sub-agent-creator

name

sub-agent-creator

description

Create, edit, convert, or audit ChatLuna sub-agent markdown files. Use when adding a new sub-agent, refining a sub-agent prompt, choosing a sub-agent's goal and output contract, restricting tools, skills, MCP, or computer permissions, pinning or omitting a model, setting maxTurns, placing agents under local `data/chatluna/agents`, writing them into a remote sandbox before `agentcli sync`, or converting Claude or OpenCode agent files into ChatLuna-compatible sub-agents.

Sub-Agent Creator

Create complete sub-agent markdown files for ChatLuna's extension-agent package. Prefer ChatLuna format unless the target file, target directory, or user request clearly calls for Claude or OpenCode compatibility.

Default Decisions

• Default new ChatLuna-native agents to local data/chatluna/agents/<agent-name>/index.md.
• If the current computer backend is not local, write the new file to the sandbox path ~/.chatluna/agents/<agent-name>/index.md instead, create the parent folder first if it does not exist, and finish with agentcli sync subagents so the local ChatLuna path is updated.
• Preserve the existing format when editing an existing sub-agent file.
• Use a short, specific agent name. Keep the job narrow.
• Omit model when you do not know the exact installed model name. Let the sub-agent inherit the parent model instead of guessing.
• Keep allowKoishiMessageTransform: false unless the agent must receive the user's Koishi message structure instead of plain text.

Authoring Workflow

1. Define the agent's exact job.

   • State the goal, the expected input, the limits, and the expected final answer.
   • Split planning, implementation, review, and UI control into separate agents when possible.

2. Choose the narrowest permission set that still works.

   • Read-only agents usually need file_read, glob, and grep.
   • Focused coding agents usually need file_read, file_edit, file_write, glob, grep, and sometimes bash.
   • Web research agents may need browser_open, web_search, browser_read_text, and browser_summarize.
   • Keep computer access denied unless the task truly needs desktop control.

3. Choose model and turn budget.

   • Use the exact requested model when the user provides one.
   • Leave model unset when the correct model is unknown.
   • Use 8-20 turns for lookup or review, 20-60 for focused implementation, and 100 only for broad multi-step work.

4. Write the prompt body.

   • Start with the role in one sentence.
   • Name the tools or capabilities the agent can rely on.
   • Describe the workflow in execution order.
   • State hard boundaries.
   • End with the required result format.

5. Review the file.

   • Remove placeholders, TODOs, and vague filler.
   • Make sure the description says when the parent should use this agent.
   • Make sure the permissions match the prompt.
   • If the file was authored in a sandbox, run agentcli sync subagents and verify the local file path after sync.

What Good Sub-Agents Look Like

• Make the description usable as catalog text. Write what the agent does and when to call it.
• Make the prompt operational, not inspirational. Tell the agent what to do first, what to avoid, and what to return.
• Keep each agent single-purpose. Do not create a giant all-in-one assistant.
• Prefer explicit output contracts such as file paths, line numbers, patches, checklists, or concise result summaries.
• Do not rely on nested delegation. The platform blocks sub-agents from using task.

ChatLuna Format

Use this format for new ChatLuna-native agents.

---
name: explore-api
description: Read-only API exploration agent. Use when you need to find routes, trace handlers, inspect request or response shapes, and report exact files and line numbers.
format: chatluna
enabled: true
hidden: false
model: openai/gpt-4.1-mini
maxTurns: 12
allowKoishiMessageTransform: false
permissions:
    skills:
        mode: deny
        allow: []
        deny: []
    mcp:
        mode: allow
        allow: []
        deny: []
    tools:
        mode: allow
        allow:
            - file_read
            - glob
            - grep
        deny: []
    computer:
        mode: deny
        allow: []
        deny: []
---

You are the Explore API sub-agent. Your job is to inspect the codebase and answer API questions with evidence.

Start with broad discovery using glob and grep. Then read only the files that matter and trace imports or call sites until the path from route entry to handler is clear.

Do not edit files, run shell commands, use MCP, or speculate beyond the code you can point to.

Return:

- the relevant endpoints or entry points
- the exact files and line numbers
- the request and response shape when present
- any uncertainty that still needs a human decision

ChatLuna Frontmatter Fields

• name: Human-facing sub-agent name.
• description: Catalog text. Include both capability and trigger.
• format: Use chatluna for native agents.
• enabled: Usually true.
• hidden: Set true only when the agent should stay out of the normal catalog.
• model: Optional exact model name. Omit to inherit from the parent agent.
• maxTurns: Hard turn budget. Default config is 100, but most specialists should use less.
• allowKoishiMessageTransform: Set true only when the prompt must preserve Koishi message elements, mentions, or multimodal transforms.
• permissions: Four permission groups: skills, mcp, tools, and computer.

