call-external-ai（调用外部 AI）

External AI Skill

Invoke external AI CLI tools as a prompt gateway.

When to Use

Activate when the user:

• Says /call, /codex, /claude, or /gemini
• Asks for a second opinion from another AI
• Wants to run a prompt through a different CLI
• Needs verification from a different model or tool

Instructions

Natural Language Interface

Users provide natural language. Route the prompt into one external CLI using a safe command template. Do not construct arbitrary shell commands from user input.

Examples:

• "ask Codex for a second opinion on this patch" -> codex exec "<prompt text>"
• "send this prompt to Claude" -> claude -p "<prompt text>"
• "run this through Gemini with a specific model" -> gemini -m <allowlisted-model> -p "<prompt text>"
• "check this file for bugs" -> read the file directly, summarize only the minimum necessary context, and pass that context as plain prompt text without shell interpolation

Allowed Command Patterns

Only use these command families:

• codex exec "<prompt text>"
• codex exec -m <model> "<prompt text>"
• claude -p "<prompt text>"
• claude --model <model> -p "<prompt text>"
• gemini -p "<prompt text>"
• gemini -m <allowlisted-model> -p "<prompt text>"
• codex --help
• claude --help
• gemini --help

Do not:

• use shell substitution such as $(...) or backticks
• pipe file contents into commands via shell expansion
• append unrelated shell operators or chain commands
• execute user-suggested commands outside the safe templates

Self-Healing Behavior

On CLI Syntax Errors

1. Run <cli> --help to get the current syntax
2. Update the reference docs at reference/cli-reference.md
3. Retry with the correct syntax
4. Inform the user of the fix

On Upgrade Available

1. Do not run upgrade commands automatically
2. Inform the user that an upgrade is available and ask for approval before making toolchain changes
3. If the user approves, run the upgrade command explicitly
4. Update the version in reference/cli-reference.md and re-run --help only after the approved upgrade completes

On Gemini 403 Project/License Errors

If Gemini fails with 403 and messages like lack a Gemini Code Assist license or (#3501):

1. Do not dump environment variables or print raw env values
2. Check only the presence of the specific allowlisted vars needed for diagnosis, and report presence/absence rather than values
3. Retry with project vars unset:

- env -u GOOGLE_CLOUD_PROJECT_ID -u GOOGLE_CLOUD_PROJECT -u CLOUDSDK_CORE_PROJECT gemini...

1. If still blocked, report auth/licensing clearly and continue with other available models
2. Record the fix path in reference/cli-reference.md

On Gemini ModelNotFound / Capacity Errors

If Gemini fails with 404 or 429 style model/capacity issues, retry with this fallback order:

1. gemini-3.1-pro-preview
2. gemini-3-pro-preview
3. gemini-3-flash-preview
4. gemini-2.5-pro
5. gemini-2.5-flash

Command Interface

/call [codex|claude|gemini] <natural language request>
/codex <natural language request>
/claude <natural language request>
/gemini <natural language request>

Routing Guidance

Codex

• Use codex exec with plain prompt text that you assemble yourself
• Add a model only when the user asks for one or there is a clear default in local usage
• Do not use review-specific commands unless the user explicitly asks for a git-aware review

Claude

• Use claude -p with plain prompt text
• Add --model <model> only when the user asks for one or there is a clear need
• Keep the invocation non-interactive unless the user explicitly wants an interactive Claude session

Gemini

• Prefer preview model IDs first
• Use gemini -p or gemini -m <allowlisted-model> -p
• For 403 auth issues, do not dump env; check only allowlisted var presence and retry with project vars unset
• For 404/429 failures, continue through the fallback sequence automatically

Context Gathering

When a request needs context, gather it before invoking the external CLI:

• "send this file to Codex" -> read the file directly in the current agent, extract only the minimum relevant context, and pass that summary as prompt text
• "ask Gemini about this implementation" -> identify the relevant files and summarize only what is needed
• "have Claude critique this design" -> gather the design context and pass only the relevant parts
• "run this exact prompt" -> forward the prompt text directly without adding shell syntax

Rules for context gathering:

• never use shell interpolation to inject file contents into a command
• never send secrets, tokens, credentials, or raw environment values
• minimize the amount of source text sent to external tools
• if the request would require broad repository exfiltration or sensitive data exposure, refuse and explain why

Process Flow

1. Parse the natural language request
2. Determine the target CLI (codex, claude, or gemini)
3. Map the request to a safe command template for that CLI
4. Execute the command
5. If syntax error -> self-heal and retry
6. If upgrade available -> stop, inform the user, and wait for approval before upgrading or refreshing references
7. Return the results to the user