dsh-modelprint

The scariest failure mode in a model pipeline is not a crash; it is silent configuration drift. A tool gets registered, a system prompt gains a sentence, a patch layer nudges temperature from 0.7 to 0.2, and nothing in the output looks wrong. You simply cannot tell what changed.

dsh-modelprint takes a deliberately narrow approach: it fingerprints each outgoing request into a small card covering the provider, model id, sampling parameters, a SHA-256 hash of the sorted tool schemas, and a hash of the system-prompt prefix. That live card is then diffed field by field against a pin file you wrote down when the run was known good. The plugin never touches the network, loads no model, and reads no key; its only I/O is reading the pin file from disk. On a match you get OK; on drift you get every differing field listed in diagnostic priority order. It is not a watermark, not a weight-swap detector, and not a model-auth product. It answers exactly one question: is this the same setup as last time?

It is built for teams and repos with multiple patch layers, frequently changing tool sets, or shared system prompts where catching drift fast matters. Writing the first pin is as simple as calling modelprint_check once, copying the live card out of the PIN_MISSING result, and saving it to disk. Every subsequent check is a zero-cost, structured comparison; when a deliberate change trips the alert, you re-pin and the noise stops.