[BlueX888]

fix(worker): close resources when startup fails before run() completes

What breaks

AgentServer.run() starts the inference executor and marks the server started (_closed = False) before bringing up the HTTP/Prometheus servers, proc pool, HTTP session and API client, but only creates _close_future afterwards. If startup fails in that window (reproduced by pre-binding the default port 8081), run() raises while _closed is already False and _close_future is still None. The CLI cleanup calls aclose(), which treats _close_future is None as "nothing was started" and returns early: the inference executor's child process and its non-daemon proc_join_thread are never closed, so the CLI process cannot exit (timeout 10s => exit 124) instead of failing cleanly.

Root cause

aclose() uses _close_future is None as the "nothing was started" signal, but run() starts the first resources well before creating that future.

Fix

• Create _close_future together with _closed = False, immediately before the first resource (the inference executor) starts, so any later startup failure runs the full shutdown path.
• Remove the _close_future is None early return from aclose(). Every resource is already individually null-checked (inference executor, HTTP server, Prometheus server, proc pool, HTTP session, API client), so aclose() closes whatever actually started and skips only the close-future result.
• Guard the final _close_future.set_result(None) with a None check.

Normal shutdown and drain semantics are unchanged: run() still waits on _close_future, repeated aclose() remains a no-op, and a never-started server still returns immediately.

Tests

New test tests/test_worker_connection.py::test_aclose_closes_resources_started_before_run_raises pre-binds the worker HTTP port and substitutes a fake inference executor whose start() launches a deliberately non-daemon join thread, then asserts that after run() raises, aclose() closes the executor, joins the thread, cleans up the HTTP server, and returns promptly.

On base f5ee4a8a: fails — AssertionError: aclose() must close the inference executor (assert False). With the fix: passes. Focused + neighboring files 22 passed, 2 skipped; startup/forkserver neighbors 16 passed, 5 skipped; ruff clean. The original repro with a real InferenceProcExecutor now reports inference alive: False and proc_join_thread alive: False, and the hang variant exits instead of timing out.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[longcw]

looks good to me. something nit:

[longcw]

nit: could you shorten the two new comments? The one in run() describes the old behaviour ("used to leak"); a single line such as # create before starting resources so aclose() can clean up a failed startup is enough. The one in aclose() can go, since the null checks below it say the same thing.

[BlueX888]

Shortened the run() comment to the requested single line and removed the aclose() comment; no code changed. New head is b8cafb717aeaa8b53f2731937c23ccb0e80d1e6f; tests/test_worker_connection.py + tests/test_worker_config.py pass (12 passed) and ruff check is clean.