dsh-model-context-catalog

When using a custom model route in DeepSeek Harness, a subtle trap lurks beneath the surface: if the route lacks an explicit context window, the engine falls back to a 262,144-token guess. Once a real request exceeds that estimate, the server may return a perfectly normal response yet the call is reclassified as a context overflow—tool calls survive but the final natural-language answer fails, and /compact compression is blocked entirely.

This plugin maintains exactly one field, contextWindow: it writes the true window size for a precise provider/model route and syncs it to llm-pi-ai through DSH\u2019s public Settings API. It does not patch internal code, does not touch credentials, does not fabricate missing routes, and never injects unverified maxTokens—its behavioural boundary is narrow and predictable.

If you run multiple custom model routes in DSH (different providers, different gateways, different plan suffixes) and keep hitting false overflow errors, this plugin was built for you: open Settings, pick the route, enter the window size, save, and it takes effect immediately.