Nodes/ERPK Collection/Claude Conversation Info
ComfyUI Node

Claude Conversation Info

Tokens, history, state

By eRepublik-Labs·Created 11 months ago·Updated 20 days ago· 1
Claude Conversation Info
  • conversation_history
  • info

Claude Conversation Info is the diagnostic sibling of Claude Conversation: you feed it a conversation_history object and it gives you a STRING report on what's actually in the thread - its state, token usage, and size. It's an output node, so you drop it at the end of a conversation chain to see what's accumulating before you run the next turn.

If you've ever stared at a long multi-turn conversation wondering "how much context have I burned and what did I even say?", this is the node that answers it. It's also genuinely useful for debugging: when a conversation starts behaving weirdly - replies getting shorter, context errors, the model "forgetting" things - the info node tells you whether the history is bloated or the system prompt is still attached, without you having to read a raw JSON blob in the console.

What it does

One required input: conversation_history, which you wire from the conversation_history output of any Claude Conversation node. One output: info, a formatted STRING describing the thread's state and token usage. That's the whole interface - there are no knobs to turn, which is exactly right for an inspection node.

Where this sits in a workflow: chain your Claude Conversation nodes as usual, then run the final conversation_history output into this node instead of (or alongside) a preview. The report shows you the message count, the stored system prompt, and the token footprint - so you can decide whether it's time to flip reset_conversation on the Conversation node, or whether auto_trim is doing its job.

How it works

The node reads the conversation state object that Claude Conversation maintains - the message list, the system prompt, and the running token counts - and formats it into readable text. It doesn't call the API itself; there's no extra generation cost for running it. That's worth repeating: this is a free inspection node, not a billable call.

Installing it

Part of the ERPK Collection (eRepublik-Labs/comfyui-nodes-erpk). ComfyUI Manager: search erpkERPK Custom Nodes → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk
pip install -r requirements.txt

Requires anthropic>=0.96.0 and an Anthropic API key (Settings → ERPK → API Keys). No models to download.

Common issues

  • "Nothing connected" / no output - you have to wire a conversation_history in; it's the only input. If it's empty, the upstream Conversation node hasn't run yet or produced no history.
  • History shows a huge token count - that's your signal to let auto_trim handle it or restart the thread. Long threads are expensive threads.
  • Info looks stale - it reflects the state object you connected at execution time. Re-run the chain if you changed a message upstream.
CategoryERPK/Claude

Inputs (1)

NameTypeDefaultDescription
conversation_historyCLAUDE_CONVERSATIONConversation state to inspect

Outputs (1)

NameTypeDescription
infoSTRING