MiniMax H3 Cache Telemetry Monitor

MiniMaxCacheMonitor ("MiniMax H3 Cache Telemetry Monitor") is the pack's dashboard - a one-input node that reads the current continuation session and spits out a plain-text report of what just happened. It's not glamorous, and it doesn't need to be: in a chaining workflow where most of the interesting state lives invisibly inside the applier's session object, this is the only way to see it without digging through console logs.

Why you'll want it

The Continuation Applier makes a lot of decisions for you, and most of them are silent unless something goes wrong. It snaps your requested context down to a phase-aligned grid length, auto-falls back from Native Masked AV to Safe Native when the geometry won't fit, and quietly increments a per-session clip counter. The monitor surfaces exactly those facts:

• which mode is active - Native Masked AV or Safe Native fallback. This is the single most useful line, because an automatic fallback is otherwise only a console warning you might miss while watching a 10-minute sample grind.
• the current clip number in the session,
• the configured protected context (your 39/90/141/192 setting) versus the last protected context actually applied - including the exact latent steps. If those two numbers disagree, the geometry snapped down, and now you know why the overlap looked shorter than you asked for,
• the number of accumulated clips the session is tracking.

So the honest use case: wire it up once, run a clip, and glance at it to confirm you're on Native Masked AV with the context you expected. It's a debugging aid, not a production output - think of it as the engine's instrument panel.

Wiring it

One input, one output - about as simple as ComfyUI gets:

• session (required) - connect the session output of the Continuation Applier. That's the only source that carries the state; there's no other way to get a session object.
• telemetry_report - a STRING with the report. It's an output you can read in the node's widget if you've enabled text display, or feed to a text/log node. It's not marked as an output node, so if you want it in your saved outputs, chain it somewhere visible - or just keep the node on the canvas and read the string in the UI.

In the pack's reference workflow, the applier's session fans out to both the Trim Prefix node and this monitor, so it costs you nothing to add it to your own graph.

Install and caveats

Same pack, same routine - Manager (search "MiniMaxH3 PrefixStream") or clone into custom_nodes, pip install -r requirements.txt, restart, Ctrl+F5. No models, no API, nothing to download.

Two honest limits. First, the report is a snapshot of the session object passed in, and the session only lives as long as your workflow run - it's not persisted across ComfyUI restarts (that's what the Save/Load latent nodes and Clip Bin are for). Second, don't expect per-layer KV-cache memory stats despite the "Cache" in the name: the "cache" here is the continuation-prefix mechanism, and the monitor reports continuation state, not VRAM telemetry. If you're hunting for a memory monitor, this isn't it - use it to verify your continuation mode and context, and you'll have what it's actually for.