@josxa/opencode-pty

opencode-pty

A plugin for OpenCode that provides interactive PTY (pseudo-terminal) management, enabling the AI agent to run background processes, send interactive input, and read output on demand.

Why?

OpenCode's built-in bash tool runs commands synchronously -- the agent waits for completion. This works for quick commands, but not for:

• Dev servers (npm run dev, cargo watch)
• Watch modes (npm test -- --watch)
• Long-running processes (database servers, tunnels)
• Interactive programs (REPLs, prompts)

This plugin gives the agent full control over multiple terminal sessions, like tabs in a terminal app.

Features

• Background Execution: Spawn processes that run independently
• Multiple Sessions: Manage multiple PTYs simultaneously
• Interactive Input: Send keystrokes, Ctrl+C, arrow keys, etc.
• Output Buffer: Read output anytime with pagination (offset/limit)
• Pattern Filtering: Search output using regex (like grep)
• Terminal Snapshots: Capture clean, parsed terminal screen state (no ANSI noise)
• Screen Diffing: Seq-based history with line-level diffs between snapshots
• Conditional Waiting: Block until screen matches a regex or stabilizes
• Exit Notifications: Get notified when processes finish (eliminates polling)
• Permission Support: Respects OpenCode's bash permission settings
• Session Lifecycle: Sessions persist until explicitly killed
• Auto-cleanup: PTYs are cleaned up when OpenCode sessions end
• Web UI: Modern React-based interface for session management
• Real-time Streaming: WebSocket-based live output updates

Setup

Add the plugin to your OpenCode config:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["@josxa/opencode-pty"]
}

That's it. OpenCode will automatically install the plugin on next run.

Updating

OpenCode automatically checks for and installs plugin updates on startup. You don't need to do anything manually!

If you ever need to force a clean reinstall, you can clear the cache:

rm -rf ~/.cache/opencode/packages/@josxa+opencode-pty*
opencode

Tools Provided

Tool	Description
pty_spawn	Create a new PTY session (command, args, workdir, env, title, notifyOnExit)
pty_write	Send input to a PTY (text, escape sequences like \x03 for Ctrl+C)
pty_read	Read PTY output buffer as chat-safe text with pagination and regex filtering
pty_snapshot	Capture parsed terminal screen as clean text with cursor, size, and hash
pty_snapshot_foreground_colors	Capture rendered foreground colors as a label grid with legend
pty_snapshot_background_colors	Capture rendered background colors as a label grid with legend
pty_snapshot_wait	Block until screen matches a regex or content stabilizes
pty_list	List all PTY sessions with status, PID, line count
pty_kill	Terminate a PTY, optionally cleanup the buffer

Slash Commands

This plugin provides slash commands that can be used in OpenCode chat:

Command	Description
/pty-open-background-spy	Open the PTY web server interface in the browser
/pty-show-server-url	Show the URL of the running PTY web server instance

Web UI

This plugin includes a modern React-based web interface for monitoring and interacting with PTY sessions.

If you instruct the coding agent to run something in background, you have to name it "session", i.e. "run xy as a background SESSION". If you name it "task" or "process" or anything else, the agent will sometimes run it as background subprocess using &.

Starting the Web UI

• Run opencode with the plugin.
• Run slash command /pty-open-background-spy.

This will start the background sessions observer cockpit server and launch the browser with web UI.

Features

• Session List: View all active PTY sessions with status indicators
• Real-time Output: Live streaming of process output via WebSocket
• Interactive Input: Send commands and input to running processes
• Session Management: Kill sessions directly from the UI
• Connection Status: Visual indicator of WebSocket connection status

REST API

The web server provides a REST API for session management:

Method	Endpoint	Description
GET	/api/sessions	List all PTY sessions
POST	/api/sessions	Create a new PTY session
GET	/api/sessions/:id	Get session details
POST	/api/sessions/:id/input	Send input to a session
DELETE	/api/sessions/:id	Kill a session (without cleanup)
DELETE	/api/sessions/:id/cleanup	Kill and cleanup a session
GET	/api/sessions/:id/buffer/plain	Get session output buffer (returns { plain: string, byteLength: number })
GET	/api/sessions/:id/buffer/raw	Get session output buffer (raw data)
DELETE	/api/sessions	Clear all sessions
GET	/health	Server health check with metrics

Session Creation

curl -X POST http://localhost:[PORT]/api/sessions \
  -H "Content-Type: application/json" \
  -d '{
    "command": "bash",
    "args": ["-c", "echo hello && sleep 10"],
    "description": "Test session"
  }'

Replace [PORT] with the actual port number shown in the server console output.

