榮民叔叔轉職攻城獅的藏書筆記·2d agoThe Same Spec, Ten Runs: What Justifies Making an Instruction a Team Default?TL;DR — Suppose a team adds a pre-submission self-check to an agent’s brief, and the next patch passes acceptance. Does that self-check…A response icon1A response icon1
榮民叔叔轉職攻城獅的藏書筆記·3d agoBefore One Agent Hands Work to Another: Authority, State, and RecoverySix Tokyo talks connect delegation, persistent workspaces, and recovery through one Kubernetes upgrade scenario.A response icon2A response icon2
榮民叔叔轉職攻城獅的藏書筆記·4d ago同一份規格，跑十次：我們憑什麼把一種交代變成團隊預設？一次成功之後，什麼證據才足以把一份檢查清單設為團隊預設？從四十份程式與公開規格，分清需求、驗收與方法比較。
榮民叔叔轉職攻城獅的藏書筆記·5d ago把工作交給另一個 Agent 之前：授權、狀態與復原從東京六場簡報與一個升級案例，追問交接後的權限、未知結果，以及什麼證據才算完成。A response icon1A response icon1
榮民叔叔轉職攻城獅的藏書筆記·Nov 5Green Is Not Done, Part 4 — A Tea Payment WalkthroughFollow one tea purchase from review comments to constraint tests, then examine what mutation scores of 28.6%, 85.7% and 100% actually prove.
榮民叔叔轉職攻城獅的藏書筆記·Nov 4After Tokyo AGNTCon: Connecting an Agent Is Only the BeginningA Tokyo trip and seven talks reveal why tool access still needs authority, recovery, and evidence of completion.A response icon1A response icon1
榮民叔叔轉職攻城獅的藏書筆記·Nov 3綠燈不是驗收（四）付款實作篇：買一瓶無糖純喫綠茶，從 Review 約束走到 mutation score一瓶無糖純喫綠茶的付款，如何從 Review 意見留下約束、寫成測試，再看懂 28.6%、85.7% 與 100% 的差別。
榮民叔叔轉職攻城獅的藏書筆記·Nov 2從東京 AGNTCon 回來：Agent 接上工具之後，工作才剛開始從受邀赴東京、午休實作到七場錄影，追問 Agent 行動的授權、失敗後的狀態與交付證據。A response icon1A response icon1
榮民叔叔轉職攻城獅的藏書筆記·Oct 29Green Is Not Done, Part 3 — The 34% SWE-Gate Found Behind a Green Build: Constraint Tests, pass^k…TL;DR — Part 3 of this four-part series asks what evidence should guide expanded agent authority. Across 75 Python repos and 303 patch…
榮民叔叔轉職攻城獅的藏書筆記·Oct 27綠燈不是驗收（三）可靠度篇：SWE-Gate 量測到的 34% — constraint tests、pass^k 與授權擴張的閘門TL;DR — 四部曲第三篇討論 reliability gate：授權要不要擴張，該依據什麼判斷。SWE-Gate 在 75 個 Python repo、303 個修補任務上量測到：功能測試通過的 644 個修補裡，有 221 個（34%）違反 reviewer…
