kaijibot 2026.6.2-2 → 2026.6.2-3

package/README.en.md

@@ -2,14 +2,13 @@
 
 > **Your AI assistant reaches out to you — not the other way around.**
 
-Fork of [OpenClaw](https://github.com/openclaw/openclaw) · Feishu + 30+ LLM Providers · Cognitive Layer Turns AI from Reactive to Proactive
+Pluggable provider/channel architecture · Cognitive layer turns AI from reactive to proactive · 30+ LLM providers
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
 [![Node.js >=22](https://img.shields.io/badge/Node.js-%3E%3D22-339933.svg)](https://nodejs.org/)
 [![TypeScript](https://img.shields.io/badge/TypeScript-6.x-3178C6.svg)](https://www.typescriptlang.org/)
-[![Vitest 450+ tests](https://img.shields.io/badge/Vitest-450%2B%20tests-6DA55F.svg)](https://vitest.dev/)
 
-**README** | **English** | [简体中文](./README.md)
+[English](./README.en.md) | [简体中文](./README.md)
 
 ## Why KaijiBot
 
@@ -41,13 +40,13 @@ Two weeks later, you bring up Rust and embedded systems. One day it tells you:
 
 How it works:
 
-- **Persona Profiling** — Every conversation teaches it something new. LLM-driven structured extraction automatically discovers domains and extracts 6 categories of typed insights (domain knowledge, behavioral patterns, preferences, tool configs, contextual facts, goals/aspirations) into your profile. Insights decay by category (behavioral patterns 60 days, tool configs 180 days). Interest lifecycle tracking: emergent → stable → declining → dormant → revived. Domain names discovered dynamically by LLM — no hardcoded keyword tables.
-- **Cross-Domain Insights** — You're interested in both A and B, it finds potential connections. You asked about something before but didn't dig deeper, it follows up from a new angle. You've gone deep enough in one area, it suggests extension directions. Domains linked via co-occurrence graphs (2-hop indirect connections supported), with edge decay. Domain cooldown prevents short-term repeated recommendations. LLM self-refine loop (generate → critique → rewrite) ensures quality. Contrastive anti-dedup framework ensures every push differs from history. Semantic freshness check blocks semantically repetitive content.
-- **Timing Gate** — It doesn't push whenever it feels like it. The PRISM model (based on signal detection theory) calculates expected value for each insight. Only pushes when expected benefit exceeds interruption cost. No late-night打扰, restraint during low trust, waits if you've been inactive.
-- **Trust Evolution** — Cautious at first, understands you better over time, eventually becomes a confident partner who can make bold recommendations. Four stages (SARA framework): orientation → exploration → rapport → partnership. Trust level determines what the system is allowed to do.
-- **Preference Learning** — You wrote a long reply? It notes you like this topic. You gave a one-word response? It tries a different direction next time. Each topic maintains Thompson Sampling parameters (Beta distribution). Implicit feedback (reply depth, response latency) is more honest than explicit feedback.
+- **Persona Profiling** — Every conversation teaches it something new. LLM-driven structured extraction automatically discovers domains and interests from conversations. Insights decay by category independently. Interest lifecycle is tracked automatically. Domain names discovered dynamically by LLM — no hardcoded keywords.
+- **Cross-Domain Insights** — You're interested in both A and B, it finds potential connections. You asked about something before but didn't dig deeper, it follows up from a new angle. You've gone deep enough in one area, it suggests extension directions. LLM self-refine loop ensures quality, semantic dedup ensures every push is fresh.
+- **Timing Gate** — It doesn't push whenever it feels like it. The PRISM model calculates expected value for each insight. Only pushes when expected benefit exceeds interruption cost. No late-night打扰, restraint during low trust, waits if you've been inactive.
+- **Trust Evolution** — Cautious at first, understands you better over time, eventually becomes a confident partner who can make bold recommendations. SARA framework drives four-stage trust evolution. Trust level determines what the system is allowed to do.
+- **Preference Learning** — You wrote a long reply? It notes you like this topic. You gave a one-word response? It tries a different direction next time. Thompson Sampling-driven preference learning. Implicit feedback is more honest than explicit feedback.
 
-Insight content is generated from your profile + LLM knowledge + real-time web search. With an Exa or Tavily API key, insights stay current.
+Insight content is generated from your profile + LLM knowledge + real-time web search. With a search API key, insights stay current.
 
