feat(opencode): add run_in_background to shell tool with auto notification - #47231
Kai-Liu001 wants to merge 3 commits into
Conversation
|
Heads up for anyone reviewing: #33310 also adds background execution to the shell tool, but stops at launch + a kill tool — there's no completion signal, so the model has to poll or the user has to paste output back. This PR covers the other half: when the command exits, its output is injected into the session and the model is re-invoked automatically. That's the "background launch + completion delivery" combination called out in the TODO at the top of Also: the typecheck/test/nix-eval workflows are stuck at "action required" for this first-time contributor — could a maintainer approve those runs? The suite passes locally ( |
|
Pushed a second commit adding the observation/control half:
Tests: 3 new cases in |
…ation Long-running commands can now be launched with run_in_background=true. The tool returns immediately with a jobId; when the command exits, its output is injected into the session as a synthetic message so the model is re-invoked automatically. Gated behind OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS. Metadata streaming and the tool-call abort race are skipped for background jobs since the tool call has already completed by the time the command exits. Timeout stays honored when passed explicitly.
Background shell commands now tee all output to a log file (logPath in the result and job metadata), so live progress can be inspected anytime with Read/Grep instead of polling. The new bash_jobs tool lists background shell jobs (id, status, command, log path) and kills one by job id, killing the whole process group. Cancelled jobs do not trigger the completion injection.
bash_jobs now lists both background shell commands (type bash, with log path) and background subagents (type task, with child session id), and kill works for both kinds. Subagent cancellation rides the existing onInterrupt wiring in the task tool.
9ef8b4f to
2bd2f8a
Compare
1 participant
Issue for this PR
Closes # (no existing issue, happy to file one if useful)
Type of change
What does this PR do?
The shell tool is synchronous, so long-running commands (dev servers, watch modes, long test suites) block the turn until they exit or time out.
This adds a
run_in_backgroundparameter. When set, the command runs via the existingBackgroundJobregistry (same one the task tool uses for background subagents), the tool returns immediately with ajobId, and when the command exits its output is injected into the session as a synthetic message so the model gets re-invoked automatically. No polling needed.Background jobs skip the live metadata streaming and the tool-call abort race, since the tool call is already done by the time the command exits. An explicit
timeoutis still honored. Gated behind the sameOPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTSflag the task tool uses. Job state stays process-local, consistent with the existing registry.How did you verify your code works?
test/tool/shell.test.tscovering the immediate return, the completed job output, the synthetic prompt delivery, and the flag-disabled rejection.sleep 20 && echo ...returned instantly, I could keep chatting while it ran, and the completion notice came back on its own ~20s later.Screenshots / recordings
Not a UI change.
Checklist