A bounded "lead" loop that faithfully ports Claude Code's
COORDINATOR_MODE: a lead model decomposes a high-level goal into a task DAG,
a work-stealing team (team_coordinate()) auto-claims and runs it, then the
lead reviews the results and either declares the goal done or adds a
follow-up round – repeating up to max_rounds.
The three LLM/execution steps are injectable (decompose_fn, review_fn,
coordinate_fn) so the loop control is unit-testable without a live model;
the defaults are ellmer structured-output calls + the real board.
Usage
team_lead(
goal,
model = NULL,
cwd = getwd(),
max_rounds = 3L,
n_workers = NULL,
permission_mode = "bypass",
worktree = FALSE,
decompose_fn = NULL,
review_fn = NULL,
coordinate_fn = NULL
)Arguments
- goal
Character(1). The high-level objective.
- model
Character. Model spec for the lead and the workers.
- cwd
Character. Working directory.
- max_rounds
Integer. Maximum decompose/review rounds (default 3).
- n_workers, permission_mode, worktree
Passed to
team_coordinate().- decompose_fn, review_fn, coordinate_fn
Injectable steps (for testing); default to ellmer structured calls + the real board.