opencode-subagentslisted
Install: claude install-skill toomingos/opencode-subagents
`run.sh` lives next to this file: `.claude/skills/opencode-subagents/run.sh` for a project
install, `~/.claude/skills/opencode-subagents/run.sh` for a global one. Use that path below.
One Bash call per agent, task on stdin. In Claude Code set `run_in_background: true`:
```bash
bash <skill-dir>/run.sh --agent <name> "<short title>" <<'EOF'
<full task prompt. End with: "Final message: what changed, where, how verified, or what blocked you.">
EOF
```
Launch returns a Claude Code task id and output file at once; keep working. On exit you get a
`<task-notification>` with the exit code. Then read the output file: it holds one `run.sh: run <run-id> …`
line and the envelope `<task id="ses_…" state="completed|error">` with `<task_result>` or
`<task_error>` (the agent's final text). Relay it faithfully, session id included. Never do the task
yourself instead.
- Many agents: one `run.sh` call per agent, all in the same message, each with `run_in_background: true`.
Each is its own Claude Code task with its own output file and its own notification; they finish
independently and in any order. You choose how many to launch; there is no built-in cap.
- Monitor one agent mid-run (foreground calls; `<run-id>` or a word from its title):
`run.sh --peek <run-id>` (state, elapsed, session, last text), `run.sh --events <run-id>` (every tool
call and text so far), `run.sh --peek list` (all runs). The raw stream is
`$TMPDIR/opencode-subagents-runs/<run-id>.jsonl`; Claude Code's Monitor