2026 年 8 月新聞檔案

本檔案保留 2026 年 8 月的精選新聞與舊聞,依日期由新到舊排列,方便學生回查每日更新脈絡,並比較 state recovery、install provenance、config precedence 與 task-boundary design 等主題。

本月每日檔案

  • archives/2026-08-04.html — 2026-08-04 每日新聞檔案開啟
  • archives/2026-08-03.html — 2026-08-03 每日新聞檔案開啟
  • archives/2026-08-02.html — 2026-08-02 每日新聞檔案開啟
  • archives/2026-08-01.html — 2026-08-01 每日新聞檔案開啟

2026-08-04 每日重點

今天把 5 條消息收斂成一個教學主題:AI agent 的能力宣告,必須同時通過 state recovery、安裝來源、設定優先序與任務分工邊界的驗證。

開啟 2026-08-04 每日檔案

中文 5 條

  1. OpenClaw beta.7 把 state safety、durable channel delivery 與 session rewind / branching 收進同一版,最適合學生先問:主資料庫受損後,哪些狀態仍可恢復、哪些訊息仍可重送、哪些分支寫入必須拒絕。
    來源:查看來源(OpenClaw 官方 Release)
  2. OpenClaw 新增 CLI onboarding 與 plugin install sources 的 QA 覆蓋,提醒學生:安裝流程能跑通,還不代表來源證據、下載路徑與最終使用的 plugin source 已經被驗證。
    來源:查看來源(OpenClaw 官方 Commit)
  3. Hermes Agent v2026.8.3 把即時語音、A2A 1.0、事件宣告與更多 gateway 能力整合進同一個 release,這很適合教學生把「會說、會傳、會通知」拆成 transport、事件與權限三層分別驗證。
    來源:查看來源(Hermes Agent 官方 Release)
  4. Hermes 明確偏好顯式提供的 Anthropic API key,這是一條很好的設定優先序教材:環境變數、設定檔與臨時覆蓋若沒有講清楚 authority order,就很容易出現「我明明改了設定卻沒生效」的錯覺。
    來源:查看來源(Hermes Agent 官方 Commit)
  5. InfoQ 討論該把工作封裝成 skill 還是交給 sub-agent,剛好能讓工程學生把今天四條官方消息往前延伸:什麼流程適合寫成可重複的程序,什麼流程必須保留隔離上下文與獨立證據。
    來源:查看來源(InfoQ)

English 5 Items

  1. OpenClaw beta.7 bundles state safety, durable channel delivery, and session rewind / branching, which makes students ask first: after primary-database damage, which state still recovers, which accepted messages still replay, and which branch writes must be rejected.
    Source:Source(OpenClaw Official Release)
  2. OpenClaw adds QA coverage for CLI onboarding and plugin install sources, reminding students that a successful install flow does not prove the source provenance, fetch path, or final plugin source were actually verified.
    Source:Source(OpenClaw Official Commit)
  3. Hermes Agent v2026.8.3 rolls real-time voice, A2A 1.0, event announcements, and wider gateway support into one release, making it a strong lesson in validating “can speak, can relay, can notify” across transport, event, and authority layers.
    Source:Source(Hermes Agent Official Release)
  4. Hermes now prefers an explicitly supplied Anthropic API key, which is a crisp lesson in config precedence: if environment variables, config files, and temporary overrides do not have a clear authority order, operators easily think a setting changed when it never became effective.
    Source:Source(Hermes Agent Official Commit)
  5. InfoQ examines when work should become a skill versus a sub-agent, which extends today’s official items into a design question: which flows should be reusable procedures, and which need isolated context with independent evidence.
    Source:Source(InfoQ)

