This repository has moved to robinv8/mino.

robinv8/mino-skills is archived. All future development, releases, and issues are now tracked in the unified monorepo at robinv8/mino.

New install command: gh copilot extension install robinv8/mino (or npx skills add robinv8/mino ...)

Migrated: 2026-04-25

English | 简体中文

Agent Skills compatible skill pack for task-driven development.

Turn a Markdown spec into executed, verified code — regardless of which AI agent you use.

A set of four engineering skills that implement the Iron Tree Protocol: an opinionated workflow for taking a Markdown requirement document all the way through execution, verification, acceptance when needed, composite aggregation, and reconciliation.

Markdown spec → /mino-task → DAG approval → /mino-run → /mino-verify → /mino-checkup → done

No GUI. No runtime. No deposition events. Just prompts that agents follow — every artifact (issue body, brief section, YAML event) is rendered from a fixed template under skills/<skill>/templates/, so different agents produce byte-identical output.

v0.1.0 ships with the following end-to-end coverage on https://github.com/robinv8/mino-skills:

Full regression evidence: reports/phase2-regression-report.md.

mino-skills/
├── skills/
│   ├── task/
│   │   ├── SKILL.md                     # Markdown → DAG → issues + briefs
│   │   └── templates/                   # brief / issue body / task_published event
│   ├── run/
│   │   ├── SKILL.md                     # Serial execution + commit + run.lock
│   │   └── templates/                   # 3 events + execution-summary section + lock
│   ├── verify/
│   │   ├── SKILL.md                     # Build/test/lint with Verify Anchor SHA
│   │   └── templates/                   # 5 outcome events + 3 brief sections
│   ├── checkup/
│   │   ├── SKILL.md                     # 7 modes incl. pre-flight & finalize
│   │   └── templates/                   # 7 events + 4 brief sections
│   └── references/
│       ├── iron-tree-protocol.md        # Execution loop specification (v1.8)
│       ├── workflow-state-contract.md   # Stage vocabulary + event whitelist
│       └── brief-contract.md            # Brief format (17 sections)
├── reports/
│   └── phase2-regression-report.md      # E2E validation evidence
├── README.md
└── LICENSE

Supported hosts: Copilot CLI, Claude Code, Cursor.

/plugin marketplace add robinv8/mino-skills
/plugin install mino@mino-skills
/plugin update mino@mino-skills    # later upgrades — finally smooth

If your agent host doesn't yet support /plugin marketplace, use this fallback.

The skills CLI installs skills to 45+ AI tools with one command.

Run inside your project directory:

Claude Code

npx skills add robinv8/mino-skills -a claude-code -y

Codex

npx skills add robinv8/mino-skills -a codex -y

Cursor

npx skills add robinv8/mino-skills -a cursor -y

Other tools

# List all supported agents
npx skills add --help

# Install to multiple agents at once
npx skills add robinv8/mino-skills -a claude-code -a codex -y

Options

cd your-project

# Claude Code
mkdir -p .claude/skills
git clone https://github.com/robinv8/mino-skills.git .claude/skills/mino

# Cursor / Codex / OpenCode (shared `.agents/skills/` path)
mkdir -p .agents/skills
git clone https://github.com/robinv8/mino-skills.git .agents/skills/mino

Any Agent Skills-compatible agent will auto-discover them:

No tool required — just copy the prompt:

cat skills/mino-task/SKILL.md
# Paste into ChatGPT, Claude, Cursor, or any AI chat

If your repo predates this protocol and already has open issues, standardize them:

/task adopt issue-12

This produces the same brief, events, and label set as a native task. Composite issues (3+ open checkboxes) are refused with the iron-tree:needs-breakdown label so you can split them first. Closed issues are refused.

Workflow position is mirrored on GitHub via labels:

Pull Requests are not adopted; they continue to merge against issues as usual.

Brief quality (v1.11+): the brief generated by adopt is structured, not a verbatim dump. The agent extracts testable acceptance criteria, derived verification steps, and inferred target files from the issue body and qualifying comments — matching the field-filling pattern of briefs produced by /mino-task PRD.md. Issue bodies on GitHub are never modified; standardization lives in .mino/briefs/issue-{N}.md only.

