comfyui-mcp 0.52.13 → 0.52.15

package/locales/pt-BR/main.json

@@ -164,7 +164,8 @@
       "minimax": "O agente em segundo plano não está respondendo — o MiniMax não conseguiu iniciar. Defina MINIMAX_API_KEY a partir de platform.minimax.io; depois Desconectar → Conectar para tentar de novo.",
       "moonshot": "O agente em segundo plano não está respondendo — o Moonshot (Kimi K3) não conseguiu iniciar. Defina MOONSHOT_API_KEY a partir de platform.kimi.ai; depois Desconectar → Conectar para tentar de novo.",
       "ollama": "O agente em segundo plano não está respondendo — o Ollama não está acessível. Suba-o com `ollama serve` e baixe o nosso modelo ajustado (`ollama pull artokun/gemma4-comfyui-mcp:e4b` — gemma4 treinado no conjunto de ferramentas do comfyui-mcp, o melhor modelo local da arena; `:12b` para cerca de 8 GB de VRAM); depois Desconectar → Conectar para tentar de novo.",
-      "pi": "O agente em segundo plano não está respondendo — a CLI do pi não conseguiu executar `pi --list-models`. Instale a partir de https://pi.dev (`curl -fsSL https://pi.dev/install.sh | sh`), configure um provedor (defina uma chave de API de provedor ou rode `pi` uma vez e `/login`); depois Desconectar → Conectar para tentar de novo."
+      "pi": "O agente em segundo plano não está respondendo — a CLI do pi não conseguiu executar `pi --list-models`. Instale a partir de https://pi.dev (`curl -fsSL https://pi.dev/install.sh | sh`), configure um provedor (defina uma chave de API de provedor ou rode `pi` uma vez e `/login`); depois Desconectar → Conectar para tentar de novo.",
+      "qwen": "O agente em segundo plano não está respondendo — a CLI do Qwen Code não pôde ser iniciada. Verifique se o Qwen Code está instalado (npm i -g @qwen-code/qwen-code) e conectado (rode `qwen` uma vez e conclua o /auth, ou defina DASHSCOPE_API_KEY); depois Desconectar → Conectar para tentar de novo."
     },
     "llamacpp_no_jinja": "Seu llama-server está rodando SEM `--jinja`, então a chamada de ferramentas está desativada — qualquer ação do agente falharia. Reinicie-o com suporte a ferramentas: `llama-server -m <model>.gguf --jinja -c 16384` (builds atuais habilitam jinja por padrão; os mais antigos precisam da flag); depois Desconectar → Conectar.",
     "llamacpp_small_context": "O contexto do seu llama-server é de {tokens} tokens (flag de inicialização -c). A carga de ferramentas do agente precisa de pelo menos 16384 — abaixo disso, turnos longos são truncados silenciosamente. Considere reiniciar com `-c 16384` ou mais.",

package/locales/ru/main.json

@@ -164,7 +164,8 @@
       "minimax": "Фоновый агент не отвечает — не удалось запустить MiniMax. Задайте MINIMAX_API_KEY с platform.minimax.io, затем «Отключиться» → «Подключиться», чтобы повторить.",
       "moonshot": "Фоновый агент не отвечает — не удалось запустить Moonshot (Kimi K3). Задайте MOONSHOT_API_KEY с platform.kimi.ai, затем «Отключиться» → «Подключиться», чтобы повторить.",
       "ollama": "Фоновый агент не отвечает — Ollama недоступна. Запустите её командой `ollama serve` и загрузите нашу дообученную модель (`ollama pull artokun/gemma4-comfyui-mcp:e4b` — gemma4, обученная на наборе инструментов comfyui-mcp, лучшая локальная модель по результатам арены; `:12b` для примерно 8 ГБ видеопамяти), затем «Отключиться» → «Подключиться», чтобы повторить.",
-      "pi": "Фоновый агент не отвечает — pi CLI не смог выполнить `pi --list-models`. Установите его с https://pi.dev (`curl -fsSL https://pi.dev/install.sh | sh`), настройте провайдера (задайте ключ API провайдера или запустите `pi` один раз и выполните `/login`), затем «Отключиться» → «Подключиться», чтобы повторить."
+      "pi": "Фоновый агент не отвечает — pi CLI не смог выполнить `pi --list-models`. Установите его с https://pi.dev (`curl -fsSL https://pi.dev/install.sh | sh`), настройте провайдера (задайте ключ API провайдера или запустите `pi` один раз и выполните `/login`), затем «Отключиться» → «Подключиться», чтобы повторить.",
+      "qwen": "Фоновый агент не отвечает — Qwen Code CLI не удалось запустить. Убедитесь, что Qwen Code установлен (npm i -g @qwen-code/qwen-code) и выполнен вход (запустите `qwen` один раз и завершите /auth или задайте DASHSCOPE_API_KEY), затем «Отключиться» → «Подключиться», чтобы повторить."
     },
     "llamacpp_no_jinja": "Ваш llama-server запущен БЕЗ `--jinja`, поэтому вызов инструментов отключён — любое действие агента завершится ошибкой. Перезапустите его с поддержкой инструментов: `llama-server -m <model>.gguf --jinja -c 16384` (в свежих сборках jinja включён по умолчанию, старым нужен флаг), затем «Отключиться» → «Подключиться».",
     "llamacpp_small_context": "Контекст вашего llama-server составляет {tokens} токенов (флаг запуска -c). Инструментам агента нужно не меньше 16384 — ниже этого длинные ходы будут молча обрезаться. Стоит перезапустить с `-c 16384` или больше.",

