mimo-subagent

name

mimo-subagent

description

Reusable MiMo CLI orchestration skill for delegating work to MiMo across any project. Use this skill whenever the user asks to use mimo, MiMo CLI, mimo run, mimo-auto, MiMo as a subagent, an external autonomous agent, or wants a task delegated to MiMo for research, coding, or summarization. Prefer this skill even when the user does not explicitly mention a skill but clearly wants work routed through MiMo.

MiMo Subagent

Purpose

Use MiMo CLI as a reusable external agent from any project directory.

AUTOMATIC: MiMo Check at Conversation Start

At the START of EVERY conversation, the agent MUST check MiMo availability:

# Check if MiMo is available
mimo --help 2>$null
if ($LASTEXITCODE -ne 0) {
    $mimoShim = "C:\Users\misha\AppData\Roaming\npm\mimo.cmd"
    if (Test-Path $mimoShim) {
        Write-Host "MiMo available via shim: $mimoShim"
    } else {
        Write-Host "WARNING: MiMo CLI not found. Install @mimo-ai/cli"
    }
} else {
    Write-Host "MiMo CLI available"
}

# Check login state
mimo providers whoami 2>$null

Display result to user:

=== MiMo Status ===
CLI: [available / not found]
Provider: [logged in / not logged in]
Model: mimo/mimo-auto (default)
====================

If MiMo is NOT available, inform the user and skip delegation features.

Work Tree Pattern

For each MiMo delegation, create an ISOLATED work tree that gets DELETED after use.

Why work trees

• MiMo works in its own space — no conflicts with main codebase
• Changes don't leak into your branches
• Report stays separate — never merged into feature/staging/master
• Clean isolation: work tree → output → delete

How to create work tree

# 1. Create isolated work tree for MiMo
$taskSlug = "mimo-task-name"
$workTree = ".mimo-worktrees/$taskSlug"
New-Item -ItemType Directory -Force -Path $workTree | Out-Null

# 2. Copy ONLY the files MiMo needs into work tree
Copy-Item "kg_engine/services/materials_kg.py" "$workTree/"

# 3. Run MiMo in the work tree context
mimo run -m "mimo/mimo-auto" "Review materials_kg.py for bugs. Work in: $workTree"

# 4. Collect output (report stays in docs/mimo-runs/)
# 5. DELETE the work tree
Remove-Item -Recurse -Force $workTree

Work tree rules

1. NEVER let MiMo work directly in the project root
2. ALWAYS create .mimo-worktrees/<task-slug>/ for each task
3. ALWAYS delete work tree after MiMo finishes
4. NEVER merge MiMo's work tree branches into main branches
5. Report goes to docs/mimo-runs/ — that's the only output that stays

Directory structure

materials-kg/
  .mimo-worktrees/          # TEMPORARY — deleted after use
    task-slug-1/             # Isolated space for MiMo task 1
    task-slug-2/             # Isolated space for MiMo task 2
  docs/
    mimo-runs/               # PERMANENT — reports stay here
      20260701/
        task-slug/
          00_prompt.txt
          20_summary.md
          sources.md

Prompt Writing Guide

Structure of a good MiMo prompt

[ROLE] You are a [specific role] reviewing/analyzing/building [what].

[TASK] Your task is to [exact action].

[CONTEXT] Here is the relevant code/data:
[paste ONLY what's needed]

[CONSTRAINTS]
- Do NOT modify any files
- Return findings as [format]
- Include source URLs for web research
- Keep response under [length] words

[OUTPUT FORMAT]
Return in this exact structure:
## Summary
[1-2 sentences]

## Findings
1. [finding]
2. [finding]

## Sources (if applicable)
- [URL]

Prompt examples by task type

Code review:

[ROLE] You are a senior Python developer reviewing code for bugs.

[TASK] Review the following module for logical errors, type issues, and edge cases.

