drive-gemini

Drive Gemini CLI

Delegate to the Gemini CLI, Google's provider-agnostic-ish coding agent. Read delegate-to-agents first for the shared Claude Code mechanism.

Flags verified against gemini 0.43.0 (gemini --help).

When to use

Coding, research, large-context analysis, or when you want Gemini's model specifically. Strong for big-context reads.

Preflight

command -v gemini && gemini --version

• Auth: set GEMINI_API_KEY or GOOGLE_API_KEY, or run gemini once interactively to complete Google OAuth. If neither is present and there's no cached OAuth, a headless -p run will fail — fall back accordingly.

One-shot / headless (PREFERRED): -p

Gemini defaults to interactive mode; a bare positional prompt opens the TUI. For automation you MUST pass -p/--prompt.

If the working directory isn't trusted (a fresh dir, /tmp, etc.), even a headless -p run is blocked by the trust gate. Add --skip-trust (or set GEMINI_CLI_TRUST_WORKSPACE=true). Already-trusted project dirs don't need it.

# headless, capture stdout
Bash(command="cd /repo && gemini -p 'Explain what src/server.ts does and list its exports'", timeout=300000)

# headless write task with auto-approved edits
Bash(command="cd /repo && gemini -p 'Add retry logic to the HTTP client and update tests' --approval-mode auto_edit", timeout=600000)

# stdin is appended to the -p prompt
Bash(command="git -C /repo diff | gemini -p 'Review this diff for bugs and security issues'", timeout=300000)

# long task: detach + log
Bash(command="cd /repo && gemini -p 'Refactor the data layer' --approval-mode auto_edit > /tmp/agent-gemini.log 2>&1", run_in_background=true)

Key flags

Flag	Effect
-p, --prompt <text>	Headless (non-interactive) run
-i, --prompt-interactive <text>	Run the prompt, then stay interactive
-m, --model <name>	Model override
--approval-mode <mode>	default (prompt), auto_edit (auto-approve edits), yolo (auto-approve all), plan (read-only)
-y, --yolo	Shorthand for approve-all
-o, --output-format <fmt>	text (default), json, stream-json
-s, --sandbox	Run tools in a sandbox
--skip-trust	Trust the workspace for this session (avoids the trust prompt)
-w, --worktree [name]	Start in a fresh git worktree
--include-directories <dirs>	Add directories to the workspace (monorepos)
-e, --extensions <list>	Restrict to specific extensions

Sessions / resume

gemini --list-sessions                 # see sessions for this project
gemini -r latest -p 'Continue: now add tests'   # resume most recent
gemini -r 5 -p '...'                    # resume by index
gemini --session-id <uuid> -p '...'     # pin a session id

Interactive (tmux)

Bash(command="tmux new-session -d -s gem -x 200 -y 50")
Bash(command="tmux send-keys -t gem 'cd /repo && gemini --skip-trust' Enter")
Bash(command="sleep 6 && tmux capture-pane -t gem -p -S -40")
Bash(command="tmux send-keys -t gem 'Add a CLI subcommand for export' Enter")
Bash(command="sleep 25 && tmux capture-pane -t gem -p -S -80")
Bash(command="tmux kill-session -t gem")

--skip-trust avoids the workspace-trust prompt. Without it, answer the trust dialog by capturing the pane and sending the right key.

Rules

1. Always -p for automation — without it, Gemini opens the TUI and hangs a non-interactive Bash call.
2. Pick the approval mode — auto_edit for building, plan for read-only review, default when you want it to ask.
3. --output-format json when you need to parse the result.
4. --skip-trust (or pre-trust) in automation to avoid the trust prompt.
5. One worktree per parallel run; clean up tmux sessions and logs.
6. Verify the diff and tests yourself before keeping changes.