package/locales/tr/main.json

@@ -164,7 +164,8 @@
       "minimax": "Arka plandaki aracı yanıt vermiyor — MiniMax başlatılamadı. platform.minimax.io adresinden aldığınız MINIMAX_API_KEY değerini ayarlayın, sonra Bağlantıyı kes → Bağlan ile yeniden deneyin.",
       "moonshot": "Arka plandaki aracı yanıt vermiyor — Moonshot (Kimi K3) başlatılamadı. platform.kimi.ai adresinden aldığınız MOONSHOT_API_KEY değerini ayarlayın, sonra Bağlantıyı kes → Bağlan ile yeniden deneyin.",
       "ollama": "Arka plandaki aracı yanıt vermiyor — Ollama'ya erişilemiyor. `ollama serve` ile başlatın ve ince ayarlı modelimizi indirin (`ollama pull artokun/gemma4-comfyui-mcp:e4b` — comfyui-mcp araç takımıyla eğitilmiş gemma4, arenadaki en iyi yerel model; yaklaşık 8 GB VRAM için `:12b`), sonra Bağlantıyı kes → Bağlan ile yeniden deneyin.",
-      "pi": "Arka plandaki aracı yanıt vermiyor — pi CLI `pi --list-models` komutunu çalıştıramadı. https://pi.dev adresinden kurun (`curl -fsSL https://pi.dev/install.sh | sh`), bir sağlayıcı yapılandırın (bir sağlayıcı API anahtarı ayarlayın ya da `pi` komutunu bir kez çalıştırıp `/login` yapın), sonra Bağlantıyı kes → Bağlan ile yeniden deneyin."
+      "pi": "Arka plandaki aracı yanıt vermiyor — pi CLI `pi --list-models` komutunu çalıştıramadı. https://pi.dev adresinden kurun (`curl -fsSL https://pi.dev/install.sh | sh`), bir sağlayıcı yapılandırın (bir sağlayıcı API anahtarı ayarlayın ya da `pi` komutunu bir kez çalıştırıp `/login` yapın), sonra Bağlantıyı kes → Bağlan ile yeniden deneyin.",
+      "qwen": "Arka plandaki aracı yanıt vermiyor — Qwen Code CLI başlatılamadı. Qwen Code'un kurulu (npm i -g @qwen-code/qwen-code) ve oturum açılmış olduğundan emin olun (`qwen` komutunu bir kez çalıştırıp /auth adımını tamamlayın ya da DASHSCOPE_API_KEY ayarlayın), sonra Bağlantıyı kes → Bağlan ile yeniden deneyin."
     },
     "llamacpp_no_jinja": "llama-server sunucunuz `--jinja` OLMADAN çalışıyor, bu yüzden araç çağırma devre dışı — aracının her eylemi başarısız olur. Araç desteğiyle yeniden başlatın: `llama-server -m <model>.gguf --jinja -c 16384` (güncel sürümlerde jinja varsayılan olarak açıktır, eski sürümlerde bu bayrak gerekir), sonra Bağlantıyı kes → Bağlan.",
     "llamacpp_small_context": "llama-server bağlamınız {tokens} belirteç (başlatma bayrağı -c). Aracının araç yükü için en az 16384 gerekir; bunun altında uzun turlar sessizce kırpılır. `-c 16384` veya daha yükseğiyle yeniden başlatmayı düşünün.",

package/locales/zh/main.json

@@ -164,7 +164,8 @@
       "minimax": "后台智能体没有响应 —— MiniMax 无法启动。请设置在 platform.minimax.io 获取的 MINIMAX_API_KEY ，然后依次“断开连接”→“连接”重试。",
       "moonshot": "后台智能体没有响应 —— Moonshot（Kimi K3）无法启动。请设置在 platform.kimi.ai 获取的 MOONSHOT_API_KEY ，然后依次“断开连接”→“连接”重试。",
       "ollama": "后台智能体没有响应 —— 连不上 Ollama 。请用 `ollama serve` 启动，并拉取我们的微调模型（`ollama pull artokun/gemma4-comfyui-mcp:e4b` —— 基于 comfyui-mcp 工具套件训练的 gemma4，竞技场里表现最好的本地模型；显存约 8 GB 可用 `:12b`），然后依次“断开连接”→“连接”重试。",
-      "pi": "后台智能体没有响应 —— pi CLI 无法运行 `pi --list-models` 。请从 https://pi.dev 安装（`curl -fsSL https://pi.dev/install.sh | sh`），配置一个服务商（设置服务商 API 密钥，或运行一次 `pi` 再执行 `/login`），然后依次“断开连接”→“连接”重试。"
+      "pi": "后台智能体没有响应 —— pi CLI 无法运行 `pi --list-models` 。请从 https://pi.dev 安装（`curl -fsSL https://pi.dev/install.sh | sh`），配置一个服务商（设置服务商 API 密钥，或运行一次 `pi` 再执行 `/login`），然后依次“断开连接”→“连接”重试。",
+      "qwen": "后台智能体没有响应 —— Qwen Code CLI 无法启动。请确认已安装 Qwen Code（npm i -g @qwen-code/qwen-code）并已登录（运行一次 `qwen` 并完成 /auth，或设置 DASHSCOPE_API_KEY），然后依次“断开连接”→“连接”重试。"
     },
     "llamacpp_no_jinja": "你的 llama-server 没有带 `--jinja` 运行，因此工具调用被禁用 —— 智能体的每个动作都会失败。请带上工具支持重新启动：`llama-server -m <model>.gguf --jinja -c 16384` （较新的版本默认启用 jinja，旧版本需要显式加这个参数），然后依次“断开连接”→“连接”。",
     "llamacpp_small_context": "你的 llama-server 上下文为 {tokens} 个 token（启动参数 -c）。智能体的工具负载需要不少于 16384，低于这个值时较长的对话会被悄悄截断。建议用 `-c 16384` 或更高重新启动。",

