2026-07-26 每日新聞檔案

今天先抓三組工程問題：OpenClaw 要把 before-tool-call policy 與 shell deny fallback 分開驗；Hermes 要把 slash command 的 target session 與 relay interactive ack lifecycle 分開驗；外部對照用 Opus 5，練習把模型升級新聞改寫成 routing、cost、tool-use 與 latency 指標。

封存時間：2026-07-26 07:00 CST

今天新增的小教學提示

先替每條消息各補兩欄：① control boundary 是 policy、approval、session target、interactive ack，還是模型路由；② first wrong write 最早會先出現在 deny fallback、寫錯聊天、漏掉 ack，還是成本／延遲失控。

中文 5 條

先讀 control boundary 與 write target，再對照 ack lifecycle 與 model routing，練習把同一天的不同敘事改寫成工程驗證表。

回到今日新聞頁
  1. OpenClaw 把 before-tool-call policy pipeline 拆開，這很適合拿來教工具執行前其實有三層責任：policy 決定能不能做、approval 決定要不要問人、diagnostics 決定失敗時能不能追到真正原因。
    來源：查看來源（OpenClaw 官方 Commit）
  2. OpenClaw 修正 shell expansion execs 會忽略 deny fallback 的問題，這是一條很好的 fail-closed 教材：展開失敗時，系統到底是中止、回退還是偷偷換路徑，必須寫得非常明確。
    來源：查看來源（OpenClaw 官方 Commit）
  3. Hermes 修正桌面版 slash commands 應該寫回使用者當前聊天，而不是偷偷開新 session；這很適合教學生分辨「命令解析成功」與「狀態寫回正確目標」其實是兩個不同層次。
    來源：查看來源（Hermes Agent 官方 Commit）
  4. Hermes relay 進到 Phase 3 interactive，把 approvals、confirms、clarify 與 React ack lifecycle 接起來；這很適合教 human-in-the-loop 不是只把 prompt 顯示出來，而是要把確認、回應與 UI acknowledgement 的生命週期都接完整。
    來源：查看來源（Hermes Agent 官方 Commit）
  5. TechCrunch 報導 Anthropic 推出 Opus 5，今天最適合拿來做外部對照的不是模型口號，而是 agent stack 要不要跟著調整 model routing、成本預算、tool-use 策略與延遲目標。
    來源：查看來源（TechCrunch）

English 5 Items

Use the English list to compare control boundaries, write targets, acknowledgement lifecycles, and model-routing implications.

查看本月檔案
  1. OpenClaw splits the before-tool-call policy pipeline, which is a clean way to teach that pre-tool control has three separate responsibilities: policy decides whether a tool may run, approval decides whether a human must confirm, and diagnostics decides whether a failure is still debuggable.
    Source：Source (OpenClaw Official Commit)
  2. OpenClaw now honors deny fallback for shell-expansion execs, making it a compact fail-closed lesson: when expansion is denied, engineers need an explicit answer about whether the system aborts, falls back, or silently changes execution path.
    Source：Source (OpenClaw Official Commit)
  3. Hermes fixes desktop slash commands so they target the user’s current chat instead of spawning a new session, which cleanly separates “the command parsed” from “the state was written back to the correct target.”
    Source：Source (Hermes Agent Official Commit)
  4. Hermes relay reaches Phase 3 interactive UX by wiring approvals, confirms, clarify, and the React acknowledgement lifecycle together, which is a strong lesson that human-in-the-loop work is not just rendering a prompt but closing the full confirmation and acknowledgement loop.
    Source：Source (Hermes Agent Official Commit)
  5. TechCrunch’s Opus 5 coverage is useful today not for the headline alone, but as an outside comparison point for whether an agent stack should change model routing, cost budgets, tool-use strategy, and latency targets after a model upgrade.
    Source：Source (TechCrunch)