npm install -g @master0071/loop-md-cli
loop-md-cli --list           # sanity check — you should see 7 platforms
loop-md-cli --all            # default ralph paradigm, emits .claude/ .opencode/ etc.
loop-md-cli --validate --all # exit code 0 means every platform matches its templates

Requires Node.js >= 18, zero runtime dependencies.

Basic Usage

loop-md-cli --all                          # all platforms
loop-md-cli --claude --opencode            # specific platforms
loop-md-cli --claude --domain coding       # specific domain
loop-md-cli --kilo --domain ralph --domain graph  # multi-domain coexistence
loop-md-cli --all --dry-run                # dry run, writes nothing
loop-md-cli --validate --all               # validate consistency
loop-md-cli --watch --claude               # watch for file changes
loop-md-cli --incremental --all            # only rewrite changed files
loop-md-cli --archive configs.zip --all    # export as ZIP
loop-md-cli --trae --domain coding \       # per-role sub-agent models
  --model-orchestrator "DeepSeek-V4-Pro" \
  --model-executor "DeepSeek-V4-Flash" \
  --model-reviewer "Doubao_1_6"

Why?

The 7 AI coding platforms each use a different agent config format; maintaining them by hand is repetitive and error-prone:

Platform	Output dir	frontmatter signature	tool name format
Claude Code	`.claude/`	`name` + `description` + `tools`	PascalCase
OpenCode	`.opencode/`	`description` + `mode` + `steps` + `permission`	snake_case
Kilo Code	`.kilo/`	same as OpenCode	snake_case
Trae IDE	`.trae/`	`name` + `description` + `tools`	lower + camelCase
CodeBuddy	`.codebuddy/`	`name` + `description` + `model:inherit` + `tools` + `permissionMode`	PascalCase
Qwen Code	`.qwen/`	`name` + `description` + `model` + `tools` + `disallowedTools` + `approvalMode`	PascalCase
Qoder	`.qoder/`	same as Claude	PascalCase

Every platform supports per-role sub-agent models via --model-orchestrator/executor/reviewer. When omitted, sub-agents inherit the main session model.

loop-md-cli solves this with a three-layer abstraction: templates + domains + renderers.

Domain-based Generation

Four built-in domains, each with dedicated templates enforcing its own engineering discipline:

Domain	Agent names	Command	Discipline
`ralph`	ralph-orchestrator / worker / reviewer	ralph-loop / ralph-graph	Kernel paradigm: TaskList-driven + backpressure breaker (loop); DAG routing + active-set execution (graph). Most general; fallback for custom domains without dedicated templates.
`graph`	graph-orchestrator / graph-worker / graph-reviewer	ralph-graph	Graph paradigm: DAG routing with active-set execution. Dynamic mode by default (AI decomposes at runtime); static via `--tasks-file`.
`coding`	coding-orchestrator / builder / reviewer	coding-loop	scope discipline + root-cause grouped fixes + zero tolerance for scope drift
`testing`	test-orchestrator / writer / coverage-reviewer	test-loop	source freeze + three signals (coverage ≥ 80% / mutation ≥ 60% / empty-assertion = 0)
`writing`	writing-orchestrator / author / reviewer	writing-loop	writing boundary + three signals (terminology drift / dead links / code examples)

loop-md-cli --opencode --domain coding
loop-md-cli --opencode --domain testing --dry-run

ralph is the kernel paradigm; coding/testing/writing are domain specializations built on ralph. Backpressure is a general kernel capability: coding/testing use a strong gate (npm test), writing uses a weak gate (npm run lint).

Custom Domains

Pass custom JSON via --domain-file, or drop it in .opencode/domains/<id>.json for auto-discovery:

loop-md-cli --opencode --domain-file ./my-domain.json
loop-md-cli --claude --domain my-domain   # auto-scans .opencode/domains/

The JSON requires engine: { type: "loop" } or engine: { type: "graph" }. The "graph" type additionally requires a tasks array defining the DAG (see below). Each command must specify kind: "entry" and agent.

Graph Engine

The built-in graph domain runs in dynamic mode by default — the generated command instructs the orchestrator to decompose $ARGUMENTS into a DAG at runtime (zero config). For a static routing table, use --tasks-file:

# Dynamic (default): AI decomposes the task at runtime
loop-md-cli --kilo --domain graph

# Static: inject a project-specific DAG from a JSON file
loop-md-cli --kilo --domain graph --tasks-file ./my-tasks.json

The tasks file is a JSON array ([{id, title, depends_on, accept_criteria}]) or {tasks: [...]}. The CLI computes the topological sort and injects a routing table. The example below shows how to define a custom graph domain with your own task DAG (team-shared):

{
  "id": "my-graph",
  "engine": { "type": "graph" },
  "agents": [
    { "role": "orchestrator", "name": "my-orchestrator" },
    { "role": "executor",     "name": "my-worker" },
    { "role": "reviewer",     "name": "my-reviewer" }
  ],
  "commands": [
    { "kind": "entry", "agent": "my-orchestrator", "name": "my-graph" }
  ],
  "tasks": [
    { "id": "t1", "title": "Setup",       "depends_on": [] },
    { "id": "t2", "title": "Build",       "depends_on": ["t1"] },
    { "id": "t3", "title": "Test",        "depends_on": ["t2"] },
    { "id": "t4", "title": "Deploy",      "depends_on": ["t2"] }
  ]
}

The tasks array defines a DAG where each task has id, title, and optional depends_on (array of task IDs). The CLI computes a topological sort and injects a routing_table into the generated command template, enabling the orchestrator to execute independent tasks in parallel.

loop-md-cli --opencode --domain-file ./my-graph-domain.json

loop is the kernel paradigm; coding/testing/writing are domain specializations built on ralph. Backpressure is a general kernel capability: coding/testing use a strong gate (npm test), writing uses a weak gate (npm run lint).

