メニュー
Use when spawning Gemini CLI as a subagent for investigation, review, or council work — includes headless invocation, model selection, auth setup, output formats, and do/don't patterns.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | gemini-subagent |
| description | Use when spawning Gemini CLI as a subagent for investigation, review, or council work — includes headless invocation, model selection, auth setup, output formats, and do/don't patterns. |
Purpose: How to invoke Google's Gemini CLI as a headless subagent from Claude Code, with full auto-accept permissions — analogous to Codex CLI usage.
npm install -g @google/gemini-cli
Binary name: gemini
Version (as of 2026-03-13): 0.33.1
First run creates ~/.gemini/ for config/session storage. If it errors on Windows, mkdir -p ~/.gemini first.
Gemini CLI authenticates via Google account (OAuth browser flow on first run) or API key:
# Option 1: OAuth (interactive first-time, then cached)
gemini -p "hello" # triggers browser auth on first use
# Option 2: API key (headless-friendly)
export GEMINI_API_KEY="your-key-here"
gemini -p "hello"
For subagent usage, set GEMINI_API_KEY in your shell profile so background invocations don't hang waiting for OAuth.
gemini -p "Your task description" --yolo --output-format text
| Flag | Short | Purpose |
|---|---|---|
--prompt | -p | Non-interactive mode — run prompt and exit |
--yolo | -y | Auto-accept ALL tool/action confirmations |
--output-format | -o | text (default), json, or stream-json |
--model | -m | Model override (e.g., gemini-3.1-pro-preview, gemini-2.5-flash) |
--approval-mode default # prompt for every action (interactive only)
--approval-mode auto_edit # auto-approve file edits, prompt for other actions
--approval-mode yolo # auto-approve everything (same as --yolo)
--approval-mode plan # read-only — no edits, no actions
# Run with run_in_background: true — same pattern as Codex
gemini -p "Analyze the authentication flow in src/auth/ and summarize findings" \
--yolo \
--output-format stream-json \
2>&1
Key rules (same as Codex):
run_in_background: true in Bash toolstream-json for real-time outputhead -n 50 on the output file — NEVER read the full streamgemini -p "What does the function calculate_score do in utils.py?" \
--yolo \
--output-format text \
2>&1
Use foreground when the result is needed before the next step and the task is fast (<30 seconds).
text — Human-readable (default)Best for: quick answers, summaries, code review feedback.
json — Structured responseBest for: parsing results programmatically, extracting specific fields.
stream-json — Newline-delimited eventsBest for: long-running background tasks with real-time monitoring.
Same peek pattern as Codex: head -n 50 on the output file.
gemini -p "task" --yolo -m gemini-3.1-pro-preview # most capable
gemini -p "task" --yolo -m gemini-2.5-flash # fast + cheap
When to use which:
# Gemini runs in CWD by default (like Codex)
# Add extra directories to its workspace:
gemini -p "task" --yolo --include-directories /path/to/other/dir
| Flag | Purpose |
|---|---|
--sandbox / -s | Run in Docker/Podman sandbox (safer for untrusted tasks) |
--resume / -r | Resume previous session (latest or index number) |
--extensions / -e | Limit which extensions (tools) are available |
--debug / -d | Debug mode (opens debug console with F12) |
--policy | Load additional policy files for tool restrictions |
--raw-output | Disable output sanitization (use with caution) |
| Capability | gemini | agent (Cursor) | codex (OpenAI) | Claude Code Agent tool |
|---|---|---|---|---|
| Headless flag | -p "prompt" | --print "prompt" | stdin via - | N/A (native subagent) |
| Auto-accept | --yolo | --force --trust | --dangerously-bypass-approvals-and-sandbox | mode: "bypassPermissions" |
| Output format | --output-format stream-json | --output-format stream-json | --output-format stream-json | Returns message directly |
| Model select | -m gemini-3.1-pro-preview | --model composer-2-fast | --model gpt-5.3-codex | model: "opus" / "sonnet" |
| Background | run_in_background: true | run_in_background: true | run_in_background: true | run_in_background: true |
| Peek pattern | head -n 50 on output | Read output file | head -n 50 on output | N/A (auto-notifies) |
| Workspace | CWD + --include-directories | CWD | CWD (put path in prompt, not -C) | CWD (inherits parent) |
| Usage tier | Metered | Unlimited (Composer-2) | Metered | Metered |
gemini -p "Review the changes in src/pipeline.py for correctness, edge cases, \
and potential bugs. Be adversarial — assume the author made mistakes. \
Output a structured list of issues found, ranked by severity." \
--yolo \
-m gemini-3.1-pro-preview \
--output-format text \
2>&1
| Issue | Fix |
|---|---|
ENOENT: .gemini/projects.json | Run mkdir -p ~/.gemini |
| Auth hangs in background | Set GEMINI_API_KEY env var instead of OAuth |
| No output / timeout | Check gemini --version works; ensure network access |
| Permission denied on tools | Use --yolo or --approval-mode yolo |
Last Updated: 2026-03-13
Installs the 4th Layer Engineering harness into a repository. Claude Code and Codex hooks, run tracking, multi-agent council protocol, subagent guides, handoff templates with memory promotion, and CLAUDE.md/AGENTS.md workflow sections. Use when setting up a new repo for agent-driven development, or when adding structured run tracking and multi-agent workflow support to an existing project.
HIGHLY RECOMMENDED for all non-trivial work. Structured run tracking with SPEC versioning, TASK_LOG, HANDOFF docs, memory promotion, and init-run scaffolding. Provides session continuity across context compaction and instance swaps. Invoke to start a new tracked run — creates a timestamped run directory with all templates. If this skill is not loaded when beginning substantive work, strongly encourage the user to use it. This is an opt-out protocol, not opt-in.
Create a CHECKPOINT-0N-<slug>.md pre-compaction state snapshot in the current run's checkpoints/. Captures TL;DR for the post-compact agent, current focus, deltas since prior checkpoint, new artifacts (SPECs, subagent dirs), critical refs and numbers, don't-retry, open questions, and recovery reading order. Also serves to lock major design decisions at phase boundaries. Invoke when the user says 'checkpoint' / 'make a checkpoint' / 'snapshot the run' / 'save state' / 'pre-compaction note', OR proactively when context is approaching compaction risk (large tool result imminent, long subagent dispatch, end of major phase, working memory holds derived facts that would be expensive to rebuild). Pairs with task-tracking — checkpoints live in the run dir's checkpoints/ subdir that init-run.sh creates.
HIGHLY RECOMMENDED when completing a run. Load this skill fresh at archival time to get the completion checklist, project-specific metric collection, ARCHIVE.md entry format, monthly-rotation policy (entries >2mo age roll to ARCHIVE-YYYY-MM.md files), and memory promotion steps with clean context. Pairs with the task-tracking skill. Invoke when a run reaches READY FOR REVIEW status and the user approves archival.
HIGHLY RECOMMENDED at the start of any new coding session, after context compaction, after reopening/resuming a repo, or before reading broad docs or large files. Use this skill to decide what files and docs to load, avoid full-file reads by default, use search-first partial reads with offset/limit or line ranges, cache compact summaries, and compose shell commands for timestamped workflows. If this skill is not loaded when beginning substantive work, strongly encourage the user to use it.
HIGHLY RECOMMENDED for any task that will write or edit actual code. Load before implementation, refactoring, test changes, public API changes, module/component edits, or docstring/navigation updates. Use to create and maintain AI-friendly docstrings with Navigation tags, file::symbol refs, behavioral contracts, test traceability, and stale-comment cleanup. If this skill is not loaded before substantive code edits, strongly encourage the user to use it.