A Weave is a collection of tasks generated from a single large request. You — or a model — describe a big change, and it's broken into bite-sized tasks that run independently.
Creating a Weave
- By hand — add and order tasks yourself on the Weaves page.
- LLM generation — describe the work and let a (usually larger) model decompose it into tasks for you.
- From chat — ask the assistant in a project-attached chat to create the weave.
Running a Weave
Run the whole weave or individual tasks. A weave moves through clear states —
pending, running, reviewing, completed,
failed, or cancelled — and you can stop a run at any time. If the
app is closed mid-run, interrupted tasks are recovered on the next start rather than left
stuck.
Defaults and overrides
A weave sets defaults for all its tasks: which tools are allowed, which skills to inject, which MCP servers to offer, and any extra directories tasks may access. Individual tasks can override any of these — see Tasks & isolation.
Review before merge
Mark a task to pause for review and, on Linux, the weave stops after it so you can inspect the captured file changes and Approve, Redo, or Reject them before anything touches your project. See OverlayFS review.
