Gemini CLI Runner

Use this skill when you need to call Gemini from the command line (any task), independent of the downstream workflow (review, drafting, etc.).

If you are already running as Gemini, don't use this runner to re-invoke yourself. Run the call in-host — inline in your own loop (the Gemini CLI may expose no native sub-agent primitive). The gemini CLI hop adds latency, a separate session, and context loss for zero gain. This runner is for reaching Gemini from a DIFFERENT host (Claude / Codex / OpenCode / …) for cross-model work. Pick the model / reasoning depth by task difficulty (quality first, not token thrift).

Session continuation is deliberately NOT wired into this runner (unlike the codex / claude / kimi runners, which expose explicit-session resume). The upstream gemini CLI only offers -r latest / -r <index> — resolution by recency or list position, which is RACY across parallel runs (no explicit session-id form) — and this runner's direct-API fallback path cannot resume a CLI session at all. For multi-turn cross-model workflows, use a runner with explicit-session support.

Preconditions

• gemini is installed: command -v gemini
• You are authenticated/configured for Gemini CLI.

Recommended: runner script (file input + model fallback)

SKILL_DIR="${SKILL_DIR:-$(for r in "${CLAUDE_CONFIG_DIR:-$HOME/.claude}" "${CODEX_HOME:-$HOME/.codex}" "$HOME/.config/opencode" "$HOME/.kimi-code"; do [ -d "$r/skills/gemini-cli-runner" ] && echo "$r/skills/gemini-cli-runner" && break; done || true)}"
bash "${SKILL_DIR}/scripts/run_gemini.sh" \
  --model gemini-3.1-pro-preview \
  --system-prompt-file /path/to/system.txt \
  --prompt-file /path/to/prompt.txt \
  --out /path/to/output.txt

Dry-run (no gemini call):

SKILL_DIR="${SKILL_DIR:-$(for r in "${CLAUDE_CONFIG_DIR:-$HOME/.claude}" "${CODEX_HOME:-$HOME/.codex}" "$HOME/.config/opencode" "$HOME/.kimi-code"; do [ -d "$r/skills/gemini-cli-runner" ] && echo "$r/skills/gemini-cli-runner" && break; done || true)}"
bash "${SKILL_DIR}/scripts/run_gemini.sh" \
  --model gemini-3.1-pro-preview \
  --system-prompt-file /path/to/system.txt \
  --prompt-file /path/to/prompt.txt \
  --out /path/to/output.txt \
  --dry-run

Notes:

• If the given --model is not recognized by your local CLI, the script retries with the default model (by omitting -m).
• Deterministic failures (usage/auth/region-ineligibility errors and similar, detected from the exit code and stderr) exit immediately with the diagnostic instead of trying the model-alias / proxy fallbacks, which would fail identically.
• Prompts are fed via stdin to avoid Argument list too long with large prompt files.
• If --system-prompt-file is provided, its contents are prepended to stdin before the prompt file (separated by a blank line).
• If no proxy env is already exported, the runner will try to bootstrap http_proxy / https_proxy / all_proxy from an interactive zsh session by invoking a user-defined proxy_on helper when available.
• Tool access is explicit:
  • Default is --tool-mode none
  • --tool-mode review maps to Gemini CLI --approval-mode plan plus --sandbox, --no-proxy-first, and --extensions none
• If --gemini-cli-home is set, the runner treats it as an isolated Gemini home root. When auth env vars are not already exported, it bootstraps the minimum Gemini auth env from $HOME/.gemini/.env so isolated review runs do not silently lose API connectivity.
• Use --dry-run to print the planned command + prompt file size/hash without calling gemini.
• --no-proxy-first still exists as an escape hatch, but read-only review/file access should normally use --tool-mode review.