 ### 🧬 Self-Evolution — Agent Decides When to Learn New Skills
 
@@ -55,56 +54,44 @@ You've done several complex Feishu knowledge base operations with KaijiBot — s
 
 > "I noticed you've been doing similar meeting minute archiving workflows recently. I wrote myself a skill — next time you say 'archive meetings', I'll execute the entire flow automatically."
 
-Or it spots a reusable pattern and tells you right away that it created a skill for itself.
-
 How it works:
 
-- **Hard Trigger Detection** — Code does one thing: detect 3+ tool calls in a conversation (noise filter). No LLM calls, no quality judgments.
-- **Agent-Driven Decision** — After detection, a system event is injected (with tool sequence and error info), triggering an Agent turn. The Agent has full conversation context and decides whether it's worth creating a skill. Not worth it? Ignored.
-- **No Cooldown, No Cap** — No code-level rate limits or complexity thresholds. The Agent sees recent suggestion history and decides frequency itself. Whatever the decision, the user is always notified.
-- **Full Lifecycle** — Dedup check before creation, usage frequency tracking after, 30-day auto-cleanup for unused skills.
+- **Hard Trigger Detection** — Code does noise filtering only, no quality judgments. After detecting complex tasks, a system event is injected.
+- **Agent-Driven Decision** — The Agent has full conversation context and decides whether it's worth creating a skill. Not worth it? Ignored.
+- **Full Lifecycle** — Dedup check before creation, usage frequency tracking after, auto-cleanup for unused skills.
 
 ### 🔄 Correction Self-Evolution — Never Makes the Same Mistake Twice
 
-Ever had an AI assistant make the same mistake every new session? Like always trying to create a blank Feishu document, getting corrected, then doing it again next session?
-
-KaijiBot doesn't do that. It has a correction memory system that ensures every mistake is made at most once.
-
-How it works:
+AI assistant making the same mistake every new session? KaijiBot won't. It has a correction memory system that ensures every mistake is made at most once.
 
-- **Dual-Path Detection** — Path A: the Agent catches its own error and reports via `record_correction` tool. Path B: when you run `/new` or `/reset`, the system pre-screens with 60 regex patterns (covering Chinese/English corrections, questioning tone, agent apologies), then passes matching transcripts to LLM for structured extraction.
-- **Jaccard Dedup + Reinforcement** — Corrections stored in `~/.kaijibot/cognitive/corrections/{userId}.json`. Jaccard similarity determines if a new error duplicates an existing record. Duplicates don't create new entries — they increment `reinforcedCount`, signaling this mistake keeps recurring and deserves higher priority. Max 50 per user, 90-day TTL.
-- **System Prompt Injection** — This is the key design. Corrections are injected into the system prompt's "Known Corrections" section, sorted by reinforcement count, up to 15 entries. The Agent sees them every single turn. Why not put them in skill files or MEMORY.md? Because skills aren't always loaded, MEMORY.md is only read at startup — only the system prompt is guaranteed to be visible every turn.
+- **Dual-Path Detection** — Agent self-reports errors, or the system automatically extracts corrections from conversations at session end.
+- **Dedup + Reinforcement** — Repeated errors don't create new records; they increase weight, ensuring high-frequency issues get priority.
+- **System Prompt Injection** — Corrections are injected into the system prompt that the Agent sees every turn, rather than in files that may not be read.
 
 Step into a pit once. That's enough.
 
-### 🔌 62 Extensions Out of the Box
+### 🔌 Pluggable Architecture
 
 Not locked into any single provider. Switch between domestic and international at will. `kaijibot onboard` wizard auto-discovers configured API keys.
 
-| China (Recommended)                                                                                                             | International                                                                        | Aggregator / Self-hosted                                                                                    |
-| ------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------- |
-| Zhipu GLM · DeepSeek · Qwen · Kimi · MiniMax · Baidu Qianfan · StepFun · Volcengine · BytePlus · Kimi Coding · Xiaomi · Alibaba | Claude · Gemini · Grok · Mistral · Perplexity · Groq · Nvidia · HuggingFace · OpenAI | OpenRouter · LiteLLM · Together · Fireworks · Cloudflare AI · Vercel AI · SGLang · vLLM · Ollama · LMStudio |
+| China (Recommended)                                | International                                    | Aggregator / Self-hosted                        |
+| -------------------------------------------------- | ------------------------------------------------ | ------------------------------------------------ |
+| Zhipu GLM · DeepSeek · Qwen · Kimi · MiniMax …     | Claude · Gemini · Grok · Mistral · Perplexity …  | OpenRouter · Together · Ollama · LMStudio · vLLM … |
 
