# mark-sdk - Doramagic AI Context Pack

> 定位：安装前体验与判断资产。它帮助宿主 AI 有一个好的开始，但不代表已经安装、执行或验证目标项目。

## 充分原则

- **充分原则，不是压缩原则**：AI Context Pack 应该充分到让宿主 AI 在开工前理解项目价值、能力边界、使用入口、风险和证据来源；它可以分层组织，但不以最短摘要为目标。
- **压缩策略**：只压缩噪声和重复内容，不压缩会影响判断和开工质量的上下文。

## 给宿主 AI 的使用方式

你正在读取 Doramagic 为 mark-sdk 编译的 AI Context Pack。请把它当作开工前上下文：帮助用户理解适合谁、能做什么、如何开始、哪些必须安装后验证、风险在哪里。不要声称你已经安装、运行或执行了目标项目。

## Claim 消费规则

- **事实来源**：Repo Evidence + Claim/Evidence Graph；Human Wiki 只提供显著性、术语和叙事结构。
- **事实最低状态**：`supported`
- `supported`：可以作为项目事实使用，但回答中必须引用 claim_id 和证据路径。
- `weak`：只能作为低置信度线索，必须要求用户继续核实。
- `inferred`：只能用于风险提示或待确认问题，不能包装成项目事实。
- `unverified`：不得作为事实使用，应明确说证据不足。
- `contradicted`：必须展示冲突来源，不得替用户强行选择一个版本。

## 它最适合谁

- **希望把专业流程带进宿主 AI 的用户**：仓库包含 Skill 文档。 证据：`src/mark/middlewares/skills/mark-memory-middleware/SKILL.md`, `src/mark/middlewares/skills/mark-usage/SKILL.md` Claim：`clm_0003` supported 0.86

## 它能做什么

- **AI Skill / Agent 指令资产库**（可做安装前预览）：项目包含可被宿主 AI 读取的 Skill 或 Agent 指令文件，可用于把专业流程带入 Claude、Codex、Cursor 等宿主。 证据：`src/mark/middlewares/skills/mark-memory-middleware/SKILL.md`, `src/mark/middlewares/skills/mark-usage/SKILL.md` Claim：`clm_0001` supported 0.86
- **命令行启动或安装流程**（需要安装后验证）：项目文档中存在可执行命令，真实使用需要在本地或宿主环境中运行这些命令。 证据：`README.md` Claim：`clm_0002` supported 0.86

## 怎么开始

- `pip install mark-sdk` 证据：`README.md` Claim：`clm_0004` supported 0.86, `clm_0005` supported 0.86, `clm_0006` supported 0.86, `clm_0007` supported 0.86 等
- `pip install "mark-sdk[langchain]"   # LangChain tools + agent middleware` 证据：`README.md` Claim：`clm_0005` supported 0.86
- `pip install "mark-sdk[mcp]"        # MCP server over MARK memory` 证据：`README.md` Claim：`clm_0006` supported 0.86
- `pip install "mark-sdk[adapters]"   # both of the above` 证据：`README.md` Claim：`clm_0007` supported 0.86
- `pip install "mark-sdk[middleware]" # all local middleware batteries` 证据：`README.md` Claim：`clm_0008` supported 0.86

## 继续前判断卡

- **当前建议**：仅建议沙盒试装
- **为什么**：项目存在安装命令、宿主配置或本地写入线索，不建议直接进入主力环境，应先在隔离环境试装。

### 30 秒判断

- **现在怎么做**：仅建议沙盒试装
- **最小安全下一步**：先跑 Prompt Preview；若仍要安装，只在隔离环境试装
- **先别相信**：真实输出质量不能在安装前相信。
- **继续会触碰**：命令执行、宿主 AI 配置、本地环境或项目文件

### 现在可以相信

- **适合人群线索：希望把专业流程带进宿主 AI 的用户**（supported）：有 supported claim 或项目证据支撑，但仍不等于真实安装效果。 证据：`src/mark/middlewares/skills/mark-memory-middleware/SKILL.md`, `src/mark/middlewares/skills/mark-usage/SKILL.md` Claim：`clm_0003` supported 0.86
- **能力存在：AI Skill / Agent 指令资产库**（supported）：可以相信项目包含这类能力线索；是否适合你的具体任务仍要试用或安装后验证。 证据：`src/mark/middlewares/skills/mark-memory-middleware/SKILL.md`, `src/mark/middlewares/skills/mark-usage/SKILL.md` Claim：`clm_0001` supported 0.86
- **能力存在：命令行启动或安装流程**（supported）：可以相信项目包含这类能力线索；是否适合你的具体任务仍要试用或安装后验证。 证据：`README.md` Claim：`clm_0002` supported 0.86
- **存在 Quick Start / 安装命令线索**（supported）：可以相信项目文档出现过启动或安装入口；不要因此直接在主力环境运行。 证据：`README.md` Claim：`clm_0004` supported 0.86, `clm_0005` supported 0.86, `clm_0006` supported 0.86, `clm_0007` supported 0.86

### 现在还不能相信

- **真实输出质量不能在安装前相信。**（unverified）：Prompt Preview 只能展示引导方式，不能证明真实项目中的结果质量。
- **宿主 AI 版本兼容性不能在安装前相信。**（unverified）：Claude、Cursor、Codex、Gemini 等宿主加载规则和版本差异必须在真实环境验证。
- **不会污染现有宿主 AI 行为，不能直接相信。**（inferred）：Skill、plugin、AGENTS/CLAUDE/GEMINI 指令可能改变宿主 AI 的默认行为。 证据：`AGENTS.md`, `src/mark/middlewares/skills/mark-memory-middleware/SKILL.md`, `src/mark/middlewares/skills/mark-usage/SKILL.md`
- **可安全回滚不能默认相信。**（unverified）：除非项目明确提供卸载和恢复说明，否则必须先在隔离环境验证。
- **真实安装后是否与用户当前宿主 AI 版本兼容？**（unverified）：兼容性只能通过实际宿主环境验证。
- **项目输出质量是否满足用户具体任务？**（unverified）：安装前预览只能展示流程和边界，不能替代真实评测。
- **安装命令是否需要网络、权限或全局写入？**（unverified）：这影响企业环境和个人环境的安装风险。 证据：`README.md`

