Messages

Messages-format JSON, streaming and client-executed tools through the native Go gateway.

POST/v1/messages

Availability and identity#

A bounded real-GPU canary passed JSON, SSE, a streamed tool call and its result continuation through isolated Better Auth/D1 accounts and the signed native Go gateway. Public inference remains unavailable until the production Go connection is enabled. Use an account-owned DarkSeek key as x-api-key or Authorization: Bearer, never both. Keys and web sessions share the same durable admission allowance.

Version and required fields#

Send anthropic-version: 2023-06-01, model, messages and a positive max_tokens within the Models deployment ceiling. Start with a user entry. System instructions belong in system, not a system-role message. Only text input and client-executed custom tools are accepted. Unknown/private fields, images, server tools, beta headers and cache hints are rejected.

Official client setup#

The official TypeScript SDK is contract-tested with version 1.8.0. Explicitly select the supported API date; newer SDK default dates are not silently accepted. Disable automatic generation retries and use a finite client timeout. SDK compatibility does not mean DarkSeek serves Claude models or every Anthropic feature.

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  apiKey: process.env.DARKSEEK_API_KEY,
  baseURL: process.env.DARKSEEK_BASE_URL, // origin, without /v1
  defaultHeaders: { "anthropic-version": "2023-06-01" },
  maxRetries: 0,
  timeout: 30000,
});
const reply = await client.messages.create({
  model: "darkseek-v4.1",
  messages: [{ role: "user", content: "Hello!" }],
  max_tokens: 128,
  thinking: { type: "disabled" },
});

Streaming lifecycle#

With stream=true, read message_start, content_block_start/delta/stop, message_delta and message_stop. Text, thinking and input_json_delta blocks retain their own indexes. The final message_delta carries the stop reason and usage; successful completion requires message_stop. There is no [DONE] marker. A malformed, interrupted, oversized or backend-error stream ends with an error event and no fabricated successful terminal event.

Tools and continuation#

Define at most 64 named custom tools with object input_schema. tool_choice supports auto, any, tool plus an existing name, or none. Tool code runs in your application, never on the gateway. Return each immediately preceding tool_use ID exactly once in the next user tool_result block before generating again. Duplicate, unmatched, pending or foreign result IDs are rejected. Tool argument buffering is bounded at 2 MiB per block; this is separate from token limits.

Thinking and limits#

thinking supports disabled or adaptive. The installed backend cannot guarantee budget_tokens or display suppression; those controls are rejected. Reasoning is not an Anthropic-verifiable signed artifact. Adaptive reasoning output and parallel tool combinations still need broader physical certification. temperature and top_p are 0–1; top_k is 1–1000; at most four nonempty stop sequences, 512 messages and 512 input blocks per entry. Shared body, response, time, concurrency and daily-attempt ceilings still apply.

Errors and safe retries#

Messages errors use type=error with an error object. AuthenticationError, RateLimitError and streaming error behavior are tested with the official SDK. Keep the fixed error code and X-Request-ID; backend diagnostics and customer credentials are not forwarded. A timeout after admission consumes an attempt. Never blindly retry a generation after an ambiguous disconnect.