Fred Bliss
cc28542508
ComfyUI starts a fresh asyncio loop per prompt execution but may cache the ShrugConnection node output (and its ShrugClient) across runs. The persistent httpx.AsyncClient's pooled connections were bound to the prior loop, so pool cleanup on the new loop raised "Event loop is closed" on every other run. Track the loop the client was created on (_http_loop) and rebuild when it differs. When discarding a stale client, best-effort schedule aclose() on the old loop if it's still alive; otherwise drop (its transports went down with it).
shrug-prompter
ComfyUI custom nodes for heylookitsanllm — a single-user MLX inference server on macOS.
Ten nodes, ~1,500 lines of Python. No multi-provider abstraction; this package is specifically tuned for heylookitsanllm's API surface (multipart vision upload, batch endpoint, Qwen3 thinking blocks, KV-cache management).
Install
Clone into ComfyUI/custom_nodes/:
cd ComfyUI/custom_nodes
git clone <this repo> shrug-prompter
Dependencies are declared in pyproject.toml and are normally already present in a ComfyUI environment. If not:
cd shrug-prompter
uv sync
Nodes
All nodes live in the shrug category.
- Shrug Connection — base URL, model, optional auth. Builds a
SHRUG_CONNobject every other node consumes. - Shrug VLM — the main chat/vision node. Optional image batch, all sampling params, Qwen3 thinking output, base64 or multipart transport.
- Shrug VLM Batch — N prompts (+ optional N images) in one call via
/v1/batch/chat/completions. 2-4× faster than N sequential calls. - Shrug ASR — speech-to-text from a ComfyUI
AUDIOinput. - Shrug Embeddings — text → tensor via
/v1/embeddings. - Shrug Template — loads a
.mdfile fromtemplates/as a prompt. Ships with 250+ style/system-prompt snippets intemplates/styles/plus a few VLM templates intemplates/vlm/. Seetemplates/README.mdfor frontmatter format. - Shrug Text Cleanup —
none/basic/standard(NFC + collapse blanks) /strict(ASCII-only). - Shrug Accumulator — graph-scoped list state (append/reset/read/pick), for loop workflows.
- Shrug Frame Pair — frame-pair extractor for video interpolation.
- Shrug Cache Control — list or clear the server's KV cache.
Configuration
Shrug Connection:
base_url— e.g.http://localhost:8080(same-host) orhttps://your-llm-host(LAN). Caddy does TLS on the LAN; any reverse proxy will do.model— themodelwidget is populated by JS fetching/v1/modelson the server. If the fetch fails, the widget falls back to free text.api_key(optional) — sent asAuthorization: Bearer <key>. Ignored by vanilla heylookitsanllm; useful for other OpenAI-compatible targets.admin_token(optional) — sent asX-Heylook-Admin-Tokenon admin endpoints only (currently just/v1/cache/clear). Matches the server'sHEYLOOK_ADMIN_TOKENenv.
Testing
uv run pytest tests/ -v
67 tests covering the client (respx-mocked), media codecs, node behavior, and template parsing.
Live smoke test against a running server:
uv run python scripts/smoke.py --base-url https://your-llm-host --model your-model-id
Development
See CLAUDE.md for architecture notes and conventions.
Repo layout:
shrug-prompter/
├── __init__.py # V3 API probe + /shrug/get_models route
├── client.py # async httpx client
├── media.py # tensor <-> bytes (image, audio)
├── nodes.py # 10 ComfyNode classes + extension
├── pyproject.toml
├── templates/ # .md prompt library
├── web/provider.js # ShrugConnection's model-dropdown widget
├── scripts/smoke.py # live LAN smoke test
└── tests/ # pytest
License
MIT.