 Switch models with a single command:
 
 ```bash
 kaijibot config set agent.model "deepseek/deepseek-chat"
-kaijibot config set agent.model "qwen/qwen-max"
 kaijibot config set agent.model "anthropic/claude-sonnet-4-20250514"
 ```
 
-### 🛠️ 21 Built-in Skills + Full Agent
-
-**Agent Loop**: Reason → Call tools → Observe → Continue reasoning. Supports streaming, context compression, parallel sub-agent dispatch.
-
-**Built-in Tools**: Code execution, web scraping, PDF operations, image/video/music generation, TTS synthesis, Canvas, file I/O, cron scheduling — 20+ total. Supports model failover and API key rotation.
+### 🛠️ Full Agent
 
-**Memory System**: Three storage backends (in-memory, LanceDB vector store, Wiki knowledge base). Semantic search over conversation history. MEMORY.md (8KB budget) is injected as context every agent turn, maintained by three complementary systems: session memory (`/new` or `/reset` triggers LLM structured summary → inline + topic), consolidation engine (daily 3 AM scan of 7-day transcripts → high-confidence content to inline sections), and memory-organize skill (manual/periodic full GC + Jaccard dedup). `memory_tidy` covers cross-dedup between inline sections and topic files. Correction memory stored independently, auto-injected into system prompt so the agent sees past corrections every turn.
+**Agent Loop**: Reason → Call tools → Observe → Continue reasoning. Supports streaming, context compression, parallel sub-agent dispatch. Built-in tools include code execution, web scraping, PDF operations, image/video/music generation, TTS synthesis, Canvas, and more.
 
-**Scheduled Tasks**: `at` (one-shot), `every` (interval), `cron` (cron expression + timezone). Supports message delivery, webhook callbacks, or silent execution. Auto-retry on failure.
+**Memory System**: Multiple storage backends with semantic search over conversation history. Three complementary systems maintain Agent context: session memory, daily consolidation, and manual organization.
 
-**Skill Marketplace**: github, weather, summarize, coding-agent, notion, obsidian, nano-pdf, taskflow, blogwatcher — 21 built-in skills, with more available from ClawHub:
+**Skill Marketplace**: Dozens of built-in skills (github, weather, summarize, coding-agent, notion, obsidian, taskflow, and more). Install additional skills from ClawHub:
 
 ```bash
 kaijibot skills install <skill-name>
@@ -112,15 +99,37 @@ kaijibot skills install <skill-name>
 
 ## 🚀 Quick Start
 
-**Prerequisites**: Node.js >= 22 (24 recommended), pnpm, git
+### macOS / Linux
 
-**Option 1: One-Click Install**
+**One-click install** (recommended — auto-detects environment, installs dependencies, runs onboard wizard):
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/Kaiji-Z/kaijibot/main/install.sh | bash
+curl -fsSL https://kaijibot.ai/install.sh | bash
+```
+
+Also available via npm:
+
+```bash
+npm install -g kaijibot
+kaijibot onboard
+```
+
+### Windows
+
+PowerShell one-click install:
+
+```powershell
+iwr -useb https://kaijibot.ai/install.ps1 | iex
+```
+
+Or via npm:
+
+```powershell
+npm install -g kaijibot
+kaijibot onboard
 ```
 
-**Option 2: Docker**
+### Docker
 
 ```bash
 git clone https://github.com/Kaiji-Z/kaijibot.git
@@ -128,19 +137,19 @@ cd kaijibot
 docker compose up -d
 ```
 
-**Option 3: Manual Install**
+### Build from Source
 
 ```bash
 git clone https://github.com/Kaiji-Z/kaijibot.git
 cd kaijibot
 pnpm install
 pnpm build
-kaijibot onboard   # Interactive wizard, auto-configures (detects OpenClaw and prompts migration)
+kaijibot onboard   # Interactive wizard, auto-configures
 # Migrating from OpenClaw? Run:
 kaijibot migrate
 ```
 
-**Start**
+### Start
 
 ```bash
 kaijibot gateway --port 18789 --verbose
@@ -153,11 +162,9 @@ After starting, find your bot on Feishu and send a message. KaijiBot automatical
 **Required**: At least one LLM provider API key + Feishu bot credentials.
 
