fix(cache): include tool_choice in ChatCompletionCache key

[ZIFeIYUuuuuuu]

Why are these changes needed?

ChatCompletionCache currently computes its cache key without including tool_choice. As a result, calls that share messages and tools but request different tool policies can reuse the same cached response. For example, a request with tool_choice="required" can return the response cached for tool_choice="none" (and vice versa). The same collision occurs for create_stream.

This is especially serious for agents because tool_choice changes the model contract: a caller requiring a tool call must not receive a completion produced under a no-tool policy.

The cache key now includes a JSON-serializable representation of tool_choice. Literal policies (auto, required, and none) are stored directly; a concrete Tool is represented by its schema, matching the existing treatment of the tools argument. Both create and create_stream pass the selected policy into the key calculation.

Related issue number

Closes #8211

Checks

• I've included any doc changes needed for https://microsoft.github.io/autogen/. No documentation behavior changed.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure the focused auto checks have passed.

Reproduction and regression coverage

Before this change, the following sequence incorrectly returned the first cached result on the second call:

await cache.create(messages, tools=[tool], tool_choice="required")
await cache.create(messages, tools=[tool], tool_choice="none")

The regression test verifies that required, none, and a concrete tool choice each produce separate non-streaming entries. It also verifies that streaming calls with required and none do not collide.

Validation

• uv run pytest packages/autogen-ext/tests/models/test_chat_completion_cache.py -q — 27 passed (3 existing deprecation warnings).
• uv run ruff format --check packages/autogen-ext/src/autogen_ext/models/cache/_chat_completion_cache.py packages/autogen-ext/tests/models/test_chat_completion_cache.py — 2 files already formatted.
• uv run ruff check packages/autogen-ext/src/autogen_ext/models/cache/_chat_completion_cache.py packages/autogen-ext/tests/models/test_chat_completion_cache.py — All checks passed.
• git diff --check — passed.

The optional model-client test modules in the full packages/autogen-ext/tests/models collection require separately installed torch, ollama, and semantic_kernel; collection stops before executing those optional tests in this local environment.

[ZIFeIYUuuuuuu]

@microsoft-github-policy-service agree