Agents & tooling

Beyond the slash commands, the plugin ships optional sub-agents, a shared skill, a non-destructive edit hook, and an optional local MCP bridge. None of them replace verification — real ml1/make runs remain authoritative.

Sub-agents

Deep workflows can delegate to a focused agent. Each loads the authority-preflight skill before acting and reads the language bundle first.

Agent	Role	Notes
`minlang-creator`	Generate or revise MinLang source using the Creator repair loop and bundle authority	Edits `.ml`/`.mlui` only — never `app/generated/` or `.env`; runs GENERATE → VALIDATE → FIX → REVALIDATE until zero violations
`minlang-gatekeeper`	Validate MinLang against bundle authority and return strict, sanitized JSON verdicts only	Reject-first; `Write`/`Edit` disabled so it cannot mutate source
`minlang-migration-architect`	Plan slice-by-slice migration from a legacy app to MinLang with a durable ledger	Inventories domains/screens/tests before proposing MinLang; prefers partial adoption over a big-bang rewrite

Agents supplement commands; they do not bypass the verification step.

Skill: authority-preflight

Every producing command and agent loads this skill first. It:

Detects project shape — minlang.json (compiler + bundle pin), the primary .ml application file, AGENTS.md/CLAUDE.md, and Makefile targets (compile, test, validate, dev).
Checks whether ml1 is on PATH. The plugin does not bundle the compiler; when it is missing the skill points to the public install scripts rather than claiming otherwise.
Loads the language bundle from the stable URL or the project’s pin, so all downstream generation and validation work against the canonical authority:
```
https://github.com/codeshift-ai-solutions/minlang-releases/releases/latest/download/minlang-language-bundle.md
```

Post-edit hook

A PostToolUse hook matches Write|Edit and runs hooks/post-edit-ml.sh. When a .ml file is touched it emits a non-destructive reminder to validate the source — it never edits or reverts your file. It is the safety net that keeps “edit the .ml, then validate” the default loop.

Optional MCP bridge

When enabled, the plugin starts a local MCP server (mcp/ml1-bridge.py) that exposes a small, allowlisted, read-only set of ml1 tools to Claude Code:

features — print the current language feature catalog.
validate — validate a .ml file.
graph — emit the project graph / compile closure.
update --check — report available compiler/bundle/dependency updates.

The bridge is a convenience for tighter loops; the shell verification commands inside each slash command remain authoritative. The MCP server is configured in the plugin’s .mcp.json and launched with python3.

Monitors

The plugin manifest reserves an experimental monitors slot (monitors/monitors.json). It currently ships empty — a placeholder for future background watchers.