model-catalog

Plugging into an OpenAI-compatible host inside the dsh ecosystem means model listings, per-token or per-million pricing, context windows, and capability flags are scattered across different endpoints in mismatched units. Manually scraping each endpoint, converting quota multipliers and per-call fees into a common scale, and hand-editing the config is tedious and error-prone, especially when you rotate gateways or add a local Ollama instance. model-catalog removes that chore: set a baseUrl, and it probes, fetches, normalizes, and emits ready-to-use configs automatically.

Three pillars drive the work. First, host-type auto-detection covers six shapes—the bare standard endpoint, rich-metadata gateways, quota-multiplier relays, capability-flag proxies, Ollama, and vLLM—with a manual override when every probe fails. Second, end-to-end pricing normalization flattens USD-per-token strings, scientific-notation numbers, multiplier-based quota pricing, and per-call fees into a single USD-per-million scale, with a traceable source chain and structured tiers for DeepSeek's peak/off-peak dynamic billing. Third, optional live capability probing fires minimal requests to verify tool calling, structured output, and streaming when metadata is silent, caching results per host and model for 24 hours and halting on 401/403 to protect your quota.

The output is three files: catalog.json (full normalized directory), dsh-models.json (a config fragment the harness ingests directly), and report.md (a human-readable summary including unknown-pricing and dynamic-billing notes). It is built for developers who already connect dsh to several API providers, relay gateways, or local inference services and want a single clean, auditable model catalog instead of re-typing fields by hand.