WebSocket Streaming

Connect to /ws for real-time updates:

const ws = new WebSocket('ws://localhost:[PORT]/ws')

ws.onmessage = (event) => {
  const data = JSON.parse(event.data)
  if (data.type === 'raw_data') {
    console.log('New output:', data.rawData)
  } else if (data.type === 'session_list') {
    console.log('Session list:', data.sessions)
  }
}

Replace [PORT] with the actual port number shown in the browser when running the slash command output.

Development

Future implementation will include:

App

• A startup script that runs the server (in the same process).
• The startup script will run bun vite with an environment variable set to the server URL
• The client will use this environment variable for WebSocket and HTTP requests

This will ease the development on the client.

Usage Examples

Start a dev server

pty_spawn: command="npm", args=["run", "dev"], title="Dev Server"
→ Returns: pty_a1b2c3d4

Check server output

pty_read: id="pty_a1b2c3d4", limit=50
→ Shows last 50 lines of output

Filter for errors

pty_read: id="pty_a1b2c3d4", pattern="error|ERROR", ignoreCase=true
→ Shows only lines matching the pattern

Send Ctrl+C to stop

pty_write: id="pty_a1b2c3d4", data="\x03"
→ Sends interrupt signal

Kill and cleanup

pty_kill: id="pty_a1b2c3d4", cleanup=true
→ Terminates process and frees buffer

Run with exit notification

pty_spawn: command="npm", args=["run", "build"], title="Build", notifyOnExit=true
→ Returns: pty_a1b2c3d4

The AI agent will receive a notification when the build completes:

<pty_exited>
ID: pty_a1b2c3d4
Title: Build
Exit Code: 0
Output Lines: 42
Last Line: Build completed successfully.
</pty_exited>

Use pty_read to check the full output.

This eliminates the need for polling -- perfect for long-running processes like builds, tests, or deployment scripts. If the process fails (non-zero exit code), the notification will suggest using pty_read with the pattern parameter to search for errors.

Capture a clean terminal snapshot

pty_read returns buffer lines in a chat-safe text form. Non-printable control bytes are escaped so the tool output does not break the UI, but TUI programs still remain hard to interpret because cursor movement and screen control sequences are only shown literally. For TUI apps (interactive UIs with cursor movement, colors, screen clearing), pty_snapshot solves this:

pty_snapshot: id="pty_a1b2c3d4"
→ Returns clean screen text with cursor position, size, seq number, and content hash

pty_snapshot: id="pty_a1b2c3d4", interleavedColors="background"
→ Returns the same text snapshot, but inserts a matching background-color run row under each visible text row

Inspect rendered colors in TUIs

When a TUI uses highlighting, inverse video, or colored selections, plain text snapshots can miss what the user would actually see. These tools expose the rendered color layers as compact label grids, one for foreground and one for background:

pty_snapshot_foreground_colors: id="pty_a1b2c3d4"
→ Returns a compact whole-screen map like `01-10: 120A` with a legend such as `A=default  B=palette:12  C=#ffcc00`

pty_snapshot_background_colors: id="pty_a1b2c3d4"
→ Returns the matching background map, collapsing repeated rows into ranges so selected rows still stand out without losing layout context

This is especially useful for ctrl+p style menus where the selected item differs mostly by highlight color rather than by text. If you want the text and highlight rows together in one response, use pty_snapshot with interleavedColors="background", which preserves horizontal alignment while making long flat spans cheaper to read.