 ```bash
-# LLM API Key (pick one, configure for the model you use)
+# LLM API Key (pick one)
 export ZAI_API_KEY="your-key"              # Zhipu GLM
 # export DEEPSEEK_API_KEY="your-key"       # DeepSeek
-# export DASHSCOPE_API_KEY="your-key"      # Qwen
-# export MOONSHOT_API_KEY="your-key"       # Kimi
 # export ANTHROPIC_API_KEY="your-key"      # Claude
 # export GOOGLE_API_KEY="your-key"         # Gemini
 
@@ -173,99 +180,7 @@ export EXA_API_KEY="your-key"
 export TAVILY_API_KEY="your-key"
 ```
 
-Config file at `~/.kaijibot/kaijibot.json`, supports hot reload. Cognitive system can be disabled via `cognitive.enabled: false` to fall back to pure OpenClaw experience. Consolidation: `memory.consolidation.enabled`, `memory.consolidation.cron` (default `0 3 * * *`), `memory.consolidation.lookbackDays` (default 7).
-
-## 🏗️ Architecture
-
-### Cognitive Insight Pipeline
-
-```
-Event Sources (timer + random jitter / persona change / info scan)
-  → PRISM Gate (pNeed × pAccept > cost threshold?)
-    → Search Insight Opportunities (cross-domain / domain depth / 3-mode exploration)
-      → Domain Cooldown + Starvation Boost → Pick Best Opportunity
-        → Semantic Freshness Check (LLM judges if semantically duplicate with recent insights)
-          → Unified Pipeline (mode routing via timestamp % 100):
-              Knowledge Mode (40%): Web search + LLM → Self-refine (critique→rewrite) → Quality retries → LLM verification
-              Pattern Mode (50%): Dialog fragment clustering → LLM behavioral insight → Partial verification
-              Extend Mode (10%): Known user domains → LLM depth suggestions
-            → Contrastive Anti-dedup + Safety-net Trigram Dedup
-              → Deliver to your Feishu conversation
-                → Collect Feedback → Thompson Sampling updates preference model
-```
-
-### Self-Evolution Pipeline
-
-```
-Agent completes task (≥3 tool calls)
-→ Code detects ≥3 tool calls (noise filter, no LLM call)
-→ Inject system event directly → Trigger Agent turn
-      → Agent sees full conversation context + recent suggestion history
-        → Worth making a skill? → Generate skill draft → Ask user or silently create
-        → Not worth it? → Ignore
-        → Frequent but worthwhile? → Silently create, mention later
-```
-
-### Correction Self-Evolution Pipeline
-
-```
-Dual-Path Detection
-  Path A: Agent self-report (record_correction tool)    Path B: Post-session LLM extraction (/new or /reset)
-              ↓                                              ↓
-        CorrectionStore.addOrReinforce (Jaccard dedup + reinforcement)
-              ↓
-        ~/.kaijibot/cognitive/corrections/{userId}.json
-              ↓
-        Next conversation → context-writer injects into system prompt
-              ↓
-        Agent sees historical corrections → Avoids repetition
-```
-
-### Memory Consolidation Flow
-
-```
-Daily at 3 AM (cron schedule)
-  → Scan last 7 days of session transcripts (grouped by userId)
-    → LLM batch extraction of structured knowledge (TypedInsights + behavioral patterns + corrections)
-      → Jaccard dedup + conflict resolution
-        → Route to cognitive stores:
-            PersonaStore (domain knowledge, preferences, goals)
-            FragmentStore (behavioral patterns)
-            CorrectionStore (high-confidence corrections)
-          → High-confidence items (≥ 0.7) written to MEMORY.md inline sections:
-              domain_knowledge / stated_preference / behavioral_pattern → ⚡ Core Memory
-              goal_or_aspiration → 🔥 Active Context
-          → Append to daily memory file (memory/YYYY-MM-DD.md)
-```
-
-### Technical Architecture
-
-Gateway provides WebSocket + HTTP dual protocol, 100+ RPC methods, compatible with OpenAI API (`/v1/chat/completions`) and MCP protocol. Plugin SDK supports 20+ lifecycle hooks. Extensions loadable via npm packages, Git repos, or bundled. Sessions isolated by channel + conversation partner.
-