package/locales/zh-TW/main.json

@@ -164,7 +164,8 @@
       "minimax": "背景代理程式沒有回應 —— MiniMax 無法啟動。請設定在 platform.minimax.io 取得的 MINIMAX_API_KEY ，然後依序「中斷連線」→「連線」重試。",
       "moonshot": "背景代理程式沒有回應 —— Moonshot（Kimi K3）無法啟動。請設定在 platform.kimi.ai 取得的 MOONSHOT_API_KEY ，然後依序「中斷連線」→「連線」重試。",
       "ollama": "背景代理程式沒有回應 —— 連不上 Ollama 。請用 `ollama serve` 啟動，並拉取我們的微調模型（`ollama pull artokun/gemma4-comfyui-mcp:e4b` —— 以 comfyui-mcp 工具套件訓練的 gemma4，競技場中表現最好的本機模型；顯示記憶體約 8 GB 可用 `:12b`），然後依序「中斷連線」→「連線」重試。",
-      "pi": "背景代理程式沒有回應 —— pi CLI 無法執行 `pi --list-models` 。請從 https://pi.dev 安裝（`curl -fsSL https://pi.dev/install.sh | sh`），設定一個供應商（設定供應商 API 金鑰，或執行一次 `pi` 再執行 `/login`），然後依序「中斷連線」→「連線」重試。"
+      "pi": "背景代理程式沒有回應 —— pi CLI 無法執行 `pi --list-models` 。請從 https://pi.dev 安裝（`curl -fsSL https://pi.dev/install.sh | sh`），設定一個供應商（設定供應商 API 金鑰，或執行一次 `pi` 再執行 `/login`），然後依序「中斷連線」→「連線」重試。",
+      "qwen": "背景代理程式沒有回應 —— Qwen Code CLI 無法啟動。請確認已安裝 Qwen Code（npm i -g @qwen-code/qwen-code）並已登入（執行一次 `qwen` 並完成 /auth，或設定 DASHSCOPE_API_KEY），然後依序「中斷連線」→「連線」重試。"
     },
     "llamacpp_no_jinja": "你的 llama-server 沒有帶 `--jinja` 執行，因此工具呼叫被停用 —— 代理程式的每個動作都會失敗。請帶上工具支援重新啟動：`llama-server -m <model>.gguf --jinja -c 16384` （較新的版本預設啟用 jinja，較舊的版本需要明確加上這個旗標），然後依序「中斷連線」→「連線」。",
     "llamacpp_small_context": "你的 llama-server 上下文為 {tokens} 個 token（啟動旗標 -c）。代理程式的工具內容需要至少 16384，低於這個值時較長的回合會被無聲截斷。建議用 `-c 16384` 或更高重新啟動。",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "comfyui-mcp",
-  "version": "0.52.13",
+  "version": "0.52.15",
   "mcpName": "io.github.artokun/comfyui-mcp",
   "description": "Local-first, agent-native control plane for ComfyUI — MCP server + autonomous sidebar agent that drives your live graph in natural language on ANY LLM: Claude/ChatGPT/Gemini on your subscription (no API key), free local models via Ollama (fully offline), or any hosted model via an OpenAI-compatible endpoint (DeepSeek, GLM, MiMo, OpenRouter). Generate images, video & audio, author and run workflows, manage models and custom nodes; a compact tool-router mode keeps even 4B models effective, and a built-in LLM Arena benchmarks them on real ComfyUI tasks. Also a Claude Code plugin with skills, slash commands, and installer packs. Local, LAN, VPS, or Comfy Cloud.",
   "homepage": "https://comfyui-mcp.artokun.io/docs",

package/plugin/.mcp.json

@@ -1,9 +1,8 @@
 {
   "comfyui": {
-    "command": "npx",
+    "command": "node",
     "args": [
-      "-y",
-      "comfyui-mcp",
+      "${CLAUDE_PLUGIN_ROOT}/scripts/launch-server.mjs",
       "--full"
     ],
     "env": {

