Coordinate a team of agents over a shared task board

Seeds a SQLite board with tasks, launches n_workers mirai daemons, and has each worker loop: atomically claim the next task, run it as a codeagent query, write the result back, repeat until the board is empty. Unlike team_run() (a fixed fan-out where worker i always gets task i), this is a work-stealing pool – a fast worker claims more tasks, so uneven task sizes are balanced automatically. Mirrors Claude Code's TeamCreate + auto-claim.

Usage

team_coordinate(
  tasks,
  model = NULL,
  n_workers = NULL,
  permission_mode = "dont_ask",
  cwd = getwd(),
  blocked_by = NULL,
  worktree = FALSE,
  backoff = 0.5,
  reclaim_timeout = 300,
  db_path = tempfile(fileext = ".sqlite"),
  security_context = NULL
)

Arguments

tasks

Character vector of task prompts.

model

Character. Model spec for each worker.

n_workers

Integer or NULL. Worker count; default cgroup-aware (min(#tasks, parallelly::availableCores())).

permission_mode

Character. Permission mode for workers (default "dont_ask"; parallel workers cannot prompt interactively so "bypass" would silently escalate privileges inherited from the parent).

cwd

Character. Working directory for workers.

blocked_by

List or NULL. Optional DAG dependencies: blocked_by[[i]] is an integer vector of 1-based task indices that must finish before task i can be claimed.

worktree

Logical. Run each worker in its own git worktree (default FALSE).

backoff

Numeric. Seconds a worker waits before retrying when a dependency is still in progress (default 0.5).

reclaim_timeout

Numeric. Seconds after which a claimed task held by a crashed worker is reclaimed back to pending (default 300).

db_path

Character. Board path (created if missing).

security_context

Internal immutable parent security snapshot.

Value

A data.frame: the final board (id, prompt, owner, status, result).