TaskWraith is a credible, differentiated local-first control plane for working with multiple coding-agent providers. It gives provider seats a shared or isolated work topology, explicit authority, and inspectable local evidence; qualified tool-capable seats also receive workspace-scoped tools and approvals.
That is stronger and more useful than “provider switching,” but it is not a claim of category leadership, universal native-tool parity, or enterprise compliance. The proof standard is live behaviour at a named version, not the size of the feature list.
- Governed orchestration: Ensemble seats can debate and hand off in one transcript; sub-threads can run isolated work in parallel and return an explicitly untrusted result. Boss/Captain authority, per-seat permissions, approval gates, and audit events keep coordination visible.
- Containment with boundaries: TaskWraith-brokered workspace operations are policy checked and auditable. Provider-native CLIs and project-level MCP loading vary, so their containment claims require version-pinned live canaries. A Read-only label is not, by itself, proof that a hostile native MCP server cannot start. Managed Cursor participates as a normal solo, Ensemble, or delegated seat: TaskWraith-mediated calls follow the selected policies and grants, while Cursor-native actions remain provider-owned and sandbox-bounded. Treat that native sandbox as a partial backstop, not a full egress seal.
- Honesty is part of the product: Cache behaviour is labelled Guaranteed only where TaskWraith controls the API request, Automatic where a provider reports observed implicit caching, and Best-effort on opaque CLI paths. Native and emulated forks are labelled separately, and returned child output is not silently promoted to trusted instruction.
| Need | Use | Why |
|---|---|---|
| One clear task, one execution context | Solo chat | Lowest cost and coordination overhead |
| Parallel, noisy, or context-isolated work | Sub-thread | Independent provider session; result returns with provenance |
| Deliberation, hand-offs, or role-separated review | Ensemble | Peers share the transcript and can challenge work in sequence or parallel |
Sub-threads are depth-one in the current product. Ensemble seats share transcript context, so extra participants add cost and noise as well as perspective.
Start with Boss + Captain + one Specialist. Add a second Specialist only when the task crosses domains, and add one Outsider when independent dissent is worth the extra turn. Keep one write-capable seat by default; give review and outsider seats Read-only posture until the task requires more. Cursor seats run with the same TaskWraith policy and grant controls for mediated tools; start outsider Cursor seats Read-only for the same least-authority reason as any other provider.
- Delivery (3–4): Boss, Captain/writer, test Specialist, optional Outsider.
- Risky change (4–5): Boss, Captain/writer, security Specialist, test Specialist, optional Outsider.
- Investigation (3–4): Boss, Captain, debugging Specialist, adversarial Outsider.
Thirty participants is a capacity ceiling, not a recommended starting panel.
The latest tagged public baseline is whatever Releases currently shows — deliberately not restated here, because a pinned number silently becomes a false claim the moment the next tag lands, which is how this line came to read v1.8.8 two releases later. This checkout is source-ahead; its compact default panels and any other Unreleased entries are implementation evidence, not shipped capability, until a new tag and matching artifacts are published. Public claims should name the release or commit they were verified against.