package/plugin/scripts/launch-server.mjs

@@ -0,0 +1,130 @@
+#!/usr/bin/env node
+/**
+ * #1447 — plugin MCP server launcher.
+ *
+ * The plugin's .mcp.json used to be `npx -y comfyui-mcp --full` directly. On a
+ * cold npx cache that downloads the whole ~900 MB dependency tree INSIDE the
+ * client's MCP handshake timeout; the client kills the attempt, nothing is
+ * persisted, and every retry pays the full cost again (the measured `_npx`
+ * cache timestamped to a manual run, never to a client attempt).
+ *
+ * This wrapper is the warm path: when `comfyui-mcp` is installed globally
+ * (`npm install -g comfyui-mcp`), it resolves `<npm root -g>/comfyui-mcp/
+ * dist/index.js` and runs it with this same node — sub-second start, no
+ * registry round-trip, and a killed handshake cannot discard an install that
+ * already happened. With no global install it falls back to the original
+ * `npx -y comfyui-mcp` behaviour, so the plugin works identically for anyone
+ * who never installs globally.
+ *
+ * STDIO IS THE MCP TRANSPORT. Everything the child says on stdout/stderr is
+ * the protocol — the wrapper inherits stdio verbatim and must never write to
+ * stdout itself. Diagnostics (there is exactly one, a spawn failure) go to
+ * stderr via console.error.
+ *
+ * Import-safe: nothing here runs on import. The resolution decisions are pure
+ * exports so the tests can call the real thing instead of reimplementing it
+ * (the #1385 lesson — a test of a copy is a test of nothing).
+ */
+
+import { execFile, spawn } from "node:child_process";
+import { existsSync } from "node:fs";
+import { join } from "node:path";
+import { pathToFileURL } from "node:url";
+
+/** `npm root -g` is normally ~100 ms; 5 s bounds a wedged npm without hanging the handshake. */
+const NPM_ROOT_TIMEOUT_MS = 5000;
+
+/**
+ * `npm root -g` stdout → the global install's entry point, or null when the
+ * output is unusable or the install is absent/incomplete (no dist/index.js).
+ * null is not an error here — it means "take the npx path".
+ */
+export function globalEntry(npmRootStdout, { exists = existsSync } = {}) {
+  const root = typeof npmRootStdout === "string" ? npmRootStdout.trim() : "";
+  if (!root) return null;
+  const entry = join(root, "comfyui-mcp", "dist", "index.js");
+  return exists(entry) ? entry : null;
+}
+
+/**
+ * Tokens the shell form may contain. The fallback's args come from the
+ * plugin's own .mcp.json (today: `--full`), never from user input — but a
+ * shell command is string concatenation, so anything outside this alphabet
+ * (whitespace, quotes, metacharacters) is REFUSED rather than passed through
+ * to be silently misparsed.
+ */
+const SHELL_SAFE_TOKEN = /^[A-Za-z0-9@._~:/=-]+$/;
+
+/**
+ * The spawn spec for the resolved server. `extraArgs` are the plugin's own
+ * args from .mcp.json (`--full`), forwarded to whichever server runs so the
+ * manifest stays the declarative source of how the server is started.
+ *
+ * With an entry point we spawn node directly — no shell, no shim resolution.
+ *
+ * The npx fallback differs by platform:
+ *  - POSIX: spawn npx with an args array, no shell.
+ *  - Windows: npx is npx.cmd, which Node refuses to spawn without a shell
+ *    since the 18.20.2/20.12.2 bat-file fix — but `shell` + an args array
+ *    triggers DEP0190 (args are concatenated UNESCAPED, and Node warns every
+ *    launch). So the Windows spec is one pre-validated command string instead:
+ *    same concatenation, but every token is checked against SHELL_SAFE_TOKEN
+ *    first, so what the shell parses is exactly what was written.
+ */
+export function serverSpec(entry, extraArgs, { platform = process.platform, node = process.execPath } = {}) {
+  if (entry) return { command: node, args: [entry, ...extraArgs], shell: false };
+  const npxArgs = ["-y", "comfyui-mcp", ...extraArgs];
+  if (platform !== "win32") return { command: "npx", args: npxArgs, shell: false };
+  for (const token of ["npx", ...npxArgs]) {
+    if (!SHELL_SAFE_TOKEN.test(token)) {
+      throw new Error(`[comfyui-mcp launcher] refusing to pass unsafe shell token to npx: ${JSON.stringify(token)}`);
+    }
+  }
+  return { command: ["npx", ...npxArgs].join(" "), args: [], shell: true };
+}
+
+/** Ask npm where global packages live. String-command form on Windows for the
+ *  same DEP0190 reason as serverSpec; the command is a fixed literal, so there
+ *  is nothing to inject. Resolves null on any failure — the npx fallback is
+ *  the original behaviour, so a failed probe degrades to exactly what the
+ *  plugin did before this wrapper existed. */
+function probeGlobalRoot() {
+  const isWin = process.platform === "win32";
+  return new Promise((resolve) => {
+    execFile(
+      isWin ? "npm root -g" : "npm",
+      isWin ? [] : ["root", "-g"],
+      { shell: isWin, timeout: NPM_ROOT_TIMEOUT_MS },
+      (error, stdout) => resolve(error ? null : stdout),
+    );
+  });
+}
+
+function run(spec) {
+  const child = spawn(spec.command, spec.args, { stdio: "inherit", shell: spec.shell });
+
+  // The client kills the WRAPPER when it shuts the server down; without
+  // forwarding, the real server would be orphaned holding stdio. Killing the
+  // child on signal makes its `exit` fire, which is what exits the wrapper.
+  for (const sig of ["SIGINT", "SIGTERM", "SIGHUP"]) {
+    process.on(sig, () => child.kill(sig));
+  }
+
+  child.on("exit", (code, signal) => {
+    // A signal death has no code; 1 keeps the client from reading it as clean.
+    process.exit(code ?? (signal ? 1 : 0));
+  });
+  child.on("error", (err) => {
+    // Honest failure, loudly: a server that never started must not look like
+    // one that exited cleanly. stderr, never stdout — see the header.
+    console.error(`[comfyui-mcp launcher] failed to start ${spec.command}: ${err.message}`);
+    process.exit(1);
+  });
+}
+
+const isMain = process.argv[1] && import.meta.url === pathToFileURL(process.argv[1]).href;
+if (isMain) {
+  const extraArgs = process.argv.slice(2);
+  const entry = globalEntry(await probeGlobalRoot());
+  run(serverSpec(entry, extraArgs));
+}

