Python: [Feature]: Python: nace.ai Drex decision models: document TypeSafeChatClient(base_url=) or add an agent-framework-nace connector with image input #9716

Problem

agent_framework_typesafe.TypeSafeChatClient (python/packages/typesafe/agent_framework_typesafe/_chat_client.py) is written for TypeSafe's System One API. nace.ai's Drex models speak the same POST /v1/systemone request and response at https://console.nace.ai, so TypeSafeChatClient(base_url="https://console.nace.ai", model="drex-max") should work for text. I have not verified that yet (the client goes through typesafe_sdk.AsyncTypeSafeClient, which may add Jev-specific headers), and nothing in the package or its README says whether another System One server is supported. Images are the larger gap: drex-max reads images as a top-level images array, while the client serializes only text, function calls and function results into state.messages, and #9411 (still open) adds images in Jev's assets shape, which Drex does not accept.

Use Case and Evidence

A screenshot or scanned-page question inside an agent loop ("which of these form types is this?", "does this page show an error?") answered with calibrated per-answer probabilities instead of generated text. The TypeSafeChatOptions question shapes (choice / noul / score) are what Drex takes; the differences are the images array, string-only choice criteria values and no box question.

Desired Outcome

Either (a) a documented, tested statement in python/packages/typesafe/README.md that base_url= plus model= is a supported way to reach other System One servers, Drex included, text only; or (b) a python/packages/nace/ subpackage (agent-framework-nace, NaceChatClient) per docs/decisions/0008-python-subpackages.md, reusing the typesafe client's option parsing and adding the images array for Content parts of type data/uri with an image media type, with unit tests on mocked HTTP and a sample. I'd prefer (b) because of the images and would start with (a) if you'd rather keep one System One client.

Alternatives Considered

Subclassing RawTypeSafeChatClient in user code: works for text, duplicates the question/answer mapping for images. Waiting for #9411: it targets Jev's assets map, which is not the Drex wire shape, so an adapter is still needed.

Proposed Direction

(b) above as a plain httpx2 client with no typesafe_sdk dependency, NACE_API_KEY / NACE_BASE_URL settings, and the same ChatOptions surface as TypeSafeChatOptions.

Scope and Non-Goals

No behavior change to the typesafe package; no box question; no streaming (System One does not stream). .NET is out of scope for this issue.

Language

Python

Contribution Intent

I am willing to implement this

Acknowledgements

• I searched existing issues and did not find a duplicate.
• I will wait for explicit maintainer agreement before starting implementation of a non-trivial change.

AI assistance disclosure (per AI_CONTRIBUTIONS.md): this issue was drafted with Claude Code, which also checked the file paths and PR references against main.

Disclosure: I work at nace.ai, which serves Drex Max. Docs: https://console.nace.ai/docs/drex/guides/drex-max

🤖 Generated with Claude Code