引擎本身不认识"销售""教育""情感陪伴"。所有业务语义——阶段、意图、公式、口吻、关系——都收进一份 DomainProfile。开箱即用的默认画像与销售域字节等价；要做别的行业，配置另一份画像即可，引擎代码一行不改。
The engine knows nothing of "sales," "education" or "companionship." All business semantics — stages, intent, formulas, voice, relationships — live in one DomainProfile. The shipped default is byte-equivalent to the sales domain; another industry is just another profile, with zero engine changes.
DomainProfile 是引擎读取业务语义的唯一入口：可配置维度（阶段 / 意图等）、承诺词、覆盖维度、对话模式、运营模式、四个方法论公式、记忆维度、成效极性、知识切片角色、阈值覆盖、评审取向、灵魂与方法论覆盖——全部在这一份结构里。
DomainProfile is the engine's only entry to business meaning: configurable dimensions (stage / intent), commitment markers, coverage dimensions, conversation modes, operation mode, four methodology formulas, memory dimensions, outcome polarity, chunk roles, threshold overrides, reviewer orientation, soul and methodology overrides — all in one structure.
默认画像的 id 是 __default__，被设计为逐字复刻销售域，并由一组"字节等价"护栏测试锁死——意味着默认行为稳定可预期，而任何新行业都从复制、改写这份画像开始。
The default profile id is __default__, designed to reproduce the sales domain verbatim and locked by byte-equivalence guard tests — so default behavior is stable and predictable, and every new industry starts by copying and editing this profile.
客户阶段、意图等级等的取值集，来自标签字典；新取值收集为候选待审，不阻塞决策。
Value sets for stage, intent etc. from the taxonomy; new values become review candidates, never block.
信任、成交就绪、情感价值、下一步动作分——四个文字表达式，由 LLM 自评。
Trust, conversion-readiness, emotional value, next-best-action — four text expressions, self-scored by the LLM.
不同业务的对话基调与节奏由画像定义，而非写死在代码里。
Each business's conversational tone and cadence is defined by the profile, not the code.
事实风险、压力风险等闸门阈值可按行业覆盖，并被钳制在合法区间。
Gate thresholds like fact/pressure risk are overridable per industry, clamped to a valid range.
什么算"好结果"因业务而异——成交、复诊、续费或仅是被信任地陪伴。
What counts as a good outcome differs — a deal, a revisit, a renewal, or simply trusted company.
知识切片在本行业里扮演的角色（卖点、政策、案例…）由画像声明。
The role each knowledge slice plays in this industry (selling point, policy, case…) is profile-declared.
独立评审在本行业更看重什么、对低风险自报是否更不信任，可配置。
What the independent reviewer weights, and whether to distrust self-reported low risk, is configurable.
整体人格基调可按行业覆盖，决定 AI 在这条业务线上"是谁"。
The overall persona can be overridden per industry — who the AI is on this business line.
下面是几类典型私域场景。代码里已带销售与情感陪伴两类示例画像；其余方向都按同样的方式——复制默认画像、改写维度与公式、配置知识库——即可适配，无需改引擎。
A few typical private-domain scenarios. Sales and companionship ship as sample profiles in code; the rest adapt the same way — copy the default, edit dimensions and formulas, configure the knowledge base — with no engine changes.
长周期培育、信任建立、成交就绪度评估。产品声明必须有可核验知识支撑，否则发送被拦截。这是默认画像直接覆盖的场景。
Long-cycle nurturing, trust-building, conversion-readiness scoring. Product claims must be grounded or the send is blocked. Covered directly by the default profile.
试听、报名、续费的阶段链路，家长决策与学员体验双线。把"成交就绪"改写为"报名就绪"，知识库换成课程与师资。
Trial → enroll → renew stages, with parent decisions and student experience in parallel. Rewrite "conversion" as "enrollment," swap the knowledge base for courses and faculty.
不追求成交，追求被信任地长期陪伴。情感价值成为核心公式，成效极性也随之改变。代码里已有最小示例画像。
Not chasing deals but trusted, long-term company. Emotional value becomes the core formula and outcome polarity shifts. A minimal sample profile ships in code.