-Agent system implements full reasoning loop: system prompt assembly (context files + cognitive mode + tool descriptions + memory search) → LLM reasoning → tool calls → observation → continued reasoning → streaming output. Supports context compression, parallel sub-agent dispatch, model failover, and API key rotation.
-
-Project scale: `src/agents/` (762 files), `src/infra/` (484), `src/gateway/` (356), `src/plugin-sdk/` (341), `src/plugins/` (256), `src/cognitive/` (10+ modules).
-
-## 📦 Extensions & Skills
-
-**62 extensions** covering all capability layers:
-
-| Category                   | Extensions                                                                                                                                 |
-| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------ |
-| **Messaging**              | feishu (Feishu/Lark)                                                                                                                       |
-| **China LLMs**             | Zhipu GLM · DeepSeek · Qwen · Kimi · MiniMax · Baidu Qianfan · StepFun · Volcengine · BytePlus · Kimi Coding · Xiaomi · Alibaba            |
-| **International LLMs**     | Claude · Gemini · Grok · Mistral · Perplexity · Groq · Nvidia · HuggingFace · OpenAI                                                       |
-| **Aggregators / Gateways** | OpenRouter · LiteLLM · Together · Fireworks · Cloudflare AI · Vercel AI · Copilot Proxy · Microsoft · Microsoft Foundry · Anthropic Vertex |
-| **Self-hosted**            | Ollama · LMStudio · SGLang · vLLM                                                                                                          |
-| **Dev Tools**              | OpenCode · OpenCode-Go · Open-Prose · OpenShell · Kilocode · Arcee · Chutes · Venice · Vydra · Runway                                      |
-| **Search / Browser**       | Exa · Tavily · Browser (Playwright)                                                                                                        |
-| **Memory**                 | Memory-Core · Memory-LanceDB · Memory-Wiki                                                                                                 |
-| **Voice / Media**          | Speech-Core · Talk-Voice · Media-Understanding · Image-Generation                                                                          |
-| **Utilities**              | Diffs · LLM-Task · Device-Pair · Webhooks · Shared · GitHub-Copilot                                                                        |
-
-**21 built-in skills**: github, gh-issues, weather, summarize, coding-agent, mcporter, skill-creator, session-logs, healthcheck, notion, obsidian, canvas, nano-pdf, taskflow, taskflow-inbox-triage, clawhub, video-frames, gifgrep, node-connect, blogwatcher, sherpa-onnx-tts.
-
-Need more? `kaijibot skills install <name>` to install from ClawHub.
+Config file at `~/.kaijibot/kaijibot.json`, supports hot reload. Cognitive system can be disabled via `cognitive.enabled: false`. For detailed configuration, see `AGENTS.md`.
 
 ## Acknowledgments
 

package/README.md

@@ -2,13 +2,11 @@
 
 > **你的 AI 助手会主动找你聊天，而不是干等着你提问。**
 
-Fork of [OpenClaw](https://github.com/openclaw/openclaw) · 飞书 + 30+ LLM 提供商 · 认知层让 AI 从被动变主动
+可插拔 provider/channel 架构 · 认知层让 AI 从被动变主动 · 30+ LLM 提供商
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
 [![Node.js >=22](https://img.shields.io/badge/Node.js-%3E%3D22-339933.svg)](https://nodejs.org/)
 [![TypeScript](https://img.shields.io/badge/TypeScript-6.x-3178C6.svg)](https://www.typescriptlang.org/)
-[![Vitest 450+ tests](https://img.shields.io/badge/Vitest-450%2B%20tests-6DA55F.svg)](https://vitest.dev/)
-[![简体中文](https://img.shields.io/badge/语言-简体中文-red.svg)]()
 
 **README** | [English](./README.en.md) | **简体中文**
 
@@ -42,13 +40,13 @@ KaijiBot 不一样。它在飞书里跟你聊了几次之后，会开始**主动
 
 它怎么做到的：
 
-- **Persona 画像** — 每次对话都在学习你。采用 LLM 驱动的结构化提取，从对话中自动发现领域、提取 6 类结构化洞察（领域知识、行为模式、偏好、工具配置、上下文事实、目标愿望）并存入画像。洞察按类别衰减（行为模式 60 天，工具配置 180 天），兴趣生命周期自动追踪（萌芽→稳定→衰退→沉寂→复苏）。域名由 LLM 动态发现，不再依赖硬编码关键词表。