[CONTEXT]
```python
[paste code here]

[CONSTRAINTS]

• Do NOT modify files
• Focus on correctness, not style
• Order findings by severity (critical → low)

[OUTPUT FORMAT]

Critical Issues

• [issue]

Warnings

• [issue]

Suggestions

• [issue]

**Research:**

[ROLE] You are a technical researcher.

[TASK] Find the latest documentation for [technology].

[CONSTRAINTS]

• Use only official documentation
• Include URLs for every claim
• Focus on version [X] and newer

[OUTPUT FORMAT]

Key Points

• [point]

Code Examples

[example]

Sources

• [URL] — [description]

**Summarization:**

[ROLE] You are a technical writer summarizing complex information.

[TASK] Summarize the following content for a [target audience].

[CONTEXT] [paste content]

[CONSTRAINTS]

• Max [N] bullet points
• No jargon without explanation
• Actionable items first

[OUTPUT FORMAT]

TL;DR

[1 sentence]

Key Takeaways

• [point]

Action Items

• [item]

### Anti-patterns (don't do this)

**BAD — too vague:**

Look at this code and tell me if it's good

**BAD — too much context:**

Here's my entire project (5000 lines). Review everything.

**BAD — no output format:**

Tell me about Neo4j

**GOOD — specific, constrained, structured:**

Review kg_engine/services/materials_kg.py lines 45-120. Focus on: error handling, type safety, edge cases. Return: Critical → Warning → Suggestion format. Do NOT modify files.

## Standard workflow
1. **Check MiMo availability** (at conversation start)
2. **Identify the goal:** research / code task / review / summary
3. **Create work tree:** `.mimo-worktrees/<task-slug>/`
4. **Copy needed files** into work tree
5. **Build prompt** using the guide above
6. **Run MiMo** in work tree context
7. **Collect report** → `docs/mimo-runs/YYYYMMDD/<slug>/`
8. **Delete work tree**
9. **Summarize** results to user

## Safety and confirmation
- **NEVER** send secrets, API keys, or credentials to MiMo
- **NEVER** send confidential code without user consent
- **NEVER** let MiMo modify files in the main project
- **ALWAYS** use work tree isolation
- **ALWAYS** delete work tree after use
- For public research → proceed when user asks
- For proprietary code → warn the user first

## How to Invoke This Skill

**This skill is a markdown file that gets loaded as instructions when called.**

### Codex (OpenAI)

Codex reads `.codex/instructions.md` at start. To invoke this skill:

/mimo-subagent

Codex loads `.agents/skills/mimo-subagent/SKILL.md` into context.

### Claude Code

Claude Code reads `.claude/CLAUDE.md` at start. To invoke:

/mimo-subagent

Claude loads the skill file as additional instructions.

### Cursor

Cursor reads `.cursor/rules/*.mdc`. To invoke, ask the agent:

Use the mimo-subagent skill to review this code

Cursor loads `.agents/skills/mimo-subagent/SKILL.md`.

### MiMo (MiMoCode)

MiMo has a `skill()` tool. To invoke:

```javascript
skill({ name: "mimo-subagent" })

MiMo loads .agents/skills/mimo-subagent/SKILL.md into context.

Any Agent via AGENTS.md

AGENTS.md has a skills table. The agent reads it and decides:

When to load: Delegating tasks to MiMo

Agent calls Read(file_path=".agents/skills/mimo-subagent/SKILL.md").

Manual (bash)

Any agent can read the file directly:

cat .agents/skills/mimo-subagent/SKILL.md

Quick Reference

Agent	Method	Command
Codex	Slash command	/mimo-subagent
Claude Code	Slash command	/mimo-subagent
Cursor	Natural language	"Use mimo-subagent skill"
MiMo	Tool call	skill({name: "mimo-subagent"})
Any	Read file	Read(".agents/skills/mimo-subagent/SKILL.md")

Built-in Slash Commands

MiMoCode has built-in skills invocable via / prefix in chat.

Available Commands

Command	Description	Example
/imagegen	Generate or edit raster images (photos, illustrations, textures, mockups, sprites)	/imagegen a 3D isometric server room
/impeccable	Design, redesign, polish frontend UI — websites, dashboards, components, theming, accessibility, UX	/impeccable redesign the login page
/mimo-subagent	Delegate tasks to MiMo as a subagent for research, coding, or summarization	/mimo-subagent review this code
/openai-docs	Get help with OpenAI products, APIs, models, Codex itself	/openai-docs which model for code gen
/plugin-creator	Create and scaffold Codex plugin directories	/plugin-creator create a plugin
/self-extend	Evolve your own capabilities — new tools, hooks, skills	/self-extend add a hook
/skill-creator	Guide for creating effective skills	/skill-creator make a new skill
/skill-installer	Install skills from curated list or GitHub repo	/skill-installer install web-search
/web-search	Multi-source web research with anti-bot resilience and cited outputs	/web-search latest GPT-5 benchmarks

How Commands Work

1. User types /command-name arguments in chat
2. MiMoCode loads the skill's SKILL.md into context
3. Skill body becomes additional instructions for that scope
4. Skill does NOT change the tool set — only behavior and guidance

Built-in Workflows (via agent tool)

These are invoked programmatically by the agent, not as slash commands:

Workflow	Description	When to use
compose	Full autonomous pipeline: brainstorm → design → implement → verify → review → report → merge	Feature, bugfix, refactor, or review tasks end-to-end
deep-research	Parallel web searches → read sources → cross-check facts → cited report	Thorough multi-source research on any topic

Invoke compose:

workflow({ operation: "run", name: "compose", args: { task: "implement feature X", type: "feature" } })

Invoke deep-research:

workflow({ operation: "run", name: "deep-research", args: "latest research on quantum computing applications" })

Skill vs Workflow

Aspect	Slash Command (/)	Workflow (workflow())
Invocation	User types in chat	Agent calls programmatically
Scope	Single skill loaded as context	Multi-phase orchestration
Subagents	No	Yes — parallel subagents
Use case	Guidance for specific domain	Complex multi-step automation

Checklist

• MiMo CLI available (checked at conversation start)
• Provider login state verified
• Work tree created: .mimo-worktrees/<slug>/
• Only needed files copied to work tree
• Prompt follows the writing guide
• MiMo run completed
• Report saved to docs/mimo-runs/
• Work tree deleted
• Results summarized to user