Permission Rules

Each permission group uses this shape:

mode: inherit | all | allow | deny
allow: []
deny: []

• inherit: Use the global default for that category.
• all: Allow everything in that category.
• allow: Allow only the listed entries.
• deny: Deny the listed entries and allow the rest of that category.

Prefer mode: allow for tight specialist agents. It is the clearest and safest way to restrict behavior.

Use the right names:

• permissions.tools: Tool names such as file_read, file_edit, file_write, glob, grep, bash, web_search, browser_read_text, and browser_summarize.
• permissions.skills: Skill names.
• permissions.mcp: MCP server names.
• permissions.computer: Backend names such as local, e2b, and open-terminal.

Important behavior:

• Sub-agents cannot use task, even if you mention it.
• Only the builtin plan and explore agents get hard-coded read-only protection. Custom read-only agents must restrict their own tools.
• When you want "no MCP", use mcp.mode: allow with an empty allow list.
• When you want "no tools", use tools.mode: allow with an empty allow list.

Recommended Permission Shapes

Use these patterns as defaults.

Read-only explorer

permissions:
    skills:
        mode: deny
        allow: []
        deny: []
    mcp:
        mode: allow
        allow: []
        deny: []
    tools:
        mode: allow
        allow: [file_read, glob, grep]
        deny: []
    computer:
        mode: deny
        allow: []
        deny: []

Focused coder with shell access

permissions:
    skills:
        mode: deny
        allow: []
        deny: []
    mcp:
        mode: allow
        allow: []
        deny: []
    tools:
        mode: allow
        allow: [file_read, file_edit, file_write, glob, grep, bash]
        deny: []
    computer:
        mode: deny
        allow: []
        deny: []

Browser or desktop operator

permissions:
    skills:
        mode: deny
        allow: []
        deny: []
    mcp:
        mode: allow
        allow: []
        deny: []
    tools:
        mode: allow
        allow: [web_search, browser_open, browser_read_text, browser_summarize]
        deny: []
    computer:
        mode: allow
        allow: [local]
        deny: []

Only grant computer when the task explicitly needs GUI or browser control.

Claude-Compatible Format

Use Claude-style frontmatter only when editing a Claude agent or when the file will live in a Claude agents directory.

Supported Claude fields:

• name
• description
• model
• maxTurns
• hidden
• tools
• disallowedTools
• skills
• mcpServers

Use a simple allow list when possible.

---
name: repo-reviewer
description: Read-only review agent. Use when you need to inspect code, trace behavior, and report issues with evidence.
model: openai/gpt-4.1-mini
maxTurns: 16
hidden: false
tools:
    - read
    - glob
    - grep
skills: []
mcpServers: []
---

You are the Repo Reviewer sub-agent. Read the relevant code, gather evidence, and return precise findings with file paths and line numbers.

Notes:

• read, write, edit, bash, glob, grep, webfetch, and task aliases are mapped into ChatLuna-compatible tool names.
• Prefer tools over disallowedTools for narrow agents.

OpenCode-Compatible Format

Use OpenCode-style frontmatter only when editing an OpenCode agent or when the file will live in an OpenCode agents directory.

Supported OpenCode fields:

• name
• description
• model
• hidden
• disable
• prompt
• tools
• permission
• mode

Example:

---
name: quick-explorer
description: Fast read-only explorer. Use when you need to search a repo, inspect files, and return exact evidence.
mode: subagent
model: openai/gpt-4.1-mini
hidden: false
disable: false
tools:
    read: true
    glob: true
    grep: true
    write: false
    edit: false
    bash: false
permission:
    edit: deny
    bash: deny
    webfetch: deny
    task: deny
---

You are the Quick Explorer sub-agent. Find the relevant files, read only what matters, and return exact evidence.

Notes:

• If prompt exists in frontmatter, ChatLuna prepends it to the markdown body.
• mode is informational here. Keep it accurate, but do not rely on it for enforcement.

Prompt Writing Rules

• Tell the sub-agent what success looks like.
• Tell it what to do first.
• Tell it what not to do.
• Tell it what format to return.
• Name concrete artifacts such as file paths, symbols, routes, schemas, or test commands.
• Avoid generic lines such as "be helpful", "think step by step", or "use your best judgment" unless they add a real constraint.

Final Checklist

• Use the correct format for the target directory or existing file.
• Match the agent name, file name, and job.
• Make the description specific enough to trigger the right agent.
• Omit model instead of inventing one.
• Use the narrowest permission set that still works.
• Keep computer off unless the task explicitly needs it.
• Keep the prompt complete. Do not leave stubs.