b0b

An AI agent pipeline that takes coding tasks from dispatch through planning, implementation, auditing, testing, and PR creation using Claude and Codex agents in isolated git worktrees.

How It Works

Each task moves through a state machine of phases, driven by a monitor loop that runs every few minutes. Agents run in tmux sessions against dedicated git worktrees, so multiple tasks can execute in parallel without interfering with each other.

flowchart LR
    planning --> plan_review
    plan_review --> implementing
    implementing --> auditing
    auditing --> fixing
    fixing --> testing
    testing --> pr_creating
    pr_creating --> reviewing
    reviewing --> pr_ready
    pr_ready --> merged

Loading

Phase	What happens
planning	Agent writes a plan; outputs PLAN_VERDICT:READY when done
plan_review	Optional human gate -- approve or reject the plan
implementing	Agent implements the plan in an isolated worktree
auditing	A different agent reviews the implementation for correctness
fixing	Agent addresses audit findings
testing	Agent runs tests and verifies the change
pr_creating	Agent creates a pull request
reviewing	Waits for CI and human review
pr_ready	CI passes, ready for merge
merged	Done

Tasks that exceed iteration limits are marked needs_split and can be broken into subtasks with auto-split.sh.

Project Structure

.clawdbot/
├── scripts/          Bash pipeline orchestration (dispatch, monitor, spawn, notify, etc.)
├── prompts/          Prompt templates with {VAR} placeholders (plan, implement, audit, test, fix, PR)
└── plans/            Design documents for the pipeline itself

.openclaw/
├── kopiclaw/         Workspace identity, persona, tools, architecture, and memory files
└── kl/               Alternate workspace variant (K&L email pipeline)

See .clawdbot/README.md for detailed script-by-script documentation.

Prerequisites

• macOS (scripts assume macOS paths and launchd)
• bash, python3 (stdlib only -- no pip dependencies), git, tmux, curl
• gh (GitHub CLI), authenticated via gh auth login
• claude (Claude Code CLI) and/or codex in your PATH

Installation

# 1. Clone
git clone <repo-url> && cd b0b

# 2. Create the state directory (lives outside the repo)
mkdir -p ~/.openclaw/workspace-kopiclaw/pipeline/{logs,plans}

# 3. Create the worktree base directory
mkdir -p ~/Projects/kopi-worktrees
# Or override later: export WORKTREE_BASE=/your/preferred/path

# 4. Set up Slack credentials (optional -- needed for notifications)
mkdir -p ~/.openclaw/credentials
echo "xoxb-your-slack-bot-token" > ~/.openclaw/credentials/slack-bot-token
chmod 600 ~/.openclaw/credentials/slack-bot-token

# 5. Authenticate GitHub CLI
gh auth login

Configuration

All configuration lives in .clawdbot/scripts/config.sh. Every variable can be overridden via environment variables.

Variable	Description	Default
CLAWDBOT_STATE_DIR	Root directory for state files, logs, and plans	~/.openclaw/workspace-kopiclaw/pipeline
WORKTREE_BASE	Where git worktrees are created for tasks	~/Projects/kopi-worktrees
CLAUDE_PATH	Path to the Claude CLI	claude
CODEX_PATH	Path to the Codex CLI	codex
SLACK_BOT_TOKEN	Slack bot token for notifications	(read from ~/.openclaw/credentials/slack-bot-token if unset)
MAX_RUNTIME_SECONDS	Agent timeout per phase	2700 (45 min)
PLANNING_TIMEOUT_SECONDS	Planning-specific timeout	1200 (20 min)
MAX_ITERATIONS	Max audit/fix cycles before giving up	4
MAX_AUTO_RETRIES	Max times a task can auto-retry from needs_split	1
MAX_SPLIT_DEPTH	Max depth of automatic task splitting	1
GH_COMMENT_DISPATCH_ENABLED	Auto-dispatch tasks from GitHub @kopi-claw mentions	true
GH_COMMENT_DEFAULT_AGENT	Which agent to use for new tasks	claude
GH_COMMENT_MAX_DISPATCHES	Max tasks dispatched per monitor cycle	3

Usage

Dispatch a task

.clawdbot/scripts/dispatch.sh \
  --task-id my-feature \
  --branch feat/my-feature \
  --product-goal "Add widget support" \
  --description "Implement widget component with tests" \
  --user-request "Original conversation context" \
  --agent claude \
  --phase planning

The --user-request flag provides the original conversation context shown to the planning agent. Without it, the agent investigates the codebase independently and may choose an approach you didn't intend.

Approve or reject a plan

.clawdbot/scripts/approve-plan.sh <task-id>
.clawdbot/scripts/reject-plan.sh <task-id> --reason "scope too broad"

Check pipeline status

.clawdbot/scripts/pipeline-status.sh   # Slack-formatted summary
.clawdbot/scripts/check-agents.sh      # Raw agent status

Run the monitor loop

The monitor is the heartbeat of the pipeline. It checks agent status, advances phases, processes GitHub mentions, and sends Slack notifications.

.clawdbot/scripts/monitor.sh

In production, run it on a schedule (e.g. every 5 minutes via launchd or cron).

Clean up completed worktrees

.clawdbot/scripts/cleanup-worktrees.sh

Running Tests

python3 -m pytest .clawdbot/scripts/

Or run individual test files:

python3 .clawdbot/scripts/test_monitor_phase_advance.py
python3 .clawdbot/scripts/test_gh_poll_process.py

License

MIT