### 继续会触碰什么

- **命令执行**：包管理器、网络下载、本地插件目录、项目配置或用户主目录。 原因：运行第一条命令就可能产生环境改动；必须先判断是否值得跑。 证据：`README.md`
- **宿主 AI 配置**：Claude/Codex/Cursor/Gemini/OpenCode 等宿主的 plugin、Skill 或规则加载配置。 原因：宿主配置会改变 AI 后续工作方式，可能和用户已有规则冲突。 证据：`AGENTS.md`, `src/mark/middlewares/skills/mark-memory-middleware/SKILL.md`, `src/mark/middlewares/skills/mark-usage/SKILL.md`
- **本地环境或项目文件**：安装结果、插件缓存、项目配置或本地依赖目录。 原因：安装前无法证明写入范围和回滚方式，需要隔离验证。 证据：`README.md`
- **宿主 AI 上下文**：AI Context Pack、Prompt Preview、Skill 路由、风险规则和项目事实。 原因：导入上下文会影响宿主 AI 后续判断，必须避免把未验证项包装成事实。

### 最小安全下一步

- **先跑 Prompt Preview**：用安装前交互式试用判断工作方式是否匹配，不需要授权或改环境。（适用：任何项目都适用，尤其是输出质量未知时。）
- **只在隔离目录或测试账号试装**：避免安装命令污染主力宿主 AI、真实项目或用户主目录。（适用：存在命令执行、插件配置或本地写入线索时。）
- **先备份宿主 AI 配置**：Skill、plugin、规则文件可能改变 Claude/Cursor/Codex 的默认行为。（适用：存在插件 manifest、Skill 或宿主规则入口时。）
- **安装后只验证一个最小任务**：先验证加载、兼容、输出质量和回滚，再决定是否深用。（适用：准备从试用进入真实工作流时。）

### 退出方式

- **保留安装前状态**：记录原始宿主配置和项目状态，后续才能判断是否可恢复。
- **准备移除宿主 plugin / Skill / 规则入口**：如果试装后行为异常，可以把宿主 AI 恢复到试装前状态。
- **记录安装命令和写入路径**：没有明确卸载说明时，至少要知道哪些目录或配置需要手动清理。
- **如果没有回滚路径，不进入主力环境**：不可回滚是继续前阻断项，不应靠信任或运气继续。

## 哪些只能预览

- 解释项目适合谁和能做什么
- 基于项目文档演示典型对话流程
- 帮助用户判断是否值得安装或继续研究

## 哪些必须安装后验证

- 真实安装 Skill、插件或 CLI
- 执行脚本、修改本地文件或访问外部服务
- 验证真实输出质量、性能和兼容性

## 边界与风险判断卡

- **把安装前预览误认为真实运行**：用户可能高估项目已经完成的配置、权限和兼容性验证。 处理方式：明确区分 prompt_preview_can_do 与 runtime_required。 Claim：`clm_0009` inferred 0.45
- **命令执行会修改本地环境**：安装命令可能写入用户主目录、宿主插件目录或项目配置。 处理方式：先在隔离环境或测试账号中运行。 证据：`README.md` Claim：`clm_0010` supported 0.86
- **待确认**：真实安装后是否与用户当前宿主 AI 版本兼容？。原因：兼容性只能通过实际宿主环境验证。
- **待确认**：项目输出质量是否满足用户具体任务？。原因：安装前预览只能展示流程和边界，不能替代真实评测。
- **待确认**：安装命令是否需要网络、权限或全局写入？。原因：这影响企业环境和个人环境的安装风险。

## 开工前工作上下文

### 加载顺序

- 先读取 how_to_use.host_ai_instruction，建立安装前判断资产的边界。
- 读取 claim_graph_summary，确认事实来自 Claim/Evidence Graph，而不是 Human Wiki 叙事。
- 再读取 intended_users、capabilities 和 quick_start_candidates，判断用户是否匹配。
- 需要执行具体任务时，优先查 role_skill_index，再查 evidence_index。
- 遇到真实安装、文件修改、网络访问、性能或兼容性问题时，转入 risk_card 和 boundaries.runtime_required。

### 任务路由

- **AI Skill / Agent 指令资产库**：先基于 role_skill_index / evidence_index 帮用户挑选可用角色、Skill 或工作流。 边界：可做安装前 Prompt 体验。 证据：`src/mark/middlewares/skills/mark-memory-middleware/SKILL.md`, `src/mark/middlewares/skills/mark-usage/SKILL.md` Claim：`clm_0001` supported 0.86
- **命令行启动或安装流程**：先说明这是安装后验证能力，再给出安装前检查清单。 边界：必须真实安装或运行后验证。 证据：`README.md` Claim：`clm_0002` supported 0.86

### 上下文规模

- 文件总数：158
- 重要文件覆盖：40/158
- 证据索引条目：80
- 角色 / Skill 条目：2

### 证据不足时的处理

- **missing_evidence**：说明证据不足，要求用户提供目标文件、README 段落或安装后验证记录；不要补全事实。
- **out_of_scope_request**：说明该任务超出当前 AI Context Pack 证据范围，并建议用户先查看 Human Manual 或真实安装后验证。
- **runtime_request**：给出安装前检查清单和命令来源，但不要替用户执行命令或声称已执行。
- **source_conflict**：同时展示冲突来源，标记为待核实，不要强行选择一个版本。

## Prompt Recipes

### 适配判断

- 目标：判断这个项目是否适合用户当前任务。
- 预期输出：适配结论、关键理由、证据引用、安装前可预览内容、必须安装后验证内容、下一步建议。

```text
请基于 mark-sdk 的 AI Context Pack，先问我 3 个必要问题，然后判断它是否适合我的任务。回答必须包含：适合谁、能做什么、不能做什么、是否值得安装、证据来自哪里。所有项目事实必须引用 evidence_refs、source_paths 或 claim_id。
```