Iron Tree Protocol v1.10 treats local .mino/events/issue-N/*.yml as the single source of truth. GitHub issue comments are a notification channel, not an event log:

• Routine successful transitions (adopt, run, verify pass) are silent — no comment.
• Halts / failures that require human action still post an immediate comment so you see the signal.
• On completion, /mino-checkup done posts one short completion notice (heading + Completion Basis + Code Ref + Code Publication State) — no inline event log. The local .mino/events/issue-{N}/ directory is the sole authoritative record; back it up yourself if you need durability.

Pre-v1.10 issues (per-event comments) continue to be readable by /mino-checkup reconcile as a fallback source. Issues completed under protocol v1.10–v1.11 still carry the legacy inline-YAML done comment; /mino-checkup reconcile falls back to that signature when the local log is missing. v1.12+ done comments contain no YAML.

mino-skills follows the standard Agent Skills convention: main is always the released branch, and the skills CLI pulls the latest content on demand.

# In your project — refresh ALL installed skills to the latest main
npx skills update -y

# Inspect what is currently installed
npx skills list

Heads-up: the CLI tracks each skill by its skill name (task, run, verify, checkup), not by its source repo. Running npx skills update robinv8/mino-skills therefore reports "no installed skills found" — use the bare npx skills update -y form instead.

The update overwrites only the files inside .claude/skills/<name>/ or .agents/skills/<name>/. Your local workflow data — .mino/briefs/, .mino/run.lock, GitHub issues, and event comments — is untouched.

Protocol upgrades are designed to be backward compatible at minor versions: state machines that are mid-flight when you update keep working without migration. Major-version bumps (e.g., v2.0) will document any required action at the top of this README.

v0.6.3 — Reply Comments + Optional Note Input + Default-Silent Status. GitHub issue comments split into two classes: status (slim, interrupt-only) and reply (conversational, content-bearing). checkup_done and verify_passed no longer post status comments. New <note> argument lets human verifiers feed prose into agent reply decisions. Protocol additions (additive, v1.13 stays): § Comment Classes, § Reply Dispatch, optional event field reply_posted. v0.6.2 — Slim-Comment Cleanup + Commit Auto-Link. Pays off v0.5.2 debt: audible GitHub comments now render exclusively from comment-*.md.tmpl files (no yaml fences, no .mino/* paths). Commit: auto-link is wired into all audible comments. Protocol gains § Slim Comment Invariant (additive; v1.13 stays). v0.6.1 — Verification Report Artifact. mino-verify now authors a human-readable evidence report at .mino/reports/issue-{N}/report.md and optionally promotes it to docs/integrations/<slug>.md as a separate commit. verify_* events gain optional report_path and promoted_doc fields. Protocol stays at v1.13. v0.6.0 — Loop Mode is now the default for /mino-task. After approval, the orchestrator drives run/verify/checkup automatically until a halt condition fires (approval-required, pending_acceptance, fail_terminal, blocked, reapproval_required, loop_budget_exhausted). New /mino-task resume <loop_id> sub-command for explicit halt resolution. New .mino/loops/ directory holds Loop entity + repo-level lease. Stepwise opt-out: invoke /mino-run, /mino-verify, /mino-checkup directly. BREAKING behavior change. v0.5.2 — GitHub comments slimmed to human-readable notifications (no inline YAML, no Local events: pointer); commits use [run] #{N} so GitHub auto-links them on the issue timeline. Protocol bumped to v1.12 (policy change, schema unchanged). v0.5.1 — Renamed slash commands to /mino-task, /mino-run, /mino-verify, /mino-checkup to avoid palette collisions. v0.5.0 — Plugin marketplace support (/plugin install mino@mino-skills).

cat > feature.md << 'EOF'
# Add dark mode

## Acceptance Criteria
- [ ] Toggle in settings
- [ ] Persists across launches
- [ ] Respects system preference by default

## Target Files
- SettingsView.swift
- Theme.swift
EOF

/task feature.md

task reads the doc, classifies it, extracts a DAG, computes a revisioned task graph, and asks for your approval before creating any issues or briefs. The generated .mino/briefs/ files are local workflow cache and should not be committed.

