The developer path

For someone writing a client, an agent or a service against ignis. The surface is OpenAI-compatible, so an OpenAI SDK works with a base URL; the pages below say where ignis adds to it, refuses a field, or behaves its own way. The Playground names the request each of its controls sends, a quick way to see what a field does before you code it.

1. Connect a client — the base URL, the key, the model id, the SDKs.
2. The API — every endpoint, the probes, keys and the error shape.
3. Chat completions — fields, streaming, tool calls, reasoning, and the agent lane.
4. The Responses API — POST /v1/responses and its WebSocket mode.
5. Examples — curl for the common requests.
6. Counting a prompt — how many tokens a body prefills, before you send it.
7. Deadlines — how long a request may run, and how a client states its own budget.
8. Switching models and idle unload — what a request sees while the server changes or reloads its model.
9. Image generation — generating and editing images with the image model. Speech to text — transcribing audio with the speech model, which runs on the CPU beside it; live transcription streams the audio as the person speaks.
10. Decisions — POST /v1/decide: an answer read from the model’s distribution instead of generated, and locating a line or an item.
11. Reasoning and the loop guard — what the reasoning channel carries, and what the server does when it loops.
12. HTTP routes — the generated route list; a running server serves the full OpenAPI document at /v1/openapi.json and a browsable page at /v1/docs/.

Tip

Every field of the OpenAI request bodies is honoured, accepted as inert, or refused with a 400 that names it in param — never dropped in silence. A 400 naming a field tells you exactly what to remove (Troubleshooting).

Tip

Prompt reuse keys on the start of the prompt. Keep the system prompt and the tool block byte-identical across a session — no timestamps, no counters — and a long conversation prefills only its new turn (The KV cache and prompt reuse).