v2: workflows bootstrap in downloader; Dockerfile adds Claude Code, hf tooling, no comfy autostart

docker/Dockerfile

@@ -1,12 +1,16 @@
 # ComfyUI Qwen-Image-Edit base image (H100 / RunPod)
 #
-# NOTE: the image plx1029/comfyui-qwen was NOT built from this file — it was
-# assembled directly on the pod with `crane append` (no Docker daemon on
-# RunPod). This Dockerfile is the reproducible equivalent: building it on any
-# machine with Docker produces the same setup.
+# NOTE: published plx1029/comfyui-qwen tags were NOT built from this file — they
+# were assembled daemonlessly (v1: crane append on the pod; v2: skopeo + umoci).
+# This Dockerfile is the reproducible equivalent: building it on any machine
+# with Docker produces the same setup.
 #
-#   docker build -t plx1029/comfyui-qwen:v1 docker/
-#   docker push plx1029/comfyui-qwen:v1
+# Build from the REPO ROOT (context needs download_missing_models.sh):
+#   docker build -t plx1029/comfyui-qwen:v2 -f docker/Dockerfile .
+#   docker push plx1029/comfyui-qwen:v2
+#
+# No secrets are baked into this image (it is public). The model-download
+# script prompts for your Hugging Face token at runtime (or reads $HF_TOKEN).
 
 FROM runpod/pytorch:1.0.2-cu1281-torch280-ubuntu2404
 
@@ -23,16 +27,36 @@ RUN git clone https://github.com/rgthree/rgthree-comfy.git /ComfyUI/custom_nodes
     cd /ComfyUI/custom_nodes/rgthree-comfy && git checkout ${RGTHREE_COMMIT} && \
     pip install comfyui-manager==4.2.2
 
+# Hugging Face hub tooling for fast model downloads (hf_transfer is best-effort:
+# hub >= 1.x ships fast Xet downloads built in and dropped the extra)
+RUN pip install -U "huggingface_hub[hf_transfer]" || pip install -U huggingface_hub && \
+    (pip install hf_transfer || true)
+
+# Claude Code CLI, on PATH for all shells, with bypassPermissions as the
+# default permission mode (this is a throwaway pod environment)
+RUN curl -fsSL https://claude.ai/install.sh | bash && \
+    ln -sf /root/.local/bin/claude /usr/local/bin/claude && \
+    mkdir -p /root/.claude && \
+    printf '{\n  "permissions": {\n    "defaultMode": "bypassPermissions"\n  }\n}\n' > /root/.claude/settings.json && \
+    printf '{\n  "bypassPermissionsModeAccepted": true\n}\n' > /root/.claude.json
+ENV PATH="/root/.local/bin:${PATH}"
+
 # Manager config baked in: security_level=weak, network_mode=personal_cloud
-COPY config.ini /ComfyUI/user/__manager/config.ini
-COPY comfy.settings.json /ComfyUI/user/default/comfy.settings.json
-COPY workflows/ /ComfyUI/user/default/workflows/
+COPY docker/config.ini /ComfyUI/user/__manager/config.ini
+COPY docker/comfy.settings.json /ComfyUI/user/default/comfy.settings.json
+COPY docker/workflows/ /ComfyUI/user/default/workflows/
+
+# Model downloader: picks workflows, finds their models in the HF repo
+# (aleph65/ComfyUI), downloads the missing ones into place. Prompts for the
+# HF token; bootstraps /workspace/workflows from the repo if absent.
+COPY download_missing_models.sh /usr/local/bin/download_missing_models.sh
 
 # Startup: `comfy` runs ComfyUI (prefers /workspace/ComfyUI when a volume is
-# attached); start-comfy.sh auto-starts it alongside the stock RunPod services.
-COPY comfy /usr/local/bin/comfy
-COPY start-comfy.sh /start-comfy.sh
-RUN chmod +x /usr/local/bin/comfy /start-comfy.sh
+# attached). ComfyUI is NOT auto-started; run `comfy` yourself, or set
+# COMFY_AUTOSTART=true in the pod env to restore the old behaviour.
+COPY docker/comfy /usr/local/bin/comfy
+COPY docker/start-comfy.sh /start-comfy.sh
+RUN chmod +x /usr/local/bin/comfy /start-comfy.sh /usr/local/bin/download_missing_models.sh
 
 EXPOSE 7865
 CMD ["/start-comfy.sh"]

docker/README.md

@@ -8,36 +8,47 @@ Everything from the manual setup is baked in — **no reinstalling on new pods**
 - comfyui-manager 4.2.2 + all pip deps (65 packages) preinstalled system-wide
 - Manager config baked: `security_level = weak`, `network_mode = personal_cloud`
 - Workflows baked into the workflow browser: `qwen-loras-very-good-nsfw.json`, `qwen-edit-nsfw.json`
-- ComfyUI **auto-starts on pod boot** on port **7865** (log: `/workspace/comfyui.log`)
+- **v2:** `huggingface_hub` + `hf_transfer` preinstalled for fast model downloads
+- **v2:** [Claude Code](https://claude.com/claude-code) CLI preinstalled (`claude`), on PATH,
+  with `bypassPermissions` as the default permission mode
+- **v2:** `download_missing_models.sh` on PATH — pick workflows, it fetches their
+  missing models from `hf.co/aleph65/ComfyUI` into place (prompts for HF token;
+  bootstraps `/workspace/workflows` from the repo if absent)
+- **v2:** ComfyUI **no longer auto-starts** — run `comfy` yourself, or set
+  `COMFY_AUTOSTART=true` in the pod env for the old behaviour (port **7865**,
+  log: `/workspace/comfyui.log`)
 
 Models are NOT in the image (keep them on the network volume, as before).
+No secrets are baked in either (the image is public) — the download script asks
+for your Hugging Face token at runtime, or reads it from `$HF_TOKEN`.
 