Track screen changes with seq-based diffs

Every snapshot gets a monotonically increasing sequence number (seq) that increments only when the screen content actually changes. Pass since to get only the lines that changed:

pty_snapshot: id="pty_a1b2c3d4", since=5
→ Returns only changed/added/removed lines since seq 5

Wait for specific screen content

Use pty_snapshot_wait to block until a condition is met instead of polling:

pty_snapshot_wait: id="pty_a1b2c3d4", search="error|Error", timeout=30000
→ Resolves when the regex matches the full rendered screen text

pty_snapshot_wait: id="pty_a1b2c3d4", searchAbsent="esc interrupt", timeout=30000
→ Resolves when that text is no longer visible on the rendered screen

pty_snapshot_wait: id="pty_a1b2c3d4", hashStableMs=2000, timeout=30000
→ Resolves when screen content is unchanged for 2 seconds

When multiple conditions are given, the first to match wins. since only changes the response to a diff against that seq and does not affect when the wait resolves.

Example: Debugging OpenCode itself

One compelling use case is running OpenCode inside OpenCode to observe TUI behavior during development. The agent can interact with the inner instance, send prompts, open menus, and watch exactly how the screen updates:

# 1. Launch OpenCode as a background TUI process
pty_spawn: command="opencode", args=["path/to/project"], title="Inner OpenCode"
→ pty_abc123

# 2. Wait for it to render, get initial screen state
pty_snapshot_wait: id="pty_abc123", hashStableMs=2000, timeout=15000
→ seq=2, shows OpenCode banner + input field

# 3. Type a prompt and submit
pty_write: id="pty_abc123", data="explain this codebase"
pty_write: id="pty_abc123", data="\n"

# 4. Watch the response stream in, frame by frame
pty_snapshot_wait: id="pty_abc123", hashStableMs=300, since=2
→ seq=15, diff shows partial response text appearing

pty_snapshot_wait: id="pty_abc123", hashStableMs=300, since=15
→ seq=28, more text streamed in

pty_snapshot_wait: id="pty_abc123", hashStableMs=3000, since=28
→ seq=46, response complete (stable for 3s)

# 5. Open the command palette
pty_write: id="pty_abc123", data="\x10"

# 6. See all available commands
pty_snapshot: id="pty_abc123", since=46
→ Shows command palette overlay with menu items

# 7. Toggle the sidebar
pty_write: id="pty_abc123", data="show sidebar\r"
pty_snapshot: id="pty_abc123", since=48
→ Sidebar appears with session info, MCP connections, context usage

This works because pty_snapshot maintains a headless terminal emulator (xterm.js) alongside each PTY session, producing the same parsed screen a human would see -- without any ANSI escape sequence noise.

Configuration

Environment Variables

Variable	Default	Description
PTY_MAX_BUFFER_LINES	50000	Maximum lines to keep in output buffer per session
PTY_WEB_HOSTNAME	::1	Hostname for the web server to bind to

Permissions

This plugin respects OpenCode's permission settings for the bash tool. Commands spawned via pty_spawn are checked against your permission.bash configuration.

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "bash": {
      "npm *": "allow",
      "git push": "deny",
      "terraform *": "deny"
    }
  }
}

[!IMPORTANT] Limitations compared to built-in bash tool:

• "ask" permissions are treated as "deny": Since plugins cannot trigger OpenCode's permission prompt UI, commands matching an "ask" pattern will be denied. A toast notification will inform you when this happens. Configure explicit "allow" or "deny" for commands you want to use with PTY.
• "external_directory" with "ask" is treated as "allow": When the working directory is outside the project and permission.external_directory is set to "ask", this plugin allows it (with a log message). Set to "deny" explicitly if you want to block external directories.

Example: Allow specific commands for PTY

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "bash": {
      "npm run dev": "allow",
      "npm run build": "allow",
      "npm test *": "allow",
      "cargo *": "allow",
      "python *": "allow"
    }
  }
}

How It Works

• Spawn: Creates a PTY using bun-pty, runs command in background
• Buffer: Output is captured into both a rolling line buffer (ring buffer) and a headless terminal emulator
• Read: Agent can read raw buffer anytime with offset/limit pagination
• Snapshot: Agent can capture the parsed visible screen (clean text, no ANSI codes) via the headless terminal
• Diff: Each content change gets a sequence number; agent can request line-level diffs between any two states
• Wait: Agent can block until screen content matches a regex or stabilizes (no polling needed)
• Filter: Optional regex pattern filters raw buffer lines before pagination
• Write: Agent can send any input including escape sequences
• Lifecycle: Sessions track status (running/exited/killed), persist until cleanup
• Notify: When notifyOnExit is true, sends a message to the session when the process exits
• Web UI: React frontend connects via WebSocket for real-time updates

Session Lifecycle

spawn → running → [exited | killed]
                      ↓
              (stays in list until cleanup=true)

Sessions remain in the list after exit so the agent can:

• Read final output
• Check exit code
• Compare logs between runs

Use pty_kill with cleanup=true to remove completely.

