Answer on an Anthropic key from the AG2 Bot, not only on an OpenAI one

[kevin9327]

What this changes

Picked with an Anthropic key, the AG2 Bot fails every run.

The harness built OpenAIConfig(model=BOT_MODEL) whatever the setup screen chose, and never read BOT_PROVIDER. The desktop catalogue gives every harness it installs Credential::AnyProvider ("Every harness on the list takes any model through an API key"), so AG2 can be paired with an Anthropic key. The desktop then writes BOT_PROVIDER=anthropic and BOT_MODEL=claude-sonnet-4-5, and leaves OPENAI_API_KEY empty. Each run died in the OpenAI client before any request went out:

openai.OpenAIError: Missing credentials. Please pass an `api_key`, ... or set the `OPENAI_API_KEY` ...

AG2 has its own Anthropic client, ag2.config.AnthropicConfig, in the package's anthropic extra. It reads ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL the way the Anthropic SDK does. The harness now builds that config when the provider is anthropic, and OpenAIConfig as before otherwise, so the OpenAI key and OpenAI-compatible endpoint choices are unchanged. The requirement becomes ag2[ag-ui,anthropic,openai].

This is a branch, not the single provider/model string #533 gave the LlamaIndex Bot, because AG2 has no LiteLLM config to route through. Google is left out: the setup screen offers no Google key, and Compose passes the harness no GOOGLE_API_KEY.

Where it runs

• New state that outlives a request? None. The config is chosen once at import.
• What happens on the second replica? The same. Each harness process picks the same config from the same environment.
• Anything serialised? No.
• Anything fanned out to a browser? No.
• New listener, port, or schedule? No. The CI step added here runs beside the harness suites already in python-harness.

Boundary and audit

• Every acting call still goes through the gateway: nothing on the server changes, and the harness still refuses a request without the server's token.
• New refusals and new failures each write a row: no new refusal or failure.
• Nothing new is trusted from the client.

Changelog

• A line in CHANGELOG.md under Unreleased.

Proof

• New agent-ag2/tests/test_main.py is agent-llamaindex/tests/test_main.py with the run posted to /, where the harness mounts AG2's ASGI app. It starts a local fake provider serving /v1/chat/completions and /v1/messages, imports the harness under the environment the desktop writes for each of the three choices, and posts an AG-UI run. It asserts HTTP 200, RUN_FINISHED, no RUN_ERROR, and that the provider got the chosen model on the chosen route.
• Linux (ubuntu-latest, Python 3.12, a throwaway workflow on my fork running the same commands as the new CI step). This branch: 3 passed. With agent-ag2/src from main (2d09a08): 1 failed, 2 passed. [an Anthropic key] fails with the OpenAIError above.
• Windows, Python 3.12: the same results.
• Versions: ag2 1.0.5, anthropic 1.7.0, openai 3.16.2.
• Not run: a real Anthropic key, and the published image.

This adds a step at the end of python-harness and an entry at the top of Unreleased, which #609, #610 and #611 also touch. Whichever lands later needs a rebase, and I'll do it.

Default endpoint regression

The default Compose environment sets ANTHROPIC_BASE_URL to an empty string. AG2's Anthropic client passed that through instead of selecting the official endpoint. The harness now explicitly uses https://api.anthropic.com for a blank override and preserves a nonblank custom URL.

The additional authenticated AG-UI regression uses the real framework and Anthropic SDK, replacing only HTTP transport. It checks the official URL, API-key header, chosen model, answer, and successful run completion. All four tests pass locally, including the original provider cases. This is SDK integration coverage; no live Anthropic account or published image was exercised.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.