package/plugin/skills/minimax-h3-video/SKILL.md

@@ -0,0 +1,270 @@
+---
+name: minimax-h3-video
+description: Build MiniMax H3 (Hailuo) local video workflows — native T2V/I2V/R2V nodes, Comfy-Org INT8 weights, turbo LoRAs for 8GB VRAM, 15-second stereo-audio clips, and the official MiniMax prompting guides (cite by link, do not copy).
+globs:
+  - "**/*.json"
+---
+
+# MiniMax H3 (Hailuo) — local video
+
+This skill teaches the **local-weights** MiniMax H3 path in ComfyUI. It is the
+pilot for `#1155` (Official vs Empirical sources) because MiniMax publishes a
+real prompting guide. **Cite that guide by URL. Do not copy it into this repo.**
+
+## Two products, two cost models — pick one
+
+They share a brand and **must not be mixed**.
+
+| Path | Nodes | Cost | VRAM | When |
+|---|---|---|---|---|
+| **Local weights** (this skill) | `MiniMaxH3ImageToVideo`, `MiniMaxH3ReferenceToVideo`, `EmptyMiniMaxH3LatentAV`, `MiniMaxH3SigmaShift`, `MiniMaxH3MemoryEfficientSageAttentionPatch` | Free after download | Yes — INT8 + turbo LoRA is the 8 GB story | User wants 4–15 s stereo clips on their GPU |
+| **Partner API** | `MinimaxHailuo03TextToVideoNode`, `MinimaxHailuo03FirstLastFrameNode`, `MinimaxHailuo03ReferenceNode`, `MinimaxTextToVideoNode`, `MinimaxImageToVideoNode`, `MinimaxHailuoVideoNode` | Paid per generation | None | User has a MiniMax / Hailuo API key and does not want local weights |
+
+API nodes do not take `MiniMaxH3SigmaShift` or Sage-attention patches. Local
+nodes do not spend API credits. If the user asked for Hailuo *cloud*, stop and
+use the API nodes + their key; do not download 40 GB of weights.
+
+`MiniMaxH3Director` is a **third-party** pack (`muse-collective-26/MiniMaxH3-Director`),
+not core. Do not require it for T2V / I2V / R2V.
+
+## License — cite, do not copy
+
+Local weights and MiniMax's own documentation sit under the
+[MiniMax H3 Community License](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE).
+`Materials` includes the Documentation. The agreement's Applicable Territory
+**excludes the United States, the EU, the UK, and South Korea**. This skill
+does **not** reproduce MiniMax's `skills/h3-prompt-writing/` SKILL.md or the
+prompting-guide prose. Linking to a public URL is the `#1155` requirement.
+
+This is not legal advice. Tell a US/EU/UK/KR user that the *local* path is
+territory-restricted and that the **paid API** is a separate product under
+MiniMax platform terms.
+
+## Prefer the Comfy-Org template over hand-wiring
+
+ComfyUI ≥ **0.30.0** (templates in the 0.33 line). These are **core**
+`comfyui-workflow-templates` graphs in the frontend **Template Library →
+Video**, not installer packs and not custom-node `example_workflows`:
+
+| Mode | Template Library card | File | Diffusion file |
+|---|---|---|---|
+| T2V / I2V / FL2VA | MiniMax H3: Text to Video / Image to Video | `video_minimax_h3_t2v.json` / `video_minimax_h3_i2v.json` | `minimax_h3_fl2va_pruned_int8_convrot.safetensors` |
+| R2V (omni-reference) | MiniMax H3: Reference to Video | `video_minimax_h3_r2v.json` | `minimax_h3_ref2va_pruned_int8_convrot.safetensors` |
+
+`list_packs action:"list_templates"` will **not** list them.
+`enqueue_workflow action:"run_template"` will **not** resolve
+`video_minimax_h3_t2v` / `_i2v` / `_r2v` — that action only loads bundled
+installer packs, and there is no `packs/minimax-h3-*` yet. Do not call it
+until a pack exists. `panel_load_workflow` needs `pack:`, a disk `path:`, or
+an inline UI `graph` — a Template Library basename is none of those.
+
+**Load path that actually works:**
+
+1. **Preferred.** Ask the user to open **Template Library → Video → MiniMax H3:
+   Text to Video** (or Image to Video / Reference to Video). Pick the local
+   `video_minimax_h3_*` cards, **not** the `api_minimax_h3_*` paid partner
+   templates.
+2. **Agent, no UI click.** Fetch the UI JSON from
+   https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_t2v.json
+   (or `_i2v` / `_r2v`; raw.githubusercontent.com is the same files), save it
+   with `save_workflow action:"save"` `filename:"video_minimax_h3_t2v.json"`,
+   then `panel_load_workflow path:"video_minimax_h3_t2v.json"`. Same pattern as
+   `video-extend` (stage on disk, then `path:`). Do not pass the GitHub URL as
+   `path:` or `pack:`.
+
+After it lands, retarget the **subgraph's exposed widgets** (prompt, duration,
+`turbo_mode`, megapixels). Official T2V/I2V graphs wrap
+`MiniMaxH3ImageToVideo` inside a subgraph (`type` is a UUID). Do not flatten
