after-actionlisted

# Post-Mortem Write like you're leaving a note for your future self who will hit the same problem in 6 months. Human language, no AI slop. Specifics over summaries. ## Boundary Require a specific incident or difficult problem plus diagnosis/fix evidence. Preserve the chronology and failed attempts. Route a standalone reusable rule to codify; route a full completed-deliverable report with implementation and evaluation to `sanitize report`. ## Structure (5 sections, no more) 1. **背景** — What broke, when, and why it mattered. One sentence + one concrete observable symptom. 2. **踩过的坑** — What you tried that DIDN'T work. Each attempt: what you did, why you thought it would work, and why it actually didn't. These are the most valuable parts — they save the next person from retrying. 3. **最终方案** — What actually fixed it. Keep it short: the key change, the critical insight, the one thing that mattered. Code snippet if it helps (≤10 lines). 4. **判断** — Pick ONE: - `临时 hack` — Works now, needs revisiting. What's the real fix? - `彻底修复` — Root cause addressed, won't recur. - `值得沉淀` — Patterns worth capturing as a Canon pattern/incident, repo-local rule, or skill. 5. **沉淀**（如果是「值得沉淀」）— Canon target first: `patterns/`, `incidents/`, `tasks/`, or `raw/update-cards/`; repo-local `.agent-state/rules/mistakes.md` is compatibility only. ## 沉淀前强制自问（Prevention Gate） 沉淀节动手前，逐条回答三问。答不出来不许写规则： 1. 触发条件：什么时候会再犯？（必须具体，如“对动力学参数做 A/B 时”） 2. 动作：该做什么？（必须可执行，如“先断言初始速度≈0”） 3.