Local Development

git clone https://github.com/shekohex/opencode-pty.git
cd opencode-pty
bun ci          # install packages from bun.lock
bun lint        # Runs Biome linting checks
bun format      # Runs Biome formatting checks
bun typecheck   # Runs TypeScript type checking
bun build:dev   # Build the React app for development
bun test        # Runs the unit tests
bun test:e2e    # Runs the e2e tests

To load a local checkout in OpenCode:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["file:///absolute/path/to/opencode-pty/index.ts"]
}

Diagrams

Sequence

Use Case 1 – Opening the PTY Monitor Web UI

sequenceDiagram
    participant User as Human User
    participant Chat as OpenCode Chat
    participant Core as OpenCode Core
    participant Plugin as PTY Plugin
    participant Manager as PTY Manager (in Plugin)
    participant WS as WebSocket Server (in Plugin)
    participant Browser as Web Browser

    Note over Plugin,WS: Plugin starts/owns the WS server and Manager

    User->>Chat: Types /pty-open-background-spy
    Chat->>Core: Slash command received
    Core->>Plugin: Dispatches to registered command handler
    Plugin->>Browser: open(server.url.origin)
    activate Browser
    Browser->>WS: Connects → ws://.../ws
    WS->>Manager: Queries for current sessions (manager.list())
    Manager-->>WS: Returns session data
    WS-->>Browser: Sends session_list + subscribes to updates
    Browser-->>User: PTY monitor UI appears (sessions + terminals)

Use Case 2 – Starting a Long-Running Background Process

sequenceDiagram
    participant User
    participant Chat as OpenCode Chat
    participant Agent as AI Agent
    participant Plugin as PTY Plugin
    participant Manager as PTY Manager
    participant PTY as bun-pty Process
    participant WS as WebSocket Server
    participant UI as PTY Web UI (optional)

    User->>Chat: "start vite dev server in background"
    Chat->>Agent: User message
    Agent->>Plugin: Calls pty_spawn(command="vite", args=["dev"], ...)
    Plugin->>Manager: spawn(options)
    Manager->>PTY: Launches real process
    activate PTY
    PTY-->>Manager: stdout/stderr chunks
    Manager->>Manager: Appends to RingBuffer
    Manager->>WS: Publishes raw_data + session_update
    alt UI already open
        WS-->>UI: Real-time terminal output
        UI-->>User: Live xterm.js view
    end
    Plugin-->>Agent: Returns session info
    Agent-->>Chat: "Dev server started (ID: pty_abc123)"
    Chat-->>User: Confirmation message

Use Case 3 – Sending Interactive Input to a Running Session

sequenceDiagram
    participant User
    participant UI as PTY Web UI
    participant WS as WebSocket Server
    participant Manager as PTY Manager
    participant PTY as bun-pty Process

    %% Variant A: Human typing in browser (most common)
    User->>UI: Types "rs<Enter>" or pastes text
    UI->>WS: Sends {type:"input", sessionId, data:"rs\n"}
    WS->>Manager: write(sessionId, data)
    Manager->>PTY: process.write(data)
    PTY-->>Manager: New output (restart message, etc.)
    Manager->>WS: Publishes raw_data
    WS-->>UI: Updates xterm.js live

    %% Variant B: AI sending input
    Note over User,UI: Alternative path – AI controlled
    Agent->>Plugin: pty_write(id, "\x03")  // e.g. Ctrl+C
    Plugin->>Manager: write(id, data)
    Manager->>PTY: process.write("\x03")

