[davidmckayv]

What this changes

Three CI checks are red on main and on every open PR. None of them is caused by product code.

• python harness regressions: the agents' requirements are unpinned, and they now resolve anthropic 2.x. That SDK sends the key as Authorization: Bearer instead of x-api-key, and dates its Messages API version later than 2023-06-01. agent-langgraph-agui/tests/test_provider_boundaries.py asserted both by their 1.x values. It now reads the key from either header and checks that the version is a date. The endpoint, key, model and body it asserts are unchanged. This PR also carries Read the Anthropic key from either header in the AG2 test #859's identical AG2 commit, so it is green on its own and merges cleanly whichever lands first. main's run stops at AG2, which is why the LangGraph failure hadn't shown there yet.

• computer (real browser): since fix(computer): install only the Chromium build the computer launches #829, CI installs Chromium with --no-shell. agent-computer/tests/enter-text.test.ts launched Playwright's default build, the separate headless shell, which is no longer on the runner, so it never started. It now launches channel: "chromium", the build the computer itself launches.

• tests (unhandled error between tests): app/tests/self-host-banner.test.tsx unregistered happy-dom while the Dismiss save still had a render posted. That render fired against a missing document, so bun reported zero failures and a red job. afterAll now awaits settleReactWork() first, as the other React tests do (see app/tests/settle-react-work.ts).

Local results: the LangGraph AG-UI suite passes (90, on anthropic 2.3.0), enter-text passes (5), and the banner tests pass three runs out of three. The flake itself only reproduced on CI.

Where it runs

• New state that outlives a request? None; only tests change.
• What happens on the second replica? Nothing; only tests change.
• Anything serialised? No.
• Anything fanned out to a browser? No.
• New listener, port, or schedule? No.

Boundary and audit

• Every acting call still goes through the gateway. No product code changes.
• New refusals and new failures each write a row. None are new.
• Nothing new is trusted from the client.

Changelog

• No line: only tests change, so a deployment behaves no differently afterwards.

[davidmckayv]

Closing as superseded: #861 fixed both tests on main (the LangGraph AG-UI test now accepts the 2.x Bearer key and API version, and enter-text launches channel chromium), and main's CI is green.