Create Embeddings

Create a message, in Anthropic's Messages format. **Point the Anthropic SDK at us and change nothing else.** Authenticate with `x-api-key` carrying your Asteria key, exactly as the SDK sends it, or with `Authorization: Bearer` as on the rest of `/v1`. `anthropic-version` is accepted and echoed. **Routing modes**, controlled by the `agent` field, exactly as on the other two dialects: - **Gateway mode** (`agent` omitted): a pure LLM proxy, no system prompt and no tools. Rate limiting and budget enforcement still apply. - **Orchestrator mode** (`agent: "asteria"`): the full Asteria agent, with RAG, web search, code execution and sub-agents. - **Custom agent mode** (`agent: "<slug>"`): a shared custom agent by slug. **Which model**: send `model: "default"` to use the organization's default chat model and its fallback chain. Pinning a real model id is brittle, since an org admin can remove it later. The model that answers need not be an Anthropic one: the dialect you speak and the provider that serves you are independent here, which is the point of this endpoint. **`max_tokens` is required**, as it is on Anthropic's own API. A request without one is refused rather than given a ceiling it did not choose. **`stream: true`** speaks the Anthropic streaming protocol: `message_start`, then `content_block_start` / `content_block_delta` / `content_block_stop` around the text, then `message_delta` carrying `stop_reason` and the final usage, then `message_stop`. `ping` arrives on a quiet run. There is **no `[DONE]` sentinel**, which is Chat Completions' shape and not this one. A run that never opened its stream ends in a bare `error` event; one that opened and then failed closes its block and ends in `error` too. Hanging up does not cancel the run: it finishes server side, is billed and is logged, exactly as if you had stayed. **This endpoint is stateless.** Nothing is stored and there is no id to fetch back; send the prior turns in `messages`. `/v1/responses` is where state lives. **`system` is a top-level field**, not a message in the array. A `system` role inside `messages` is refused, as it is upstream. **Content blocks**: `text`, `image` and `document` are carried, and `cache_control` is honoured as a caching boundary at the end of the message that marks it (at most four per request, as upstream). `tool_use`, `tool_result`, `thinking` and `redacted_thinking` are refused by name rather than accepted and dropped, as are the `tools`, `tool_choice` and `thinking` request fields and `citations` on a document: there is no client-side tool loop on any `/v1` dialect, and a `200` that silently discarded a tool result is worse than a clear refusal. Anything carried with a documented degradation says so in an `X-Asteria-Warning` header.