-- **跨域洞察** — 你同时关注 A 和 B，它发现二者有潜在联系。你之前问过但没深入的问题，它从新角度跟进。你在某个领域钻得够深了，它推荐延伸方向。领域间通过共现图谱建立关联（支持 2-hop 间接连接），图谱边同样带衰减。域冷却机制避免短期内重复推荐同一领域。LLM 自我精炼循环（生成→评审→改写）保证质量，对比反重复框架确保每次推送都不同于历史，语义新鲜度检查拦截语义重复。
-- **时机门控** — 不是想发就发。基于信号检测论的 PRISM 模型计算每条洞察的期望价值，只有当预期收益超过打扰成本时才推送。凌晨不打扰，信任度低时克制，你最近活跃度低就先等等。
-- **信任演化** — 刚认识时谨慎试探，聊多了之后越来越懂你，最终变成可以大胆推荐的深度伙伴。四个阶段（SARA 框架）：定向 → 探索 → 融洽 → 伙伴，信任等级决定系统被允许做什么。
-- **偏好学习** — 你回复长了、追问了"为什么"，它记下你喜欢这个话题。你敷衍了，它下次换一个方向。每个话题维护一组 Thompson Sampling 参数（Beta 分布），隐式反馈（回复深度、响应延迟）比显式反馈更诚实。
+- **Persona 画像** — 每次对话都在学习你。LLM 驱动的结构化提取，从对话中自动发现领域和兴趣。洞察按类别独立衰减，兴趣生命周期自动追踪。领域由 LLM 动态发现，不依赖硬编码关键词。
+- **跨域洞察** — 你同时关注 A 和 B，它发现二者有潜在联系。你之前问过但没深入的问题，它从新角度跟进。你在某个领域钻得够深了，它推荐延伸方向。LLM 自我精炼保证质量，语义去重确保每次推送都有新意。
+- **时机门控** — 不是想发就发。PRISM 模型计算每条洞察的期望价值，只有预期收益超过打扰成本时才推送。凌晨不打扰，信任度低时克制，你最近活跃度低就先等等。
+- **信任演化** — 刚认识时谨慎试探，聊多了越来越懂你，最终变成可以大胆推荐的深度伙伴。SARA 框架驱动四个阶段的信任演化，信任等级决定系统被允许做什么。
+- **偏好学习** — 你回复长了、追问了"为什么"，它记下你喜欢这个话题。你敷衍了，它下次换一个方向。Thompson Sampling 驱动的偏好学习，隐式反馈比显式反馈更诚实。
 
-洞察内容结合你的画像 + LLM 知识 + 实时网络搜索生成。配了 Exa 或 Tavily API Key，洞察会紧跟时事。
+洞察内容结合你的画像 + LLM 知识 + 实时网络搜索生成。配了搜索 API Key，洞察会紧跟时事。
 
 ### 🧬 自我进化 — Agent 自主判断何时学新技能
 
@@ -56,56 +54,44 @@ KaijiBot 不一样。它在飞书里跟你聊了几次之后，会开始**主动
 
 > "我注意到你最近几次都在做类似的会议纪要归档流程，我给自己写了个技能，以后你说'归档会议'我就自动执行整个流程。"
 
-或者它发现了一个可复用的模式，当场告诉你它给自己写了个技能。
-
 它怎么做到的：
 
-- **Hard Trigger 检测** — 代码层只做一件事：检测你这次对话用了 3 个以上工具（噪音过滤）。不调 LLM、不做质量判断。
-- **Agent 自主决策** — 检测到后注入一条系统事件（含工具序列和错误信息），触发 Agent turn。Agent 拥有完整对话上下文，自己判断是否值得做成技能。不值得就忽略。
-- **无冷却无上限** — 没有代码级的频率限制或复杂度门槛。Agent 看到近期建议历史，自己决定频率。无论是否创建技能，都会告知用户结果。
-- **完整生命周期** — 创建前去重检查、创建后跟踪使用频率、30 天不用自动清理。
+- **Hard Trigger 检测** — 代码层只做噪音过滤，不做质量判断。检测到复杂任务后注入系统事件。
+- **Agent 自主决策** — Agent 拥有完整对话上下文，自己判断是否值得做成技能。不值得就忽略。
+- **完整生命周期** — 创建前去重检查、创建后跟踪使用频率、长期不用自动清理。
 
 ### 🔄 纠错自进化 — 同样的错误不犯第二次
 
-你有没有遇到过这种情况：AI 助手每次新建会话都犯同样的错。比如每次都试图创建一个空白的飞书文档，被你骂了之后改正，下次会话又忘了，照犯不误。
-
-KaijiBot 不会这样。它有一套纠错记忆系统，保证同样的错误只犯一次。
-
-它怎么做到的：
+AI 助手每次新建会话都犯同样的错？KaijiBot 不会。它有一套纠错记忆系统，保证同样的错误只犯一次。
 
-- **双路径检测** — 错误来源有两条路：Path A 是 Agent 自己发现了错误，主动调用 `record_correction` 工具上报；Path B 是你执行 `/new` 或 `/reset` 开启新会话时，系统用 60 个正则模式预筛（覆盖中英文纠正表达、质疑语气、Agent 认错等），命中后再交 LLM 分析上一轮对话，自动提取结构化纠错记录。
-- **Jaccard 去重 + 强化** — 纠错记录存入 `~/.kaijibot/cognitive/corrections/{userId}.json`，用 Jaccard 相似度判断新错误是否和已有记录重复。重复的不会新建，而是增加 `reinforcedCount`，说明这个错误反复出现，优先级更高。每人最多 50 条，TTL 90 天。
-- **系统提示注入** — 这是关键设计。纠错记录被注入到系统提示的 "Known Corrections" 区域，按强化次数排序，最多 15 条。Agent 每一轮对话都能看到它们。为什么不放在技能文件或 MEMORY.md 里？因为技能不一定被读取，MEMORY.md 只在启动时加载，只有系统提示是 Agent 每一轮必定看到的。
+- **双路径检测** — Agent 自报错误，或会话结束时系统自动从对话中提取纠错记录。
+- **去重 + 强化** — 重复的错误不会新建记录，而是增加权重，确保高频问题优先被看到。
+- **系统提示注入** — 纠错记录注入到 Agent 每一轮都能看到的系统提示中，而不是放在可能不被读取的文件里。
 
 同样的坑，踩过一次就够了。
 
-### 🔌 62 个扩展开箱即用
+### 🔌 可插拔架构
 
 不绑死任何一家。国内国际随意切换，`kaijibot onboard` 向导自动发现已配置的 API Key。
 
-| 国内（推荐）                                                                                                               | 国际主流                                                                             | 聚合 / 自部署                                                                                               |
-| -------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------- |
-| 智谱 GLM · DeepSeek · 通义千问 · Kimi · MiniMax · 百度千帆 · 阶跃星辰 · 火山引擎 · BytePlus · Kimi Coding · 小米 · Alibaba | Claude · Gemini · Grok · Mistral · Perplexity · Groq · Nvidia · HuggingFace · OpenAI | OpenRouter · LiteLLM · Together · Fireworks · Cloudflare AI · Vercel AI · SGLang · vLLM · Ollama · LMStudio |
+| 国内（推荐）                                     | 国际主流                                           | 聚合 / 自部署                                         |
+| ------------------------------------------------ | -------------------------------------------------- | ----------------------------------------------------- |
+| 智谱 GLM · DeepSeek · 通义千问 · Kimi · MiniMax … | Claude · Gemini · Grok · Mistral · Perplexity …    | OpenRouter · Together · Ollama · LMStudio · vLLM …    |
 
 切换模型只需一行：
 
