Start an opencode run in the background as a directly-spawned child process (no tmux, no shared visibility into the orchestration layer) and return a task_id immediately. Poll with opencode_status, then read opencode_result once done.

Stop a running task: sends SIGTERM to the task's whole process group (opencode and any subprocess it spawned), escalating to SIGKILL after a grace period if it hasn't exited. A finished task's status is unaffected and returns a note instead of an error. Poll opencode_status afterward; the task moves to status 'cancelled' once its exit event lands.

Block on a running task's real exit event (or a timeout, whichever comes first) and return its status once settled. The closest analog to the built-in Agent tool's auto-resume behavior available over plain MCP request/response, without a poll loop. Capped internally at 45s so the call returns cleanly instead of hitting Claude Code's own MCP tool-call timeout; if status is still 'running' when it returns, call opencode_wait again.

Return structured status for a dispatched task: running | done | crashed | cancelled | unknown, plus exit code and log path once finished. Backed by the child process's real exit event, not log string-matching.

Return the final assistant message and metadata (tokens, cost, session id) for a finished task, parsed from opencode's own --format json event stream. message is only the model's last turn (after all tool calls finish); narration includes intermediate step narration too, in order. Errors politely if the task is still running.

List all known tasks (this server process's lifetime) with their statuses, newest first.