/run issue-8

run picks the next eligible task from the DAG, resolves the canonical Task Key, increments the attempt counter, makes changes, and hands off to verification.

Triggered automatically by run, or call directly:

/verify issue-8

Runs build, tests, linters. Results:

• ✅ pass → advances to checkup
• ❌ retryable → feeds Failure Context back to run (max 3 retries)
• 🚫 terminal → blocks the task
• ⏸️ manual acceptance → stops for human review, then continue with /mino-checkup accept issue-8

/checkup reconcile
/checkup accept issue-8
/checkup aggregate issue-1

checkup handles pre-flight checks, brief reconciliation, recording manual acceptance, aggregating composite parents, and printing a centralized Pending Acceptance list before a task can reach done.

Etymology — Iron for the iron-clad guarantees the protocol enforces (immutable event log, deterministic state machine, idempotent publish, audit-trail by construction); Tree for the data structure every workflow takes — a DAG of composite parents and child tasks linked by depends_on. Not a reference to 铁树开花; the protocol is engineered, not miraculous.

┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│   task      │────▶│    run      │────▶│   verify    │
│  define     │     │  execute    │     │   validate  │
└─────────────┘     └─────────────┘     └──────┬──────┘
                                               │
                          ┌────────────────────┼────────────────────┐
                          │                    │                    │
                          ▼                    ▼                    ▼
                    ┌──────────┐       ┌──────────┐       ┌──────────┐
                    │   pass   │       │ retryable│       │ terminal │
                    │  checkup │       │   run    │       │  blocked │
                    └──────────┘       └──────────┘       └──────────┘
                          │
                          ▼
                    ┌──────────┐
                    │   done   │
                    └──────────┘

• Self-correction: verify failures feed Failure Context back to run for a different approach
• Serial execution: DAG nodes run one at a time (v1), respecting depends_on. Enforced by .mino/run.lock (file lock, V3 will revisit parallel runs).
• Approval gates: Human must approve the DAG before any execution begins
• Manual acceptance: if automation cannot prove correctness, verify stops at pending_acceptance and checkup accept records the human decision
• Shared visibility: detailed manual checklists stay in local briefs, while issue labels/comments make acceptance status visible to collaborators
• Revision-aware approval: published work stays executable only while Spec Revision matches Approved Revision
• Canonical identity: Task Key is the protocol identity; issue-8 is a user-facing locator after publish
• Verify Anchor SHA: every verify result binds to the HEAD SHA at start, immune to mid-flight code drift
• External-event awareness: if an issue is closed outside the workflow, checkup reconcile records an External Event section instead of silently syncing to done
• Template-driven artifacts: all events, briefs, and issue bodies are rendered from templates/*.tmpl files — agent-agnostic and diff-stable

• skills/references/iron-tree-protocol.md — the execution loop
• skills/references/workflow-state-contract.md — stage vocabulary
• skills/references/brief-contract.md — brief format

• Agent Skills compatible agent (Claude Code, Cursor, Copilot, Goose, Gemini CLI, etc.)
• gh CLI for GitHub issue creation
• .mino/briefs/ directory (created automatically on first use, local-only and not committed)

• Agent Skills specification
• skills CLI — install skills to 45+ AI tools
• Mino — the macOS GUI app this skill set was extracted from

# Remove from current project
npx skills remove robinv8/mino-skills

# Remove globally (from all projects)
npx skills remove robinv8/mino-skills -g

# Remove from a specific agent globally
npx skills remove robinv8/mino-skills -a claude-code -g

# Project-level
rm -rf .claude/skills/mino
rm -rf .agents/skills/mino

# Or globally
rm -rf ~/.claude/skills/mino
rm -rf ~/.agents/skills/mino

Contributions welcome. Workflow:

1. Fork this repository
2. Edit a SKILL.md or add a new skill
3. Open a PR describing the motivation and impact

For questions, contact hello@robinren.me.

MIT