Overstory

Multi-agent orchestration for AI coding agents.

No longer maintained. Overstory is no longer actively maintained and this repository is archived (read-only) — I am not reviewing issues or merging pull requests here anymore. Thank you to everyone who contributed. The code stays MIT-licensed, so feel free to fork and carry it forward. Active development has moved to Warren, a self-hostable control plane for sandboxed cloud agents and overstory's spiritual successor — start there for anything new.

Overstory turns a single coding session into a multi-agent team by spawning worker agents in isolated git worktrees, coordinating them through a custom SQLite mail system, and merging their work back with tiered conflict resolution. New projects spawn Claude agents headless and surface them through a web UI (ov serve); tmux attach is the opt-in escape hatch for live steering. A pluggable AgentRuntime interface lets you swap between 11 runtimes — Claude Code, Pi, Gemini CLI, Aider, Goose, Amp, or your own adapter.

Warning: Agent swarms are not a universal solution. Do not deploy Overstory without understanding the risks of multi-agent orchestration — compounding error rates, cost amplification, debugging complexity, and merge conflicts are the normal case, not edge cases. Read STEELMAN.md for a full risk analysis and the Agentic Engineering Book (web version) before using this tool in production.

Install

Requires Bun v1.0+ and git. tmux is optional — only needed if you want to spawn workers with --no-headless or attach to a coordinator/worker pane directly. At least one supported agent runtime must be installed:

• Claude Code (claude CLI)
• Pi (pi CLI)
• GitHub Copilot (copilot CLI)
• Codex (codex CLI)
• Gemini CLI (gemini CLI)
• Cursor CLI (agent CLI)
• Sapling (sp CLI)
• OpenCode (opencode CLI)
• Aider (aider CLI)
• Goose (goose CLI)
• Amp (amp CLI)

bun install -g @os-eco/overstory-cli

Or try without installing:

npx @os-eco/overstory-cli --help

Development

git clone https://github.com/jayminwest/overstory.git
cd overstory
bun install
bun link              # Makes 'ov' available globally

bun test              # Run all tests
bun run lint          # Biome check
bun run typecheck     # tsc --noEmit

Quick Start

# Initialize overstory in your project
cd your-project
ov init

# Install hooks into .claude/settings.local.json
ov hooks install

# Start a coordinator (persistent orchestrator)
ov coordinator start

# Open the web UI — primary operator surface for the swarm
ov serve   # then open http://localhost:7321

ov serve is where you watch the fleet, read the mail bus, and inspect per-agent timelines. New projects spawn Claude workers headless by default, so the UI sees them with full structured-event fidelity.

Other common commands:

# Spawn an individual worker agent (coordinator usually does this for you)
ov sling <task-id> --capability builder --name my-builder

# Force a worker into tmux when you want to attach mid-session
ov sling <task-id> --capability builder --name my-builder --no-headless
tmux attach -t ov-my-builder

# Inspect state from the CLI (also visible in the UI)
ov status
ov dashboard          # live TUI alternative to the web UI
ov mail check --inject
ov nudge <agent-name> # send a follow-up to a stalled agent

Commands

Every command supports --json where noted. Global flags: -q/--quiet, --timing, --project <path>. ANSI colors respect NO_COLOR.

Core Workflow

Coordination

Messaging

Task Groups

Merge

Observability

Infrastructure

Architecture

Overstory uses instruction overlays and tool-call guards to turn agent sessions into orchestrated workers. Each agent runs in an isolated git worktree; new projects spawn Claude workers as headless subprocesses (stream-json over stdout) and surface them through ov serve's web UI, with tmux available as an opt-in for live attach. Inter-agent messaging is handled by a custom SQLite mail system (WAL mode, ~1-5ms per query) with typed protocol messages and broadcast support. A FIFO merge queue with 4-tier conflict resolution merges agent branches back to canonical. A tiered watchdog system (Tier 0 mechanical daemon, Tier 1 AI-assisted triage, Tier 2 monitor agent) ensures fleet health. See CLAUDE.md for full technical details.

Runtime Adapters

Overstory is runtime-agnostic. The AgentRuntime interface (src/runtimes/types.ts) defines the contract — each adapter handles spawning, config deployment, guard enforcement, readiness detection, and transcript parsing for its runtime. Set the default in config.yaml or override per-agent with ov sling --runtime <name>.

Claude Code agents can run in headless mode (the default for new projects — -p --output-format stream-json subprocess, NDJSON events parsed by ClaudeRuntime.parseEvents, surfaced through ov serve's web UI) or tmux mode (escape hatch for live attach — operator can tmux attach to watch and steer mid-session). ov init writes runtime.claudeHeadlessByDefault: true for new projects; legacy projects upgrading from earlier overstory versions keep tmux until they edit config. Override per-spawn with ov sling --no-headless (force tmux) or --headless (force headless). Sapling is statically headless; Pi, Codex, and Cursor have no buildDirectSpawn and reject --headless.

