Delegate a task to one CLI and return its normalized result.

cli is an adapter id (see capabilities); model is optional (the adapter's default otherwise). safety_mode is read_only | propose | write | yolo (default read_only); write and yolo also need a trusted workspace (trust_workspace=true or a configured allowlist). With mode="async" a job id is returned; poll job_status / job_result. session_id resumes a prior session where the CLI supports it.

Ask the same prompt of several targets in parallel and return every voice.

targets is a list of {cli, model} objects (or cli / cli:model strings). Omit it, pass an empty list, or pass "all" to fan out to every installed + authenticated CLI at its default model (capped at max_targets); the result's skipped list explains any adapter left out. Optional stances (parallel to targets) steer each voice: for | against | neutral, and cannot be combined with the auto-expanded panel. synthesize=true adds a server-side combined answer (off by default, so the orchestrator can synthesize the voices itself). With mode="async" a job id is returned.

Return the status and progress of a background job.

Return the result of a finished background job (or a still-running notice).

Review a diff or a set of files across one or more targets (read-only). Provide diff or paths.

Ask one target to produce an ordered implementation plan for a goal (read-only).

List every known CLI: whether it is installed, its auth status, and its available models.

Health-probe each adapter (binary, version, auth, runtime) and diagnose unavailable targets.

Adapters with no non-interactive auth check (e.g. Antigravity) are unknown from the cheap probe; by default (live=true) doctor verifies each installed unknown with a minimal real round trip -- the only trustworthy signal absent a whoami. Pass live=false for a metadata-only check with no model calls (capabilities is the always-cheap snapshot).

List the available role personas that can steer a delegation.