@runminton/pi-async-bash0.1.2

Async Bash

Long bash commands no longer hold a tool call hostage.

What installing it changes

One changed behaviour and one new tool:

1. A bash call still running after 63 seconds returns a task ID instead of blocking. The command keeps running and the session is notified when it ends.
2. bash_task is added, so the session can check or stop those tasks.

flowchart TB
  subgraph before["Before: the tool call owns the wait"]
    b1["bash({ command, timeout })"] --> b2["pi runs the command"]
    b2 --> b3["the tool call waits for the exit code"]
    b3 --> b4["one result: output, exit code, wall time"]
  end
  subgraph after["After: the wait can be released once"]
    a1["bash({ command, timeout })"] --> a2["pi runs the same command"]
    a2 --> a3{"still running after 63 s?"}
    a3 -->|"no"| a4["same result as before"]
    a3 -->|"yes"| a5["result: status running, task_id, log path"]
    a5 --> a6["the turn continues"]
    a2 -.->|"same process, never restarted"| a7["command exits"]
    a7 --> a8["completion message wakes the session that started it"]
  end

The bash interface is unchanged

Tool name	bash
Parameters	command and timeout, same schema and validation. There is no background parameter, so the model never guesses how long a command will take
Shell	Same shellPath, shellCommandPrefix, session environment (PI_SESSION_ID and friends) and tool_call policies
Foreground	Same native execution code, wrapped only to copy output into a log. Same output text, same truncation to 2000 lines / 50 KB, same exit code and error marking
Timeout	Still an execution limit. A timeout of 63 seconds or less never hands off

Three things about the tool definition do differ:

• The description ends with one sentence that tells the model about the handoff.
• The declared output schema accepts a normal result or a running result.
• In the TUI, the bash tool no longer supplies a result renderer, so pi's generic renderer draws the result body instead of the built-in bash result widget.

What gets added

Addition	Effect
bash_task tool	status and kill for tasks owned by, or inherited by, the current session
Footer status	TUI only, for example BG 1 · 2m 10s, refreshed every second
Completion message	When a handed-off command ends, one message reaches the session that started it and may start a turn
Context note	While tasks exist, a hidden async-bash-state message joins model requests, capped at 16 KiB
Session entry	Each handoff appends an async-bash-task entry to the branch
Files	<agent dir>/extension-data/async-bash/ (0700) gets one .log per bash call, short calls included, and one .json per handed-off task. Nothing is deleted automatically

Two behaviours change only after a handoff:

• Escape stops the turn, not the background command. Use bash_task with kill.
• Quitting pi stops tasks owned by that process.

What it does not do

• It writes no settings. pi install registers the package; the extension reads shellPath and shellCommandPrefix and changes neither.
• It touches no project files, starts no daemon and opens no network connection.
• It makes no model calls of its own. A completion message starts an ordinary turn, using your configured model like any other turn.
• It does not adopt tasks across processes, and it does not compose another extension's private spawn hook, SSH or sandbox backend.

Install

pi install npm:@runminton/pi-async-bash

Or load it for one invocation without changing your settings:

pi -e npm:@runminton/pi-async-bash

Installing modifies your pi configuration explicitly through pi install. Loading with -e does not.

Verified on pi 1.3.0, Node 22.23.2 and macOS. Implementation reuses pi's local shell backend; Windows is untested.

Do not load another extension that overrides bash at the same time. pi 1.3.0 reports duplicate tools as a conflict and selects a definition by load order; it does not compose two execution backends. This extension honours the standard shellPath, shellCommandPrefix and tool_call restrictions.

With a version of pi-compact-tool that no longer overrides bash execution, async-bash executes and compact renders; both load orders work. Older compact versions still register bash and must be updated together.

Usage

Normal bash parameters are unchanged:

bash({ command: "npm run e2e", timeout: 600 })

For work whose result you need, run the command in the foreground. Do not wrap it in nohup ... & and then wait with sleep; tail: the shell returns after launching the process, so this tool cannot track its completion. The tool already handles long commands and keeps their logs.

