agent-track

.track

The conductor's podium for AI-assisted development.

.track lets you orchestrate multiple Claude Code sessions (or any AI agents) working on the same codebase in parallel. It passively observes agent activity via hooks — zero token overhead, zero agent compliance required — and gives you a live dashboard to see everything at once.

Why

You have 3 Claude Code sessions open. Each one is productive in isolation. But:

• Which agent is working on what?
• Are two agents editing the same file?
• Did anyone run the tests?
• Is the original plan being followed?

.track answers all of these automatically. Agents don't need to do anything special — hooks capture everything passively.

Install

pipx install agent-track    # recommended
# or
pip install agent-track
# or
uv add agent-track

This gives you the track command.

Quick Start

track init                                              # set up .track/ + hooks
track create --title "Fix auth bug" --desc "token refresh fails after 2h"  # create + auto-claim
track list                                              # see all tickets
track board --last 10                                   # see what's happening
track serve                                             # web dashboard at localhost:7777

That's the entire agent workflow. Everything else — registration, heartbeats, file tracking, conflict detection — happens automatically via hooks.

How It Works

Passive Observation

.track installs Claude Code hooks that fire on every session start, tool use, and session end. No agent compliance needed:

Split Storage

• .track/ (in your repo, git-tracked) — tickets, board, config
• ~/.track/projects/{key}/ (ephemeral, per-machine) — agents, sessions, locks, activity logs

This means tickets are shared across worktrees and collaborators, while runtime state never causes merge conflicts.

Agent API

Agents only need 3 commands:

# Create a ticket (auto-claimed to the current session)
track create --title "Fix auth bug" --desc "See docs/plan.md:45-72"

# Create a ticket for another agent to pick up
track create --title "Add retry logic" --no-claim

# Mark work as done
track update T-0001 --status review

Agent identity is auto-detected from the active session. No --agent flags needed.

Board

track board --last 10                          # read
track board -m "blocked on auth module"        # post
track board -m "found the bug" --ticket T-0001 # post with ticket ref

Features

• Passive observation — all tracking via Claude Code hooks, zero token overhead
• Auto-registration — agents get a NATO phonetic alias on session start
• File conflict detection — warns when two agents edit the same file
• Sensitive file protection — blocks/warns on access to .env, *.pem, etc.
• Session summaries — auto-generated on session end (files modified, test runs, errors)
• Todo capture — agent's internal task list tracked with diffs
• Web dashboard — kanban board + agent panel + file conflicts + board at localhost:7777
• Zero dependencies — pure Python stdlib
• Git-friendly — tickets are markdown, all state is plain text
• Worktree-safe — resolves all worktrees to the same project identity

Commands

For agents (minimal)

For humans (full control)

Legacy (still supported)

Ticket Lifecycle

backlog → claimed → in-progress → review → done

Valid transitions are enforced. Use --force to override.

CLAUDE.md Integration

Add this to your project's CLAUDE.md:

## Agent Protocol

Your session is **automatically tracked** via hooks. No registration, heartbeats, or deregistration needed.

Before starting work:
- `track board --last 10`
- `track list`

Create a ticket (auto-claimed to you):
- `track create --title "Fix auth bug" --desc "See docs/plan.md:45-72"`

Create for another agent: `track create --title "..." --no-claim`

While working:
- Reference ticket IDs in git commits: `T-0001: fix token refresh`
- Post to the board: `track board -m "message" --ticket T-NNNN`

When done:
- `track update T-NNNN --status review`

Rules:
- One ticket at a time
- Check the board before starting work
- Reference ticket IDs in commits
- Never modify `.track/` files directly

A full example is available at src/agent_track/data/CLAUDE.md.example.

Dashboard

track serve          # start at http://localhost:7777
track stop           # stop

The dashboard shows:

• Kanban board — tickets across all statuses
• Agent panel — active agents with heartbeat, current ticket, model
• File ownership — who's editing what, conflict warnings
• Message board — recent agent communication
• Agent todos — live view of each agent's internal task list (on ticket detail)

Platform

macOS and Linux. Windows is not supported (fcntl.flock, os.fork are Unix-only). WSL works.

License

MIT