🤖 Pulsar ACP Agent

Bring your own Agent Client Protocol (ACP)-compatible AI coding agent into Pulsar.

Pulsar ACP Agent running an ACP-compatible coding agent inside Pulsar.

Highlights:

• Run ACP-compatible coding agents such as GitHub Copilot CLI, Mistral Vibe, and Gemini CLI from a Pulsar dock panel.
• Attach files, selections, and images to prompts.
• Review permission prompts, tool output, diffs, plans, and session history inline.
• Configure and switch between multiple agents.

This package is being written by AI coding agents, with human guidance.

Currently tested with GitHub Copilot CLI and Mistral Vibe.

• Install
• Configure agents
  • Example: Copilot CLI
  • Selecting agents
• Use
  • The panel
  • Agent details
  • Attaching files and images
  • Permissions
  • Session configuration
  • Slash commands
  • Tool calls, diffs, and plans
  • Following the agent
  • Sessions
  • Find in thread
  • Working directory
  • Status bar
• Develop
• Architecture
• Testing
• Supported ACP features
• Security

In Pulsar, open Settings → Install, search for pulsar-acp-agent, and click Install. Or from a terminal:

ppm install pulsar-acp-agent

Package page: https://packages.pulsar-edit.dev/packages/pulsar-acp-agent

Install Copilot CLI and authenticate once:

copilot login

Defaults:

• A GitHub Copilot agent (copilot --acp --stdio) is seeded the first time the package activates.
• send host context: enabled

The agent picker in the header (top-left) shows the active agent and lets you switch between configured agents or open Edit agents…. Switching stops the current agent and clears the conversation. See Agent details for the connected agent's live runtime info.

Agents are stored in your Pulsar config file (config.cson) under the pulsar-acp-agent namespace and are managed from the picker or by editing the file directly:

"pulsar-acp-agent":
  activeAgentId: "copilot"
  agents:
    copilot:
      name: "GitHub Copilot"
      command: "copilot --acp --stdio"
    gemini:
      name: "Gemini CLI"
      command: "gemini --experimental-acp"
  version: 1

• agents maps a stable id to an agent { name, command }.
• activeAgentId selects which agent launches; it persists across reloads.
• command is a full command line for an executable that speaks ACP over stdio. It can also launch the agent without a global install, e.g. npx @google/gemini-cli --experimental-acp.

Run Pulsar ACP Agent: Edit Agents (also in the picker and the Packages menu) to open the config file. Changes apply on the next Restart or Switch.

If Pulsar cannot find the executable, set its full absolute path in command. On Linux this may be something like:

/home/you/.local/bin/copilot --acp --stdio

Wrap a path that contains spaces in double quotes, e.g. "C:\Program Files\agent\agent.exe" --acp --stdio.

By default, Pulsar ACP Agent also sends a short host-context hint once per session so the agent knows the conversation is happening through Pulsar, while also making clear that the agent cannot directly control Pulsar's UI. Disable Send host context in package settings if you do not want this extra context included in prompts.

Open the panel with Ctrl+Alt+A or from the command palette (Pulsar ACP Agent: Toggle Panel). Pulsar ACP Agent: Focus Panel opens and focuses the panel without toggling it closed. Type your prompt in the composer and press Enter to send or Shift+Enter for a newline; Stop cancels the current turn.

Once connected, click More... in the header to show or hide the agent's self-reported name, version, advertised capabilities, and metadata as key/value rows, plus agent actions such as Restart. The same live row shows context-token usage when the agent reports it.

Attach prompt content from the composer's Attach to prompt button. Its menu offers:

• Image — pick a PNG, JPEG, GIF, or WebP image up to 5 MiB (you can also drag-and-drop or paste images straight into the input); shown when the agent advertises image prompt support.
• Current selection — attach the current editor selection.
• Current file — attach the active editor file.

Selection and File are shown when the agent advertises embedded-context support, and are also available from the editor right-click menu and the commands Pulsar ACP Agent: Add Active File to Prompt and Pulsar ACP Agent: Add Selection to Prompt. Attachments appear as removable chips above the input and are sent inline as ACP resource blocks (the active file is re-read when you send, so unsaved edits are included). The file must be inside the open project.

A Permissions pill in the footer toggles between Ask (default) and Allow all. In Ask mode, the agent's permission requests are shown inline so you can approve or deny each one.

In Allow all mode, ACP permission prompts are auto-approved for the session using allow_once. If no allow_once option is offered, the prompt is shown normally. The toggle is session-local and resets when the view is closed, restarted, or switched to a different session. It does not sandbox or restrict what the agent process can do; it only skips the confirmation dialog.

When the agent exposes session configuration options (such as model, custom agent, or reasoning effort), the footer shows a dropdown for each. Selecting a value applies it to the active session; the agent stays authoritative, so the dropdowns reflect whatever it reports back. Options are per-session.

