Copilot Queue (Pi Extension)

Queue user feedback ahead of time and let the model consume it via an ask_user tool.

This extension is inspired by TaskSync-style workflows: you preload responses, then configured providers can pull them during long runs.

What it does

• Registers tool: ask_user
• Registers command: /copilot-queue
• Keeps a FIFO queue of responses
• Supports autopilot prompt cycling (1→2→3→1…)
• Activates queue/autopilot only on configured providers (defaults to github-copilot)
• Injects a provider-targeted ask_user loop policy into the system prompt on each new run
• Adds a hidden ask_user protocol reminder, and reinforces compatible provider payloads (tool_choice: required for OpenAI-style requests, tool_choice: { type: "any" } for Anthropic-style requests when thinking is not enabled) when ask_user is available
• While a configured provider is actively running, normal interactive input is captured into queue by default (instead of triggering a new turn)
• Interactive capture can be toggled with /copilot-queue capture on|off (on by default)
• Tracks session elapsed time, ask_user call count, other tool-call count, and direct-reply misses in status line
• Emits session hygiene warnings at configurable thresholds (default: 120 minutes, 50 ask_user calls)
• Persists state in session entries
• Shows queue/autopilot/session state in Pi status line when the current provider is configured for Copilot Queue
• Status line output follows the active Pi theme, and can be disabled with copilotQueue.showStatusLine

When ask_user is called:

1. If queue has items → returns next queued response
2. Else if autopilot is enabled and has prompts → returns next autopilot prompt (cycling)
3. Else in interactive UI for a configured provider → waits for /copilot-queue add <message>, /copilot-queue done, or /copilot-queue stop (optionally with timeout)
4. Else → returns fallback response (continue by default)

When the current model provider is not configured for Copilot Queue, queue/autopilot is bypassed, ask_user uses manual/fallback behavior only, and the extension status line is hidden.

Install

Option 1: Direct with Pi (npm or git)

Install from npm:

pi install npm:pi-copilot-queue

Install from git:

pi install git:github.com/ayagmar/pi-copilot-queue

Pinned to a specific release tag:

pi install git:github.com/ayagmar/[email protected]

Project-local install (writes to .pi/settings.json):

pi install -l git:github.com/ayagmar/pi-copilot-queue

Install from local path:

pi install /absolute/path/to/pi-copilot-queue

Then reload in Pi:

/reload

Option 2: With pi-extmgr (/extensions)

Install extmgr once:

pi install npm:pi-extmgr

Then in Pi (GitHub source):

/extensions install git:github.com/ayagmar/pi-copilot-queue
/reload

Or install the extension file directly from GitHub (index.ts/entrypoint path):

/extensions install https://github.com/ayagmar/pi-copilot-queue/blob/master/src/index.ts
/reload

You can also install a local extension file directly:

/extensions install /absolute/path/to/pi-copilot-queue/src/index.ts
/reload

Settings

By default, Copilot Queue is active only for github-copilot.

You can override that in ~/.pi/agent/settings.json:

{
  "copilotQueue": {
    "providers": ["github-copilot", "openai"],
    "showStatusLine": true
  }
}

Single-provider shorthand also works:

{
  "copilotQueue": {
    "provider": "github-copilot"
  }
}

Settings are global and are read from ~/.pi/agent/settings.json.

Defaults:

• providers = ["github-copilot"]
• showStatusLine = true

Use an empty array to disable provider interception entirely:

{
  "copilotQueue": {
    "providers": []
  }
}

Hide the extension status line while keeping provider management enabled:

{
  "copilotQueue": {
    "showStatusLine": false
  }
}

You can also manage the global settings from inside Pi:

/copilot-queue providers
/copilot-queue providers github-copilot openai
/copilot-queue providers global openai anthropic
/copilot-queue providers off
/copilot-queue providers global off
/copilot-queue settings

• providers with no arguments shows the current active list.
• Passing one or more provider names writes ~/.pi/agent/settings.json.
• global is still accepted as an alias.
• off writes an empty global provider list.
• Tab completion covers the full command surface, including top-level commands, provider routing, session subcommands, timeout presets, and common provider names.

Usage

Queue messages

/copilot-queue add continue with the refactor
/copilot-queue add now add tests for edge cases
/copilot-queue list
/copilot-queue clear
/copilot-queue done
/copilot-queue stop

Done / stop waiting

/copilot-queue done
/copilot-queue stop

Both commands request an explicit stop. If ask_user is currently waiting, it is released with stop. Otherwise the next ask_user call will immediately receive stop.

After an explicit stop, managed-provider runs stay in normal direct-reply mode until you queue a new message or enable autopilot. That re-arms the ask_user loop immediately.

Quick settings UI

/copilot-queue settings

The settings UI lets you adjust all current Copilot Queue settings from one place, including:

• managed providers
• busy input capture
• status line visibility
• empty-queue wait timeout
• fallback response
• warning thresholds
• autopilot on/off
• autopilot prompt add/clear

Interactive capture while busy (configured providers only)

/copilot-queue capture on
/copilot-queue capture off

• on (default): while a run is active, interactive input is queued for ask_user.
• off: keep normal steering behavior (input is not auto-queued).

Wait timeout (for empty queue in UI mode)

/copilot-queue wait-timeout 0
/copilot-queue wait-timeout 60

• 0 disables timeout (default): wait indefinitely.
• >0 makes waiting ask_user return fallback after <seconds>.

Session counters and hygiene warnings

/copilot-queue session status
/copilot-queue session reset
/copilot-queue session threshold 120 50

• When enabled, the status line includes elapsed time + ask_user call count, and shows missed ask_user runs when they happen.
• /copilot-queue session status also reports completed managed-provider runs, runs that used ask_user, direct replies that skipped ask_user, compliance rate, other tool-call count, the last missed direct reply preview, and the non-ask_user tool count from that missed run.
• Warnings are advisory only (no forced stop).
• Default thresholds are 120 minutes and 50 ask_user calls.

Fallback message

/copilot-queue fallback continue

Autopilot (cycling prompts)

/copilot-queue autopilot add continue with implementation
/copilot-queue autopilot add now write tests
/copilot-queue autopilot on
/copilot-queue autopilot list
/copilot-queue autopilot off
/copilot-queue autopilot clear

Recommended instruction snippet for your model

When provider is managed by Copilot Queue, use the ask_user tool whenever feedback is required. Keep calling ask_user after every completed step instead of ending with a direct reply. Only stop when the user explicitly replies with stop, end, terminate, or quit.

Development

pnpm install
pnpm run check

Release commands:

pnpm run release:patch
pnpm run release:minor
pnpm run release:major

Quick run:

pi -e ./src/index.ts