// Create new persistent Codex agents for this repo when asked to add a specialist agent, define a reusable subagent role, scaffold `.codex/agents/*.toml`, or wire the agent to local skills and MCP servers.
Browser-first QA skill for local web apps. Use when exploring an application, smoke-testing routes, checking whether a page loads, validating a user flow, reproducing a UI bug, or confirming a fix. Prefer the local `agent-browser` CLI for screenshots, snapshots, console checks, navigation, and basic interactions.
Repo-local bridge skill for Next.js frontend work. Use when a dedicated agent is handling App Router UI changes, layouts, metadata, route-aware rendering, or Server/Client Component boundaries.
Repo-local bridge skill for React TSX reviews. Use when a dedicated agent is checking component quality, hooks usage, accessibility, and performance risks after frontend changes.
Repo-local bridge skill for shadcn/ui work. Use when a dedicated agent is composing UI, extending the design system, or adding shadcn components in a Next.js frontend.
Research the current best implementation pattern before making code changes. Use when Codex needs up-to-date guidance for a feature, integration, bugfix, or upgrade, especially around Next.js, AI SDK, Gemini, Drizzle, Bun, Docker, or other stack pieces that may have changed recently.
Turn plain-language requests into transient query or mutation scripts for the local seeded Postgres database in this repo. Use when Codex needs to inspect, clean up, backfill, bulk update, or analyze demo_orders or other local seed data, preferably through DOCKER MCP code-mode and otherwise through js_repl without requiring a checked-in script.
| name | agent-creator |
| description | Create new persistent Codex agents for this repo when asked to add a specialist agent, define a reusable subagent role, scaffold `.codex/agents/*.toml`, or wire the agent to local skills and MCP servers. |
Use this skill when the user wants a new reusable agent added to this repository, not just a one-off spawned subagent for the current turn.
Create repo-local agent definitions under .codex/agents that match the patterns already used in this codebase.
.codex/agents/<slug>.toml.name inside the file.name, description, model, model_reasoning_effort, sandbox_mode, and developer_instructions.nickname_candidates when the role is collaborative or likely to be spawned often.[[skills.config]] only when the agent has a repeatable workflow that should load a repo-local skill automatically.[mcp_servers.*] blocks only when the role truly depends on an MCP server.gpt-5.4-mini with medium reasoning for focused specialists.gpt-5.4 when judgment quality matters more than speed, such as reviews or research-heavy roles.read-only for QA, exploration, docs, and review agents.workspace-write only for builder or repair agents that are expected to edit files.developer_instructions short, imperative, and role-specific..codex/agents and copy its style.scripts/init_agent.mjs.nickname_candidates, MCP blocks, or linked skills..codex/skills/<skill-name> in the same turn.node .codex/skills/agent-creator/scripts/init_agent.mjs \
--slug qa-helper \
--name "QA Helper" \
--description "Browser-first QA assistant for lightweight smoke tests." \
--model gpt-5.4-mini \
--reasoning medium \
--sandbox read-only \
--nicknames "QA,Helper" \
--instructions "Act as a lightweight QA helper.\nCheck key routes, capture evidence, and avoid code edits unless asked."
.codex/agents/<slug>.toml.--skill-path .codex/skills/<skill-name> to add an enabled [[skills.config]] block.--dry-run to print the TOML without writing files.--force is passed.Return the new agent file path, the main behavior choices you encoded, and any assumptions you made while filling in missing details.