 ## Using it on RunPod
 
-1. Create a template with container image `plx1029/comfyui-qwen:v1`.
+1. Create a template with container image `plx1029/comfyui-qwen:v2`.
 2. Expose HTTP port **7865** (plus 8888/22 as usual).
 3. Attach your network volume at `/workspace`.
+4. In a terminal: `download_missing_models.sh` to fetch models, then `comfy` to start.
 
-On boot the image:
+When ComfyUI starts it:
 - uses `/workspace/ComfyUI` if it exists (your current volume — models already there);
-- otherwise seeds a fresh `/workspace/ComfyUI` from the baked-in copy (download models per the main README);
-- SSH / Jupyter / nginx from the stock RunPod start script still work normally.
+- otherwise seeds a fresh `/workspace/ComfyUI` from the baked-in copy;
+- SSH / Jupyter / nginx from the stock RunPod start script work normally either way.
 
-Manual control: `COMFY_AUTOSTART=false` env var disables auto-start; run `comfy` in a
-terminal to start it by hand. `COMFY_PORT` overrides the port.
+`COMFY_PORT` overrides the port. `COMFY_AUTOSTART=true` restores auto-start on boot.
 
 ## How the image was built
 
-No Docker daemon exists on RunPod pods, so the image was assembled *on the pod
-itself* with [`crane`](https://github.com/google/go-containerregistry): the exact
-pip-package delta vs the base image (identified via dist-info RECORDs) plus the
-pinned git clones were tarred into a single OCI layer and appended onto the base
-image manifest — the 20 GB base was never re-uploaded.
+No Docker daemon exists on RunPod pods, so images are assembled daemonlessly:
+v1 with [`crane`](https://github.com/google/go-containerregistry) (pip-delta +
+git clones appended as one OCI layer); v2 with `skopeo` + `umoci` (unpack v1,
+apply changes in a chroot, repack the delta as a new layer) — the 20 GB base is
+never re-uploaded.
 
 `Dockerfile` in this directory is the reproducible equivalent for rebuilding on
-any machine with real Docker:
+any machine with real Docker (build from the repo root — the context needs
+`download_missing_models.sh`):
 
 ```bash
-docker build -t plx1029/comfyui-qwen:v1 /workspace/docker
-docker push plx1029/comfyui-qwen:v1
+docker build -t plx1029/comfyui-qwen:v2 -f docker/Dockerfile .
+docker push plx1029/comfyui-qwen:v2
 ```

docker/start-comfy.sh

@@ -1,9 +1,10 @@
 #!/bin/bash
-# Image CMD: start ComfyUI in the background (auto-restart), then hand off to
-# the stock RunPod /start.sh (SSH, Jupyter, nginx) in the foreground.
-# Set COMFY_AUTOSTART=false in the pod env to disable the auto-start.
+# Image CMD: hand off to the stock RunPod /start.sh (SSH, Jupyter, nginx).
+# ComfyUI does NOT auto-start — run `comfy` yourself when you're ready.
+# Set COMFY_AUTOSTART=true in the pod env to restore the old auto-start
+# behaviour (background loop with auto-restart).
 
-if [[ "${COMFY_AUTOSTART:-true}" != "false" ]]; then
+if [[ "${COMFY_AUTOSTART:-false}" == "true" ]]; then
     LOG=/workspace/comfyui.log
     [[ -d /workspace && -w /workspace ]] || LOG=/var/log/comfyui.log
     (
@@ -14,6 +15,9 @@ if [[ "${COMFY_AUTOSTART:-true}" != "false" ]]; then
         done
     ) &
     echo "ComfyUI starting in background (log: $LOG, port ${COMFY_PORT:-7865})"
+else
+    echo "ComfyUI not auto-started. Run 'comfy' to start it (port ${COMFY_PORT:-7865})."
+    echo "Run 'download_missing_models.sh' to fetch models for your workflows."
 fi
 
 exec /start.sh

download_missing_models.sh

@@ -173,6 +173,17 @@ def main():
     dry_run = "--dry-run" in argv
 
     cleanup_debris()
+
+    if not os.path.isdir(WORKFLOWS_DIR):
+        print(f"{C_YELLOW}{WORKFLOWS_DIR} not found — downloading workflows from hf.co/{REPO_ID} ...{C_RESET}")
+        from huggingface_hub import snapshot_download
+        snapshot_download(
+            repo_id=REPO_ID,
+            allow_patterns=["workflows/*.json"],
+            local_dir=WORKSPACE,
+            token=os.environ.get("HF_TOKEN"),
+        )
+
     files = sorted(glob.glob(os.path.join(WORKFLOWS_DIR, "*.json")))
     if not files:
         sys.exit(f"No workflow JSONs found in {WORKFLOWS_DIR}")