robsonrung OpenCode Runner
Use this skill as a safe OpenCode handoff guide. This package intentionally does not ship an executable OpenCode wrapper.
Native routing
Read shared/references/host-model-execution.md before choosing an external route. If the current host can dispatch the exact selected model with the required effort, isolation, and receipt policy in a native subagent or task thread, use that route. agents/openai.yaml is host UI metadata; it does not dispatch a native model.
Routing limit
OpenCode has no runner envelope in this library. It cannot be an approved
implementation or review route because there is no model_receipt for the
launcher to validate. Use it only for a manually approved comparison. Report
the requested provider and model as configuration, not as proof of the model
that served the run.
Safety Model
OpenCode can read project files and send prompt context to the configured model provider. Treat every run as an external model call.
Before running OpenCode:
- Confirm the user asked for OpenCode or that the active workflow explicitly selected it.
- Confirm the working directory and prompt scope.
- Prefer read only review or planning prompts.
- Avoid attaching secrets, credentials, private keys, tokens, or unrelated files.
- Use the host agent's normal shell approval, sandbox, and logging flow for the command.
If the user asks for unattended editing, explain that this skill does not provide an unattended wrapper. Use a normal host approved command instead.
Manual Invocation Pattern
After the checklist above passes, run the local OpenCode CLI directly from the target repository. Keep the command narrow and pass a concise prompt. Capture stdout and stderr in the normal host tool output.
opencode run "<concise prompt>"
opencode run -m <provider/model> "<concise prompt>"
opencode run --agent plan "<concise review prompt>"
For review tasks, ask OpenCode for analysis only; the plan agent variant above keeps the run read only. For implementation tasks, ask OpenCode for a patch plan first, then apply changes through the primary host agent.
Output Contract
Return a short summary with:
- OpenCode command scope
- Main result
- Files or areas discussed
- Any uncertainty, failure, or missing prerequisite
When acting as a seat in a cross-runner workflow (for example models-consensus or council), return this same summary to the orchestrating skill verbatim.
Missing Prerequisite
If OpenCode is not installed or authenticated, report the missing prerequisite and stop. Do not provide install commands from this skill.