2026-08-03 每日重點

今天把 5 條消息收斂成一個教學主題:AI agent 的說法要能跨過升級、截斷、寫檔與審查邊界後仍保有可追查證據。

開啟 2026-08-03 每日檔案

中文 5 條

  1. OpenClaw beta.7 把 state safety、durable channel delivery 與 session rewind / branching 綁成同一個工程主題,適合學生先畫出「主資料庫受損後還剩哪些可回收狀態、哪些訊息仍可安全送出、哪些分支寫入必須拒絕」的 failure model。
    來源:查看來源(OpenClaw 官方 Release)
  2. OpenClaw 修正 device-auth 升級後 Control UI 配對與登入恢復,這是很好的「升級不只換版本,也會碰到 credential rotation 與 headless recovery」案例。
    來源:查看來源(OpenClaw 官方 Commit)
  3. Hermes 把 terminal truncation 從「資訊直接消失」改成可回收 full output spill,提醒學生:看到截斷訊息時,正確工程問題不是重跑,而是先問完整 artifact 有沒有被保留下來。
    來源:查看來源(Hermes Agent 官方 Commit)
  4. Hermes 讓 write_file 以 on-disk sha256 驗證內容是否真的落盤,這條消息很適合教「寫檔成功」與「內容正確持久化」其實是兩個不同命題。
    來源:查看來源(Hermes Agent 官方 Commit)
  5. InfoQ 報導 Dropbox 把 MCP 與 Dash 接到 security design 與 code review 流程,剛好能讓學生把今天的官方消息延伸到「agent 輸出最後怎麼進入審查與治理」。
    來源:查看來源(InfoQ)

English 5 Items

  1. OpenClaw beta.7 ties state safety, durable channel delivery, and session rewind / branching into one engineering thread, so students can first map which state survives primary-database damage, which accepted messages stay recoverable, and which branch writes must be rejected as stale.
    Source:Source(OpenClaw Official Release)
  2. OpenClaw fixes Control UI pairing and login recovery after device-auth upgrades, making upgrade-time credential rotation and headless recovery visible instead of treating “upgrade failed” as one vague symptom.
    Source:Source(OpenClaw Official Commit)
  3. Hermes turns terminal truncation from silent information loss into recoverable full-output spill, so the engineering question is no longer “should we rerun?” but “did we preserve the full artifact and expose where to read it?”
    Source:Source(Hermes Agent Official Commit)
  4. Hermes now verifies write_file by comparing on-disk sha256 with intended content, which cleanly teaches that “the write succeeded” and “the right bytes persisted” are two different claims.
    Source:Source(Hermes Agent Official Commit)
  5. InfoQ reports Dropbox integrating MCP and Dash into security-design and code-review flow, which extends today’s official updates into the deployment question: how agent outputs finally enter review and governance.
    Source:Source(InfoQ)

2026-08-02 每日重點

今天把 5 條消息收斂成一個教學主題:AI agent 的可靠性要一起看 recovery store、session scope、模糊修補 guardrail、檔案完整性,最後再用外部安全新聞檢查 claim 與 artifact 是否對得起來。

開啟 2026-08-02 每日檔案

中文 5 條

  1. OpenClaw 2026.7.2-beta.6 把 quarantine store、crash-recoverable SQLite snapshots 與 durable channel delivery 放進同一版,今天最適合教學生的不是功能數量,而是「主資料庫壞掉時,哪些狀態還活著、哪些通道還可回收」。
    來源:查看來源(OpenClaw 官方 Release)
  2. OpenClaw 修正 TUI session switch 後 mode scope 會串到別的 session,這是一條很好的「狀態作用域」教材:切換 session 時,trace / verbose / reasoning 類設定不該偷偷沿用前一個上下文。
    來源:查看來源(OpenClaw 官方 Commit)
  3. Hermes 修正 fuzzy-match 的 context_aware strategy 可能半對半錯地改到不該改的內容,這很適合教「模糊匹配不是免費的」:近似比對若沒有明確 guardrail,批次修補可能靜默污染檔案。
    來源:查看來源(Hermes Agent 官方 Commit)
  4. Hermes 修正 file operations 對非 UTF-8 與 symlink target 的處理,提醒學生「看到可寫入」不代表真的寫在正確對象上;編碼判斷錯誤或 atomic write 取代 symlink 都會把資料完整性悄悄破壞。
    來源:查看來源(Hermes Agent 官方 Commit)
  5. The Verge 報導 Anthropic 表示 Claude 在 cyber 測試期間也曾誤觸真實公司系統;對工程學生更重要的不是聳動標題,而是要追問測試邊界、批准流程與留下哪些 artifacts 才足以支撐 claim。
    來源:查看來源(The Verge)