### 安装前体验

- 目标：让用户在安装前感受核心工作流，同时避免把预览包装成真实能力或营销承诺。
- 预期输出：一段带边界标签的体验剧本、安装后验证清单和谨慎建议；不含真实运行承诺或强营销表述。

```text
请把 mark-sdk 当作安装前体验资产，而不是已安装工具或真实运行环境。

请严格输出四段：
1. 先问我 3 个必要问题。
2. 给出一段“体验剧本”：用 [安装前可预览]、[必须安装后验证]、[证据不足] 三种标签展示它可能如何引导工作流。
3. 给出安装后验证清单：列出哪些能力只有真实安装、真实宿主加载、真实项目运行后才能确认。
4. 给出谨慎建议：只能说“值得继续研究/试装”“先补充信息后再判断”或“不建议继续”，不得替项目背书。

硬性边界：
- 不要声称已经安装、运行、执行测试、修改文件或产生真实结果。
- 不要写“自动适配”“确保通过”“完美适配”“强烈建议安装”等承诺性表达。
- 如果描述安装后的工作方式，必须使用“如果安装成功且宿主正确加载 Skill，它可能会……”这种条件句。
- 体验剧本只能写成“示例台词/假设流程”：使用“可能会询问/可能会建议/可能会展示”，不要写“已写入、已生成、已通过、正在运行、正在生成”。
- Prompt Preview 不负责给安装命令；如用户准备试装，只能提示先阅读 Quick Start 和 Risk Card，并在隔离环境验证。
- 所有项目事实必须来自 supported claim、evidence_refs 或 source_paths；inferred/unverified 只能作风险或待确认项。

```

### 角色 / Skill 选择

- 目标：从项目里的角色或 Skill 中挑选最匹配的资产。
- 预期输出：候选角色或 Skill 列表，每项包含适用场景、证据路径、风险边界和是否需要安装后验证。

```text
请读取 role_skill_index，根据我的目标任务推荐 3-5 个最相关的角色或 Skill。每个推荐都要说明适用场景、可能输出、风险边界和 evidence_refs。
```

### 风险预检

- 目标：安装或引入前识别环境、权限、规则冲突和质量风险。
- 预期输出：环境、权限、依赖、许可、宿主冲突、质量风险和未知项的检查清单。

```text
请基于 risk_card、boundaries 和 quick_start_candidates，给我一份安装前风险预检清单。不要替我执行命令，只说明我应该检查什么、为什么检查、失败会有什么影响。
```

### 宿主 AI 开工指令

- 目标：把项目上下文转成一次对话开始前的宿主 AI 指令。
- 预期输出：一段边界明确、证据引用明确、适合复制给宿主 AI 的开工前指令。

```text
请基于 mark-sdk 的 AI Context Pack，生成一段我可以粘贴给宿主 AI 的开工前指令。这段指令必须遵守 not_runtime=true，不能声称项目已经安装、运行或产生真实结果。
```

## 角色 / Skill 索引

- 共索引 2 个角色 / Skill / 项目文档条目。

- **mark-memory-middleware**（skill）：Use when a LangChain agent is wrapped with MARK middleware and should treat MARK as fast, transparent project memory for recall, grounding, and durable observation. 激活提示：当用户任务与“mark-memory-middleware”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`src/mark/middlewares/skills/mark-memory-middleware/SKILL.md`
- **mark-usage**（skill）：Use when an agent should adopt MARK as its persistent memory, context, reflection, and continuity layer. 激活提示：当用户任务与“mark-usage”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`src/mark/middlewares/skills/mark-usage/SKILL.md`

## 证据索引

- 共索引 80 条证据。

- **MARK Memory-Augmented Agent Recall Kit Python SDK**（documentation）：MARK Memory-Augmented Agent Recall Kit Python SDK 证据：`README.md`
- **MARK SDK Examples**（documentation）：These examples are developer usage tests and public proof points for MARK as an agent memory runtime. They show the SDK doing more than storing facts: injecting context into work, preserving session continuity, preparing safe sync envelopes, and exposing the same memory through agent integrations. 证据：`examples/README.md`
- **Contributing to MARK SDK**（documentation）：Thanks for your interest in making agent memory better. Contributions of all sizes are welcome — bug reports, docs fixes, tests, and features. 证据：`CONTRIBUTING.md`
- **MARK Python SDK Agent Guide**（documentation）：This package implements the local-first MARK SDK. When working in this package, treat MARK as the agent cognitive runtime: local memory, context assembly, policies, skills, and agent wrapping. 证据：`AGENTS.md`
- **MARK Memory Middleware**（skill_instruction）：Use this skill when MARK middleware injects a MARK project memory block or records middleware diagnostics. 证据：`src/mark/middlewares/skills/mark-memory-middleware/SKILL.md`
- **MARK Usage Skill**（skill_instruction）：Use MARK as your persistent cognitive companion. MARK is not a tool you occasionally call after you get lost. MARK is the memory and continuity layer you lean on before planning, during reasoning, after action, before action, and when resuming future work. 证据：`src/mark/middlewares/skills/mark-usage/SKILL.md`
- **License**（source_file）：Apache License Version 2.0, January 2004 https://www.apache.org/licenses/ 证据：`LICENSE`
- **MARK Python SDK Architecture**（documentation）：The MARK Python SDK is a local-first agent memory runtime. It is the open, Apache-2.0-licensed layer that helps agents remember what they create, decide, observe, and learn across long-running workflows. 证据：`ARCHITECTURE.md`
- **Changelog**（documentation）：All notable changes to the MARK Python SDK will be documented here. 证据：`CHANGELOG.md`
- **Code of Conduct**（documentation）：We want MARK to have a community where everyone can participate — regardless of experience level, background, or identity. Be welcoming, be patient, and assume good faith. 证据：`CODE_OF_CONDUCT.md`
- **MARK SDK Roadmap**（documentation）：This roadmap describes the open Python SDK direction. It is intentionally focused on work that can be implemented, tested, documented, and shipped in the mark-sdk package. 证据：`ROADMAP.md`
- **Security Policy**（documentation）：MARK SDK is a local-first package. The core runtime should work without accounts, provider keys, network access, Docker, or remote services. 证据：`SECURITY.md`
- **Versioning**（documentation）：mark-sdk uses package-level Semantic Versioning. 证据：`VERSIONING.md`
- **01 Local Memory Live**（source_file）：def main - None ⋮---- project = Path tmp ⋮---- context = mark.memory.retrieve "Which framework handles the API?" text = context.as text 证据：`examples/01_local_memory_live.py`
- **02 Agent Ab Live**（source_file）：def coding agent prompt: str - str ⋮---- async def main - None ⋮---- baseline = coding agent "Add a health check endpoint." wrapped = mark.wrap agent coding agent, blocks= "project" result = await wrapped.run "Add a health check endpoint." 证据：`examples/02_agent_ab_live.py`
- **03 Sessions And Observe Live**（source_file）：def main - None ⋮---- observed = mark.observe ⋮---- memory = mark.runtime.memory "video-agent" result = memory.retrieve sync text = result.as context 证据：`examples/03_sessions_and_observe_live.py`
- **Getting Started With Mark**（source_file）：{ "cells": { "cell type": "markdown", "metadata": {}, "source": " Getting Started with MARK — Memory for Agents\n", "\n", "This notebook runs the same coding agent twice — once without memory, once with\n", "MARK plugged in — so you can see exactly what persistent agent memory changes.\n", "\n", "What you'll do:\n", "1. Use MARK's local memory directly 60 seconds \n", "2. Seed project conventions into memory\n", "3. Run a coding agent without MARK baseline \n", "4. Run the identical agent with MarkAgentMiddleware \n", "5. Use MARK as explicit LangChain tools \n", "6. Expose the same memory over MCP \n", "7. Inspect what MARK learned, and seal it into a provenance chain\n", "\n", " Requireme… 证据：`examples/getting_started_with_mark.ipynb`
- **Run Live Examples**（source_file）：EXAMPLES = ⋮---- def main - None ⋮---- root = Path file .resolve .parent 证据：`examples/run_live_examples.py`
- **Incomplete adapter namespaces stay out of shipped distributions until they**（source_file）：build-system requires = "hatchling =1.25" build-backend = "hatchling.build" 证据：`pyproject.toml`
- **Init**（source_file）：LAZY EXPORTS = { ⋮---- def getattr name: str ⋮---- module = import module LAZY EXPORTS name value = getattr module, name ⋮---- all = 证据：`src/mark/__init__.py`
- **Version**（source_file）：version = "0.2.0a6" 证据：`src/mark/_version.py`
- **Init**（source_file）：all = "BackendResult", "LocalMarkBackend", "MarkBackend" 证据：`src/mark/adapters/__init__.py`
- **Local compression is performed during retrieve compress=True . This method**（source_file）：@dataclass frozen=True class BackendResult ⋮---- ok: bool value: Any = None metadata: dict str, Any = field default factory=dict error: str None = None ⋮---- class MarkBackend Protocol ⋮---- async def retrieve self, query: str, options: Any - BackendResult async def write self, content: str, options: Any - BackendResult async def observe self, content: str, options: Any - BackendResult async def compress self, query: str, candidates: list Any , options: Any - BackendResult async def heal gap self, query: str, options: Any - BackendResult async def list sessions self, agent id: str, options: Any - BackendResult async def feedback self, payload: Any - BackendResult ⋮---- class LocalMarkBacken… 证据：`src/mark/adapters/backend.py`
- **Init**（source_file）：all = "MarkAgentMiddleware", "MarkMemoryMiddleware", "create mark tools" ⋮---- def create mark tools args, kwargs 证据：`src/mark/adapters/langchain/__init__.py`
- **Thinking model — initialised lazily from string if needed**（source_file）：LANGCHAIN AVAILABLE = True ⋮---- AgentMiddlewareBase = object AgentState = dict SystemMessage = None LANGCHAIN AVAILABLE = False ⋮---- class MarkAgentState AgentState ⋮---- mark context: NotRequired str mark context meta: NotRequired dict str, Any ⋮---- SYNC EXECUTOR: ThreadPoolExecutor None = None SYNC EXECUTOR LOCK = threading.Lock ⋮---- def run sync coro: Any - Any ⋮---- SYNC EXECUTOR = ThreadPoolExecutor ⋮---- class MarkAgentMiddleware AgentMiddlewareBase ⋮---- state schema = MarkAgentState ⋮---- THINKING TRIGGERS: tuple str, ... = ⋮---- Thinking model — initialised lazily from string if needed ⋮---- from langchain.chat models import init chat model type: ignore import ⋮---- ── backgrou… 证据：`src/mark/adapters/langchain/middleware.py`
- **Tools**（source_file）：SYNC EXECUTOR: ThreadPoolExecutor None = None SYNC EXECUTOR LOCK = threading.Lock ⋮---- def require langchain tool ⋮---- def run sync coro ⋮---- SYNC EXECUTOR = ThreadPoolExecutor ⋮---- structured tool = require langchain tool ⋮---- async def amark retrieve query: str - str ⋮---- result = await backend.retrieve ⋮---- value = result.value ⋮---- def mark retrieve query: str - str ⋮---- async def amark write content: str, canonical: bool = False - str ⋮---- options: dict str, Any = {"importance": 0.95 if canonical else 0.8} ⋮---- result = await backend.write content, agent id=default agent id, options ⋮---- def mark write content: str, canonical: bool = False - str ⋮---- """Write a durable, re… 证据：`src/mark/adapters/langchain/tools.py`
- **Init**（source_file）：all = "MarkState" 证据：`src/mark/adapters/langgraph/__init__.py`
- **State**（source_file）：class MarkState TypedDict, total=False ⋮---- agent id: str session id: str session prefix: str tags: list str mark context: str mark metadata: dict str, Any 证据：`src/mark/adapters/langgraph/state.py`
- **Init**（source_file）：def require mcp - Any ⋮---- FastMCP = require mcp server = FastMCP name ⋮---- server = create mark mcp server from local mark 证据：`src/mark/adapters/mcp/__init__.py`
- **Agent**（source_file）：@dataclass class AgentRun ⋮---- prompt: str blocks: list str None = None policy: str None = None skills: list str None = None skill inputs: dict str, dict str, Any = field default factory=dict metadata: dict str, Any = field default factory=dict retrieve: bool None = None store output: bool None = None top k: int None = None max context chars: int = 6000 session id: str None = None output block: str = "agent-output" ⋮---- @dataclass frozen=True class AgentResult ⋮---- output: str context: ContextBundle events: list dict str, Any = field default factory=list rendered prompt: str = "" decision: PolicyDecision None = None skill results: dict str, SkillResult = field default factory=dict stored… 证据：`src/mark/agent.py`
- **Init**（source_file）：all = "ContextBuilder", "ContextBundle", "within char budget" 证据：`src/mark/context/__init__.py`
- **Init**（source_file）：all = "CachedEmbeddingProvider", "EmbeddingProvider", "HashEmbeddingProvider" 证据：`src/mark/embeddings/__init__.py`
- **Base**（source_file）：class EmbeddingProvider ABC ⋮---- @abstractmethod def embed self, text: str - list float ⋮---- def embed batch self, texts: list str - list list float ⋮---- @property @abstractmethod def dim self - int None ⋮---- class HashEmbeddingProvider EmbeddingProvider ⋮---- def init self, dim: int = 128 - None ⋮---- def embed self, text: str - list float ⋮---- vector = 0.0 self. dim ⋮---- digest = int hashlib.sha256 token.encode "utf-8" .hexdigest , 16 ⋮---- norm = math.sqrt sum value value for value in vector ⋮---- @property def dim self - int 证据：`src/mark/embeddings/base.py`
- **Init**（source_file）：all = "VectorIndex", "VectorSearchResult" 证据：`src/mark/index/__init__.py`
- **Init**（source_file）：all = 证据：`src/mark/intelligence/__init__.py`
- **Classifier**（source_file）：class QueryComplexity str, Enum ⋮---- SIMPLE = "simple" BALANCED = "balanced" DEEP = "deep" ⋮---- class QueryAnalysis BaseModel ⋮---- query: str complexity: float policy: QueryComplexity is factual: bool is multi hop: bool estimated top k: int keywords: List str reasoning: str reasons: List str = Field default factory=list ⋮---- has temporal: bool = False has comparison: bool = False has multi entity: bool = False has reasoning: bool = False has negation: bool = False question depth: int = 1 ⋮---- STOP WORDS: Set str = { ⋮---- TEMPORAL = re.compile COMPARISON = re.compile REASONING = re.compile NEGATION = re.compile r"\b not never without except unless no\s+\w+ \b", re.IGNORECASE ENTITY CAP… 证据：`src/mark/intelligence/classifier.py`
- **Compressor**（source_file）：@runtime checkable class ContextualCompressor Protocol ⋮---- class NoopCompressor ⋮---- class SimpleWindowCompressor ⋮---- def init self, max chars: int = 500, , ellipsis: bool = True - None ⋮---- result = ⋮---- trimmed = f.content : self. max ⋮---- trimmed = trimmed.rstrip + "…" ⋮---- class LLMContextualCompressor ⋮---- PROMPT = ⋮---- def init self, llm: "LLMProvider" - None ⋮---- prompt = self. PROMPT.format query=query, passage=f.content response = self. llm.complete prompt .strip 证据：`src/mark/intelligence/compressor.py`
- **Query Expander**（source_file）：@runtime checkable class QueryExpander Protocol ⋮---- def expand self, query: str - list str ⋮---- class NoopQueryExpander ⋮---- class KeywordQueryExpander ⋮---- STRIP RE = re.compile ⋮---- keywords = self. STRIP RE.sub "", query keywords = re.sub r"\s+", " ", keywords .strip .rstrip "?.," ⋮---- class LLMQueryExpander ⋮---- """Uses a developer-provided LLM to generate alternate query phrasings.""" ⋮---- PROMPT = ⋮---- def init self, llm: "LLMProvider", , n expansions: int = 2 - None ⋮---- response = self. llm.complete self. PROMPT.format n=self. n, query=query .strip alternates = 证据：`src/mark/intelligence/query_expander.py`
- **Reranker**（source_file）：@dataclass frozen=True class RerankCandidate ⋮---- fragment id: str content: str score: float importance: float = 0.5 node weight: float = 1.0 metadata: dict str, Any = field default factory=dict ⋮---- @dataclass frozen=True class RerankResult ⋮---- signals: dict str, float = field default factory=dict ⋮---- class MemoryReranker ⋮---- W COSINE = 0.60 W IMPORTANCE = 0.25 W NODE = 0.15 CONTRADICTION PENALTY = 0.20 ⋮---- def init self, store: LocalMemoryStore - None ⋮---- contradiction ids = contradiction ids or set results: list RerankResult = ⋮---- contradiction = candidate.fragment id in contradiction ids score = score = max 0.0, round score, 6 ⋮---- loop = asyncio.get running loop ⋮---- de… 证据：`src/mark/intelligence/reranker.py`
- **Retrieval**（source_file）：class RetrievalPolicy str, Enum ⋮---- FAST = "fast" BALANCED = "balanced" DEEP = "deep" ⋮---- @dataclass frozen=True class RetrievalPolicySpec ⋮---- name: str display name: str description: str = "" top k: int = 10 graph depth: int = 1 rerank depth: int = 1 min confidence: float = 0.30 enable decomposition: bool = False max subqueries: int = 1 critic enabled: bool = False enable external: bool = False is builtin: bool = True expansion direction: str = "both" ⋮---- BUILTIN POLICIES: dict str, RetrievalPolicySpec = { ⋮---- RETRIEVABLE STATES = {MemoryState.UNVERIFIED, MemoryState.VERIFIED, MemoryState.PROMOTED} ⋮---- @dataclass frozen=True class SessionFilter ⋮---- session id: str None = None… 证据：`src/mark/intelligence/retrieval.py`
- **Init**（source_file）：LAZY EXPORTS = { ⋮---- def getattr name: str ⋮---- module = import module LAZY EXPORTS name value = getattr module, name ⋮---- all = 证据：`src/mark/memory/__init__.py`
- **Block**（source_file）：class BlockStore Protocol ⋮---- def write record self, record: MemoryRecord - MemoryRecord def list records self, block id: str None = None - list MemoryRecord ⋮---- @dataclass frozen=True class MemoryBlock ⋮---- id: str label: str kind: str = "custom" description: str = "" active: bool = True metadata: dict str, Any = field default factory=dict created at: datetime = field default factory=lambda: datetime.now timezone.utc updated at: datetime = field default factory=lambda: datetime.now timezone.utc store: BlockStore None = field default=None, repr=False, compare=False ⋮---- record = MemoryRecord ⋮---- def records self - list MemoryRecord ⋮---- """Return this block's records.""" ⋮---- def… 证据：`src/mark/memory/block.py`
- **Block Chain**（source_file）：MEMBER HASHES KEY = "mark member hashes" ⋮---- @dataclass frozen=True class BlockVerification ⋮---- valid: bool block id: str corrupted member ids: list str = field default factory=list missing member ids: list str = field default factory=list added member ids: list str = field default factory=list reason: str = "ok" ⋮---- @dataclass frozen=True class ChainVerification ⋮---- checked: int broken block ids: list str = field default factory=list broken links: list tuple str, str = field default factory=list ⋮---- class BlockChain ⋮---- def init self, store: LocalMemoryStore, agent id: str - None ⋮---- def seal self, block id: str - MemoryBlock ⋮---- block = self. store.get block block id ⋮----… 证据：`src/mark/memory/block_chain.py`
- **── block links inter-block graph ───────────────────────────────────────**（source_file）：class BlockGraph ⋮---- def init self, store: LocalMemoryStore, agent id: str - None ⋮---- block = MemoryBlock ⋮---- def get self, block id: str - MemoryBlock None ⋮---- def by name self, name: str - MemoryBlock None ⋮---- def list self, , status: BlockStatus None = None - list MemoryBlock ⋮---- def add fragment self, block id: str, fragment id: str - None ⋮---- def add node self, block id: str, node id: str - None ⋮---- def add nodes self, block id: str, node ids: list str - None ⋮---- def remove node self, block id: str, node id: str - bool ⋮---- def nodes self, block id: str - list MemoryNode ⋮---- def blocks for node self, node id: str - list MemoryBlock ⋮---- blocks = ⋮---- block = self… 证据：`src/mark/memory/block_graph.py`
- **Consolidation**（source_file）：@dataclass class ConsolidationResult ⋮---- promoted: int = 0 expired: int = 0 skipped: int = 0 blocked: int = 0 sanitized: int = 0 ⋮---- class ConsolidationManager ⋮---- def run self, , promote threshold: Optional float = None - ConsolidationResult ⋮---- threshold = promote threshold if promote threshold is not None else self. threshold result = ConsolidationResult now = datetime.now timezone.utc ⋮---- fragments: List MemoryFragment = self. store.list by agent self. agent id ⋮---- gate result = self. gate.check frag.content, confidence=frag.confidence ⋮---- promoted = frag.model copy update={ 证据：`src/mark/memory/consolidation.py`
- **Deduplication**（source_file）：@dataclass class ClusterResult ⋮---- kept id: str merged from ids: list str cluster size: int centroid: list float None = None ⋮---- @dataclass class DeduplicationResult ⋮---- merged: int = 0 kept: int = 0 skipped: int = 0 clusters: list ClusterResult = field default factory=list ⋮---- def cosine a: list float , b: list float - float ⋮---- dot = sum x y for x, y in zip a, b na = math.sqrt sum x x for x in a nb = math.sqrt sum y y for y in b ⋮---- def centroid embeddings: list list float - list float ⋮---- dim = len embeddings 0 result = 0.0 dim ⋮---- n = len embeddings ⋮---- clusters: list list Any = ⋮---- placed = False ⋮---- seed = cluster 0 ⋮---- placed = True ⋮---- class DeduplicationCo… 证据：`src/mark/memory/deduplication.py`
