Bridge Input (CONDITIONING)

After CLIP Text Encode does its job, what travels toward the KSampler isn't text anymore - it's conditioning, a bundle of embeddings plus metadata that the sampler actually consumes. Bridge Input (CONDITIONING) is ComfyBridgeIO's named tap on that specific stage of the pipeline. If your agent-driven workflow does anything clever with conditioning - regional prompting, ControlNet, merging positive and negative - this is the node that marks where those streams sit.

Like its MODEL/CLIP/VAE siblings it's a passthrough: optional value input socket (CONDITIONING), single value output socket (CONDITIONING). You wire a CLIP Text Encode (or a ControlNet apply node, or a conditioning combiner) into the input and take the output to the sampler's positive/negative input. The client doesn't serialize conditioning into the patch step - it's an in-memory structure - so the node's value is structural: the name widget (STRING, default "input_name") is reported by bridge.get_schema(), giving the agent a labeled map of where each conditioning stream enters the sampler.

That labeling pays off in the agent context precisely because conditioning is where most "clever" workflows get complicated. The KB's writeup on the plumbing layer makes the general point - a graph too big to read is a graph nobody (human or agent) can debug. Two named CONDITIONING bridges for positive and negative turn an unreadable noodle into something an agent can reason about: "positive comes from this node, negative from that one, swap the negative bridge's upstream and the whole vibe changes." When you're having an agent iterate on prompts, that's the difference between it fumbling and it acting.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/cbpark84/ComfyBridgeIO

Restart ComfyUI, or ComfyUI Manager → search ComfyBridgeIO → Install. No model files, no heavy deps. Optional Python client (pip install -e in the repo; aiohttp, websockets, Pillow). The README's clone URL points at nicekate/ComfyBridgeIO - stale; the repo is cbpark84 as above.

Common issues

The standard unwired-input failure: no conditioning connected, sampler gets nothing, and you get a type error at the KSampler that doesn't tell you the bridge is the culprit. Wire the Text Encode into it. One more thing worth knowing: if you tag a CONDITIONING bridge and also wire its upstream from a Text Encode whose prompt comes from a STRING bridge, you've got the full chain the agent can drive - but the STRING bridge's name and this bridge's name are different contracts, and the client matches each independently. Keep them distinct and you'll never confuse the two.