Custom Templates

Three-level Fallback

When rendering <domain>-<role>, lookup proceeds in order:

<domain>-<role>.md (e.g. my-domain-orchestrator.md) — domain-specific
ralph-<role>.md — the most general kernel paradigm
throw (avoid silently generating 0 files)

A custom domain always runs (falling back to ralph); domain specialization requires dedicated templates.

Template Format

---
name: { { name } }
description: { { description } }
mode: subagent
temperature: 0.2
steps: 30
permission:
  edit: allow
  bash:
    "*": allow
---

## Role

You are **{{name}}**, the Loop executor agent.

Placeholders

Placeholder	Source	Applies to
`{{name}}`	agent/command name from the domain definition	agent
`{{description}}`	description from the domain definition	agent + command
`{{agent}}`	orchestrator name the loop binds to	command
`{{backpressure}}`	the domain's backpressure breaker config	orchestrator
`{{engine_type}}`	`engine.type` from the domain definition	agent
`{{executor_name}}`	executor agent name (from domain, by role)	command
`{{reviewer_name}}`	reviewer agent name (from domain, by role)	command
`{{routing_table}}`	DAG topological sort + entry points (graph)	command (graph)

Built-in Domain Templates

Domain	Template files	Discipline
`ralph`	`ralph-orchestrator.md` / `ralph-worker.md` / `ralph-reviewer.md` / `ralph-loop.md` / `ralph-graph.md`	TaskList + backpressure (loop); DAG routing + active-set (graph)
`coding`	`coding-orchestrator.md` / `coding-executor.md` / `coding-reviewer.md` / `coding-loop.md`	scope + root-cause
`testing`	`testing-orchestrator.md` / `testing-executor.md` / `testing-reviewer.md` / `testing-loop.md`	source freeze + signals
`writing`	`writing-orchestrator.md` / `writing-executor.md` / `writing-reviewer.md` / `writing-loop.md`	writing boundary

Supported Platforms

Platform	CLI flag	Output dir	Renderer family
Claude Code	`--claude`	`.claude/`	named
Qoder	`--qoder`	`.qoder/`	named
Qwen Code	`--qwen`	`.qwen/`	qwen
OpenCode	`--opencode`	`.opencode/`	mode
Kilo Code	`--kilo`	`.kilo/`	mode
CodeBuddy	`--codebuddy`	`.codebuddy/`	codebuddy
Trae IDE	`--trae`	`.trae/`	trae

Output Examples

The same ralph domain template produces different frontmatter per renderer family. The following are real outputs of loop-md-cli --all (lightly trimmed for brevity).

named family (Claude Code / Qoder) — .claude/agents/ralph-reviewer.md:

---
name: ralph-reviewer
description: Ralph Loop reviewer. Read-only quality gate emitting machine-routable verdict/issues.
tools: Read, Grep, Glob, Bash
---

mode family (OpenCode / Kilo) — .opencode/agents/ralph-orchestrator.md (trimmed):

---
description: Ralph Loop orchestrator. Maintains the task list, delegates to executor/reviewer, stops on gate decisions.
mode: subagent
temperature: 0.3
steps: 30
permission:
  edit: deny
  bash:
    "*": deny
    "test *": allow
  read: allow
---

codebuddy family — .codebuddy/agents/ralph-orchestrator.md:

---
name: ralph-orchestrator
description: Ralph Loop orchestrator. Maintains the task list, delegates to executor/reviewer.
model: inherit
permissionMode: default
---

trae family — .trae/agents/ralph-reviewer.md:

---
name: ralph-reviewer
description: Ralph Loop reviewer. Read-only quality gate emitting machine-routable verdict/issues.
tools: Read, Grep, Glob
---

qwen family — .qwen/agents/ralph-reviewer.md:

---
name: ralph-reviewer
description: Ralph Loop reviewer. Read-only quality gate emitting machine-routable verdict/issues.
tools: Read, Grep, Glob, Bash
disallowedTools: [Write, Edit]
approvalMode: plan
---

Note the qwen family's disallowedTools: [Write, Edit] is a YAML flow sequence (array) and must not be quoted on render, or it would parse as a string. Only free-text fields (name/description/model) containing dangerous characters like : or # are automatically quoted/escaped.

Full CLI Reference

Option	Short	Description
`--help`	`-h`	Show help
`--version`	`-V` / `-v`	Show version
`--validate`	—	Validate platform config against templates
`--watch`	`-w`	Watch for changes, regenerate automatically
`--incremental`	`-i`	Incremental generation (only rewrite changed files)
`--archive`	`-o`	Export as a ZIP archive
`--all`	`-a`	Generate/validate/export all platforms
`--list`	`-l`	List supported platforms
`--dry-run`	`-n`	Dry run, write nothing
`--domain`	`-d`	Use the given domain (repeatable: `-d ralph -d graph`)
`--domain-file`	`-D`	Path to a custom domain file
`--tasks-file`	`-t`	External tasks DAG file (JSON), injects a static routing table for graph
`--model-orchestrator`	—	Orchestrator sub-agent model (e.g. `--model-orchestrator "DeepSeek-V4-Pro"`)
`--model-executor`	—	Executor sub-agent model
`--model-reviewer`	—	Reviewer sub-agent model