Gemini sub-agent bridge (Antigravity agy)

The old gemini CLI can no longer sign in with personal Google accounts ("This client is no longer supported for Gemini Code Assist for individuals… migrate to Antigravity"). The Antigravity CLI agy is installed and signed in with the user's Google AI Pro account, so this bridge uses it — subscription only; never use or ask for a GEMINI_API_KEY. Always call through the wrapper:

pwsh -NoProfile -File "$HOME\.claude\skills\gemini\scripts\gemini.ps1" -Mode <mode> -PromptFile <brief.md> [options]

Runs take 10 s – several minutes (deep research up to ~60 min): timeout: 600000 or run_in_background: true.

Modes

gemini-4-argon-high is Argon's top level (effort is part of the model name; agy --effort conflicts with it). Override with -Model (agy models). agy also lists Claude and GPT-OSS models — don't route Claude work through it; use the Agent tool. Other options: -Out <file> · -Session <GEMINI_SESSION_ID> · -Include <dirs> · -TimeoutSec.

Implementation notes (already handled by the script): agy --mode plan is a "plan and wait for approval" mode, not read-only, so read-only modes run without --mode and are told not to write, not to stop for approval and not to run shell commands (headless runs auto-deny them). Prompts go as an argument (agy ignores stdin); briefs over ~20K chars are passed via a temp file.

Output contract

Final answer text, then GEMINI_WARNING= (blocked actions, if any), GEMINI_ERROR= (if any), GEMINI_TOKENS=, GEMINI_SESSION_ID=, GEMINI_STATUS=ok|error (exit 1 on error or empty answer).

Briefing

• deepresearch: the research question and the decision it feeds; scope in/out; sub-topics you already know matter; recency ("as of <date>"); preferred source types (papers, official docs, filings); target length and language of the report. Always pass -Out. Afterwards spot-check the citations that carry the conclusion (open 2–3 URLs, or send a verifier/researcher), and surface the report's "Contradictions & open questions" to the user.
• research: the exact question, recency requirement, source types, answer length. Gemini's search can lag the live web; it is told to flag what it could not confirm — pass those gaps on.
• digest: exact paths, the question, answer format, "file:line for every claim". Prefer this over reading dozens of files into Claude's own context.
• opinion: same neutral peer brief you'd give gpt-6-astra (problem, evidence, options, uncertainty) — never your conclusion to rubber-stamp. Run in parallel with astra for multi-model consensus.
• task/bulk: same detailed step/acceptance or rule/example format as the codex skill.

Rules

1. Gemini's answer is a claim: verify code changes and spot-check cited sources before relying on them.
2. Briefs go to Google: no secrets, credentials or private data the user hasn't cleared.
3. Never pass --dangerously-skip-permissions; write access only via -Write inside the intended folder.
4. On GEMINI_STATUS=error with an auth message, ask the user to open Antigravity / run agy once and sign in; don't loop retries.