agy-delegate

name	agy-delegate
description	Use when the user wants coding/implementation work delegated to Google Antigravity CLI (agy) agents instead of implemented directly — when they say "delegate to agy", "agy delegate", "hand this to antigravity", "fan this out to agy", or when kimi-delegate is unavailable/quota-exhausted and parallel local-agent delegation is still wanted.

Agy Delegate

Turn Claude into a manager of Google Antigravity CLI (agy) agents. Claude does the judgment work — decomposition, briefs, review. agy does the implementation volume. Same doctrine as kimi-delegate; the executor and its contract differ.

The contract you depend on (verified on agy v1.1.1, 2026-07-10)

agy runs inside the debian proot-distro container. The Termux wrapper ~/.local/bin/agy handles that transparently for any path under $HOME.

agy -p "FULL BRIEF" --dangerously-skip-permissions --print-timeout 15m \
    [--model "Gemini 3.1 Pro (High)"]

• -p/--print = non-interactive; result on stdout; exit code meaningful.
• -p BINDS the next token as the prompt (verified live 2026-07-10): the prompt must immediately follow -p, all other flags AFTER it. A flag placed between --print and the prompt is swallowed as the prompt text — the agent then researches your flag instead of doing the task.
• Verified non-TTY safe via the proot path (GitHub issues #76/#318 do NOT reproduce here). agy SELF-UPDATES silently — if behavior shifts, run scripts/agy_parallel.sh --verify (3 contract checks, costs credits).
• --dangerously-skip-permissions = auto-approve: trusted repos only, prefer worktree isolation (the script's default).
• Model names contain spaces — always quote. agy models lists them. Route bulk work to "Gemini 3.5 Flash (High)" (default), hard briefs to "Gemini 3.1 Pro (High)" or "Claude Opus 4.6 (Thinking)" via brief frontmatter.
• No JSON output mode exists; structure is enforced by the schema gate below.
• Requires Google Pro/Ultra subscription; runs burn credits. On FAILED(quota): fall back to kimi-delegate, then native subagents.

Workflow

1. Scope & decompose (Claude): disjoint file ownership per agent; 2–4 agents.
2. Write briefs: copy assets/brief-template.md; every section is required (the script lints). Optional frontmatter per brief: model:, timeout:, schema: (path to JSON schema for the output gate).
3. Launch: bash scripts/agy_parallel.sh --repo /path/to/repo briefs/*.md Worktree per brief on branch agy/<name>; logs in .agy-runs/<ts>/; --max-parallel 3 default; exit = #failed agents.
4. Schema gate (when a brief declares schema:): the agent is instructed to end with one JSON object; validate_output.py checks it. Default salvage mode: valid fields land in <name>.partial.json with _missing/_invalid reports — status PARTIAL(schema) keeps the work reviewable instead of discarding it. --schema-mode strict|warn to change.
5. Review & integrate (Claude): read logs + git -C <repo> diff base..agy/<name>, run tests yourself, merge, report. Never trust agent self-reports.

Statuses

OK · PARTIAL(schema) · FAILED(exit|timeout|quota|schema|schema-file|empty-brief|bad-timeout|worktree) — the process exit code counts every FAILED item (agent failures and pre-launch failures alike); PARTIAL never fails the run.