メニュー
Use Google Vertex AI high-end coding models (Gemini 3.5 Flash and Gemini 3.1 Pro) directly on local files. Bypasses standard LLM token limits and bills to GCP. Supports direct coding and fully autonomous interactive agent loops.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | vertex-coder |
| description | Use Google Vertex AI high-end coding models (Gemini 3.5 Flash and Gemini 3.1 Pro) directly on local files. Bypasses standard LLM token limits and bills to GCP. Supports direct coding and fully autonomous interactive agent loops. |
This skill allows any agent (e.g. Claude Code, Codex, Antigravity, or other local AI models) to delegate heavy code generation, debugging, testing, and multi-file editing tasks to premium Google Vertex AI models (gemini-3.1-pro-custom-tools, gemini-3.1-pro, or gemini-3.5-flash) on behalf of the user, executing directly on your macOS.
It features two distinct modes of execution:
vertex_direct_coder.py): High-speed, single-file generation/modification.vertex_interactive_agent.py): A fully autonomous, multi-step ReAct tool-use loop that can explore the directory, read files, write multiple files, run bash commands, catch test failures, and auto-correct itself until completion.Before invoking this tool, ensure the following are available on your system:
gcloud) must be installed and authenticated:
gcloud auth application-default login
fair-geography-494614-q0 (active billing).google-genai and google-oauth2:
Interpreter: ~/.config/dt/venv/bin/python3 (or any active virtualenv)Use this when you want to quickly generate or modify a specific single file.
~/.gemini/config/skills/vertex-coder/scripts/vertex_direct_coder.py <file_path> "<prompt_or_instruction>" [model_name]
gemini-3.1-pro-custom-tools (Rank 1 Default: Specialized variant of Gemini 3.1 Pro heavily optimized for mixing bash executions with defined custom tool calling)gemini-3.1-pro (Ultimate reasoning, maximum coding capabilities, default fallback)gemini-3.5-flash (High speed, outstanding coding, default for fast tasks)# Create a new flask app file with gemini-3.5-flash
~/.gemini/config/skills/vertex-coder/scripts/vertex_direct_coder.py app.py "Create a modular flask application with routes for a user dashboard" gemini-3.5-flash
# Modify an existing file with gemini-3.1-pro
~/.gemini/config/skills/vertex-coder/scripts/vertex_direct_coder.py app.py "Add JWT-based user authentication and error-handling middleware" gemini-3.1-pro
Use this when the task requires exploring the workspace, designing a complete feature, running tests (e.g., pytest), catching failures, and editing multiple files iteratively until everything works.
The interactive agent is equipped with robust, premium local workspace tools:
line_replace [PREFERRED]: Finds and replaces specific content in a file using explicit 1-indexed line numbers. Supports ellipsis ... line matching for large blocks, dramatically minimizing code rewrite size and latency.add_dependency: Safely adds npm (Node.js) or pip (Python) packages to your project automatically with auto-detection.save_memory & read_memories: Stores and retrieves custom architectural rules, preferences, tech stacks, or project context inside a local persistent .agent_memory.json file.grep_search: Powerful regular expression search (utilizing Python's compiled re engine) to scan the codebase efficiently.list_dir: Lists files and subdirectories.read_file: Reads entire contents of any file.write_file: Creates new files or overwrites existing ones (mainly for new file creation).run_command: Runs terminal/bash commands locally to test, run servers, or build codebases (60-second timeout).list_global_skills & use_global_skill: Allows Vertex Coder to dynamically query and execute sub-agents using any available global agent skills on your system.load_superpower_skill: Dynamically loads a global skill's full instructions (SKILL.md) directly into the agent's current context, allowing it to follow its methodologies (e.g. test-driven-development, doubt-driven-development, graphify, security-and-hardening) natively while retaining full local file and CLI access.The Interactive Agent utilizes advanced, highly optimized systems instructions inspired by leading coding tools (Claude Code, Windsurf, Devin, v0):
.agent_memory.json on startup to seamlessly inject design parameters, user preferences, and technology constraints into the active session without requiring manual tool invocations.pytest, vitest), builders, or linters immediately after code changes to confirm correctness before finalizing.add_dependency before writing any files importing them.node_modules or local packages if documentation or types are lacking.path/file.ext:line_number format for direct navigation.The interactive agent natively integrates Jesse Vincent's Superpowers and other expert software engineering guidelines (such as test-driven-development, spec-driven-development, doubt-driven-development, security-and-hardening, code-review-and-quality, and graphify codebase analysis).
These guidelines can be loaded in two ways:
-s / --skills argument.load_superpower_skill(skill_name) tool.~/.gemini/config/skills/vertex-coder/scripts/vertex_interactive_agent.py "<complex_task_prompt>" [model_name] [--skills skill1 skill2 ...]
# Create a math library using TDD (Test-Driven Development) and verify with pytest
~/.gemini/config/skills/vertex-coder/scripts/vertex_interactive_agent.py "Create a simple matrix multiplication function in matrix_mul.py" gemini-3.5-flash --skills test-driven-development
# Execute a complex feature with robust clean-code reviews and planning files
~/.gemini/config/skills/vertex-coder/scripts/vertex_interactive_agent.py "Add JWT auth and validation to our API" gemini-3.1-pro-custom-tools --skills planning-and-task-breakdown security-and-hardening code-review-and-quality
(Note: If you omit the second parameter, it defaults to gemini-3.1-pro-custom-tools for ultimate agentic reasoning and custom tool execution).
