cacr92/wereply

Overview

This skill provides practical guidance and patterns for integrating the DeepSeek chat API into an application. It covers HTTP client configuration, streaming and non-streaming response handling, robust error classification and retry strategies, and secure API key management. The guidance focuses on connection pooling, streaming SSE processing, and parsing generated suggestions for a chat reply assistant.

How this skill works

The skill explains how to configure a reusable HTTP client with connection pooling and timeouts, authenticate using a Bearer API key, and send chat completion requests. It shows both blocking JSON responses and Server-Sent Events (streaming) handling, including chunk parsing and invoking callbacks for incremental UI updates. Error classification and exponential backoff retries are provided to make calls resilient, and secure API key storage via system keychain is recommended.

When to use it

• Integrating DeepSeek for automated reply suggestions or chat completions.
• When you need streaming partial outputs to update a UI in real time.
• When you want efficient HTTP usage with connection pooling for high throughput.
• When you need resilient calls with retry/backoff handling for transient failures.
• When you must securely store and retrieve DeepSeek API keys.

Best practices

• Reuse a single HTTP client instance across the app to leverage connection pooling and reduce latency.
• Set sensible timeouts: separate connect timeout and overall request timeout to avoid stuck requests.
• Enable streaming only when you need partial results; handle SSE parsing and the [DONE] sentinel safely.
• Classify errors (401, 429, 5xx, timeouts) and only retry retryable cases with exponential backoff and an upper cap.
• Store API keys in the system keychain with an environment variable fallback for CI or containerized deployments.

Example use cases

• Generate three short WeChat reply suggestions from chat context and a style prompt, then parse numbered or newline-separated outputs.
• Stream model output to a frontend: emit chunks as they arrive so the UI shows progressive suggestions.
• Implement a background task that calls DeepSeek with a pooled HTTP client and retries on transient network errors.
• Securely read the DeepSeek API key from keychain in a desktop app and fall back to environment variables for automated tests.