fix(acp): start a fresh agent when the last turn's own tool closed the session
Merged
SA
17d ago+48+48+01234567890123456789-28-28-01234567890123456789
What and why
- When a step timed out and the next step started right away, the old agent could run a tool of its own as it stopped.
ownToolthen closed the session and dropped it from the attempt. The next step was still waiting on that session's last turn, then prompted the closed agent and failed withMODEL_PROVIDER_FAILED("ACP connection closed"). runStepnow gets its session fromready()inpackages/acp/src/executor.ts. After the last turn stops,ready()checks that the session is still the attempt's. If not, it starts a fresh agent.- The executor tests waited on the scripted agent, a child process, with
vi.waitFor's 1s default and one fixed 500ms sleep. Oneeventuallyhelper now gives every wait 15s, and the sleep waits for the first prompt instead. - The fixture wrote its tool log line only after it reported the tool, and the executor stops the agent on that report. It now writes the line first. This is the line the Node 22 job of #1452 failed on.
- The test for this case now starts the next step at once, as the runner does. Before, it waited for the log line, which hid the race in the full file. Run alone, it failed on main.
Issue
Fixes #1456
Verified
Ran it locally: yes
packages/acp,npx vitest run tests/executor.test.ts: the file alone, the one test alone with-t, and under load (12 runs in parallel on 8 cores).- The new test shape fails without the
ready()change (3 of 3 runs) and passes with it.
| main | this branch | |
|---|---|---|
The test alone, -t "ran once its step ended" | 2 of 3 runs fail: ACP connection closed | 3 of 3 pass |
| The file under load | 24 of 24 runs fail, 2 to 5 tests each | 36 of 36 runs pass, 31/31 each |
Reviewed with sol-loop: 2 rounds with gpt-6.1-sol on high. Sol verdict: approved
š¤ Generated with Claude Code