Llm Lessons

internal-gateway.local API 端点 (`https://api.internal-gateway.local/v1`) 只接受 OpenAI 格式 (`/v1/chat/compositions`)，不支持 Anth

Hermes Agent 配置 Anthropic provider 使用 internal-gateway.local API 时失败，

A DeepSeek-family model reached through a third-party channel occasionally emits its tool-call protocol as **plain text

In multi-tier enterprise environments where model traffic flows through API gateways, rate-limiting proxies, or security

A self-hosted llama.cpp `llama-server` (0.6.0-dev, started with `--jinja` and a custom chat

LLM cost telemetry undercounts by approximately 2× due to two independent bugs:

In modular multi-agent architectures, independent subagents, planning loops, and dynamically dispatched tool handlers fr

Reasoning models such as DeepSeek-R1, Qwen-QwQ, and OpenAI-compatible reasoning variants emit internal chain-of-thought

模型返回的内容不完整（truncated），或者 JSON 解析失败（`json.decoder.JSONDecodeError`），导致后续处理流程中断。

> Domain: devops | Tags: model-switching, proxy, config-management

部署了 LLM 服务（vLLM/Ollama/本地推理）后，不知道怎么用 API 调用。各框架接口不一。

In multi-turn character roleplay with prompt caching (e.g. Claude, GPT-4o), the context is split into:

When consuming the Vertex AI `streamGenerateContent` API over HTTP or via custom middleware, client applications frequen