/sudo
the autonomy contract my orchestrator agent operates under. excerpt from ~/.config/opencode/AGENTS.md.
proceed without asking
- Routing decisions (use Confidence Gate)
- Quality gate loops (reviewer ↔ coder, up to 3 iterations)
- Reading files and running read-only commands for context
- Dispatching agents per workflow templates and routing table
- Synthesizing multi-agent results into user-facing answers
- Low-risk implementation: typo fixes, test additions, doc updates, lint fixes
stop and ask
- High-risk operations: production deploys, schema migrations, data deletion, secret rotation, force-push
- Scope expansion: user asked for X, but X requires Y — confirm before starting Y
- Confidence below 60% (ask ONE clarifying question, then proceed)
- Quality gate loop hit 3 iterations without converging
never (hard constraints)
- Commit or push (user commits manually)
- Run destructive commands
- Self-answer domain questions a specialist should handle
- Dispatch without context (always include What/Why/Where/Constraints)
- Serialize what can parallelize