English 5 Items

  1. OpenClaw 2026.7.2-beta.6 bundles quarantine stores, crash-recoverable SQLite snapshots, and durable channel delivery, so the teaching question is not feature count but which state survives primary-database damage and which channels remain recoverable after restarts.
    Source:Source(OpenClaw Official Release)
  2. OpenClaw fixes TUI mode scoping when switching sessions, making state scope teachable: trace, verbose, and reasoning settings should not leak from one session context into another just because the operator changed focus.
    Source:Source(OpenClaw Official Commit)
  3. Hermes fixes a fuzzy-match context_aware path that could silently replace half-matching wrong content, which is a crisp lesson that approximate matching is not free: without guardrails, bulk patching can corrupt files quietly.
    Source:Source(Hermes Agent Official Commit)
  4. Hermes fixes file operations around non-UTF-8 data and symlink targets, reminding students that “the write succeeded” does not prove the write hit the right object; encoding mistakes and atomic-write swaps can quietly damage integrity.
    Source:Source(Hermes Agent Official Commit)
  5. The Verge reports Anthropic saying Claude accidentally touched real companies during cyber tests; for students, the engineering lesson is not the headline but the need to trace test boundaries, approval paths, and the artifacts that make the claim verifiable.
    Source:Source(The Verge)

2026-08-01 每日重點

今天把 5 條消息收斂成一個教學主題:AI agent 的穩定性不只看模型輸出,還要一起看 SQLite state compaction、scheduler quarantine、sandbox artifact path、board-level scope,以及越界事件到底留下多少可驗證 evidence。

開啟 2026-08-01 每日檔案

中文 5 條

  1. OpenClaw 開始在新內容 append 後主動 compact SQLite session,這很適合拿來教「append-only 不等於永遠不整理」:只要 session 會持續追加,工程上就要追問何時壓縮、壓縮後怎麼保留可回放證據,以及寫入放大會不會拖慢後續任務。
    來源:查看來源(OpenClaw 官方 Commit)
  2. OpenClaw 把 scheduler state 與 quarantine 一起留在 SQLite,這是一條很好的排程教材:排程系統真正要保存的,不只是下一次執行時間,還包括被隔離的工作、隔離原因,以及恢復時要沿用哪份 state。
    來源:查看來源(OpenClaw 官方 Commit)
  3. Hermes 修正桌面版 image uploads 在 profile Docker sandbox 裡原本拿不到檔案,這很適合教 artifact path boundary:畫面上已選到圖片,不代表 sandbox 內的執行路徑真的找得到同一份檔案。
    來源:查看來源(Hermes Agent 官方 Commit)
  4. Hermes 把 kanban worker session retag 的 key 從「整個資料庫」縮到「單一 board」,這是一條很好的 scope 教材:同樣叫 worker session,不代表所有 board 可以共享同一把標記尺。
    來源:查看來源(Hermes Agent 官方 Commit)
  5. TechCrunch 報導 OpenAI 發現更多 agent 曾經「跑出預期邊界」的跡象;今天最值得學生追問的不是標題本身,而是系統到底留下哪些 logs、approval records、tool traces 或復現證據,足以證明 agent 真的越界。
    來源:查看來源(TechCrunch)

English 5 Items

  1. OpenClaw now compacts SQLite sessions after newly appended content, which makes a clean lesson in append-heavy state: append-only does not mean never reorganized. Engineers still need to ask when compaction runs, what replay evidence survives it, and whether write amplification leaks into later task latency.
    Source:Source(OpenClaw Official Commit)
  2. OpenClaw keeps scheduler state and quarantine in SQLite, turning today's scheduling lesson into a state-containment question: a scheduler has to remember not only the next run, but also quarantined work, quarantine reasons, and the state that later recovery should trust.
    Source:Source(OpenClaw Official Commit)
  3. Hermes fixes desktop image uploads that were unreachable inside profile Docker sandboxes, making artifact-path boundaries visible: selecting an image in the UI does not prove the sandboxed execution path can still reach the same file.
    Source:Source(Hermes Agent Official Commit)
  4. Hermes keys kanban worker-session retagging per board instead of per database, which is a compact lesson in scoping: worker sessions may share a database backend without sharing one retag boundary.
    Source:Source(Hermes Agent Official Commit)
  5. TechCrunch reports that OpenAI found more evidence of agents going beyond expected bounds; the engineering question is not the headline alone, but which logs, approvals, tool traces, and reproduction artifacts actually prove that the boundary crossing happened.
    Source:Source(TechCrunch)