| name | opencode-runtime |
| description | Spawn a Haiku worker to run a task via OpenCode. The worker calls oc-run.sh which executes opencode run headlessly. No tmux, no server, no custom UI. |
| user-invocable | false |
OpenCode Runtime
Spawn a Haiku subagent that relays your task to OpenCode via bash. Progress shows in Claude's native agent panel.
How to delegate
Agent(
subagent_type="swarm-code:opencode-worker",
model="haiku",
prompt="<your full task with all context inline>"
)
The Haiku worker will:
- ACK via SendMessage
- Run
bash oc-run.sh "<prompt>" → calls opencode run --model <m> "<prompt>"
- Return the result via SendMessage
When to use
- Code review of a specific diff or file excerpt
- Writing an implementation plan with clear requirements
- Answering a specific technical question with context
- Any analytical task over ~50 tokens where you'd repeat work OpenCode can do
When NOT to use
- Tasks that need live file editing (do those yourself)
- Short questions you can answer in one sentence
- Tasks requiring iterative back-and-forth with the user mid-task
Prompt requirements
Workers CANNOT read files. Include all context inline:
## Task
Review this function for bugs.
## Code
<paste the actual code here>
## Output Format
Bullet list: [SEVERITY] file:line — description. Max 5 findings.
Model setup
Run /swarm-code:init once per project to configure the OpenCode model.
The worker automatically uses the configured model — no manual override needed.
