[apucacao]

TL;DR: I keep catching my agent trying to shove handler concerns into Vega.

Decision

A caller passes base_url for a running opencode serve. The handler owns every HTTP call. create_opencode_handler does not accept an httpx client or transport.

Boundary

The caller may start opencode serve, write its runtime config, prepare the workspace and skills, and pass base_url. The handler owns the OpenCode protocol. That covers the HTTP client, GET /event and its SSE reads, POST /session, POST /session/{id}/message with the evaluated model, abort, event reduction, filtering to this session's id, racing a failed message POST, and opening a separate session for each judge. A caller that passes only base_url gets all of it.

Vega is audited against this boundary in observability #2547 and the two PRs on top of it. turnkey.py has no httpx, no transport, no placeholder host, no OpenCode routes, no event handling, and no skip_judges workaround. test_turnkey.py keeps a loopback fake of opencode serve that answers /event, /session, and /message. It plays the server, not the client.

Why

Vega was opening its own httpx.AsyncClient and a custom transport so it could start opencode serve on the first request. That made Vega a special caller. A customer starts the server and passes base_url. Vega should do the same, and it should not manage httpx.

What this PR does not do

• It does not add a connect callback or any other hook for starting the server.
• It does not start opencode serve.
• It does not change session lifecycle, judge sessions, or template rendering.
• It does not change Vega. Observability #2547 does that.

How did you test this change?

uv run ruff check, uv run ruff format --check, uv run mypy packages/opencode-agents/src, and uv run pytest packages/opencode-agents/tests (45 passed). The tests talk to a loopback server through base_url and the handler's own client.

  

[jeffdupont]

The cancelled-read ordering test lost its assert and now passes with the await line_task fix removed. The loopback fake can't see that ordering, so the check needs to live client-side, e.g. patch _next_line to record when it finishes and assert that happens before the response closes. Also, serve_forever(poll_interval=0.01) takes the suite from 20s to about 1s.

[apucacao]

The cancelled-read check is on the client now. A patched _next_line records when it finishes, and the test requires that before GET /event closes. The loopback fake no longer tries to see that order.

I took await line_task out and that assert failed, then put the await back. The loopback server uses serve_forever(poll_interval=0.01).

d9bf7f92337fa397b52ee0640bd718f12640de9a — CI is green, including Bugbot.

[jeffdupont]

Thanks, the cancelled-read test now fails with the fix removed, and the suite is back to about 1s. LGTM.