mimo-subagent

name	mimo-subagent
description	Delegate substantial research, coding, review, or summarization to MiMoCode as a cheap external subagent. Use when MiMo is requested or offloading would save host context and cost.
compatibility	Requires MiMoCode CLI and Git 2.5+; Windows PowerShell 5.1+ or Bash 3.2+ on Linux, macOS, WSL, or Git Bash; provider use may require network access.

MiMo Subagent

Use the expensive host model as the orchestrator and MiMoCode as the external worker. Keep requirements, judgment, integration, and the final answer in the host. Send bounded execution-heavy work to MiMo.

Delegation rule

Use MiMo when any of these are true:

• The user explicitly asks for MiMo, MiMoCode, mimo run, or this skill.
• Research, code generation, review, log reading, or summarization would consume substantial host context.
• Several independent tasks can run as separate MiMo workers.
• A cheap first pass followed by host review is appropriate.

For an eligible task, call the bundled launcher directly. Do not replace MiMo with the host's built-in deep-research feature or ordinary host subagents.

Skip MiMo when the task is trivial, the user opts out, MiMo is unavailable, or the material contains secrets that should not be sent to the configured model provider. Never claim that MiMo ran when only the host agent did the work.

Read the focused references

• Read references/modes.md when choosing or chaining plan, build, and compose.
• Read references/host-agents.md when invocation, approvals, parallelism, or installation differs across host agents.
• Read references/models.md when selecting a model or troubleshooting provider availability.

Launchers

Use the script matching the current operating system:

# Windows PowerShell 5.1+ or PowerShell 7+
powershell -NoProfile -File scripts/run_mimo.ps1 `
  -Prompt "Inspect the authentication flow and report correctness risks." `
  -Mode plan `
  -TaskSlug auth-review

# Linux, macOS, WSL, or Git Bash
bash scripts/run_mimo.sh \
  --prompt "Inspect the authentication flow and report correctness risks." \
  --mode plan \
  --slug auth-review

Resolve these paths relative to this SKILL.md, not relative to the user's current directory. Run the launcher's check once before the first delegation in a task:

powershell -NoProfile -File scripts/run_mimo.ps1 -Check

bash scripts/run_mimo.sh --check

Do not repeatedly print availability status. Report it only when MiMo is missing, unusable, or the user asks.

Choose a mode

Mode	Use for	File writes
plan	Research, exploration, diagnosis, summaries, implementation plans	No
build	Focused implementation, fixes, tests, refactors	Yes, in a worktree
compose	Large spec-to-ship work with multiple phases or internal workers	Yes, in a worktree
auto	Launcher selects plan without --write, otherwise build	Depends

Pass -Write on Windows or --write on POSIX for any task that may edit files. The launcher rejects plan plus write access.

MiMo locks its agent mode after the first message. Build and Plan can switch in an interactive TUI, but Compose is isolated. In headless orchestration, switch modes by starting a new run for the next phase. See references/modes.md.

Build a worker contract

Give MiMo a compact contract with:

1. Goal - one measurable outcome.
2. Scope - files, directories, URLs, or questions it owns.
3. Constraints - compatibility, safety, style, and exclusions.
4. Verification - commands or evidence that prove completion.
5. Return - concise summary, changed files, tests, risks, and sources.

For write tasks, include:

Work only inside the current isolated worktree. Do not modify remotes, global
configuration, credentials, or files outside this worktree. Implement the
requested change, run focused verification, and finish with a concise summary
of changes, tests, and remaining risks.

For research tasks, include:

Do not edit project files. Use primary sources where possible. Return claims
with source URLs, note uncertainty, and keep raw exploration out of the final
summary.

Do not dump the whole conversation into MiMo. Pass only the context needed for the delegated result.

Standard workflow

1. Classify the task as read-only or write-capable.
2. Check MiMo once, select plan, build, or compose, and apply the model policy from references/models.md.
3. Write the worker contract.
4. Run the platform launcher and wait for process completion.
5. Read run.json, 10_events.jsonl, and 20_final.md from the returned run directory.
6. For write tasks, inspect 30_changes.patch, 40_changed_files.txt, and the retained worktree.
7. Review MiMo's claims and changes with the host model. Run relevant tests.
8. Apply accepted changes to the user's working tree. Do not blindly apply a patch.
9. Remove the worktree only after integration or explicit rejection:

git worktree remove --force "<worktree-path>"

10. Give the user the host's synthesis, including MiMo failures or uncertainty.

The launcher saves outputs under:

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

Unattended permissions

The launchers pass MiMoCode's native --dangerously-skip-permissions flag by default. This prevents MiMo's own tool approval prompts during a headless run. Use -AllowPrompts or --allow-prompts only when the user wants interactive permission handling.

This flag does not override the host agent's sandbox or approval system. If the host asks every time, configure a one-time narrow allow rule for the bundled launcher command prefix. Do not silently disable all host protections. See references/host-agents.md.

Permission bypass is appropriate only inside a trusted, isolated worktree. It removes MiMo's confirmation boundary; it does not sandbox network access, secret reads, plugins, or processes.

Worktree contract

For every write-capable run, the launchers:

• Require a Git repository.
• Create a unique detached worktree under .mimo-worktrees/.
• Run MiMo with that worktree as --dir.
• Stage changes only inside the worktree for complete diff capture.
• Save a binary-capable patch and changed-file list outside the worktree.
• Keep changed worktrees by default for inspection and recovery.
• Remove an unchanged worktree automatically.

Use -Cleanup or --cleanup only when a preserved patch is sufficient and the caller explicitly wants the worktree removed. The launcher never merges, cherry-picks, pushes, or edits the user's current branch.

Read-only plan runs use the current directory and do not create worktrees.

Parallel workers

For independent work, start one launcher process per bounded task. Each write run receives a unique worktree and output directory. Do not send two workers to the same worktree. Wait for all workers, then let the host agent compare, verify, and integrate their outputs.

Prefer a few strong partitions over a large swarm. MiMo workers still consume provider tokens and can duplicate effort.

Completion gate

Do not finish until:

• MiMo's process exit code is recorded.
• The requested artifact or evidence exists.
• Write-task diffs have been reviewed by the host agent.
• Relevant verification has run, or the missing verification is disclosed.
• Sources are present for research claims that depend on the web.
• Retained worktrees are either integrated, intentionally kept, or removed.