 ```bash
 kaijibot config set agent.model "deepseek/deepseek-chat"
-kaijibot config set agent.model "qwen/qwen-max"
 kaijibot config set agent.model "anthropic/claude-sonnet-4-20250514"
 ```
 
-### 🛠️ 21 个内置技能 + 完整智能体
-
-**Agent 循环**：推理 → 调用工具 → 观察 → 继续推理，支持流式输出、上下文压缩、子智能体并行派生。
-
-**内置工具**：代码执行、网页抓取、PDF 操作、图片/视频/音乐生成、TTS 语音合成、Canvas 画布、文件读写、cron 定时任务，共 20+。支持模型故障转移和 API Key 轮换。
+### 🛠️ 完整智能体
 
-**记忆系统**：三种存储后端（内存、LanceDB 向量库、Wiki 知识库），语义搜索历史对话。MEMORY.md（8KB 预算）作为 Agent 每轮对话的上下文注入，由三个系统协同维护：会话记忆（`/new` 或 `/reset` 时 LLM 结构化摘要 → inline + topic）、记忆整合引擎（每日凌晨 3 点扫描 7 天会话 → 高置信度内容写入 inline sections）、记忆整理技能（手动/定期全量 GC + Jaccard 去重）。`memory_tidy` 工具覆盖 inline sections 与 topic 文件的交叉去重。纠错记忆独立存储，系统提示自动注入，确保 Agent 每次对话都能看到过去的纠错。
+**Agent 循环**：推理 → 调用工具 → 观察 → 继续推理，支持流式输出、上下文压缩、子智能体并行派生。内置代码执行、网页抓取、PDF 操作、图片/视频/音乐生成、TTS 语音合成、Canvas 画布等工具。
 
-**定时任务**：`at`（一次性）、`every`（间隔）、`cron`（cron 表达式 + 时区），支持消息投递、webhook 回调或静默执行，失败自动重试。
+**记忆系统**：多存储后端，语义搜索历史对话。会话记忆、每日整合、手动整理三个系统协同维护 Agent 上下文。
 
-**技能市场**：github、weather、summarize、coding-agent、notion、obsidian、nano-pdf、taskflow、blogwatcher 等 21 个内置技能，更多从 ClawHub 安装：
+**技能市场**：数十个内置技能（github、weather、summarize、coding-agent、notion、obsidian、taskflow 等），更多从 ClawHub 安装：
 