+that interior unless you are hand-building.
+
+Hand-building the subgraph is slower and easy to get wrong.
+
+Comfy tutorial (wiring, not MiniMax's prompt formula):
+https://docs.comfy.org/tutorials/video/minimax/minimax-h3
+
+## Models (Comfy-Org INT8 pack)
+
+All from `huggingface.co/Comfy-Org/MiniMax-H3`. Download with
+`download_model` `action:"download"`.
+
+| File | Folder | Role |
+|---|---|---|
+| `minimax_h3_fl2va_pruned_int8_convrot.safetensors` | `diffusion_models/` | T2V / I2V / first-last-frame |
+| `minimax_h3_ref2va_pruned_int8_convrot.safetensors` | `diffusion_models/` | R2V only — **different UNet** |
+| `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors` | `text_encoders/` | Qwen3-VL-32B encoder, `CLIPLoader` **type=`minimax`** |
+| `minimax_h3_video_vae_fp16.safetensors` | `vae/` | Visual VAE |
+| `minimax_h3_audio_vae_fp32.safetensors` | `vae/` | Stereo audio VAE (32 kHz) |
+
+### Turbo LoRAs (4–8 steps instead of ~20)
+
+The Comfy-Org T2V template already switches these on with `turbo_mode`:
+
+| Steps | File | Source |
+|---|---|---|
+| 8 | `minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors` | `lightx2v/Minimax-h3-Turbo` |
+| 4 | `minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors` | `Comfy-Org/MiniMax-H3` `loras/` |
+
+Kijai conversions live at `Kijai/MiniMax-H3_comfy` (`loras/`) and experimental
+W4A8 at `Kijai/MiniMax-H3-experimental`. Same job (low-step / low-VRAM). Prefer
+the Comfy-Org / lightx2v filenames the template already names; only switch to a
+Kijai file if that is what is on disk.
+
+4-step is faster and softer; **6–8 steps** is the usual sharpness compromise.
+
+## Output spec
+
+| Knob | Value |
+|---|---|
+| Duration | 4–15 seconds |
+| Frame rate | **24 fps** (`CreateVideo.fps`) |
+| Audio | Native stereo, decoded by the audio VAE, muxed in `CreateVideo` |
+| Short edge | 768 px native; cap **768×1344**, multiple of **32** |
+| Preview size | `ResolutionSelector` megapixels **0.4** → 864×480 at 16:9 |
+| Full 768p | megapixels **~0.98** → **1344×768** at 16:9 |
+
+Duration → frame `length` (Comfy-Org template math, 17-frame blocks):
+
+```
+max(5, round(seconds * 24)) + (5 - (max(5, round(seconds * 24)) % 17)) % 17
+```
+
+That is the `17k+5` grid. Do not invent a WAN-style `4n+1` length.
+
+## Node graph (local T2V / I2V)
+
+From the Comfy-Org T2V subgraph (core nodes, not the Markdown notes):
+
+```
+ResolutionSelector (aspect, megapixels, multiple=32) → width, height
+
+UNETLoader (fl2va int8)
+  ├─ LoraLoaderModelOnly (turbo LoRA) ─┐
+  └────────────────────────────────────┤ ComfySwitchNode (turbo_mode)
+                                       ▼
+                         BasicGuider + BasicScheduler + KSamplerSelect(res_multistep)
+                                       ▼
+CLIPLoader (type=minimax, qwen3vl 32b) → MiniMaxH3ImageToVideo
+VAELoader (video vae) ─────────────────→   prompt, width, height, length
+optional first_frame / last_frame ─────→   → CONDITIONING + LATENT
+                                       ▼
+                         SamplerCustomAdvanced → LATENT
+                                       ├─ VAEDecode (video vae) → IMAGE
+                                       └─ VAEDecodeAudio (audio vae) → AUDIO
+                                       ▼
+                         CreateVideo (fps=24) → SaveVideo
+```
+
+`MiniMaxH3ImageToVideo` **is** T2V when both image sockets are empty, I2V with
+`first_frame`, FL2VA with both frames. Do not add a second T2V-only node.
+
+R2V replaces the UNet with **ref2va** and the conditioner with
+`MiniMaxH3ReferenceToVideo`. Do not load fl2va into an R2V graph.
+
+### Local-only helpers
+
+| Node | Role |
+|---|---|
+| `EmptyMiniMaxH3LatentAV` | Empty audio-video latent when you are not using `MiniMaxH3ImageToVideo`'s built-in latent |
+| `MiniMaxH3SigmaShift` | Flow-matching shift on the local UNet |
+| `MiniMaxH3MemoryEfficientSageAttentionPatch` | Core Sage patch; or KJNodes `Patch Sage Attention KJ` (`sage_attention=auto`) between `UNETLoader` and `BasicGuider` |
+
+Sage roughly doubles speed. H3 has mixed dtypes — console lines about falling
+back to pytorch attention on some layers are expected.
+
+## Sampler defaults (Comfy-Org template)
+
+| Mode | Sampler | Scheduler | Steps |
+|---|---|---|---|
+| Base (no turbo) | `res_multistep` | `simple` | **20** |
+| Turbo on | `res_multistep` | `simple` | **4–8** (template default turbo steps widget) |
+
+Guider is `BasicGuider` (CFG-distilled checkpoint — do not crank CFG). Seed via
+`RandomNoise`.
+
