Drive OpenCode sessions

OpenCode is client/server. The Grok chat is the coordinator; OpenCode sessions are the workers. Never do lane work as bare Grok Shell when an OpenCode session should own it.

1. Pick where it runs (by capability, not habit)

• Box (default): pure repo, gh, CI/CD, web flows, headless checks. Run opencode2 run --agent computa … (or the lane's configured primary agent) on the shared box. Wrap standalone runs in the lane's watchdog script when one exists (e.g. /workspace/opencode-lane/bin/opencode-run-watchdog.sh).
• User's Mac (capability-signaled only): work needing Mac-only tools — Simulator, Paper MCP, Figma desktop MCP, aws CLI + SSO, Desktop-only Executor tools, local signed-in apps. Drive OpenCode on the Mac (headless against the Mac opencode serve is fine; TUI not required). Scope that session to the Mac-only slice, collect proof, then continue remaining work on the box.
• Never burn the Mac for box-capable work, and never use bare Grok Shell + machineId as a substitute for a Mac OpenCode session (only tiny diagnostics the user explicitly asked for outside OpenCode).

2. Start visible sessions

• Open the session in the matching project directory/worktree so the user can see it in OpenCode /sessions.
• Title every session you start <Bot name>: <short title> (e.g. kewl-aid: invite accept QA, OpenCode Sr: aws SSO handoff). Set/rename at session start.
• Use the lane's locked model and primary agent from the bot's persona/config; don't launch disabled agents (e.g. build/plan) or other models.
• One task-owned worktree per independent accepted slice; workers on the same task share its checkout and ownership rules.

3. Prompt like an owner

Preserve the user's outcome and explicit decisions, not every mechanism suggested during discussion. Use links to authoritative context rather than copying whole procedures: