AnthropicClient sends `temperature: null` when temperature is unset (400 from API) — and the MCP queue then silently drops every episode

[drcollect]

Summary

With the MCP server configured for llm.provider: anthropic and temperature
left unset (the schema default is None), every extraction request fails with
HTTP 400 from the Anthropic API:

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

Because the episode queue catches all exceptions, logs them, and marks the task
done, every add_memory call still returns success — we submitted a
20-document batch that reported "queued for processing" 20 times and landed 0
episodes. Only comparing get_episodes counts against submissions revealed it.

Environment

• graphiti-core 0.29.2 (also present in current main @ 993e081a)
• mcp_server from the same checkout (package mcp-server 1.0.2)
• anthropic SDK 0.111.0, Python 3.10.21, FalkorDB backend
• Model: claude-haiku-4-5-20251001

Reproduction

1. config.yaml with llm.provider: anthropic, a valid ANTHROPIC_API_KEY,
   and no temperature key (schema default None).
2. Call the add_memory MCP tool with any episode.
3. Tool returns Episode ... queued for processing; the server log shows the
   400 above, three retries, then
   services.queue_service - ERROR - Error processing queued episode ....
4. get_episodes for the group: the episode never lands.

Sample failing request IDs (2026-08-24): req_011CeMSZvQydGf5GLXqoYTGN,
req_011CeMSZwJouHufv2wPfzPDT, req_011CeMSZy2zEZLgDLdP5N9uL.

Cause

mcp_server/src/config/schema.py defaults temperature to None
("defaults to None for reasoning models"). The factory comment says downstream
clients omit temperature when it is None — the OpenAI clients do, but
AnthropicClient passes it through to messages.create(...) verbatim, and the
API rejects temperature: null.

Suggested fix

In AnthropicClient, omit the temperature parameter when
config.temperature is None (mirroring the OpenAI clients), or use the SDK's
NOT_GIVEN sentinel.

Separately but compounding: consider surfacing per-episode failures somewhere a
caller can see them (e.g. failure counters in get_status, or a failed-episode
list). A queue that converts hard failures into log lines while the tool call
reports success makes this class of bug invisible to MCP clients.

Workaround

Set temperature: 0.0 explicitly in the llm section of config.yaml.