What problem does it solve? Getting a second-engine opinion on code or research normally means pasting context into another tool, losing track of which findings came from which model, and flooding your main conversation with the delegate's raw output. This Skill routes advisory and coding jobs to the Gemini CLI through an isolated relayer subagent, stamps every result with provenance, and keeps the delegate's I/O out of the main context. ## Core Features & Use Cases - Advisory delegation: Run review, adversarial-review, and research verbs against Gemini through one chokepoint that resolves the model, warns on secrets, and degrades loudly instead of silently falling back. - Worktree-staged coding tasks: The task verb with --write stages Gemini's edits in a throwaway git worktree and returns a diff for review, never touching the live tree. - Dual-engine transport: Choose between gemini-print (API key) and agy-print (Google OAuth), with automatic credential detection and loud cross-engine fallback under auto mode. - Use Case: Before merging a risky refactor, spawn a Gemini adversarial-review pass that reads the changed files itself, returns a structured findings report stamped with reviewer_engine and reviewer_model, and iterate over multiple rounds using the converge loop pattern. ## Quick Start Ask the agent to run a Gemini review of a specific file or directory by delegating through the gemini-relayer subagent with a prompt naming the paths to read.