Complete agentic engineering supersystem — one command orchestrates the entire engineering chain. Use when the user asks for "the full arsenal", "use everything you have", "/Apeiron", "/delegate-team", "delegate to a backend", "spawn a multi-agent team", "make a CLI to scaffold skills", or any natural-language request that should drive a Claude Code or Codex session through the full brainstorm → plan → execute → verify → ship pipeline. Composes four companion frameworks (superpowers, Waza, unslop-preflight, autoresearch) plus the original `dt` gateway CLI, multi-agent team (MMAS) framework, skill scaffolder, the /Apeiron orchestrator (47-case selftest routing), and the v2.7.0 arsenal: autopilot.sh (7-stage GOD command), apeiron-uni (smart universal wrapper that detects the calling runtime), and 1890 auto-discovered skills across 3 sources.
Local-first governance kernel for AI coding agents — gives Claude Code, Codex, Cursor, Antigravity, and Gemini CLI a shared memory, an approval inbox for new rules, generated AGENTS.md / CLAUDE.md / cursor-rule.mdc, Claude + git hooks, and a deterministic policy guard. Adds a JSON-first memory layer (rules + preferences + workflows + project-notes + skills), an episodic memory archive, and MCP tools for search/read/capture/sync of past sessions. Use when the user asks to "remember this rule", "save this as a memory", "what did we do last time", "search past episodes", "propose a new rule", "make agent follow our standards automatically", or any request to give an agent persistent local memory across sessions. Bundled inside delegate-team v2.5.0 — installed via `./install.sh --kernel`. Triggers on: "agent kernel", "ak", "remember this", "save this rule", "memory tool", "episodic memory", "approval inbox", "guard policy", "rule inbox".
Route heavy coding tasks to premium multi-model backends (Codex GPT-5.5/5.4 high, GLM-5.2 max, Qwen 3.7 max, Kimi K2.7 code-max, MiniMax M3 high-thinking). Mirrors vertex-coder's structure but shells out to local CLI binaries (codex, opencode) instead of Vertex AI. Two modes: direct (single-file) and interactive (multi-file autonomous with tool use). Use when delegating heavy code generation, debugging, or refactoring tasks to the strongest available model. Triggers on: "delegate to aonios agent", "use the strongest model", "premium model coding", "opencode glm", "codex gpt-5 high", "minimax m3 thinking", "aonios agent direct", "aonios agent interactive". Do NOT use for trivial edits (do them locally) or when the user explicitly forbids external delegation.
The ONE-COMMAND orchestrator that wires the full delegate-team arsenal for any task in any project. Use when the user says "use everything you got", "use full engineering arsenal", "go all-in", "/Apeiron <task>", "/ship <task>", "use all capabilities", "full stack", "end to end", "everything intelligently". Composes 38 curated integrations (superpowers, Waza, unslop, autoresearch, agent-kernel, ...) and routes tasks through 7-stage autopilot chain (PREWARM → BRAINSTORM → PLAN → EXECUTE → REVIEW → QUALITY-GUARD → REPORT). Auto-discovers 1890+ skills across 3 sources (apeiron, agents, claude) with intelligent dedupe. Includes apeiron-uni — the smart universal wrapper that detects the calling runtime (Apeiron / codex / claude / gemini / opencode / mmx / shell) and dispatches the right flow. Skip this skill only when the user wants a SPECIFIC tool (e.g. "run /autoresearch with metric X" or "delegate to aonios-agent").
Local MiniMax-powered coding agent. Uses the official `mmx` CLI (https://github.com/MiniMax-AI/cli) to call MiniMax models (M3 flagship, M2.7, M2.7-highspeed) for direct single-file edits and multi-file autonomous work with full tool use. Sibling backend to vertex-coder inside delegate-team; mirrors its structure (direct + interactive + tools registry) but routes through mmx instead of an SDK. The Python scripts in this directory are local-agent helpers — Apeiron (M3) can also invoke `mmx text chat` directly via bash for fast ad-hoc calls. Three MiniMax models supported. Use when delegating coding work to MiniMax models. Triggers on: "use minimax coder", "MiniMax M3", "MiniMax M2.7", "minimax direct", "minimax interactive", "mmx text chat". Do NOT use for trivial edits (do them locally).
Apeiron Multi-Agent System (MMAS) — spawn a team of specialized AI agents under your direct command. Each agent has a specific power + model (Atlas orchestrator, Forge deep worker, Scout searcher, Oracle debugger, Librarian docs-finder, Reviewer validator, Visionary media-analyzer, Sentinel security-auditor). A watchdog polls every 30s and reports status back to you. Inspired by oh-my-openagent's specialized agent team + hermes-agent's self-improving learning loop. Use when you want parallel agents working on a complex task while you stay the boss. Triggers on: "spawn team", "multi-agent", "MMAS", "parallel agents", "team of agents", "boss mode", "watchdog", "/apeiron-team". Do NOT use for single-file edits (do them locally).