Add FMAPI tool calling contract tests for DatabricksOpenAI

[dhruv0811]

Summary

End-to-end FMAPI tool calling integration tests for DatabricksOpenAI and ChatDatabricks (LangGraph), mirroring user code patterns from app-templates. Both #269 (strict field) and #333 (empty assistant content) were caught by customers — these tests ensure those CUJs don't regress.

Also includes a bug fix for Gemini models on FMAPI (details below).

Tests

Models are dynamically discovered via workspace_client.serving_endpoints.list(), filtered to databricks-* + llm/v1/chat, then probed for tool calling support. Tests retry up to 3 times.

OpenAI (Agents SDK + McpServer):

• Single-turn, multi-turn, streaming via Runner.run / Runner.run_streamed

LangChain (LangGraph create_react_agent):

• Single-turn, multi-turn, streaming via agent.invoke / agent.ainvoke / agent.stream / agent.astream

Gated behind RUN_FMAPI_TOOL_CALLING_TESTS=1.

Bug fix: Gemini FMAPI tool calling compatibility

Gemini FMAPI doesn't conform to the OpenAI API spec in two ways during tool calling:

1. Request side — Rejects tool messages where content is a list of content blocks (e.g. [{"type": "text", "text": "hello"}]). The Agents SDK always produces this format when using MCP tools. Fix: _flatten_list_content_in_messages() flattens to a plain string before sending.

2. Response side (streaming) — Returns delta.content as a list instead of a string in streaming responses. The Agents SDK crashes with ValidationError: Input should be a valid string, input_type=list. Fix: _GeminiStreamWrapper / _AsyncGeminiStreamWrapper intercept stream chunks and flatten list content.

Reproduce request-side issue:

# Tool result with content as LIST — FAILS on Gemini FMAPI
curl -s -X POST "$HOST/serving-endpoints/databricks-gemini-2-5-flash/invocations" \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"model":"databricks-gemini-2-5-flash","messages":[
    {"role":"user","content":"Echo hello"},
    {"role":"assistant","content":null,"tool_calls":[{"id":"echo","type":"function","function":{"name":"echo","arguments":"{\"msg\":\"hello\"}"}}]},
    {"role":"tool","tool_call_id":"echo","content":[{"type":"text","text":"hello"}]}
  ],"tools":[{"type":"function","function":{"name":"echo","description":"Echo","parameters":{"type":"object","properties":{"msg":{"type":"string"}},"required":["msg"]}}}],"max_tokens":100}'
# → 400: "Expecting 'content' to be a String"

# Same request with content as STRING — WORKS
# (change "content":[{"type":"text","text":"hello"}] to "content":"hello")

Note: Gemini 2.5 Pro + LangChain

Gemini 2.5 Pro is a reasoning model that consumes 200-600 reasoning tokens from the max_tokens budget before producing output. LangChain tests use max_tokens=1000 for this model (vs 200 for others) to accommodate the reasoning overhead.

Known model issues (skipped)

Model	Issue
gpt-5-nano	Too small for reliable tool calling
gpt-oss-20b, gpt-oss-120b, llama-4-maverick	Hallucinates tool names
gemini-3-flash, gemini-3-pro, gemini-3-1-pro	Requires thought_signature (Gemini 3.x)
gemma-3-12b (LangChain only)	Outputs raw tool call text instead of executing tools

Test plan

• Full CI run via ai-oss runner: https://github.com/databricks-eng/ai-oss-integration-tests-runner/actions/runs/22597021048

[bbqiu]

this looks great! left two comments

will there be a separate PR for the payloads that langgraph agents will generate?