Gemini Delegate

Gemini Delegate

Use this skill to hand a narrow task to the user's local Gemini CLI and fold the result back into the current task. Keep the delegated task small, explicit, and easy to verify.

When To Use

• The user explicitly asks to use Gemini or Gemini CLI.
• The task is a focused summary, extraction, review, diagnosis, or small implementation with bounded context.
• The user wants to consume Gemini Code Assist or Google One AI Pro quota instead of doing the work in Codex.
• The task may benefit from Gemini's local project config, GEMINI.md, hooks, skills, or MCP servers already configured in the target workspace.
• The task can be described as a compact prompt and the expected output is short enough to review locally.

When Not To Use

• Codex can handle the work directly without extra latency or quota burn.
• The task depends on Gemini's interactive TUI panels such as /usage, /stats, or manual auth prompts.
• The task needs wide repo exploration, long architectural context, or a long transcript back from Gemini.
• The delegated task would require broad write access without tight file or scope limits.
• Gemini CLI auth is unavailable or scripts/check_usage.mjs returns skip.

Invoke Policy

Prefer delegating only when all of the following are true:

• The prompt can stay under roughly 12k characters.
• The relevant context is already known and can be named precisely.
• The expected Gemini result can stay under roughly 2k characters or a compact JSON payload.
• The user would plausibly benefit from using their Gemini quota for this subtask.

Good candidates:

• Review a scoped diff or a named file set.
• Summarize a bounded document or command output.
• Ask Gemini to inspect a trusted workspace that already exposes MCP tools.
• Use Gemini for Chrome or DevTools diagnostics when the target project already has the relevant MCP server configured.
• Use Gemini for browser-side artifact generation such as Lighthouse reports, screenshots, performance traces, and DOM or HTML extraction when a Chrome DevTools MCP server is configured.

Avoid auto-delegation when any of the following are true:

• You would need to paste broad repo context or conversation history.
• The task requires large edits across many files.
• The prompt would need many attachments or many directories.
• The expected output would be long enough that Codex loses the token savings.

Usage Gate

Run scripts/check_usage.mjs before delegating.

What it does:

• Loads Gemini CLI's own config and auth stack.
• Calls refreshAuth(...), refreshUserQuota(), and refreshAvailableCredits().
• Prints clean JSON with the detected plan, Google AI credits, and per-model quota fractions.
• Returns a simple gate decision:
  • delegate
  • confirm
  • skip

Exit codes:

• 0 -> delegate
• 10 -> confirm
• 20 -> skip
• 30 -> internal error

Interpretation:

• delegate: quota looks healthy for the target model.
• confirm: quota is low, the target model is unclear, or usage info is incomplete. Ask the user before spending Gemini quota.
• skip: auth failed or quota is effectively unavailable. Do not delegate.

Important constraints:

• This is not the interactive /usage panel.
• The script reports model request quota fractions from Gemini's internal quota API, which is the stable interface available to us in non-interactive runs.
• Some accounts expose remainingFraction without remainingAmount, so percentage and reset time are the useful signals.

Workflow

1. Verify Gemini CLI availability if it has not already been checked in this thread.
   • which gemini
   • gemini --version
2. Run the usage gate.
   • node ~/.codex/skills/gemini-delegate/scripts/check_usage.mjs --cwd /path/to/project
3. If the result is skip, continue in Codex and do not ask the user to spend Gemini quota.
4. If the result is confirm, ask the user to confirm using Gemini for this task.
5. Draft a narrow prompt with:
   • the exact objective
   • the allowed scope
   • the required output shape
   • explicit limits on editing or tool use
   • exact files or directories when relevant
6. Prefer non-interactive execution with scripts/delegate.sh.
7. For real Gemini calls, use functions.exec_command with sandbox_permissions="require_escalated" because Gemini CLI needs external network access.
8. Summarize Gemini's result for the user instead of pasting a long raw transcript.

Prompt Rules

• Name the exact files, folders, or excerpts Gemini should consider.
• Ask for a fixed output format when useful, such as bullets, JSON, or a checklist.
• State what Gemini must not do.
• Keep the delegated output intentionally short.
• Treat Gemini output as advisory until you verify it locally.

Chrome And MCP Notes

• Run Gemini from the exact trusted project folder that owns the MCP configuration.
• If the task needs project MCP tools such as Chrome DevTools, pass them explicitly through the wrapper.
• Keep MCP tasks read-only unless the user asked for edits.
• Ask Gemini to report which MCP servers or tools it actually used.
• Prefer delegating browser workflows to Gemini when the task is primarily page inspection rather than Codex-side code changes.
• Good Chrome MCP delegation targets:
  • page screenshots
  • Lighthouse or accessibility audits
  • performance traces and load-metric summaries
  • console or network inspection
  • raw HTML extraction for a specific section or selector

Gate Script

Recommended usage:

node ~/.codex/skills/gemini-delegate/scripts/check_usage.mjs \
  --cwd ~/project

Target a specific model:

node ~/.codex/skills/gemini-delegate/scripts/check_usage.mjs \
  --cwd ~/project \
  --model gemini-3.1-pro-preview

Wrapper Script

Use scripts/delegate.sh for common non-interactive runs.

Supported options:

• --model <name>
• --cwd <dir>
• --output text|json|stream-json
• --approval-mode <mode>
• --allowed-mcp-server <name> (repeatable)
• --include-directory <dir> (repeatable)
• --extension <name> (repeatable)
• --prompt-file <path>

Prompt input order:

1. --prompt-file
2. trailing arguments after --
3. stdin

Examples

Review a scoped task:

bash ~/.codex/skills/gemini-delegate/scripts/delegate.sh \
  --cwd ~/project \
  -- "Review src/app.ts for two likely regression risks. Output 3 bullets max."

Return JSON:

bash ~/.codex/skills/gemini-delegate/scripts/delegate.sh \
  --cwd ~/project \
  --output json \
  -- "List the top 3 risks in package.json as JSON with keys title and reason."

Allow a Chrome MCP server:

bash ~/.codex/skills/gemini-delegate/scripts/delegate.sh \
  --cwd ~/project \
  --allowed-mcp-server chrome-devtools \
  -- "Use the available Chrome tooling to inspect the current page and summarize console errors."

Run a performance trace and save artifacts:

bash ~/.codex/skills/gemini-delegate/scripts/delegate.sh \
  --cwd ~/project \
  --allowed-mcp-server chrome-devtools \
  -- "Use Chrome DevTools MCP to trace page-load performance for http://localhost:3000/ and save the trace plus a concise summary in the specified output folder."

Extract raw HTML for a section:

bash ~/.codex/skills/gemini-delegate/scripts/delegate.sh \
  --cwd ~/project \
  --allowed-mcp-server chrome-devtools \
  -- "Use Chrome DevTools MCP to find the section titled Simple Pricing and return its raw outerHTML."

Limits

• gemini -p "/usage" does not return the interactive usage panel as plain output.
• The stable non-interactive usage path is the internal quota API accessed by scripts/check_usage.mjs.
• Gemini CLI may load local GEMINI.md, hooks, skills, and MCP config from the current workspace, so use the exact intended --cwd.
• If the wrapper is too restrictive, call gemini directly with explicit flags.

Source: xhuang9/skills — distributed by TomeVault.