DeepSeek API Chat Streaming

Previous page: Overview

Chat Streaming

Official reference:

• Create Chat Completion
• Thinking Mode
• Multi-round Conversation

Main path

• src/deepseekApi/providers/deepseek/DeepSeekProvider.ts
• src/deepseekApi/providers/deepseek/features/Chat.ts
• src/VsCodeApi/webviews/handlers/Chat/Streaming.ts

Flow

1. ChatHandler prepares messages and configuration.
2. createDeepSeekProvider(config) creates the provider.
3. The provider always opens an SSE request for chat responses.
4. Each chunk is normalized as content or reasoning.
5. streaming.ts publishes events to the webview, where the UI renders accumulated deltas progressively rather than jumping per transport chunk.
6. When finished, the accumulated content is returned for history.

DeepSeek contract

• The main endpoint is POST /chat/completions.
• messages accepts system, user, assistant, and tool roles.
• stream: true sends deltas through Server-Sent Events and closes with data: [DONE].
• In thinking mode, reasoning arrives as reasoning_content, separate from content.
• finish_reason may indicate stop, length, content_filter, tool_calls, or insufficient resources.
• usage may include prompt, completion, cache hit/miss, and reasoning tokens.

Cancellation

cancelGeneration must trigger an AbortController shared by the active request. The handler should publish a controlled error or close event depending on the exact cancellation point.

Errors

Errors should arrive as streamError with a useful message. Do not leak the API key or full sensitive response bodies.

Review Error Codes before changing HTTP error mapping.

---

Next page: Models and Configuration