简体中文

Model Catalog — Model Catalog Auto-Discovery (dsh plugin)

This plugin is designed for the dsh ecosystem (a plugin-based DeepSeek harness where everything is a plugin): once you configure an OpenAI-compatible API host in dsh (official API, relay gateway, local inference service, etc.), the plugin automatically pulls model information from that host's endpoints — model listings, pricing (input/output/cache prices per million tokens), inference parameters (context length, max output, capability flags: tool calling / structured output / vision / parallel tools, etc.) — normalizes it, and generates ready-to-use model configs, saving you the manual work.

Core value:

• Host-type auto-detection, no protocol details to configure;
• End-to-end unit normalization (per-token USD / per-million USD / multipliers / per-call);
• Traceable pricing (captured_at + source endpoint), structured dynamic pricing;
• Optional lightweight live probing for missing capabilities;
• Three outputs: full catalog, dsh config fragment, human-readable report.

---

Features

• Host type auto-detection: standard compatible (minimal fields) / rich metadata (/models with pricing and parameter lists) / multiplier-priced gateway (quota system) / capability-flag proxy (/model/info capability booleans) / Ollama / vLLM; fails with an error when all endpoint probes fail, and supports manual --kind.
• Normalized model records: id / context window / max output / pricing family (input, output, cache_read, cache_write, internal_reasoning, in USD/1M) / capability flags / source provenance.
• Pricing source chain: host endpoint → user override config → external price mirror → built-in default table → marked unknown (with warning).
• Built-in default table: model facts and pricing fallback for the official API (DeepSeek); dynamic pricing (peak/off-peak billing from 2026-08-16) expressed as structured tiers.
• Capability probe verification (switchable): sends minimal requests to live-test capabilities missing from metadata (tool calling / structured output / streaming), results cached per (baseUrl, model).
• Cache and concurrency safety: classification/probe/mirror results cached with TTL; file locking + atomic writes prevent concurrent processes from corrupting each other; corrupted caches auto-recover.
• Interactive config generation: the pick command lists models and prices; select one to generate a dsh config fragment.
• Self-contained: manifest + entry factory + tool/event interfaces, directly loadable by the harness (see integration docs).

How It Works

…