MiMo Subagent Skill

Portable Agent Skill for delegating substantial research, coding, review, and summarization to MiMoCode.

The host agent remains the orchestrator. MiMo performs bounded worker tasks, while the host reviews results, verifies changes, and decides what to integrate.

Features

• Works with Codex, Claude Code, OpenCode, Cline, Aider, and other Agent Skills-compatible hosts.
• Supports MiMoCode plan, build, and compose agents.
• Uses isolated Git worktrees for every write-capable run.
• Enables MiMoCode's native unattended permission mode by default.
• Discovers the MiMo executable dynamically on Windows, Linux, macOS, WSL, and Git Bash.
• Preserves the user's current branch, index, and working tree.
• Captures JSON events, final text, patches, changed files, and a run manifest.
• Never merges, cherry-picks, pushes, or switches model providers automatically.

Requirements

• MiMoCode CLI
• Git 2.5 or newer
• Windows PowerShell 5.1+ or Bash 3.2+

Install MiMoCode through its official installer or npm:

npm install -g @mimo-ai/cli
mimo --help

Provider access and cost depend on the user's MiMoCode configuration. Inspect the available catalog with:

mimo models

Install

Clone this repository into a skills directory supported by your host:

git clone https://github.com/Bobr2610/mimo-subagent.git \
  "$HOME/.agents/skills/mimo-subagent"

For project-local installation, place it under:

<project>/.agents/skills/mimo-subagent/

Hosts with a different skills directory can use the same repository contents; SKILL.md and every bundled path are location-independent.

Use

Ask the host agent directly:

Use MiMo to inspect the authentication flow and return a concise risk report.

Delegate this implementation to MiMo in an isolated worktree, run the tests,
and bring the patch back for review.

The host should read SKILL.md, choose a mode, build a bounded worker contract, and invoke the platform launcher.

Mode	Purpose	Writes
plan	Research, diagnosis, summaries, implementation plans	No
build	Focused implementation, fixes, tests, refactors	Worktree
compose	Large multi-phase or spec-driven work	Worktree
auto	plan without write access, otherwise build	Depends

Direct launcher examples:

powershell -NoProfile -File scripts/run_mimo.ps1 `
  -Prompt "Inspect the parser and report correctness risks." `
  -Mode plan `
  -TaskSlug parser-review

bash scripts/run_mimo.sh \
  --prompt "Inspect the parser and report correctness risks." \
  --mode plan \
  --slug parser-review

Write-capable run:

powershell -NoProfile -File scripts/run_mimo.ps1 `
  -Prompt "Implement the requested fix and run focused tests." `
  -Mode build `
  -Write `
  -TaskSlug parser-fix

bash scripts/run_mimo.sh \
  --prompt "Implement the requested fix and run focused tests." \
  --mode build \
  --write \
  --slug parser-fix

Run Artifacts

Each run writes:

docs/mimo-runs/YYYYMMDD/<task-slug>-<unique-id>/

The directory contains the prompt, JSON event log, final response, patch, changed-file list, and run.json. Changed worktrees are retained by default for review. Use --cleanup or -Cleanup only when the saved patch is enough.

Permission Model

The launchers pass --dangerously-skip-permissions to MiMoCode by default so a headless worker does not stop for repeated tool approvals. This does not bypass the host agent's sandbox or approval system.

Use unattended mode only in trusted repositories. MiMoCode may run commands, read files, and access configured network providers.

Tests

Windows:

powershell -NoProfile -ExecutionPolicy Bypass `
  -File tests/test_run_mimo.ps1

POSIX:

bash tests/test_run_mimo.sh

The test suites use a fake MiMo executable and temporary Git repositories. They do not require provider credentials.

License

MIT