List configured worker agents and their capabilities (adapter type, transport, native session resume support, usage accounting quality) plus per-agent SCORE STATS (avg_score, success_rate, repair_rate, out_of_scope_rate, blocked_rate, avg_tokens, median_duration_sec, last_tasks). Use the stats to route work: high-score agents for hard units, cheaper/lower-score agents for well-specified mechanical units. Use this before run() to pick a valid agent_id.

Submit a complete work unit to a worker agent and return its task_id.

Requirements:

• agent_id: exact id from agents(); every task is explicitly bound to one agent.

• goal: full objective, constraints and context in ONE prompt (complete work unit; do not ping-pong per file/line).

• workspace: directory path (canonicalized to an absolute path by the kernel).

• scope_files: RELATIVE paths the worker may modify (absolute or escaping paths are rejected; enforced by diff audit).

• forbid: extra prohibitions.

• verify: commands executed by the kernel after the turn; the worker cannot pass by editing them. Prefer list form; string form is arg-split (no shell) unless config verify_allow_shell is on. Exit code 0 = pass.

• mode: adapter-specific permission mode (zcode: plan|build|edit|yolo).

• session_ref: task_id of a previous task on the SAME agent+workspace to continue its native session. Cross-agent/workspace reuse is rejected.

• timeout_sec: per-turn timeout; on timeout the task becomes blocked (never silently resubmitted; the lock is held and a watcher keeps observing).

• wait_sec: bounded block until finish (default: return immediately).

• budget_tokens: hard token budget; exceeded -> task failed (budget_exceeded).

• permission_policy: "deny" (default) or "allow" (auto-allow within this task).

• repair_allowed: one pre-authorized targeted repair attempt (default true).

• idempotency_key: resubmitting the same key returns the original task.

Returns a snapshot dict with ok/status/code. A run() success only means SUBMITTED, not done — follow with wait()/inspect().

Bounded blocking wait for a task to reach a terminal or blocked state. Does not busy-poll. Returns the snapshot dict; if it still runs when the timeout expires, wait_timed_out=true — call again later. 'submitted'/'running' is NOT completion; only status succeeded/failed/blocked/cancelled/interrupted is.

Read task details on demand (logs/diffs stay on disk; read slices here). what: status | summary | diff | verify | log | raw | usage | diagnose.

• diagnose returns a compact SANITIZED report (secrets/emails/blobs redacted, size-bounded) with a stable code and actionable hints — safe to share.

Request cancellation and report whether the stop was CONFIRMED by the native runtime (confirmed=true only on an observed terminal/exit). A cancel request alone is not a stop.