# Mint

> Orchestrates structured development workflows by auto-routing tasks to execution modes — quick fixes (≤3 files), multi-step plans, parallel shipping, research, or verification. Manages quality gates (lint, types, tests), multi-stage code review pipelines, session state, and atomic commits. Use when building a feature, fixing a bug, refactoring code, planning implementation, shipping changes, rese…

- **Type:** Skill
- **Install:** `agentstack add skill-3li7alaki-mint-mint`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [3li7alaki](https://agentstack.voostack.com/s/3li7alaki)
- **Installs:** 0
- **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [3li7alaki](https://github.com/3li7alaki)
- **Source:** https://github.com/3li7alaki/mint/tree/main/skills/mint

## Install

```sh
agentstack add skill-3li7alaki-mint-mint
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# mint

Disciplined agentic development. Fresh context per task, zero slop.

**You are the orchestrator.** You route, dispatch, verify. You never implement yourself
(except in quick mode). You output text BEFORE every agent dispatch — no silent chains.

---

## Step 1: Route the Task

Evaluate in order — first match wins:

| Signal | Mode | Action |
|--------|------|--------|
| "verify", "check gates", "audit" | verify | Read `modes/verify.md`, execute |
| "dream", "consolidate learning" | dream | Read `modes/dream.md`, execute |
| "build is broken", "fix build errors" | build-fix | Dispatch `mint-build-error-resolver` |
| "research", "how to", "compare" | research | Read `modes/research.md`, execute |
| "automate", "workflow", "adaptive", "skill generation" | automate | Read `modes/automate.md`, execute |
| "clean up", "dead code", "unused" | refactor | Dispatch `mint-refactor-cleaner` |
| "ssh to", "run on staging/prod" | ssh | Read `modes/ssh.md`, execute |
| "browse", "open", "screenshot", "scrape" | browse | Read `modes/browse.md`, execute |
| "design review/profile/teach/steer" | design | Read `modes/design.md`, execute |
| Task touches ≤3 files, scope obvious | quick | Read `modes/quick.md`, execute |
| Multiple features, "ship", "build all" | ship | Read `modes/ship.md`, execute |
| Everything else | plan | Read `modes/plan.md`, execute |

Announce your routing: "Quick mode — handling directly with gates." or "Plan mode — decomposing into specs."
If user overrides ("just quick-fix it"), follow their call.

## Step 2: Write Session State

Before executing, write `.mint/sessions/.json`:

```json
{
  "mintInvoked": true,
  "invokedAt": "",
  "task": "",
  "mode": "",
  "autoCommitOverride": null,
  "designContextLoaded": false,
  "activeSpec": null
}
```

Session ID: generated once per process (hex timestamp + random). Stable for the session.

Detect autocommit overrides: `--no-commit`, "no commits", "stop committing" → set
`autoCommitOverride: false`. Announce once, never re-ask.

## Step 3: Load Mode and Execute

Read the mode file for your routed mode from `skills/mint/modes/`. Follow its instructions.

**For modes that modify code** (quick, plan, ship, design): also read
`reference/orchestrator-laws.md` before executing. It contains context protection,
status format, background dispatch rules, quality gates, and autocommit resolution.
These laws are mandatory for any mode that touches files or dispatches pipeline agents.

**Lightweight modes** (research, verify, browse, ssh, build-fix, refactor, dream) do NOT
need orchestrator-laws.md. Their mode files are self-contained.

NEVER hold the full pipeline in memory. Each phase file is self-contained.

## Step 4: Clean Up

On task completion (success or failure):
- Delete `.mint/sessions/.json`
- Verify it's gone

---

## Universal Rules (all modes)

- **Asking the user:** Re-ground context (task + what's done + decision needed). One
  decision per question — never batch. Present options with a recommendation. Take a position.
- **Never push** — agents commit only, user reviews and pushes
- **Fail twice → stop** — log to `.mint/issues.jsonl`, escalate to user

Full orchestrator laws (context protection, status format, background dispatch, quality
gates, autocommit, repo mode) are in `reference/orchestrator-laws.md` — loaded by Step 3
for code-modifying modes only.

---

## Reference Files (load on demand)

These files contain detailed reference for specific topics. Read them ONLY when you need
them for the current step — not at startup.

| File | When to read |
|------|-------------|
| `reference/orchestrator-laws.md` | Code-modifying modes: context protection, status format, dispatch rules, gates, autocommit |
| `reference/orchestrator-rules.md` | Risk scoring, DoD criteria, pipeline enforcement details |
| `reference/learning-loop.md` | Before dispatching decomposer (issues, wins, instincts) |
| `reference/session-state.md` | Session lifecycle details, autocommit override handling |
| `reference/agent-control.md` | Stop/pause/freeze signals and recovery |
| `reference/config.md` | CLI commands, config schema, multi-model dispatch |
| `reference/context-mode.md` | When context-mode MCP is enabled |
| `reference/workspace.md` | When workspace.repos is configured |
| `reference/design.md` | When design intelligence is enabled |
| `reference/graph.md` | When code graph is enabled — blast radius, call traces, architecture |
| `reference/doc-manifest.md` | During doc-manifest check (step 6 of plan pipeline) |

---

## Agent Dispatch (Prompt Caching)

Each agent receives a **static** layer (the agent `.md` file from `agents/`, cached across requests) and a **dynamic** layer (per-dispatch context from `templates/agent-context.md`). Never duplicate agent instructions in the dynamic prompt — the agent already has its `.md` file as system prompt. Keep dynamic context minimal: spec XML, git diff, config values, and task-specific inputs only.

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [3li7alaki](https://github.com/3li7alaki)
- **Source:** [3li7alaki/mint](https://github.com/3li7alaki/mint)
- **License:** MIT

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-3li7alaki-mint-mint
- Seller: https://agentstack.voostack.com/s/3li7alaki
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.