+## Prompting — read the vendor guide, do not paste it here
+
+Write the prompt **in the MiniMax H3 node**, not a generic `CLIPTextEncode`.
+
+**Official MiniMax guides** (read these; do not copy them into graphs as a
+system prompt dump):
+
+- T2VA / I2VA / FL2VA / L2VA:
+  https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md
+- Full-reference / R2V:
+  https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md
+- Vendor skill (install separately if the user wants it — we do not bundle it):
+  https://github.com/MiniMax-AI/MiniMax-H3 (`npx skills add … --skill h3-prompt-writing`)
+
+H3-Context-IR (the hosted prompt rewriter) is **not** in the open weights.
+Local ComfyUI has no IR node. Either write the structured prompt yourself from
+the guide, or call MiniMax's Context-IR API and paste `content.prompt` into the
+local node.
+
+Comfy-Org's own template notes (safe to follow, not MiniMax docs):
+
+1. One block covering **look, scene, timed shots, camera, and audio** (dialogue,
+   SFX, score).
+2. Time shots (`[0s-1.5s] Shot 1: …`).
+3. R2V: name each input in connection order (`<Picture 1>`, `<Video 1>`,
+   `<Audio 1>`) and say what job each one does (identity, motion, voice).
+4. R2V caps (vendor model card, not a guess): ≤9 images, ≤3 videos, ≤3 audio
+   clips, ≤12 files mixed; each AV clip 2–15 s.
+
+## 15-second clips and chaining
+
+One H3 shot is **at most ~15 s**. Longer pieces are concatenated clips, not a
+bigger `length`.
+
+1. Generate clip N (up to 15 s).
+2. Confirm the file with `get_image` `action:"list_outputs"` (`kind:"video"`) —
+   video nodes often skip `/history`.
+3. Stage the last frame (or the whole clip) with `upload_image` `action:"stage"`.
+4. Clip N+1: `MiniMaxH3ImageToVideo.first_frame` = last frame of N, **or** R2V
+   with `<Video 1>` as a continuation reference.
+5. Concat with ffmpeg (`director` skill) or an editor.
+
+This is **not** WAN Pusa (`video-extend`). Pusa LoRAs and `flowmatch_pusa` do
+not apply to H3.
+
+## VRAM
+
+| Card | Practical setup |
+|---|---|
+| **24 GB+** | INT8 fl2va + Qwen3-VL + both VAEs; 1344×768; 10–15 s; Sage optional |
+| **12–16 GB** | Same INT8 pack; drop megapixels toward 0.4–0.6; turbo LoRA on; Sage |
+| **8 GB** | INT8 + turbo LoRA + Sage + short preview (0.2–0.4 MP, 4–6 s). Minutes per clip. Kijai W4A8 if INT8 still OOMs. |
+
+Always `clear_vram` before switching to H3 from WAN / LTX / a checkpoint.
+
+## Gotchas
+
+- **`CLIPLoader` type must be `minimax`.** `qwen_image` / `flux` will load the
+  wrong encoder layout.
+- **fl2va vs ref2va.** T2V/I2V templates on ref2va (or R2V on fl2va) are garbage
+  or a load error.
+- **Turbo off, 4 steps.** The switch defaults off and base steps are 20. Four
+  steps without the LoRA is mush.
+- **API node in a local graph.** Costs money and ignores the UNet you downloaded.
+- **WAN frame math.** H3 is 24 fps and `17k+5`, not 16 fps `4n+1`.
+- **Verify video on disk**, then stage — never guess `input/` paths.
+- **ffmpeg** is required for `CreateVideo` / `SaveVideo` / `VHS_VideoCombine`.
+- Desktop/Cloud ComfyUI lags nightly. Missing `MiniMaxH3*` nodes → update to
+  ≥0.30.0 (0.33 templates) before hunting custom packs.
+
+## See also
+
+- `video-extend` — WAN Pusa temporal continuation (different family)
+- `director` — multi-clip concat after you have 15 s H3 shots
+- `prompt-engineering` — generic CLIP syntax; **H3 does not use it**
+- `triton-sageattention` — installing Sage on Windows
+
+No bundled `packs/minimax-h3-*` installer yet — that is why
+`enqueue_workflow action:"run_template"` cannot load these graphs. Use the
+Template Library (or the GitHub fetch → `save_workflow` →
+`panel_load_workflow path:` path above) + `download_model` against
+`Comfy-Org/MiniMax-H3`.
+
+## Sources
+
+- **Official:** MiniMax prompting guides https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md (T2VA/I2VA/FL2VA/L2VA) and https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md (full-reference / R2V); vendor repo https://github.com/MiniMax-AI/MiniMax-H3; ComfyUI tutorial + templates https://docs.comfy.org/tutorials/video/minimax/minimax-h3 (wiring, duration grid, INT8 filenames). MiniMax's own `skills/h3-prompt-writing` is linked, not copied — Community License includes Documentation.
+- **Empirical:** local vs partner-API node split and 8 GB turbo-LoRA note from issue #1167 / the reporter's rig; Sage mixed-dtype fallback from the Comfy tutorial; chaining last-frame→next-clip from observed ComfyUI I/O (stage + list_outputs), not a vendor extender.