When the agent advertises slash commands, typing / at the start of the composer opens a completion popup of its commands, each with a description. Filter by typing, navigate with the arrow keys, and select with Enter, Tab, or a click. Commands that take no argument are sent immediately; commands that take an argument insert /name and show the agent's hint so you can finish the line and press Enter. Commands are per-session and update live as the agent reports them.

Tool calls, diffs and terminal output are rendered inline in the conversation, with long output collapsed behind a Show more toggle.

When the agent reports an execution plan, a collapsible Plan bar appears above the input showing progress and the current step. It stays pinned while the conversation scrolls and can be dismissed; completed plans are snapshotted into the conversation.

Toggle the Follow button in the footer to have Pulsar open and scroll to each file the agent works on, keeping the editor in step with its progress without stealing focus from the panel. Following is off by default and turns off again as soon as you navigate to a different file yourself.

After connection, a Sessions list is accessible via the history icon in the panel header when the agent advertises sessionCapabilities.list; switching is enabled when the agent also advertises loadSession. If the agent advertises sessionCapabilities.delete, a trashcan button appears on hover to permanently remove a session. Use the + icon in the header to start a new session.

Press Ctrl+F / Cmd+F while the chat is focused (or click the magnifier in the header) to open Find in thread for the active session. The bar at the top of the chat shows a search box, Aa (match case), .* (regex), and a scope toggle (All = everything including tool output, diffs and terminal; Chat = only your prompts and the agent's final answers), plus ▲ / ▼ and a n/N counter.

Enter goes to the next match, Shift+Enter to the previous; F3 / Shift+F3 also work, and Escape or × closes the bar and clears highlights. Matches are highlighted and the active match is outlined; navigating to a match inside a collapsed tool output expands it. Search is live while the agent streams and is capped for large threads. To keep the UI responsive, regex mode accepts only simple patterns that cannot cause catastrophic backtracking.

New sessions use the first open project folder as their working directory. Multi-root workspaces are not supported yet: only the first open project folder is sent to the agent and allowed for file access. Loaded sessions keep their recorded working directory when it is still inside that project.

When Pulsar's status bar service is available, it shows a tile with the agent name when known, otherwise Agent, plus a state dot: a hollow ring while connecting, then solid for ready, working, or error, and the theme's warning color while awaiting your confirmation of a permission request. Click the tile to reveal the panel.

git clone https://github.com/hovancik/pulsar-acp-agent.git
cd pulsar-acp-agent
npm install
ppm link          # symlink the checkout into Pulsar
npm run typecheck
npm run build
npm run watch

Pulsar loads lib/main.js. Rebuild after editing src/, then reload Pulsar.

• src/main.ts registers commands, opener, dock item, deserializer, and status-bar service consumer.
• src/agent-view.ts renders the panel UI.
• src/agent-session.ts manages the ACP session via @agentclientprotocol/sdk.
• src/agent-config.ts holds the pure agent-registry logic (normalize, migrate, resolve) split out so it can be unit-tested without loading atom.
• src/util.ts holds pure helpers (parseCommandLine, flattenInfoRows, TerminalRecord) split out so they can be unit-tested without loading atom.
• src/search.ts holds pure search helpers (escapeRegExp, buildSearchRegExp, isChatMessageBody) split out so they can be unit-tested without loading atom.

The SDK is ESM-only, so esbuild bundles it and zod into lib/main.js.

npm test runs Node's built-in runner over test/*.test.mjs. Run npm run build first because tests import the built bundles (lib/util.js, lib/agent-config.js, lib/search.js), not src/.

Beyond ACP, this package also adds:

The agent runs as a separate command-line process that Pulsar launches with your user account. Pulsar does not sandbox it, so it has the same access to your machine as any program you run yourself.

The limits below only constrain the requests an agent makes through Pulsar over ACP. They do not restrict what the agent does in its own process:

• fs/read_text_file and fs/write_text_file are served only when the target path resolves inside the session working directory; the agent reads and writes there without a per-action prompt.
• Writes to open files with unsaved changes are refused.
• The ACP terminal capability is accepted: the agent can run commands through Pulsar, with their merged output shown in the panel. Commands default to the session working directory; an agent-requested absolute cwd is allowed only inside the project. Commands run as separate processes with your user account. There is no per-command approval prompt and no sandbox — a launched command can do anything your account can. Stop cancels the current turn; terminals are killed when the agent releases/kills them, when you switch or restart sessions, or when you close the panel.

These are guard rails for a cooperating agent, not a security boundary: a malicious agent can read or write any file your account can, or run any command, without going through Pulsar at all. Use Pulsar ACP Agent only with agents you trust; for stronger isolation, run it inside a container or VM.