Call before any coding, adaptation, debugging, or design task. Returns prohibitions you must follow, related items, and a skill catalog (name/description/path only). Use feedback_token with rsi_feedback before you wrap up. If decisions is non-empty, call rsi_conflicts / rsi_knowledge_review immediately; do not ask the user to run rsi in a terminal. 任务开始前调用。返回必须遵守的 prohibitions、相关 items，以及技能目录。用 feedback_token 在收工前调用 rsi_feedback。有 decisions 时立刻调 rsi_conflicts / rsi_knowledge_review，不要让用户去终端跑 rsi。

Navigate the knowledge tree by domain module. Empty path lists top-level modules with counts; pass a path (e.g. backend or backend/api) to drill down and list member items (title/brief/id), plus graph edges (related / conflicts) touching this node's items. Use rsi_recall to pull full content after you find the node you want. 按领域板块浏览知识树：空 path 列顶层板块与计数；传 path（如 backend 或 backend/api）下钻列出成员条目（标题/简述/id）并附本节点条目的图谱关联（related）与冲突（conflicts）；需要正文用 rsi_recall。

Submit feedback on a recall: explicit rating or IDE implicit action (accepted/applied/modified/copied/referenced/ignored/rejected). On rejected, a comment is extracted as a prohibition. 对召回结果提交反馈：显式评分或 IDE 隐式行为（accepted/applied/modified/copied/referenced/ignored/rejected）；rejected 时填 comment 将提炼为禁止项。

Add a knowledge item (title + content required). 添加知识条目（title + content 必填）。

Search the knowledge base (FTS5 + CJK bigram by default; embedding is optional). 搜索知识库（默认纯 FTS5 + CJK bigram；embedding 为可选增强）。

Delete a knowledge item (deletion right). 删除指定知识条目（删除权）。

Update an existing memory in place by id (title/content/tags/roles/domain/description/trigger_condition/expiry_condition) — prefer updating over re-adding duplicates. action=merge combines near-duplicates: keep the more complete one, archive the rest (recoverable, not hard-deleted) and record a supersedes edge. Preserves id and location. 按 id 原地更新已有记忆（改 title/content/tags/roles/domain/description/前提与失效条件），改优于新建重复。action=merge 合并近重复：保留更完整者、其余归档（可恢复非硬删）并写 supersedes 边；保留 id 与所在目录。

Approve pending norms: approve moves to official dirs and injects prohibitions; reject archives. Does not create new bodies. Approved conventions are recall-only (no rsi-convention rule files). 审批 pending 规范：approve 搬到正式目录并注入禁止项，reject 搬到 archive。不新建正文。约定获准后只进召回，不会写成 rsi-convention 规则文件。

Harness improvement proposals: list/approve/reject/rollback/generate and config snapshots snapshot_list/snapshot_switch/snapshot_export. Harness 改进提案：list/approve/reject/rollback/generate 及配置快照 snapshot_list/snapshot_switch/snapshot_export。

List or resolve suspected conflicts. Three kinds: (1) learned memory vs handwritten rules (contradiction/stale/overlap) — user_wins keeps your rule (learned memory no longer injected), memory_wins keeps learned memory (edit the rule file yourself), coexist keeps both and stops reminding; (2) knowledge pairs (version/doc_code/incoherent) — keep_item keeps item_id and archives the peer source_url, keep_peer keeps the user_rule_path side, coexist keeps both; (3) explain (conflict_id required) expands both sides read-only. Call explain first when the user wants detail; do not close the card. 查询疑似冲突并对冲突做出裁决。三类：① 学习记忆 vs 手写规则（contradiction/stale/overlap）——user_wins=以你的规则为准（学习记忆不再注入）、memory_wins=以学习记忆为准（返回规则文件位置请手动修改）、coexist=两者共存不再提醒；② 知识对（version/doc_code/incoherent）——keep_item=保留 item_id 侧并归档对侧 source_url、keep_peer=保留 user_rule_path 侧、coexist=两侧转/保持可用且不再提醒；③ explain（conflict_id 必填）=只读展开两侧摘录与各 option 影响，不改库。用户要展开说明时先 explain，不要关掉抉择卡。

Memory usage stats: period=day/week/month summaries of reach, adoption, prohibition follow-through, and proposal pass rate. 记忆使用统计：period=day/week/month 汇总触达/采纳/禁止项遵循/提案通过率。

Value/impact panel for evangelism: injected cost (chars + est tokens), recall value (hit rate + on-demand exposure vs full corpus), behavioral gains (adoption, compliance, proposal pass), a time curve, and an A/B token-savings calculator. Honesty: RSI cannot see the host's real token usage, so 'tokens saved' is ONLY produced from a user-supplied before/after comparison (ab={baseline:{tokens,tasks},treatment:{tokens,tasks}}); without it no savings number is invented. export=true writes a one-page markdown to ~/.rsi/exports/. 价值/影响力面板（供宣讲）：注入成本（字符+token 估算）、召回价值（命中率与按需暴露 vs 全库）、行为收益（采纳/合规/提案通过率）、时间曲线，以及一个 A/B 省 token 计算器。诚实原则：RSI 看不到宿主真实 token 用量，‘省了多少’只在用户喂入用前/用后对照（ab 参数）时才产出，否则不编造任何省量；export=true 导出一页纸 markdown。

Teaching / audit / closeout learn / promote / reading-pack actions. The host agent writes the lesson (Catch→Teach→Fix) then teach_record; RSI does not write the lesson. pack_list / pack_open / pack_done walk reading packs. Unavailable actions return not_in_phase — follow message. 教学/审计/收工学习/晋升/阅读包。Agent 自己 Catch→Teach→Fix 后 teach_record；不要让 RSI 代写 lesson。pack_list / pack_open / pack_done 串联阅读包。未开放的 action 返回 not_in_phase，按 message 改用其它工具。

Memory health findings: scan/list/explain/resolve. scan attributes a rejected or low-rated recall to every retrieved id, opens incorrect when comment polarity contradicts the memory, and opens stale_source when a changed source file no longer contains the title phrase. resolve is archive, revise, keep, or snooze. Scan never archives by itself. 记忆健康发现：scan/list/explain/resolve。scan 把拒绝或低分记到本次召回的每一条记忆上；评语极性与记忆相反时开 incorrect，来源已变且标题短语不在文件中时开 stale_source。resolve 取 archive、revise、keep、snooze。扫描本身不会归档。

Open or search memory files. index lists entries, open reads one id, reindex rebuilds cache, graph returns mermaid text. 打开或检索记忆文件。index 列目录，open 按 id 读全文，reindex 重建缓存，graph 返回 mermaid 文本。