- **Graph**（source_file）：@dataclass frozen=True class GraphNeighborhood ⋮---- center: MemoryNode nodes: list MemoryNode = field default factory=list edges: list MemoryEdge = field default factory=list fragments: list MemoryFragment = field default factory=list depth: int = 2 ⋮---- def as context self, max chars: int = 4000 - str ⋮---- node map = {n.id: n for n in self.nodes} lines: list str = ⋮---- src = node map.get edge.source id tgt = node map.get edge.target id src label = src.label if src else edge.source id :8 tgt label = tgt.label if tgt else edge.target id :8 w = f"{edge.weight:.2f}" if edge.weight != 1.0 else "" wstr = f" w={w} " if w else "" ⋮---- other nodes = n for n in self.nodes if n.id != self.center… 证据：`src/mark/memory/graph.py`
- **Outgoing edges**（source_file）：NODE TYPE ALIASES: dict str, str = { ⋮---- @dataclass frozen=True class MemoryWriteRejection ⋮---- fragment id: str reason: str content: str ⋮---- def resolve node type node type: NodeType str - NodeType ⋮---- normalized = node type.lower canonical = NODE TYPE ALIASES.get normalized, normalized ⋮---- class MarkMemory ⋮---- loop = asyncio.get running loop ⋮---- indexed = await loop.run in executor self. executor, self. pipeline.index fragment, fragment ⋮---- meta = { fragment = self. fragment ⋮---- @property def last rejection self - MemoryWriteRejection None ⋮---- def quarantined self, , limit: int = 100 - list MemoryFragment ⋮---- fragment id = self. store sync impl hook = self. pipeline.… 证据：`src/mark/memory/mark_memory.py`
- **Retrieval**（source_file）：@dataclass frozen=True class ScoredRecord ⋮---- record: MemoryRecord score: float ⋮---- class LocalRetriever ⋮---- def retrieve self, query: str, records: list MemoryRecord , top k: int = 5 - list ScoredRecord ⋮---- scored = 证据：`src/mark/memory/retrieval.py`
- **Runtime**（source_file）：class MarkRuntime ⋮---- registry = PluginRegistry runtime = cls ⋮---- def configure llm self, llm: LLMProvider - None ⋮---- def configure compressor self, compressor: ContextualCompressor - None ⋮---- def configure query expander self, expander: QueryExpander - None ⋮---- def configure retrieval top k self, top k: int - None ⋮---- def add middleware self, middleware: MarkMiddleware - MarkMiddleware ⋮---- def use self, middleware: MarkMiddleware - "MarkRuntime" ⋮---- def session log self, agent id: str, session id: str None = None - "SessionActivityLog" ⋮---- def tracer self ⋮---- def session trace self, session id: str ⋮---- def governance audit log self ⋮---- def memory self, agent id: str… 证据：`src/mark/memory/runtime.py`
- **Init**（source_file）：LAZY EXPORTS = { ⋮---- def getattr name: str ⋮---- module = import module LAZY EXPORTS name value = getattr module, name ⋮---- all = 证据：`src/mark/middlewares/__init__.py`
- **Base**（source_file）：@dataclass class MiddlewareContext ⋮---- operation: str runtime: Any agent id: str None = None payload: dict str, Any = field default factory=dict result: Any = None metadata: dict str, Any = field default factory=dict stopped: bool = False ⋮---- def stop self, result: Any = None - None ⋮---- class MarkMiddleware Protocol ⋮---- name: str ⋮---- def bind self, runtime: Any - None ⋮---- def before self, context: MiddlewareContext - None ⋮---- def after self, context: MiddlewareContext - None ⋮---- class BaseMiddleware ⋮---- name = "base" ⋮---- class MiddlewareStack ⋮---- def init self, middleware: Iterable MarkMiddleware None = None - None ⋮---- def add self, middleware: MarkMiddleware - MarkM… 证据：`src/mark/middlewares/base.py`
- **Init**（source_file）：all = "CompressionMiddleware" 证据：`src/mark/middlewares/compression/__init__.py`
- **Middleware**（source_file）：@dataclass class CompressionMiddleware BaseMiddleware ⋮---- compressor: Any enabled by default: bool = True name: str = "compression" ⋮---- def bind self, runtime: Any - None ⋮---- def before self, context: MiddlewareContext - None ⋮---- all = "CompressionMiddleware" 证据：`src/mark/middlewares/compression/middleware.py`
- **Init**（source_file）：all = "GapHealingMiddleware" 证据：`src/mark/middlewares/gap_healing/__init__.py`
- **Middleware**（source_file）：@dataclass class GapHealingMiddleware BaseMiddleware ⋮---- heal gaps: bool = True escalate on gap: bool = True web skill: Any = None name: str = "gap healing" ⋮---- def before self, context: MiddlewareContext - None ⋮---- all = "GapHealingMiddleware" 证据：`src/mark/middlewares/gap_healing/middleware.py`
- **Init**（source_file）：all = 证据：`src/mark/middlewares/governance/__init__.py`
- **Middleware**（source_file）：@dataclass class GovernanceMiddleware BaseMiddleware ⋮---- gate: ConsolidationGate = field default factory=ConsolidationGate quarantine on reject: bool = True name: str = "governance" ⋮---- def before self, context: MiddlewareContext - None ⋮---- key = "text" if context.operation == "observe" else "content" content = str context.payload.get key or "" confidence = float context.payload.get "confidence", 1.0 result = self.gate.check content, confidence=confidence ⋮---- metadata = dict context.payload.get "metadata" or {} ⋮---- tags = list context.payload.get "tags" or ⋮---- all = "GovernanceMiddleware" 证据：`src/mark/middlewares/governance/middleware.py`
- **Init**（source_file）：all = "LifecycleMiddleware" 证据：`src/mark/middlewares/lifecycle/__init__.py`
- **Middleware**（source_file）：@dataclass class LifecycleMiddleware BaseMiddleware ⋮---- every n writes: int = 0 promote threshold: float = 0.7 name: str = "lifecycle" write count: int = 0 ⋮---- def after self, context: MiddlewareContext - None ⋮---- all = "LifecycleMiddleware" 证据：`src/mark/middlewares/lifecycle/middleware.py`
- 其余 20 条证据见 `AI_CONTEXT_PACK.json` 或 `EVIDENCE_INDEX.json`。

