pac-collab

Multi-project sandbox coordinator for concurrent AI agent sessions.

Multiple independent agents can work on different tasks across entirely different repositories at the same time. pac-collab gives each task a lock (so two agents never grab the same task in the same project) and an isolated git worktree (so their files never touch), all tracked in one shared SQLite registry.

How it works

• Shared state — a single coordinator.db lives in the OS user data directory (located via platformdirs), so every invocation of the package, from any process or repository, sees the same master registry. WAL journal mode + 10s busy timeout make it safe under concurrency.
• The lock — a partial unique index on (project_id, task_id) WHERE status = 'ACTIVE'. Acquiring inserts the ACTIVE row inside a BEGIN IMMEDIATE transaction; two concurrent acquires of the same task cannot both succeed. Finished rows stay as history.
• File isolation — every task gets its own worktree at ~/.pac-collab/sandboxes/{project_id}/task_{task_id}, on its own branch agent/{feature_name}-{task_id} cut from origin/main (configurable via base_ref / --base-ref for repos whose default branch differs).

Install

pip install pac-collab

MCP server (primary interface for agents)

The package ships an MCP server (stdio) exposing every coordinator operation as a tool, so agent harnesses call the coordinator natively instead of shelling out:

pac-collab-mcp                     # console script
python -m pac_collab.mcp_server    # PATH-independent equivalent

Register it in any MCP client (.mcp.json / Claude Code / plugin):

{
  "mcpServers": {
    "pac-collab": {
      "command": "python",
      "args": ["-m", "pac_collab.mcp_server"]
    }
  }
}

Tools: acquire_sandbox(base_repo_path, project_id, task_id, feature_name, base_ref="origin/main"), heartbeat(project_id, task_id), release_sandbox(project_id, task_id, outcome), list_sandboxes(project_id=None), prune_stale_sandboxes(timeout_minutes=60) — identical semantics to the Python API below. Every tool call opens its own SQLite connection, so concurrent sessions can each run their own server instance against the same shared registry.

Python API

from pac_collab import SandboxCoordinator

with SandboxCoordinator() as coord:
    grant = coord.acquire_sandbox(
        base_repo_path="C:/repos/backend-api",
        project_id="backend-api",
        task_id="T-101",
        feature_name="rate-limiter",
    )
    # {"status": "ALLOCATED", "sandbox_path": "...", "branch_name": "agent/rate-limiter-T-101"}
    # or {"status": "REJECTED", "reason": "Task already active in this project"}

    coord.heartbeat("backend-api", "T-101")          # call periodically while working
    coord.release_sandbox("backend-api", "T-101", "SUCCESS")  # commit + push + destroy worktree

    coord.list_sandboxes()                # everything, all projects
    coord.list_sandboxes("backend-api")   # one project
    coord.prune_stale_sandboxes(60)       # reap ACTIVE rows with heartbeats older than 60 min

CLI

Every command prints JSON; failed operations exit non-zero.

pac-collab acquire --base-repo-path C:/repos/backend-api --project-id backend-api --task-id T-101 --feature-name rate-limiter
pac-collab heartbeat --project-id backend-api --task-id T-101
pac-collab release  --project-id backend-api --task-id T-101 --outcome SUCCESS
pac-collab list     [--project-id backend-api]
pac-collab prune    [--timeout-minutes 60]

Semantics worth knowing

• release --outcome SUCCESS commits (agent: completed {task_id}) and pushes before destroying the worktree. If the commit or push fails, the sandbox is kept ACTIVE and intact — no work is lost, and the error is returned so the agent can retry.
• release --outcome FAILED discards everything: worktree removed, local branch deleted, row marked FAILED.
• prune marks stale sessions STALE_FAILED and removes their worktrees even if the base repo has vanished — the lock is always released.
• Identifiers are slugified ([^A-Za-z0-9._-] → -) for branch and directory names; the lock itself always uses the raw ids.
• Statuses: ACTIVE → SUCCESS | FAILED | STALE_FAILED.

Development / publishing

python -m build
twine upload dist/*

Publish with the eatopia PyPI token (in ~/.pypirc, username __token__) — same account as pac-pincers.