Offload — Development skill for Claude Code

offload

`offload` is a skill for CLI coding agents. It sends work to `agy` (the Antigravity CLI, running Gemini models) instead of running it in whatever agent you're already sitting in.

Whatever agent loads this skill stays the orchestrator: Claude Code, Codex CLI, anything that can read a `SKILL.md` file and run shell commands. Anything except `agy`. `agy` is the worker here, not the orchestrator too. The orchestrator never repeats what a worker says about its own work. It checks the work, or checks what a second worker said about it.

Five worker roles split the labor:

Role	Job
scout	Reads the repo, reports which files a task would touch. Can't write.
gate-author	Turns the orchestrator's acceptance criteria into a test file.
implementer	Does the task.
reviewer	Judges a finished diff against the criteria. Can't write.
researcher	Investigates a bounded question with concrete evidence. Can't write.

The orchestrator still decides how the work splits and what "correct" means for each piece. That judgment has no gate behind it, so it stays put. Everything downstream goes to a worker and then gets checked: a red run and a read before a gate-author's test is trusted, an exit code for the implementer's result, a grep against a quoted line before a reviewer's verdict is trusted, and direct evidence checks and sampling for research findings.

What this does not do

This is not a general multi-agent framework. It won't do open-ended research fan-out with no acceptance criteria. Every worker it dispatches answers to a gate or an evidence-backed verification protocol: a machine-gate command, written criteria a reviewer judges the diff against, or a bounded read-only research/audit assignment with evidence citations subject to risk-based verification.

It does not sandbox a writing worker's filesystem access. `agy`'s `--add-dir` grants a worker access to a directory. It does not confine the worker to it. Nothing in this ski