Instruction file imported from akhil-naidu/react-todo (.cursor/rules/main-agent-orchestration.mdc). Copyright stays with the author.

Main agent orchestrates

This session (Grok 4.6) is the only decision layer. Subagents research, implement, and review. They do not choose architecture, scope, skill routing, or the next stage.

Main agent must

• Pick the skill path: brainstorming → OpenSpec propose → writing-plans → subagent-driven-development
• Approve designs and plans; resolve conflicts between plan text and review findings
• Dispatch research with Superpowers parallel agents (one domain per agent)
• Dispatch one implementer at a time with a file brief, not the whole plan
• Adjudicate reviewer verdicts; run self-improvement after a ship

Subagents may

• Explore the repo, query Context7, implement one tasked slice, run that slice's checks, write a report

Subagents must not

• Change the plan, skip review, start the next roadmap stage, or archive an OpenSpec change
• Inherit this chat's history for implementation (resume: self is not an implementer)

Dispatch

Specify a model on every Task call. Keep this session on Grok 4.6. Use explore/docs-researcher for research; a fresh implementer per plan task; a separate reviewer after each task. Never run two implementers on the same branch.