从一次邮箱登录出发，讲清 Session 与 JWT 的取舍，再拆解 OAuth 授权码流程、OIDC 身份校验、redirect_uri、CSRF、state 与 PKCE 的防护边界，以及 Web、App 与 CLI 的接入方式。
网站和面板同时卡死、重启就好，监控却显示 CPU 和内存都正常。从 journalctl --list-boots 和历史 syslog 里翻出一周 32 次 OOM，定位到前端进程，再用 swap、容器上限和 Node 堆上限分层设防。
Build a Vite project from scratch and understand configuration, assets, and chunking through module requests, hot updates, and production builds; then dissect core Hooks and integrations with React, SSR, TanStack Start, and Workers through runnable plugin and dual-runtime compilation experiments.
Build a runnable hotel assistant with slot filling, short replies, conditional cancellation, action confirmation, mock orders, and SQLite state, plus an HTTP interface, support for switching intent recognizers, code downloads, and a complete reading guide to the series.
From request IDs, state versions, and idempotency to atomic commits, time budgets, caching, and error feedback, this article uses a retry sequence diagram and real local HTTP checks to explain how the service runs reliably.
Using a hotel assistant to distinguish knowledge retrieval, personal booking lookup, write operations, and clarification, this article explains query rewriting, the division of labor between Workflows and Agents, task dependencies, confirmation, and failure paths.
Trains a frozen-encoder classification head on eight BANKING77 categories, distills it into a character-feature student, and compares real classification quality with inference latency, with the training script, student weights, and a quantized-storage example included.
Explains accuracy, Macro-F1, and confusion matrices, then covers out-of-scope rejection, state matching, and multi-turn task success, using actual prediction results to illustrate data splitting, error attribution, latency, and cost.
Distinguish out-of-scope requests, label ambiguity, missing slots, and service failures; use a tutorial experiment to select score and score-gap thresholds, examine coverage and errors, and design specific clarification strategies with limits on follow-ups.
Model compound requests as independent tasks with parameters, dependencies, and conditions. The examples cover multiple tasks of the same type, missing conditions, and skipped branches, with a task graph schema, validator, and runnable simulation included.