package/plugin/skills/model-registry/SKILL.md

@@ -70,6 +70,22 @@ Comfy-Org repackages everything: `huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repac
 | Qwen-Image-Edit 2511 Lightning LoRA | `huggingface.co/lightx2v/Qwen-Image-Edit-2511-Lightning` | `loras/` |
 | Qwen-Image 2512 Turbo LoRA | `huggingface.co/Wuli-art/Qwen-Image-2512-Turbo-LoRA` | `loras/` |
 
+## MiniMax H3 video (see `minimax-h3-video` skill)
+
+Comfy-Org INT8 pack: `huggingface.co/Comfy-Org/MiniMax-H3`. Local weights are
+under the MiniMax H3 Community License (territory-restricted). Prefer the
+filenames the Comfy-Org templates already name.
+
+| File | Source | Target |
+|---|---|---|
+| `minimax_h3_fl2va_pruned_int8_convrot.safetensors` | `huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors` | `diffusion_models/` |
+| `minimax_h3_ref2va_pruned_int8_convrot.safetensors` | same repo, `diffusion_models/` | `diffusion_models/` |
+| `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors` | same repo, `text_encoders/` | `text_encoders/` |
+| `minimax_h3_video_vae_fp16.safetensors` | same repo, `vae/` | `vae/` |
+| `minimax_h3_audio_vae_fp32.safetensors` | same repo, `vae/` | `vae/` |
+| `minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors` | `huggingface.co/lightx2v/Minimax-h3-Turbo/resolve/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors` | `loras/` |
+| `minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors` | `huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors` | `loras/` |
+
 ## Z-Image (see `z-image-txt2img` skill)
 
 | File | Source | Target |

package/plugin/skills/rgthree/SKILL.md

@@ -71,6 +71,28 @@ All of the following are node properties — set them with `panel_set_property`:
 | `showNav` | bool | `true` | Per-row jump-to-group arrow. |
 | `showAllGraphs` | bool | `true` | Include groups that live inside subgraphs. |
 
+### matchTitle does not rebuild the toggle list on the first write
+
+`panel_set_property` stores `matchTitle` (and `matchColors` / `sort` / …) and
+the reply `from`/`to` is truthful. Fast Groups nodes **do not implement
+`onPropertyChanged`**. The toggle list is rebuilt by rgthree's
+`refreshWidgets()` on a service tick (~8 ms after add, then every ~500 ms),
+and leftover-row removal increments the index while splicing — a 22-group
+list can stall at 13 with non-matching `Enable …` rows still present. A
+never-drawn node can also come back as `widgets:{}` — that means the list
+has **not been built yet**, not that there are no matching groups.
+`panel_query_graph` also keys widgets by name, and every toggle is named
+`RGTHREE_TOGGLE_AND_NAV`, so a built list collapses to one key.
+
+**Do this:**
+
+1. Set `matchTitle` **immediately** after `panel_add_node` (before the first
+   unfiltered refresh paints every group).
+2. Re-read with `panel_query_graph {ids:[<id>], fields:'detail'}`.
+3. If `widgets` is empty **or** the canvas still shows `Enable` rows that do
+   not match the regex, **set `matchTitle` again**. Do **not** delete and
+   re-add the node — that is slower and still needs a second set.
+
 ### Recipe — make pipeline stages toggleable
 
 1. `panel_create_group` per stage, with a **prefixed title** (`STAGE 1 — …`) so one
@@ -94,6 +116,9 @@ All of the following are node properties — set them with `panel_set_property`:
    unwired. Add nodes one at a time, not as a parallel batch.
 
 4. `panel_set_property` → `matchTitle` = `^STAGE`, and `sort` = `alphanumeric`.
+   Set them immediately after the add, then re-read the node. If `widgets` is
+   empty or leftover `Enable` rows remain, set `matchTitle` again — do not
+   delete and re-add.
 
 5. Toggle, then verify with `panel_graph_outline` — it tags nodes `[bypass]` / `[mute]`.
 
@@ -182,4 +207,4 @@ Turning a LoRA **off** (`lora_N.on = false`) is usually safer than clearing it a
 ## Sources
 
 - **Official:** https://github.com/rgthree/rgthree-comfy
-- **Empirical:** frontend-only allowlist and properties-not-widgets notes verified against the installed pack and the panel guard.
+- **Empirical:** frontend-only allowlist and properties-not-widgets notes verified against the installed pack and the panel guard. `#1808` matchTitle rebuild: Fast Groups have no `onPropertyChanged`; `refreshWidgets()` leftover removal is `removeWidget(index++)` in `fast_groups_muter.ts`; first unfiltered tick is scheduled from `addFastGroupNode` in `fast_groups_service.ts`.