Ask Opencli

Ask via OpenCLI (Grok / Gemini)

Use opencli to query Grok or Gemini through the user's already-logged-in Chrome session. Zero API cost, fully driven by the existing browser session.

opencli 项目：https://github.com/jackwener/opencli 安装：npm install -g @jackwener/opencli（需要 Node ≥20） 首次运行：opencli doctor 会自动引导安装 Chrome Browser Bridge 扩展 登录：用装了扩展的 Chrome profile 打开 grok.com 和 gemini.google.com 登录一次 必须环境变量：export OPENCLI_BROWSER_COMMAND_TIMEOUT=300（写进 shell rc） 更深度的多 AI 并行调研工作流：use the multi-ai-research skill instead

When to trigger:

• User explicitly asks to query Grok or Gemini ("问一下 grok", "用 gemini 查", "让 grok 看看")
• User wants a second opinion / external advisor from Grok or Gemini
• User says "ask grok"/"ask gemini" or mentions these two as the information source

When NOT to trigger:

• User wants to use native Gemini CLI (gemini -p) → use ask-gemini skill instead
• User wants Claude's own opinion → answer directly
• User wants other AI (Doubao, Yuanbao, ChatGPT, etc.) → use other tools

Model selection

If the user doesn't specify which model, default to Grok for real-time / social / news / tech-trend questions, and Gemini for long-form reasoning, research, coding, and document analysis.

If still ambiguous, ask the user which one to use before sending.

⚠️ 常见错误命令（直接 fail fast）

完整子命令速查（运行 opencli grok --help 或 opencli gemini --help 查看）：

Grok:
  ask <prompt>                    # 唯一命令

Gemini:
  ask <prompt>                    # 最常用
  new                             # 开新对话
  deep-research <prompt>          # Deep Research 模式
  deep-research-result [query]    # 取 Deep Research 结果
  image <prompt>                  # 图片生成

⚠️ Critical: Two-layer timeout (from 2026-04-08 debugging)

opencli has two independent timeout layers, both must be set for grok to wait long enough:

1. Inner: --timeout 300 — grok.com web response wait time
2. Outer: OPENCLI_BROWSER_COMMAND_TIMEOUT=300 environment variable — entire browser command execution timeout (default 60s, hardcoded in runtime.js:25)

The outer takes priority over the inner. Setting only --timeout 300 still gets killed at 60s. You MUST set both.

Source code: .omx/reference/opencli/dist/src/runtime.js:25:

export const DEFAULT_BROWSER_COMMAND_TIMEOUT = parseEnvTimeout('OPENCLI_BROWSER_COMMAND_TIMEOUT', 60);

Recommended: add export OPENCLI_BROWSER_COMMAND_TIMEOUT=300 to ~/.zshrc so every grok/gemini call has room to breathe.

Commands

Grok (MANDATORY env var)

OPENCLI_BROWSER_COMMAND_TIMEOUT=300 opencli grok ask "{{PROMPT}}" --timeout 300 -f json

Or if env var already exported in shell:

opencli grok ask "{{PROMPT}}" --timeout 300 -f json

Flags:

• --new true — start a fresh chat (avoids polluting prior conversation context)
• --web true — use the hardened grok.com consumer web flow (better error messages)
• --timeout 300 — MUST match or be lower than OPENCLI_BROWSER_COMMAND_TIMEOUT
• -f json — structured output (recommended over --format plain for parsing)

Gemini

opencli gemini ask "{{PROMPT}}" --format plain

Gemini is more stable than grok and usually doesn't hit the 60s timeout. No env var required but adding it doesn't hurt:

OPENCLI_BROWSER_COMMAND_TIMEOUT=300 opencli gemini ask "{{PROMPT}}" --format plain

Flags:

• --new true — start a fresh chat
• --timeout 90 — extend timeout for long answers (default 60s)

Parallel dispatch (for multi-AI workflows)

To call grok and gemini in parallel from Claude Code, use the Bash tool with run_in_background=true for both calls in the same assistant turn. Claude Code will launch them concurrently and wait for task-completion notifications.

For deep cross-validation workflows (N internal sub-agents + 2 external AIs, confidence-tiered synthesis), use the multi-ai-research skill instead — it covers the full orchestration pattern.