If the command is still running after 63 seconds, the call returns:

Still running after 63 seconds. Task: <task_id>
Log: <absolute_log_path>
You will be notified when it finishes. Do not poll.
Continue work that does not depend on this result. If none remains, end your turn and resume when notified.

Then, on demand:

bash_task({ task_id: "...", action: "status" })
bash_task({ task_id: "...", action: "kill" })

New task IDs are 8 random hexadecimal characters, for example a3f8c921, regenerated on collision. Existing UUID tasks remain readable and controllable.

Action	Behaviour
status	Returns the status, log location and a bounded log tail
kill	Stops the managed process group and returns after observing the result; a finished task keeps its original result

timeout is the execution limit for the whole command. A background handoff does not restart it. The 63 seconds are a waiting threshold, not an execution timeout.

bash_task no longer supports wait, and bash gains no synchronous waiting parameter. After a running result, continue work that does not depend on it; when nothing is left, end the turn and resume when notified.

Completion messages are delivered at the next available turn boundary, even while the agent continues calling tools. They do not interrupt a running tool batch. They use steering delivery rather than a follow-up queue, which would wait for the agent's entire tool loop to stop.

Footer

In the TUI, current-session background tasks appear through pi's native status API. The footer itself is not replaced and compact is not required:

BG 2 · 3m 12s / 1m 05s

Each duration counts from when its command actually started, including the foreground 63 seconds. Entries are ordered by start time and refresh every second; IDs are not shown. A finished task is removed immediately, and the status disappears when nothing is running. Foreground commands and unknown tasks from an earlier runtime are never counted.

Narrow terminals drop spaces inside durations first, then report omitted entries as +N, for example BG 2 · 3m12s / +1. The native footer still applies its single-line truncation to all extension statuses.

Switching sessions only shows tasks the current session owns or inherits by reference. The compact tool row shows Backgrounded to record that a handoff happened at that point; the footer shows what is running now, and the completion notification carries the final result.

Lifecycle

• Before a task ID is returned, Escape cancels the running command.
• After a task ID is returned, Escape does not stop that task; use kill.
• /tree, session switching, fork/clone and reload do not stop tasks.
• Completion notifications wake only the owning session. Returning to another branch adds state but does not replay the automatic notification.
• fork/clone inherits the same task ID without copying the process; an inheriting session can query and stop it explicitly.
• A normal pi exit stops tasks managed by the current process.
• After a crash, task survival and cleanup are not promised. A new runtime marks executions it cannot confirm as unknown; it does not guess an exit code or adopt or stop anything based on a stale PID.
• Long-running services usually never exit on their own. A running task does not mean a service is ready to accept requests.

Automatic handoff releases only that one bash call. When the same batch contains other blocking tools, the current turn may still have to wait.

Output and state

Completion notifications and status include at most 100 log lines and 16 KiB. Restored context has a separate 16 KiB total budget and does not repeatedly inject full logs.

Statuses are running, exited, killed, timed_out, failed and unknown. A non-zero exit is still exited, with the exit code kept separately. Querying a failed task with status is not itself a tool error; kill on a failing terminal state is marked as an error.

Inside codemode:

• Short bash calls return the original structured fields such as output and exit_code.
• Long bash calls return task_id, status: "running", log_path and similar fields, with a null exit code.
• bash_task returns task status and a bounded output; scripts should check status and exit code.

There is no automatic log cleanup, PTY, automatic retry, task list or cross-process task adoption. Termination is guaranteed only for the managed process group, excluding daemon descendants that deliberately leave it.

Development

npm ci
npm run check
npm test

Tests use a local deterministic provider, real AgentSessions, real CLI/RPC and real child processes. No external model API is called. The suite performs several real 63-second handoffs and usually takes four to five minutes.

RPC tests print a temporary artifact directory that keeps each process's events-<pid>.jsonl, stderr-<pid>.log, session and command logs.

Local implementation notes are in TICKETS.md, the acceptance contract in SPEC.md, and research evidence in INVESTIGATION.md.

License

MIT