Each task in a Weave is a single, isolated logical change. Because no context is carried between tasks, the context window for any one task stays small — small enough that a locally-hosted model can handle it reliably.
This is the whole trick behind Weave: intelligence moves out of the model's enormous context window and into the structure of the work. The better the decomposition, the smaller the model you can get away with.
What a task carries
- A prompt describing the change, and an order within the weave.
- An agent (model/provider) to run on — tasks in one weave can use different agents.
- Optional per-task overrides of tools, skills, MCP servers, and allowed paths.
Isolation controls
- Pause for review — stop the weave after this task so you can review its writes. See OverlayFS review.
- Disable network — cut off network access for this task's shell commands (file tools never get network regardless).
- Allowed paths — grant access to directories beyond the project when a task genuinely needs them.
After a run
Each task records its output, status, and metrics — tool-call count and input/output token usage — plus the exact command every tool ran (including the full sandbox arguments) in its raw view. You can re-run a task with the same or a modified prompt at any time.