How It Works

Instruction overlays + tool-call guards + the ov CLI turn your coding session into a multi-agent orchestrator. A persistent coordinator agent manages task decomposition and dispatch, while a mechanical watchdog daemon monitors agent health in the background.

Orchestrator (multi-repo coordinator of coordinators)
  --> Coordinator (persistent orchestrator at project root)
        --> Supervisor / Lead (team lead, depth 1)
              --> Workers: Scout, Builder, Reviewer, Merger (depth 2)

Agent Types

Key Architecture

• Agent Definitions: Two-layer system — base .md files define the HOW (workflow), per-task overlays define the WHAT (task scope). Base definition content is injected into spawned agent overlays automatically.
• Messaging: Custom SQLite mail system with typed protocol — 8 message types (worker_done, merge_ready, dispatch, escalation, etc.) for structured agent coordination, plus broadcast messaging with group addresses (@all, @builders, etc.)
• Worktrees: Each agent gets an isolated git worktree — no file conflicts between agents
• Merge: FIFO merge queue (SQLite-backed) with 4-tier conflict resolution
• Watchdog: Tiered health monitoring — Tier 0 mechanical daemon (tmux/pid liveness), Tier 1 AI-assisted failure triage, Tier 2 monitor agent for continuous fleet patrol
• Tool Enforcement: Runtime-specific guards (hooks for Claude Code, extensions for Pi) mechanically block file modifications for non-implementation agents and dangerous git operations for all agents
• Task Groups: Batch coordination with auto-close when all member issues complete
• Session Lifecycle: Checkpoint save/restore for compaction survivability, handoff orchestration for crash recovery
• Token Instrumentation: Session metrics extracted from runtime transcript files (JSONL)

Project Structure

overstory/
  src/
    index.ts                      CLI entry point (Commander.js program)
    types.ts                      Shared types and interfaces
    config.ts                     Config loader + validation
    errors.ts                     Custom error types
    json.ts                       Standardized JSON envelope helpers
    commands/                     One file per CLI subcommand (38 commands)
      agents.ts                   Agent discovery and querying
      coordinator.ts              Persistent orchestrator lifecycle
      supervisor.ts               Team lead management [DEPRECATED]
      dashboard.ts                Live TUI dashboard (ANSI via Chalk)
      hooks.ts                    Orchestrator hooks management
      sling.ts                    Agent spawning
      group.ts                    Task group batch tracking
      nudge.ts                    Agent nudging
      mail.ts                     Inter-agent messaging
      monitor.ts                  Tier 2 monitor management
      merge.ts                    Branch merging
      status.ts                   Fleet status overview
      prime.ts                    Context priming
      init.ts                     Project initialization
      worktree.ts                 Worktree management
      watch.ts                    Watchdog daemon
      log.ts                      Hook event logging
      logs.ts                     NDJSON log query
      feed.ts                     Unified real-time event stream
      run.ts                      Orchestration run lifecycle
      trace.ts                    Agent/task timeline viewing
      clean.ts                    Worktree/session cleanup
      doctor.ts                   Health check runner (13 check modules)
      inspect.ts                  Deep per-agent inspection
      spec.ts                     Task spec management
      errors.ts                   Aggregated error view
      replay.ts                   Interleaved event replay
      stop.ts                     Agent termination
      costs.ts                    Token/cost analysis
      metrics.ts                  Session metrics
      ecosystem.ts                os-eco tool dashboard
      update.ts                   Refresh managed files
      upgrade.ts                  npm version upgrades
      discover.ts                 Brownfield codebase discovery via coordinator-driven scout swarm
      orchestrator.ts             Multi-repo coordination (PersistentAgentSpec)
      completions.ts              Shell completion generation (bash/zsh/fish)
      serve.ts                    HTTP + WebSocket surface for the web UI
      serve/                      REST handlers, WebSocket broadcaster, static SPA fallback
    canopy/
      client.ts                   Canopy client (prompt rendering, listing, emission)
    agents/                       Agent lifecycle management
      manifest.ts                 Agent registry (load + query)
      overlay.ts                  Dynamic CLAUDE.md overlay generator
      identity.ts                 Persistent agent identity (CVs)
      checkpoint.ts               Session checkpoint save/restore
      lifecycle.ts                Handoff orchestration
      hooks-deployer.ts           Deploy hooks + tool enforcement
      copilot-hooks-deployer.ts   Deploy hooks config to Copilot worktrees
      guard-rules.ts              Shared guard constants (tool lists, bash patterns)
      mail-poll-detect.ts         Bash mail-poll pattern detector (runtime backstop)
      scope-detect.ts             Soft FILE_SCOPE violation detection (builder/merger)
    worktree/                     Git worktree + tmux management
    mail/                         SQLite mail system (typed protocol, broadcast)
    merge/                        FIFO queue + conflict resolution + sentinel-file lock + dry-run prediction
    watchdog/                     Tiered health monitoring (daemon, triage, health)
    logging/                      Multi-format logger + sanitizer + reporter + color control + shared theme/format
    metrics/                      SQLite metrics + pricing + transcript parsing
    doctor/                       Health check modules (13 checks)
    utils/                        Shared utilities (bin, fs, pid, time, version)
    insights/                     Session insight analyzer + quality-gate runner (success/partial/failure)
    runtimes/                     AgentRuntime abstraction (registry + adapters: Claude, Pi, Copilot, Codex, Gemini, Sapling, OpenCode, Cursor, Aider, Goose, Amp)
    tracker/                      Pluggable task tracker (beads + seeds backends)
    mulch/                        mulch client (programmatic API + CLI wrapper)
    e2e/                          End-to-end lifecycle tests
  agents/                         Base agent definitions (.md, 9 roles) + skill definitions
  templates/                      Templates for overlays and hooks

