dsh-ctx-probe

When you run a local llama.cpp or Ollama model through DSH, llm-pi-ai falls back to a 262144-token contextWindow for any model that does not declare one. If your server was actually started with -c 65536, the built-in 80 percent pressure compaction can never reach its trigger, and the first oversized request hard-fails with a 400 overflow. Manually editing the setting is easy to forget, and a server restart with a different window size forces another round of edits. dsh-ctx-probe was built for exactly this: before every model call it queries the server for the real runtime context window (meta.n_ctx, not the training-time n_ctx_train) and aligns the DSH contextWindow to that value. Shrink the server and the plugin tightens before overflow; grow the server and it widens automatically. Both directions, zero numbers to type.

The probe strategy covers four endpoints, namely llama.cpp /v1/models, /slots, /props, and Ollama /api/show, adopting the first usable result. Loopback addresses are re-probed on every request, so a server restart is picked up on the very next message; remote addresses use a roughly ten-minute TTL cache with single-flight and exponential backoff to avoid probe storms. All errors are swallowed on the plugin side, so a failed probe can never break the agent loop.

If you are wiring local inference servers into DSH and would rather not keep tuning the context window by hand, this plugin keeps working after a one-time Web Host restart. It is fully self-contained, does not touch node_modules, depends on no @deepseek-ai package, and differs from manual-catalog or scheduled-sync alternatives by probing per request and syncing both ways.