Menu
Run the harness-mem review daily action when the user invokes hm-review.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Run the harness-mem distill daily action when the user invokes hm-distill.
Run the harness-mem distill daily action when the user invokes hm-distill.
Evidence answerer for harness-mem memory admission and review. Use when grill-before-distill or review needs concrete support from packet evidence, observations, memory search, files, docs, tests, config, or command output. Does not write candidates or confirmed truth.
Architecture, product-boundary, roadmap, and long-lived-rule consultation for harness-mem memory admission and review. Use when a candidate's truth depends on design judgment rather than only evidence lookup. Does not write memory.
Standard memory admission mode for harness-mem (grill-me role). Runs automatically on distill — depth scales by risk, not a forced heavy loop every time. Three modes: deep interrogation, light checklist, lookback for stale truth. No writes. More than a review-stage helper: admission plus lookback.
Conversation autopilot for harness-mem. Use proactively when starting or resuming a task, when the user asks about previous decisions, when the user asks to remember a rule, when recent work becomes stable enough to preserve, or when a memory conflict appears.
| name | hm-review |
| description | Run the harness-mem review daily action when the user invokes hm-review. |
This is a user-invocable harness-mem Daily command. Follow the action below through the configured harness-mem MCP server; do not replace it with terminal maintenance commands.
通过 MCP 列出当前项目的待审候选记忆。AI 自动做质量判断并处理低风险候选，最后给用户一份复核摘要；用户只在结果不对时纠错。
Input: 可指定项目名（/hm-review bazi-apps）。省略则用 active project。
Core UX
不要把 raw candidates 直接做成逐条单选题或多选表单。用户不是分类器。
正确体验是：
1. 拉取 pending 候选
2. AI 自动判定每条候选的处理动作：confirm / reject / keep_pending / migrate
3. AI 直接执行低风险 reject / keep_pending / migrate 判断
4. 对高置信、低风险 confirm 可直接确认；对会改变长期行为的高风险 confirm 先列入待确认摘要
5. 最后给用户处理结果和待确认项，用户只在不满意时纠错
Steps
确认项目
get_project_status(project_root=<当前工作区>, host_client="grok") 读取 active project，并幂等检查项目 Hook拉所有 status=pending 的候选
list_candidates：
project_name=<project>status="pending"limit=100search_memory 找 pending；search_memory 默认只返回 truth 层（auto_confirmed / user_confirmed），不能作为审核列表来源AI 预审分组
每条候选先判断它属于哪类，而不是直接问用户是否纳入：
默认拒绝这些类型：
TeamCreate / SendMessage / TeamDelete / agent idle 之类工具运行问题，除非项目本身就在开发这些工具。/plan-eng-review、/plan-ceo-review、/plan-design-review、/plan-devex-review 等跨项目 review workflow，除非用户明确要记录为全局工作流记忆。自动处理策略
默认由 AI 直接处理，不要打开逐条单选/多选 UI，不要让用户逐条决定，也不要提供 "All Confirm / Only rule / All Reject" 这种选项菜单。
AI 应按以下规则执行：
执行 MCP 操作
govern_memory(action="decide", arguments={kind:"rule", ...})govern_memory(action="decide", arguments={kind:"memory", ...})govern_memory(action="decide", arguments={kind:"relation", ...})keep_pending 不调用工具。最终复核摘要
最后只给用户看处理结果，不再让用户逐条选择：
已自动处理：confirmed A，rejected B，kept pending C。
需要你最终确认：D 条高风险确认建议。
如果有处理不对，告诉我编号，我会改。
总结呈现
confirmed Nrejected Mstill pending K/hm-distill 应避免同类候选。Notes