pi-bg-run

Background process management for Pi coding agent.

Spawn long-running commands (ML training, builds, tests, experiments) and get automatically notified when they complete. Monitor jobs via widget, TUI panel, or tool calls.

Install

pi install npm:pi-bg-run

Tools

bg_run

Run a shell command in the background.

Parameters:

• command (string, required) — Shell command to run
• label (string, optional) — Human-readable label

Returns job ID and log path. Agent is automatically notified on completion.

bg_list

List all background jobs with status and elapsed time.

bg_kill

Terminate a running job. Sends SIGTERM, escalates to SIGKILL after timeout.

Parameters:

• job_id (string, required) — Job ID to kill

Interactive Panel

Press /bg in Pi TUI to open the interactive jobs panel:

• ↑/↓ — Navigate jobs
• Enter — View log tail
• k — Kill (press twice to confirm)
• q / Esc — Close panel

Widget

The inline widget shows a compact single-line status above the editor:

 ✶ bg: build 2m30s, train 5m12s · 3 done · 1 failed

• Running: Star spinner (✶) + job names with elapsed time
• Completed: Collapsed summary (N done · N failed · N killed)
• Auto-dismiss: Widget clears 10s after all jobs finish
• No stale data: Only counts recent jobs within TTL window

Configuration

Add to .pi/settings.json or ~/.pi/agent/settings.json:

{
  "bgRun": {
    "maxConcurrentJobs": 10,
    "completedTtlMs": 604800000,
    "widgetRefreshMs": 3000,
    "killTimeoutMs": 10000
  }
}

Setting	Default	Description
maxConcurrentJobs	10	Maximum simultaneous background jobs
completedTtlMs	604800000 (7 days)	How long completed job data is kept before garbage collection
widgetRefreshMs	3000	Base widget refresh interval (auto-adapts to 500ms when jobs are running)
killTimeoutMs	10000 (10s)	Time before SIGKILL escalation after SIGTERM

Notification Delivery Model

When a background job completes, pi-bg-run uses pi.sendMessage() with deliverAs: "steer" and triggerTurn: true. This means:

• Agent idle: Notification triggers a new agent turn immediately.
• Agent running a tool call: Notification is queued and delivered after the current tool call finishes, before the next LLM call. There is no delay beyond the current tool call's remaining duration.
• Multiple jobs completing while agent is busy: Each notification is delivered as a separate steer message, one after each tool call boundary.

Known limitation: Notifications cannot interrupt a running tool call. If the agent is executing a long-running tool (e.g., a slow bash command), all pending notifications wait until that tool call completes. This is an architectural constraint of Pi's message delivery model, not pi-bg-run itself.

Storage

Job data is stored under /tmp/bg-run/ keyed by project path hash — all sessions working on the same project share the same directory:

/tmp/bg-run/<project-hash>/
├── jobs.json          # Job metadata (sidecar)
├── bg_abc12345.log    # Job output logs
└── bg_def67890.log

Garbage collection runs on session start:

• Removes log files and their entries from jobs.json for completed jobs older than completedTtlMs
• Removes empty project directories
• Skips directories with running jobs

Architecture

src/
  index.ts           # Extension entry point
  types.ts           # Shared types
  config.ts          # Settings loader
  core/
    job-manager.ts   # Orchestrator (DI)
    process-spawner.ts
    process-monitor.ts
    process-killer.ts
    persistence.ts   # Sidecar JSON + GC
  ui/
    notifier.ts      # Batched notifications (deliverAs: steer)
    widget.ts        # Star spinner, single-line compact display
    panel.ts         # TUI panel component
  tools/
    bg-run.ts
    bg-list.ts
    bg-kill.ts

Development

npm install
npm run build         # Vite production build
npm run deploy        # Build + deploy to ~/.pi/agent/extensions/
npm test              # Run all tests
npm run test:unit     # Unit tests only
npm run test:integration  # Integration tests only
npm run test:watch    # Watch mode

License

MIT