Use Case 4 – Reading Output / Logs On Demand

sequenceDiagram
    participant User
    participant Chat
    participant Agent as AI Agent
    participant Plugin as PTY Plugin
    participant Manager as PTY Manager
    participant Buffer as RingBuffer

    User->>Chat: "show me the last 200 lines of the dev server"
    Chat->>Agent: User question
    Agent->>Plugin: pty_read(id, offset?, limit=200, pattern?)
    Plugin->>Manager: read / search request
    Manager->>Buffer: read(offset, limit) or search(pattern)
    Buffer-->>Manager: Matching / paginated lines
    Manager-->>Plugin: Raw or formatted lines
    Plugin-->>Agent: Text response
    Agent-->>Chat: "Here are the logs:\n\n1 | [vite] ... \n..."
    Chat-->>User: Logs displayed in chat

Use Case 5A – Killing / Cleaning Up a Session via Web UI

sequenceDiagram
    participant User
    participant UI as PTY Web UI
    participant HTTP as HTTP Server
    participant Manager as PTY Manager
    participant PTY as bun-pty Process
    participant WS as WebSocket Server

    Note over PTY: Assuming PTY is active (running process)<br>notifyOnExit = false (no chat notification,<br>but WS/UI always gets status updates)

    activate PTY

    User->>UI: Clicks "Kill" / "×" on session
    UI->>HTTP: DELETE /api/sessions/:id  (or /cleanup)
    HTTP->>Manager: kill(id, cleanup?)
    Manager->>PTY: Sends SIGTERM (if running)
    PTY-->>Manager: onExit event (code, signal)
    deactivate PTY
    Manager->>WS: Publishes session_update (status: killed/exited)
    WS-->>UI: UI updates → shows "exited" or removes entry

Use Case 5B – Killing / Cleaning Up a Session via Agent

sequenceDiagram
    participant User
    participant Chat as OpenCode Chat
    participant Agent as AI Agent
    participant Plugin as PTY Plugin
    participant Manager as PTY Manager
    participant PTY as bun-pty Process
    participant WS as WebSocket Server
    participant UI as PTY Web UI (optional)

    Note over PTY: Assuming PTY is active (running process)<br>notifyOnExit = false (no chat notification,<br>but WS/UI always gets status updates)

    activate PTY

    User->>Chat: "kill the dev server"
    Chat->>Agent: User request
    Agent->>Plugin: pty_kill(id, cleanup=true)
    Plugin->>Manager: kill(id, true)
    Manager->>PTY: SIGTERM + remove from list (if cleanup)
    PTY-->>Manager: onExit event (code, signal)
    deactivate PTY
    Manager->>WS: Broadcast session_update (status: killed/exited)
    alt UI open
        WS-->>UI: UI updates → shows "exited" or removes entry
    end
    Plugin-->>Agent: Success response
    Agent-->>Chat: "Session killed"
    Chat-->>User: Confirmation in chat

Use Case 6 – Automatic Exit Notification

sequenceDiagram
    participant PTY as bun-pty Process
    participant Manager as PTY Manager
    participant Plugin as PTY Plugin
    participant Chat as OpenCode Chat
    participant Agent as AI Agent
    participant User

    %%{init: {'sequence': {'messageAlign': 'left'}}}%%

    activate PTY
    Note over PTY: Long-running process (dev server, tests, etc.)
    PTY-->>Manager: Process exits → exitCode
    deactivate PTY
    alt notifyOnExit was true when spawned
        Manager->>Plugin: Triggers exit notification
        Plugin->>Chat: Sends formatted message via SDK<br><pty_exited><br>ID: pty_abc123<br>Exit: 0<br>Lines: 342<br>Last: Server running at http://localhost:5173<br></pty_exited>
        Chat-->>User: Notification appears in chat
        Chat->>Agent: Triggers agent with exit message
    end
    Manager->>WS: Publishes final session_update (status: exited)
    alt UI open
        WS-->>UI: UI shows red "exited" badge / stops live output
    end

License

MIT

Contributing

Contributions are welcome! Please open an issue or submit a PR.

Credits

• OpenCode - The AI coding assistant this plugin extends
• bun-pty - Cross-platform PTY for Bun
• xterm.js - Headless terminal emulator powering pty_snapshot