T8 节点来源／运行版本诊断

What it is, and when you'd actually use it

ComfyUI's worst practical failure is the one where everything looks right and behaves like an older version of itself. Custom nodes live in custom_nodes/ directories in one shared Python environment, updates are git pulls, and nothing verifies that what's on disk is what the running process loaded. You pull, restart, and either the fix is live or you're staring at an old stack trace wondering which of the two copies in your directory tree is answering.

This node answers that question for this pack. Type in a node class name, run it, and it reports where the running instance's implementation came from: the loaded module path, the SHA-256 of that file on disk right now, the SHA it saw when the diagnostic module was imported, and whether the file has changed since. If the same node is being provided by more than one loaded module path, it says so instead of quietly picking one.

How it works, and what it won't tell you

The input is a single string, node_id, defaulting to MiniMaxH3AudioConditioningT8. It resolves that against the classes this pack registers - by class name, by the class's own node ID, or by the ID declared in the node's schema - then pulls each match's module path out of the live interpreter.

Two honest limits are baked into the output, and they're the reason I'd trust this over "just check the folder":

• A matching file hash is not proof that the loaded bytecode matches. The report literally carries memory_bytecode_matches_disk: not_proven_by_file_hash. If the file changed after import, you get a warning telling you the running instance may still be the old implementation and to restart and check again.
• It only knows about this pack's nodes. The lookup runs against this pack's own registered classes, so typing in a node class from a different custom-node pack returns "not found" even when that node is working fine. The default value is a node from this pack for that reason.

The other useful field is the disk package version - read from pyproject.toml - reported next to running_package_version, which stays unknown_unless_runtime_declares_it. Translation: the version on disk does not tell you the version in memory.

Crucially, it's read-only. It doesn't delete directories, restart your service or block sampling, and the node is marked as an output node purely so its text preview shows up when you run it.

Inputs and outputs

node_id (STRING) is the only input. Outputs are 说明 (STRING, the human-readable summary, also shown in a scrollable text panel) and report_json (STRING) with the structured version: formal project root, disk package version, per-match module path, both hashes, warnings, and explicit mutation: false.

Wire the JSON output into a Save Text node if you're filing an issue - it's far more useful than a screenshot of a red node.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Full restart, then refresh the browser. Manager finds the pack under the title MiniMax H3 Audio T8. The diagnostic nodes are the lightest part of this pack - no models, no extra dependencies, and the pack's requirements.txt is deliberately empty so installing it won't reinstall Torch under you.

When it earns its keep

You updated the pack and the behaviour didn't change. Check the disk SHA, then check the duplicate-provider warning. Two installs of the same pack - one under minimax-h3-audio-T8, one cloned to a different folder name - is a real and common cause, and it's exactly the pattern the pack's own notes about Manager versus GitHub installs invite.

You're about to report a bug. The full error, the workflow, the Core version and this report is the combination the pack asks for. Without the source path, maintainers are guessing which build you're on.

A node is red and you don't know why. This won't resurrect a failed import, but it does separate "the node never registered" from "the node registered from an unexpected path," which is where most of these hunts actually end.