合规是命门：调高事实风险与产品准确度阈值，把"绝不替客户做医疗判断"写进灵魂与策略，超界事项一律请示。
Compliance is everything: raise fact-risk and product-accuracy thresholds, bake "never make medical calls for the client" into soul and policy, escalate anything out of scope.
目标是把线上对话变成到店。开启辅助模式后，识别到"想到店参观"的客户可由 AI 主动引荐真人顾问名片完成临门一脚。
The goal is turning chats into visits. With assist mode on, customers signaling "want to visit" can be handed a human advisor's card by the AI for the last step.
法律、财税、保险等强专业域：知识切片角色重写为"条款 / 案例 / 风险提示"，超出职权的判断走幕后请示，再由 AI 转述结论。
Law, tax, insurance and other expert domains: rewrite chunk roles as "clause / case / risk note," route out-of-mandate calls through the principal, and relay the verdict via the AI.
把数字分身拆成两条正交的轴，互不耦合：触达轴决定对不同关系采取什么运营模式，口吻轴决定 AI 用什么语气说话。两者可以任意组合。
The digital twin splits into two orthogonal, decoupled axes: the reach axis sets the operation mode per relationship; the voice axis sets how the AI speaks. They combine freely.
联系人的关系类型来自全局标签字典的三个 canonical 值。画像可为每种关系映射不同的运营模式——比如对客户主动跟进，对同行只被动回复，对朋友自然寒暄。
A contact's relationship type comes from three canonical values in the global taxonomy. The profile maps each to its own operation mode — proactively follow up customers, only react to peers, chat naturally with friends.
每个联系人都可写一段自然语言的专属指令。它在系统提示词的最末位注入，被声明为最高优先级，与灵魂、策略冲突时以它为准——让 AI 对这个人"换一种说法"，而不必新建一套人格。
Each contact can carry a natural-language instruction. It is injected last in the system prompt, declared highest priority, overriding soul and policy on conflict — so the AI "talks differently" to this person without a whole new persona.
两轴正交：你可以对一位朋友关系的人，配置主动问候的触达模式，同时叠加一段更随性口语的专属口吻——互不干扰。
Orthogonal axes: a friend can get a proactive-greeting reach mode plus a more casual bespoke voice — independently.
关系类型不是只能手工打标。AI 在对话中识别出关系类型建议，但建议要经过字典校验、落入待审队列，运营确认后才生效——AI 的判断从不直接改写联系人。
Relationship type isn't manual-only. The AI proposes a type from conversation, but the suggestion is dictionary-validated, queued for review, and takes effect only after operator confirmation — the AI never rewrites the contact directly.
决策时向 Agent 注入"关系类型建议"的引导，让它在思考里给出判断。
At decision time the agent is prompted to suggest a relationship type in its reasoning.
网关从决策结果里提取关系建议，进入校验前不做任何写入。
The gateway extracts the suggestion; nothing is written before validation.
必须命中标签字典的合法取值，越界直接拒绝——AI 不能发明新关系类型。
Must hit a valid taxonomy value; out-of-range is rejected — no inventing new types.
写入建议队列、标记待审，运营在收件箱确认后才作用于联系人。
Written as pending; only applied to the contact after operator confirmation in the inbox.
为什么这样设计：关系类型会改变运营模式，是高影响动作。让 AI 识别（省去运营逐个手标），但把"是否生效"留给人——既通用化又不失控。这与"AI 自创标签先落候选待审"是同一条铁律。
Why: relationship type changes the operation mode — a high-impact action. The AI detects (sparing manual tagging) but humans decide whether it takes effect — universal yet controlled. Same rule as "AI-proposed tags land as candidates first."
默认全自治模式下，客户永远只跟 AI 对话。辅助模式是账号级、默认关闭的受控例外：当 AI 判定客户契合人类预先标注的引荐条件，它会主动把真人专属顾问的名片推送过去，自己退为辅助答疑——发起方与转述方仍是 AI。
By default the customer only ever talks to the AI. Assist mode is an account-level, off-by-default exception: when the AI judges a customer fits human-annotated referral conditions, it proactively pushes a human advisor's card and steps back to support — the AI is still the initiator.