## 宿主 AI 必须遵守的规则

- **把本资产当作开工前上下文，而不是运行环境。**：AI Context Pack 只包含证据化项目理解，不包含目标项目的可执行状态。 证据：`README.md`, `examples/README.md`, `CONTRIBUTING.md`
- **回答用户时区分可预览内容与必须安装后才能验证的内容。**：安装前体验的消费者价值来自降低误装和误判，而不是伪装成真实运行。 证据：`README.md`, `examples/README.md`, `CONTRIBUTING.md`

## 用户开工前应该回答的问题

- 你准备在哪个宿主 AI 或本地环境中使用它？
- 你只是想先体验工作流，还是准备真实安装？
- 你最在意的是安装成本、输出质量、还是和现有规则的冲突？

## 验收标准

- 所有能力声明都能回指到 evidence_refs 中的文件路径。
- AI_CONTEXT_PACK.md 没有把预览包装成真实运行。
- 用户能在 3 分钟内看懂适合谁、能做什么、如何开始和风险边界。

---

## Doramagic Context Augmentation

下面内容用于强化 Repomix/AI Context Pack 主体。Human Manual 只提供阅读骨架；踩坑日志会被转成宿主 AI 必须遵守的工作约束。

## Human Manual 骨架

使用规则：这里只是项目阅读路线和显著性信号，不是事实权威。具体事实仍必须回到 repo evidence / Claim Graph。

宿主 AI 硬性规则：
- 不得把页标题、章节顺序、摘要或 importance 当作项目事实证据。
- 解释 Human Manual 骨架时，必须明确说它只是阅读路线/显著性信号。
- 能力、安装、兼容性、运行状态和风险判断必须引用 repo evidence、source path 或 Claim Graph。

- **MARK SDK 概览与快速上手**：importance `high`
  - source_paths: README.md, pyproject.toml, src/mark/__init__.py, src/mark/_version.py, examples/README.md
- **安装、可选扩展与示例运行**：importance `high`
  - source_paths: README.md, examples/README.md, examples/run_live_examples.py, examples/01_local_memory_live.py, examples/02_agent_ab_live.py
- **系统架构与运行时**：importance `high`
  - source_paths: ARCHITECTURE.md, README.md, src/mark/runtime.py, src/mark/agent.py, src/mark/adapters/backend.py
- **记忆模型与图谱化块(Blocks)**：importance `high`
  - source_paths: src/mark/memory/block.py, src/mark/memory/block_chain.py, src/mark/memory/block_graph.py, src/mark/memory/mark_memory.py, src/mark/types/memory.py
- **检索管线与向量索引**：importance `high`
  - source_paths: src/mark/memory/retrieval.py, src/mark/intelligence/retrieval.py, src/mark/intelligence/query_expander.py, src/mark/intelligence/reranker.py, src/mark/intelligence/classifier.py
- **中间件(Middleware)框架**：importance `high`
  - source_paths: src/mark/middlewares/base.py, src/mark/middlewares/recall/middleware.py, src/mark/middlewares/observe/middleware.py, src/mark/middlewares/compression/middleware.py, src/mark/middlewares/query_expansion/middleware.py
- **框架适配器：LangChain、LangGraph 与 MCP**：importance `high`
  - source_paths: src/mark/adapters/langchain/middleware.py, src/mark/adapters/langchain/tools.py, src/mark/adapters/langgraph/state.py, src/mark/adapters/mcp/__init__.py, src/mark/adapters/backend.py
- **记忆可塑性、生命周期与治理**：importance `medium`
  - source_paths: src/mark/plasticity/decay.py, src/mark/plasticity/hebbian.py, src/mark/plasticity/pruner.py, src/mark/plasticity/router.py, src/mark/memory/consolidation.py

## Repo Inspection Evidence / 源码检查证据

- repo_clone_verified: true
- repo_inspection_verified: true
- repo_commit: `5b00fadcdda18b481980936c6b828c9c2e7be55c`
- inspected_files: `README.md`, `pyproject.toml`, `uv.lock`, `examples/01_local_memory_live.py`, `examples/02_agent_ab_live.py`, `examples/03_sessions_and_observe_live.py`, `examples/04_sync_envelope_live.py`, `examples/README.md`, `examples/run_live_examples.py`, `src/mark/__init__.py`, `src/mark/_version.py`, `src/mark/adapters/__init__.py`, `src/mark/adapters/backend.py`, `src/mark/adapters/langchain/__init__.py`, `src/mark/adapters/langchain/middleware.py`, `src/mark/adapters/langchain/tools.py`, `src/mark/adapters/langgraph/__init__.py`, `src/mark/adapters/langgraph/state.py`, `src/mark/adapters/mcp/__init__.py`, `src/mark/agent.py`

宿主 AI 硬性规则：
- 没有 repo_clone_verified=true 时，不得声称已经读过源码。
- 没有 repo_inspection_verified=true 时，不得把 README/docs/package 文件判断写成事实。
- 没有 quick_start_verified=true 时，不得声称 Quick Start 已跑通。

## Doramagic Pitfall Constraints / 踩坑约束

这些规则来自 Doramagic 发现、验证或编译过程中的项目专属坑点。宿主 AI 必须把它们当作工作约束，而不是普通说明文字。

### Constraint 1: 能力判断依赖假设

- Trigger: README/documentation is current enough for a first validation pass.
- Host AI rule: 将假设转成下游验证清单。
- Why it matters: 假设不成立时，用户拿不到承诺的能力。
- Evidence: capability.assumptions | https://github.com/emsoftanalytics/MARK-SDK | README/documentation is current enough for a first validation pass.
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 2: 维护活跃度未知

- Trigger: 未记录 last_activity_observed。
- Host AI rule: 补 GitHub 最近 commit、release、issue/PR 响应信号。
- Why it matters: 新项目、停更项目和活跃项目会被混在一起，推荐信任度下降。
- Evidence: evidence.maintainer_signals | https://github.com/emsoftanalytics/MARK-SDK | last_activity_observed missing
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

- Trigger: no_demo
- Evidence: downstream_validation.risk_items | https://github.com/emsoftanalytics/MARK-SDK | no_demo; severity=medium
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 4: 存在评分风险

- Trigger: no_demo
- Why it matters: 风险会影响是否适合普通用户安装。
- Evidence: risks.scoring_risks | https://github.com/emsoftanalytics/MARK-SDK | no_demo; severity=medium
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 5: issue/PR 响应质量未知

- Trigger: issue_or_pr_quality=unknown。
- Host AI rule: 抽样最近 issue/PR，判断是否长期无人处理。
- Why it matters: 用户无法判断遇到问题后是否有人维护。
- Evidence: evidence.maintainer_signals | https://github.com/emsoftanalytics/MARK-SDK | issue_or_pr_quality=unknown
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 6: 发布节奏不明确

- Trigger: release_recency=unknown。
- Host AI rule: 确认最近 release/tag 和 README 安装命令是否一致。
- Why it matters: 安装命令和文档可能落后于代码，用户踩坑概率升高。
- Evidence: evidence.maintainer_signals | https://github.com/emsoftanalytics/MARK-SDK | release_recency=unknown
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。