Adding a model (capabilities) - Kyma API

A model that is missing tools, files, or output formats — or that returns the wrong HTTP status — makes the caller’s agent invent a story. Declare the facts; verify them; then ship. The full checklist lives in the repo: docs/MODEL-ADD-CHECKLIST.md. Guard: bun run test:model-add-checklist.

What `/v1/models` must tell you

Field	Meaning
`supports_tools`	OpenAI-style `tools[]` on chat completions. Language models only.
`supports_vision`	Image in the prompt.
`supports_structured_outputs`	`response_format` / JSON mode.
`input_modalities`	`text`, `image`, `audio`, `video`, `file`. PDF is `file`.
`context_window` / `max_output_tokens`	Creator limits.
`supported_parameters`	Includes `stream` for chat; `tools` only when tools are real.

If a flag is absent from a language model in the registry, Kyma currently treats tools / JSON / caching as on. New rows must set the booleans explicitly.

Errors the caller can trust

Status	When
400	This model cannot do what you asked (tools, JSON, PDF, STT timestamps).
401	Bad key.
402	Not enough credits.
404	Unknown model id.
429	Rate limit.
4xx	STT heard no speech — not 502.
5xx	Every route failed.

Never a 200 with an empty billed transcript. See Error handling.