agent-systems-handbook

用 connector-first、最少 token 的方式审阅 Gmail 客户支持线程。适用于 Codex 需要通过 Codex Gmail connector 读取 Gmail、把清洗后的消息导入本地 SQLite 问题队列、先执行确定性的清洗和分类、只把模型调用保留给 JSON-only 的消息理解和草稿字段生成，并支持 dashboard 审阅、客户审批与排队回复处理的时候。

审阅客户来信，以本地政策或 FAQ 文档为依据起草安全的投诉和咨询回复。

Compare two structured agent-run artifacts to estimate cache efficiency, explain likely cache breaks, and produce a local benchmark report. Use when a user wants to understand whether a prompt layout, tool manifest, or history shape is helping or hurting prompt-cache reuse.

Persistent daily news monitoring backed by local SQLite and Markdown reports. Use when a user asks Codex to track named publications, fetch the last N hours of news, summarize recent articles by topic, deduplicate articles across runs, or maintain a personal newsroom that survives across sessions.

Scan local cleanup targets, apply readable cleanup rules, produce a preview report, and execute approved cleanup actions with logs. Use when a user asks Codex to clean up their computer, empty old Trash items, find duplicated Downloads files, review local storage clutter, or propose safe file cleanup actions before making changes.

Preview-first local file organizer. Scan a user-named folder, classify files into category subfolders using readable rules, write a preview Markdown report and JSON plan, execute confirmed moves with persistent SQLite state, and reverse moves with undo. Use when a user asks Codex to organize Downloads, sort a messy folder into Invoices/Receipts/School/Images/Software/PDFs subfolders, propose a folder structure before moving anything, or undo a previous organization run.

Capture local or explicitly provided web knowledge sources into cited Markdown notes. Use when a user asks Codex to watch a research folder, register local folders for later scans, summarize new or modified local Markdown/TXT/PDF/DOCX files, capture a provided URL, maintain SQLite state for personal knowledge capture, or generate searchable source-grounded daily notes.

Persistent product price tracking for natural-language product requests. Use when a user asks to watch, track, monitor, compare, or report prices for a product, especially with a target price or threshold such as "Watch MacBook Pro M3 14-inch and tell me if it drops below $1200." Supports source discovery, Playwright/browser product checks, SQLite history, threshold comparison, and Markdown price reports.

Plan and inspect prompt-cache behavior for long-running Claude agent loops. Use when a user wants to split stable tool, system, and history context into cacheable layers, compare captured cache metadata, estimate cost impact from supplied pricing inputs, or keep durable memory outside the cached prefix.

Review a local evidence bundle or assistant transcript for credible harm signals, produce a redacted escalation memo and checklist, and stop before any external reporting or authority contact. Use when a user asks Codex to prepare a safety, legal, trust-and-safety, or incident handoff from local evidence.