness, so an explicit temperature=0.0 is still sent.

Verification

Reproduction against the real AnthropicClient with a mocked
messages.create, before → after:

config	before	after
temperature unset (None)	sent as None	omitted
temperature=0.0	sent as 0.0	sent as 0.0
temperature=0.7	sent as 0.7	sent as 0.7

Tests (tests/llm_client/test_anthropic_client.py, new
TestAnthropicClientTemperature) cover all three rows.

• Full file: 11 passed before, 14 passed after.
• Red/green: reverting only anthropic_client.py while keeping the new tests
  fails exactly test_temperature_omitted_when_unset (13 passed, 1 failed).
  The two guard cases stay green on both sides, which is what they are for.
• pytest tests/llm_client tests/test_text_utils.py tests/utils: 297 passed,
  2 skipped (the skips are the _int tests, no ANTHROPIC_API_KEY).
• ruff==0.14.11 (the version pinned in .github/workflows/lint.yml)
  check and format --check clean on both files.

Deliberately not in scope

The issue reports two things; this PR fixes only the first.

• OpenAIGenericClient, GeminiClient, GroqClient also pass
  temperature=self.temperature through unconditionally, so the factory's
  "downstream clients omit temperature when it is None" comment is not true
  for them either. Whether that is actually a bug depends on how each of those
  APIs treats an explicit null, which I have not verified the way I verified
  Anthropic's — so I left them alone rather than guess. Happy to follow up if
  you'd like them covered.
• Surfacing per-episode queue failures (the issue's second half — a queue
  that turns hard failures into log lines while the tool call reports success).
  That is an MCP server API-shape decision, not a bug fix.

🤖 Generated with Claude Code

[linhongyu510]

Verified at 6e51b5c. The change is correctly limited to request construction: temperature is omitted only for None, while explicit 0.0 and 0.7 remain present, with no changes to retries, token accounting, tool parsing, or exception mapping. Independent validation: the full Anthropic client module passes 14 tests; tests/llm_client tests/test_text_utils.py tests/utils passes 319 tests with 2 expected integration skips; Ruff check/format and git diff --check pass. I also performed a red/green check by reverting only anthropic_client.py while retaining the new test: test_temperature_omitted_when_unset fails exactly because the baseline still forwards the key, then all three temperature tests pass again on this HEAD. The separate queue-failure observability concern is appropriately left out of this minimal fix.