Co-pilot

Orchetree’s Co-pilot is implemented as agent trees in the built-in Showcase Project. Requests route through an orchestration tree into focused Specialist Trees, and deeper work can hand off to expert trees instead of collapsing into an opaque direct model call.

Action Cards

For workspace changes, the Co-pilot proposes structured Action Cards. A card identifies the action, target, editable parameters, and reason. Review and approve the card before any mutation is applied.

Typical actions include creating a tree or node and attaching a connector or reviewed skill.

Prompt Highlights

Prompt coaching can identify a specific span, explain the issue, and propose a rewrite. Apply or reject each suggestion from the prompt editor; suggestions do not silently overwrite prompt blocks.

Testing

The Co-pilot can propose representative, edge-case, and adversarial inputs for a node or tree. Test executions create ordinary run and trace records, which keeps model output, latency, errors, and suggested changes inspectable.

Optional background prompt analysis starts with a local heuristic pass. It only invokes the Prompt Coaching Specialist Tree when that pass finds a potential issue.

Auditability

Co-pilot executions appear in the Showcase Project’s Runs area. Workspace mutations still pass through normal role checks, skill review, and explicit approval.

See Product tour, Runs and traces, and Roles and governance.