Bacchus: Multi-Agent Coordination

Bacchus coordinates parallel agent work on codebases using isolated jj workspaces.

Quick Start

For fully autonomous multi-agent work, start Claude in yolo mode:

claude --dangerously-skip-permissions

Then: plan -> import -> orchestrate

# 1. Initialize tasks file
bacchus task init

# 2. Edit .bacchus/tasks.yaml with your task breakdown

# 3. Import to SQLite
bacchus task import --epic-id <EPIC>

# 4. Check ready tasks
bacchus task list --ready

Workflow Commands

Handle Commands (Token-Saving)

Use handles to reduce token overhead when searching symbols:

Example:

bacchus symbols --search "validate" --handle  # Returns $sym1
bacchus handle expand $sym1 --limit 5         # Get first 5 results
bacchus handle filter $sym1 --kind function   # Returns $sym2 (filtered)

Planning Tasks

Create .bacchus/tasks.yaml:

version: 1

tasks:
  # Implementation tasks
  # type = PM workflow (bug_fix, feature, refactor, test, docs, infra, generic)
  # archetype = Agent specialization (frontend, backend, data, test, infra, review, security, generic)
  - id: TASK-001
    title: "Add login endpoint"
    type: feature           # PM workflow type
    archetype: backend      # Agent specialization
    priority: 1
    status: open
    depends_on: []
    footprint:
      creates: ["src/api/login.rs"]
      modifies: ["src/api/mod.rs::*"]

  - id: TASK-002
    title: "Add login form component"
    type: feature
    archetype: frontend
    priority: 2
    depends_on: [TASK-001]
    footprint:
      creates: ["src/components/LoginForm.tsx"]

  # Review tasks (depend on implementation)
  - id: TASK-001-SECURITY
    title: "Security review of login endpoint"
    type: feature           # It's a feature task (the review itself)
    archetype: security     # But needs security expertise
    priority: 3
    depends_on: [TASK-001]

  - id: TASK-002-REVIEW
    title: "Code review of login form"
    type: feature
    archetype: review
    priority: 3
    depends_on: [TASK-002]

Archetype System

Bacchus uses archetypes to provide specialized prompts for agents. The planner explicitly assigns an archetype to each task based on the required expertise. Archetypes are defined in archetypes.yaml.

Type vs Archetype:

• type: PM workflow category (what kind of work: bug_fix, feature, refactor, test, docs, infra, generic)
• archetype: Agent specialization (what expertise: frontend, backend, data, test, infra, review, security, generic)

Archetype Commands

# List available archetypes
bacchus archetype list

# Show archetype details
bacchus archetype show frontend

# Get the prompt for an archetype
bacchus archetype prompt security

# Select best archetype for a task
bacchus archetype select TASK-001

Discovering Archetypes

Archetypes are defined in archetypes.yaml. Use CLI commands to explore:

bacchus archetype list                    # List all archetypes
bacchus archetype show frontend           # See details for one

Customizing Archetypes

Copy to your project to customize:

cp ~/.claude/skills/bacchus/archetypes.yaml .bacchus/archetypes.yaml

Project-level .bacchus/archetypes.yaml takes precedence over the skill default.

Orchestrator Mode

When orchestrating, spawn agents using the Task tool with archetype-specific prompts.

Prerequisites

Before spawning agents, ensure jj is initialized:

# Check if jj is initialized
if ! jj root 2>/dev/null; then
  # Initialize jj in colocated mode (works with existing git repo)
  jj git init --colocate
  jj config set --repo user.name "Bacchus"
  jj config set --repo user.email "bacchus@localhost"
  jj bookmark create main -r @
  jj describe -m "Initial commit"
  jj new
fi

Spawning Agents

1. Get the archetype for a task:

bacchus archetype select TASK-001

2. Spawn agent with Task tool in background, using --dangerously-skip-permissions for autonomous execution:

Task tool:
  subagent_type: "general-purpose"
  run_in_background: true
  prompt: |
    Run with: --dangerously-skip-permissions

    [Archetype prompt from bacchus archetype prompt <type>]

    ## Your Task

    Task ID: {task_id}
    Title: {title}
    Description: {description}

    ## Setup

    Start your session and claim the task:
    ```bash
    bacchus session start agent --task-id "{task_id}"
    bacchus claim "{task_id}" agent-{unique_id}
    ```

    ## Work in Workspace

    IMPORTANT: Never `cd` into the workspace. Use `jj -R` flag instead.

    ```bash
    jj -R .bacchus/workspaces/{task_id} status
    jj -R .bacchus/workspaces/{task_id} describe -m "Your commit message"
    jj -R .bacchus/workspaces/{task_id} diff
    ```

    ## Complete

    When done:
    ```bash
    bacchus release {task_id} --status done
    ```

Note: Agents run in yolo mode (--dangerously-skip-permissions) to work autonomously without permission prompts.

Session Management

Sessions enable stop hooks that prevent premature exit:

# Agent mode - blocks until task is closed
bacchus session start agent --task-id TASK-42

# Orchestrator mode - blocks while work remains
bacchus session start orchestrator --max-concurrent 3

# Check status
bacchus session status

# Clear session to allow exit
bacchus session stop

Orchestrator Loop

Each iteration:

1. Check status: bacchus task list && bacchus task list --ready && bacchus list
2. Select archetypes: bacchus archetype select <task_id> for each ready task
3. Spawn agents for ready tasks (up to max_concurrent) with appropriate archetype prompts
4. Monitor progress: Check for completed/failed agents
5. Handle releases: Tasks marked ready_for_release get merged by orchestrator
6. Cleanup stale work: bacchus stale --minutes 30 --cleanup

When session ends (all tasks complete):

git push  # Push all completed work to remote

Note: jj git export runs automatically after each task completes, keeping local git in sync.

Conflict Resolution

If the orchestrator detects merge conflicts:

1. Task is marked needs_resolution
2. Agent resolves: jj -R .bacchus/workspaces/{task_id} resolve
3. Agent marks resolved: bacchus resolve {task_id}

Force Exit

To exit without completing:

bacchus session stop

Example: Full Workflow

# User asks to implement authentication with multiple agents

# 1. Plan the work - create tasks with types and reviews
bacchus task init
# Edit .bacchus/tasks.yaml with implementation + review tasks

# 2. Import tasks
bacchus task import --epic-id AUTH

# 3. Start orchestrator session
bacchus session start orchestrator --max-concurrent 3

# 4. Check ready tasks and their archetypes
bacchus task list --ready
bacchus archetype select AUTH-001  # Shows: backend archetype

# 5. Spawn agents with appropriate archetypes
# (use Task tool with archetype prompts)

# 6. Monitor and wait for completion
bacchus task list
bacchus list  # Active claims

# 7. All done - push to remote and end session
git push
bacchus session stop