| name | taskfleet |
| description | Operate a local Taskfleet task DAG, claims, leases, worktrees, gates, workflow pipelines, linked objectives, integration, and recovery through its packaged MCP stdio server. Use when orchestrating or observing multi-agent repository work with Taskfleet. |
Taskfleet
Taskfleet resolves an explicit TASKFLEET_CONFIG, the nearest ancestor
taskfleet.toml, or an external enrollment created by the human command
/fleet enable external. It never auto-enables an unconfigured repository.
Optionally set TASKFLEET_BIN when taskfleet is not on PATH.
Discover the packaged server contract before calling it:
tools = await taskfleet.list_tools()
Call MCP tools as async methods, for example:
await taskfleet.taskfleet_task_query(ready=True)
await taskfleet.taskfleet_task_claim(owner="worker-1", limit=1)
Default linked-objectives ritual
After claiming (or before implementing) any task, expand linked objectives.
Taskfleet does not encode domain roles; connectors put a shared value on an
arbitrary dossier path (conventionally meta.bundle) and optional free-form
meta.role labels.
- Claim or inspect a task.
- Call
taskfleet_task_related(task="<uri>") (optional path overrides
meta.bundle).
- Use the returned set to plan the full objective package before coding.
- For each ready dependency, call
taskfleet_task_context / receipt tools.
- Implement only this claim's role; keep the lease alive with heartbeats.
- Advance only after committed-tree gates pass. Aggregator tasks that list the
package in
depends_on stay unclaimable until those edges are
candidate or done.
If task.related fails closed because the seed task has no shared path value,
do not invent siblings — ask the operator or re-ingest with a bundle key.
Pipeline runs
Workflows are domain-agnostic pipelines. After claim or spawn, read
execution.active_step.instruction and execution.active_step.args (merged as
workflow defaults < step defaults < meta.args). Create runs with
taskfleet_task_spawn(workflow="...", args={...}). Repeat with
taskfleet_task_rerun(task="...") — this always creates a new uri and respects
max_runs (0 = unlimited). Never reset a finished uri in place.
Taskfleet remains the source of execution truth. Use Prime Agent subagents for
implementation, keep claims alive with taskfleet_task_heartbeat, and advance
only after committed-tree gates pass. Use taskfleet_reconcile() after an
interruption or restart.