辅助模式默认 false，必须由账号显式打开；还可对单个客户用 assist_mode_override 强制开 / 关。
Assist mode defaults to false and must be enabled per account; a single customer can be forced on/off via assist_mode_override.
顾问名片库里录入触发提示（如"明确要签约 / 到店参观"），作为候选清单注入 AI 的决策上下文。
The advisor card library carries trigger hints (e.g. "clearly wants to sign / visit"), injected as a candidate list into the AI's decision context.
提示是软引导而非硬过滤门——由 AI 在对话语境中自主选择是否、何时引荐，运营 Agent 因此更简单。
Hints are soft guidance, not a hard filter — the AI chooses whether and when to refer, keeping the ops agent simpler.
经 MCP 工具 message_send_namecard 发送真人名片，由客户与顾问对接完成临门一脚。
Sends the human card via the MCP tool message_send_namecard, letting the customer connect with the advisor for the final step.
被引荐的台前顾问 ≠ 幕后决策源（领导），两者解耦。名片是 AI 主动引荐的"发送物"，对话始终是 AI 在说。全自治模式（默认）下"客户永远只跟 AI 对话"的红线一字不动。
The front-stage advisor is not the behind-the-scenes principal — the two are decoupled. The card is something the AI sends; the conversation is always the AI speaking. The default-mode red line — customers only ever talk to the AI — stays untouched.
"无人工接管"不等于"AI 什么都自己拍板"。遇到超出职权的事，AI 向幕后决策源（领导）请示，拿回结论后用自己的口吻向客户转述——客户从不直接面对真人，对话始终是 AI 在说。这是请示，不是接管。
"No human handoff" doesn't mean "the AI decides everything." For out-of-mandate matters the AI consults a behind-the-scenes principal, then relays the verdict to the customer in its own voice — the customer never faces a human. This is consultation, not handoff.
客户提出的事项超出 AI 被授予的决策权限——价格让步、特殊承诺、政策外请求等，须由决策源拍板。
The matter exceeds the AI's granted authority — price concessions, special promises, off-policy requests — and needs the principal's call.
评审闸门判定动作高风险被拦下，需要人来确认是否放行，而不是 AI 自行强推。
A review gate flags the action as high-risk and blocks it, needing human confirmation rather than the AI forcing it through.
对话陷入僵局或消息长期未送达，AI 主动上报，避免客户被"晾在那里"。
The conversation stalls or a message stays undelivered; the AI raises it proactively so the customer isn't left hanging.
请示状态在 pending → resolved 之间流转；占位等待期间 AI 给出的兜底回复不含任何转接类措辞——红线在代码层就被守住。
Escalation flows pending → resolved; while waiting, the AI's holding reply contains no handoff wording at all — the red line is enforced in code.
系统按两层键隔离：workspace 是租户边界，account 是单个微信号。几乎所有数据查询都带双键 scope，从根上防止跨租户、跨账号串数据。多账号是一等公民，由专门的调度器轮转。
Two isolation keys: workspace is the tenant boundary, account is a single WeChat number. Nearly every query is scoped by both keys, preventing cross-tenant and cross-account leakage by construction. Multi-account is a first-class citizen, rotated by a dedicated scheduler.
画像、灵魂、提示词包、知识库、请示通道全部按 workspace 隔离。不同客户、不同业务线之间互不可见，每个租户都有自己的一套配置与数据。
Profiles, souls, prompt packs, knowledge bases and escalation channels are all isolated per workspace. Different clients and business lines can't see each other; each tenant owns its config and data.
一个 workspace 下可挂多个微信号，由多账号调度器轮转运营。Webhook 按 account 路由入站消息，冷启动联系人按账号归属处理，互不串台。
A workspace can hold many WeChat accounts, rotated by the multi-account scheduler. Webhooks route inbound by account and cold-start contacts are handled by ownership — no crosstalk.
无论换到哪个行业、对哪种关系、用什么口吻，同一套红线与闸门都不会松动。这正是 WeAgent 敢全自治的底气。
Whatever the industry, relationship or voice, the same red lines and gates never loosen. That's what lets WeAgent run fully autonomously.