Babysitter + Accomplish Integration Guide

What is Accomplish?

Accomplish AI is an Electron desktop application for AI-assisted software development. It spawns OpenCode as a subprocess to execute AI coding tasks, providing a rich desktop UI around the agent interaction -- including task management, permission dialogs, and a conversation view.

Babysitter integrates with Accomplish through the OpenCode plugin system. When babysitter is installed, Accomplish gains:

• **Structured orchestration** -- multi-phase task execution with deterministic replay
• **Breakpoint approval gates** -- human-in-the-loop decisions resolved conversationally via Accomplish's ask-user-question MCP tool
• **Progress visibility** -- phase progress, iteration counts, and pending effects visible in the agent conversation

---

Prerequisites

1. **Babysitter SDK** installed globally or in your project: ``bash npm install -g @a5c-ai/babysitter # or npm install @a5c-ai/babysitter ``

2. **Accomplish AI** installed and set up on your machine. The installer auto-detects Accomplish by checking for the platform-specific data directory: - **macOS**: ~/Library/Application Support/Accomplish/ - **Windows**: %APPDATA%/Accomplish/ - **Linux**: ~/.config/Accomplish/

3. **OpenCode** -- bundled with Accomplish; no separate installation needed.

---

Installation

Automatic (recommended)

The babysitter plugin installer auto-detects Accomplish when present:

babysitter plugin:install babysitter --global

This performs two installations in one step: 1. Installs the OpenCode plugin to ~/.opencode/plugins/babysitter/ (standard OpenCode) 2. Detects Accomplish and copies the plugin to <accomplishDataDir>/opencode/plugins/babysitter/

Accomplish-only install

To install exclusively to the Accomplish data directory without touching the standard OpenCode paths:

cd artifacts/generated-plugins/opencode
node bin/install.cjs --accomplish

Manual install

Copy the plugin bundle manually:

# Determine your Accomplish data directory
# macOS:   ~/Library/Application Support/Accomplish/opencode/plugins/babysitter/
# Windows: %APPDATA%/Accomplish/opencode/plugins/babysitter/
# Linux:   ~/.config/Accomplish/opencode/plugins/babysitter/

cp -r artifacts/generated-plugins/opencode/* <accomplishDataDir>/opencode/plugins/babysitter/

After any installation method, restart Accomplish to pick up the plugin.

---

How It Works

Plugin auto-discovery

Accomplish spawns OpenCode with its data directory as the working context. OpenCode discovers plugins by scanning .opencode/plugins/ for JS modules. The babysitter plugin registers itself via index.js, which exports hooks for the OpenCode plugin system:

• session.created -- initializes babysitter session state
• session.idle -- handles iteration orchestration
• shell.env -- injects AGENT_SESSION_ID and other env vars
• tool.execute.before / tool.execute.after -- pre/post tool hooks

Hook lifecycle within Accomplish

Accomplish spawns OpenCode subprocess
  |
  v
OpenCode loads babysitter plugin (index.js)
  |
  v
session.created hook fires
  |-- Creates session state file
  |-- Detects ACCOMPLISH_TASK_ID from env
  |-- Stores task correlation in session metadata
  |
  v
User invokes /babysitter skill (or agent triggers it)
  |
  v
babysitter run:create
  |-- Creates .a5c/runs/<runId>/ directory
  |-- Writes run-status JSON to IPC directory
  |-- Session state updated with runId
  |
  v
babysitter run:iterate (loop)
  |-- Replays journal, executes process
  |-- On pending effects: agent handles them
  |-- On breakpoints:
  |     |
  |     v
  |   Agent uses ask-user-question MCP tool
  |   User approves/rejects in Accomplish conversation
  |     |
  |     v
  |   babysitter task:post resolves the breakpoint
  |
  v
Run completes or fails

Session binding

When running inside Accomplish, the adapter detects ACCOMPLISH_TASK_ID and stores it in the session state's metadata field:

// SessionState.metadata
{
  accomplishTaskId: "task_abc123"
}

This correlation allows the Accomplish daemon to match babysitter runs to its own task records.

---

Using the /babysitter Skill in Accomplish

Once the plugin is installed, the /babysitter skill is available in Accomplish's agent conversation. It orchestrates complex multi-step tasks:

1. **Invoke**: Type /babysitter in the Accomplish conversation, or let the agent invoke it for multi-step tasks 2. **Run creation**: The skill creates a structured run with phased execution 3. **Progress reporting**: Each iteration updates Accomplish's UI with current phase and progress 4. **Breakpoints**: Human approval gates appear as native Accomplish dialogs -- approve, reject, or provide feedback without leaving the app 5. **Completion**: Results and artifacts are available in .a5c/runs/<runId>/

Additional skills

• /accomplish-status -- Manually trigger a status update to the Accomplish UI for a specific run

---

Environment Variables

---

How Breakpoints Work

When the orchestration reaches a breakpoint (human approval gate), the agent resolves it conversationally using Accomplish's ask-user-question MCP tool:

1. babysitter run:iterate returns with a pending breakpoint effect 2. The agent reads the breakpoint details (title, question, options) 3. The agent calls the ask-user-question MCP tool to present the decision to the user 4. The user responds in the Accomplish conversation UI 5. The agent posts the resolution: babysitter task:post <runDir> <effectId> --status ok --value-inline '{"approved": true}' 6. The next iteration picks up the resolved breakpoint and continues

This works without any changes to the Accomplish codebase. The user experience is conversational rather than a dedicated dialog, but fully functional.

---

Troubleshooting

Plugin not detected by Accomplish

**Symptoms**: /babysitter skill not available in Accomplish conversation; no babysitter hooks firing.

**Steps**: 1. Verify the plugin directory exists: - macOS: ls ~/Library/Application\ Support/Accomplish/opencode/plugins/babysitter/ - Windows: dir %APPDATA%\Accomplish\opencode\plugins\babysitter\ - Linux: ls ~/.config/Accomplish/opencode/plugins/babysitter/ 2. Verify index.js and plugin.json exist in that directory 3. Restart Accomplish completely (quit and relaunch, not just close the window) 4. Check Accomplish logs for plugin loading errors

Breakpoints not being presented to the user

**Symptoms**: Babysitter run reaches a breakpoint but the agent doesn't ask for approval.

**Steps**: 1. Check that the run has pending breakpoints: babysitter task:list .a5c/runs/<runId> --pending 2. Verify the agent is handling breakpoint effects by calling ask-user-question MCP tool 3. If using the /babysitter skill, ensure the skill instructions are loaded correctly 4. Check that ACCOMPLISH_TASK_ID is set: the agent uses this to correlate the session

Run stuck or state inconsistent

**Steps**:

# Rebuild derived state from journal
babysitter run:rebuild-state .a5c/runs/<runId>

# Repair journal integrity
babysitter run:repair-journal .a5c/runs/<runId>

# Full diagnostic
babysitter-harness doctor --run-id <runId>

ACCOMPLISH_TASK_ID not set

If ACCOMPLISH_TASK_ID is not in the environment, babysitter still works normally -- it just cannot correlate runs with specific Accomplish tasks. Runs execute, breakpoints are resolved conversationally, and results are available in .a5c/runs/.

Verify Accomplish is setting this variable when spawning OpenCode. This is an Accomplish-side configuration.

---

Architecture Reference