Configuration

Gateway Providers

Route agent API calls through custom gateway endpoints (z.ai, OpenRouter, self-hosted proxies). Configure providers in .overstory/config.yaml:

providers:
  anthropic:
    type: native
  zai:
    type: gateway
    baseUrl: https://api.z.ai/v1
    authTokenEnv: ZAI_API_KEY
  openrouter:
    type: gateway
    baseUrl: https://openrouter.ai/api/v1
    authTokenEnv: OPENROUTER_API_KEY
models:
  builder: zai/claude-sonnet-4-6
  scout: openrouter/openai/gpt-4o

How it works: Model refs use provider/model-id format. Overstory sets ANTHROPIC_BASE_URL to the gateway baseUrl, ANTHROPIC_AUTH_TOKEN from the env var named in authTokenEnv, and ANTHROPIC_API_KEY="" to prevent direct Anthropic calls. The agent receives "sonnet" as a model alias and Claude Code routes via env vars.

Environment variable notes:

• ANTHROPIC_AUTH_TOKEN and ANTHROPIC_API_KEY are mutually exclusive per-agent
• Gateway agents get ANTHROPIC_API_KEY="" and ANTHROPIC_AUTH_TOKEN from provider config
• Direct Anthropic API calls (merge resolver, watchdog triage) still need ANTHROPIC_API_KEY in the orchestrator env

Validation: ov doctor --category providers checks reachability, auth tokens, model-provider refs, and tool-use compatibility.

ProviderConfig fields:

Troubleshooting

Recovering a dead lead (or any agent that exited mid-task)

If a lead exits without sending merge_ready (process termination, watchdog kill, manual ov stop) and the task was already closed, both ov nudge and ov sling would normally refuse to re-engage:

• ov nudge <name> reports No active session for agent "..." (state: completed). The agent's process is gone, so there's nothing to send keystrokes to.
• ov sling <task-id> --capability lead reports Task "..." is not workable (status: closed).

To re-dispatch a fresh lead against the same task, pass --recover:

ov sling <task-id> --capability lead --recover --name <fresh-name>

--recover bypasses the workable-status check so the new lead can pick up where the dead one left off (the task remains closed; the new lead reads the spec and proceeds). The terminal-state nudge error itself includes a copy-paste hint to this exact form.

Coordinator died during startup

This error means the coordinator tmux session exited before the TUI became ready. The most common cause is slow shell initialization.

Step 1: Measure shell startup time

time zsh -i -c exit   # For zsh
time bash -i -c exit  # For bash

If startup takes more than 1 second, slow shell init is likely the cause.

Step 2: Common slow-startup causes

Step 3: Configure shellInitDelayMs in .overstory/config.yaml:

runtime:
  shellInitDelayMs: 3000

• Default: 0 (no delay)
• Typical values: 1000–5000 depending on shell startup time
• Values above 30000 (30s) trigger a warning
• Inserts a delay between tmux session creation and TUI readiness polling

Step 4: Optimization examples

Lazy-load nvm (add to ~/.zshrc or ~/.bashrc):

# Lazy-load nvm — only activates when you first call nvm/node/npm
export NVM_DIR="$HOME/.nvm"
nvm() { unset -f nvm node npm npx; [ -s "$NVM_DIR/nvm.sh" ] && . "$NVM_DIR/nvm.sh"; nvm "$@"; }
node() { unset -f nvm node npm npx; [ -s "$NVM_DIR/nvm.sh" ] && . "$NVM_DIR/nvm.sh"; node "$@"; }
npm()  { unset -f nvm node npm npx; [ -s "$NVM_DIR/nvm.sh" ] && . "$NVM_DIR/nvm.sh"; npm  "$@"; }
npx()  { unset -f nvm node npm npx; [ -s "$NVM_DIR/nvm.sh" ] && . "$NVM_DIR/nvm.sh"; npx  "$@"; }

Reduce oh-my-zsh plugins (edit ~/.zshrc):

# Before: plugins=(git zsh-autosuggestions zsh-syntax-highlighting node npm python ruby rbenv pyenv ...)
# After: keep only what you use regularly
plugins=(git)

Part of os-eco

Overstory is part of the os-eco AI agent tooling ecosystem.

Contributing

Contributions are welcome! See CONTRIBUTING.md for guidelines.

License

MIT