 ```bash
 kaijibot skills install <skill-name>
@@ -113,15 +99,37 @@ kaijibot skills install <skill-name>
 
 ## 🚀 快速开始
 
-**前置要求**：Node.js >= 22（推荐 24）、pnpm、git
+### macOS / Linux
 
-**方式一：一键安装**
+**一键安装**（推荐，自动检测环境、安装依赖、运行 onboard 向导）：
 
 ```bash
-curl -fsSL https://raw.githubusercontent.com/Kaiji-Z/kaijibot/main/install.sh | bash
+curl -fsSL https://kaijibot.ai/install.sh | bash
+```
+
+也支持 npm 全局安装：
+
+```bash
+npm install -g kaijibot
+kaijibot onboard
+```
+
+### Windows
+
+PowerShell 一键安装：
+
+```powershell
+iwr -useb https://kaijibot.ai/install.ps1 | iex
+```
+
+或 npm：
+
+```powershell
+npm install -g kaijibot
+kaijibot onboard
 ```
 
-**方式二：Docker**
+### Docker
 
 ```bash
 git clone https://github.com/Kaiji-Z/kaijibot.git
@@ -129,7 +137,7 @@ cd kaijibot
 docker compose up -d
 ```
 
-**方式三：手动安装**
+### 从源码构建
 
 ```bash
 git clone https://github.com/Kaiji-Z/kaijibot.git
@@ -137,12 +145,12 @@ cd kaijibot
 # 国内镜像加速
 pnpm install --registry https://registry.npmmirror.com
 pnpm build
-kaijibot onboard   # 交互式向导，自动配置（首次运行会检测 OpenClaw 并提示迁移）
+kaijibot onboard   # 交互式向导，自动配置
 # 从 OpenClaw 迁移？运行：
 kaijibot migrate
 ```
 
-**启动**
+### 启动
 
 ```bash
 kaijibot gateway --port 18789 --verbose
@@ -155,11 +163,9 @@ kaijibot gateway --port 18789 --verbose
 **必需**：至少一个 LLM 提供商的 API Key + 飞书机器人凭证。
 
 ```bash
-# LLM API Key（任选一个，模型用哪个就配哪个）
+# LLM API Key（任选一个）
 export ZAI_API_KEY="your-key"              # 智谱 GLM
 # export DEEPSEEK_API_KEY="your-key"       # DeepSeek
-# export DASHSCOPE_API_KEY="your-key"      # 通义千问
-# export MOONSHOT_API_KEY="your-key"       # Kimi
 # export ANTHROPIC_API_KEY="your-key"      # Claude
 # export GOOGLE_API_KEY="your-key"         # Gemini
 
@@ -175,99 +181,7 @@ export EXA_API_KEY="your-key"
 export TAVILY_API_KEY="your-key"
 ```
 
-配置文件位于 `~/.kaijibot/kaijibot.json`，支持热重载。认知系统可通过 `cognitive.enabled: false` 关闭，退化为纯 OpenClaw 体验。整合配置：`memory.consolidation.enabled`、`memory.consolidation.cron`（默认 `0 3 * * *`）、`memory.consolidation.lookbackDays`（默认 7）。
-
-## 🏗️ 架构概览
-
-### 认知洞察流程
-
-```
-事件源（定时器 + 随机抖动 / 画像变更 / 信息扫描）
-  → PRISM 门控（pNeed × pAccept > 成本阈值？）
-    → 搜索洞察机会（跨域连接 / 领域深度 / 三模式探索）
-      → 域冷却 + 饥饿加成 → 选最佳机会
-        → 语义新鲜度检查（LLM 判定与近期洞察是否语义重复）
-          → 统一管线生成洞察（模式路由 via timestamp % 100）：
-              知识模式（40%）：Web 搜索 + LLM → LLM 自我精炼（评审→改写）→ 质量重试 → LLM 验证
-              模式（Pattern）模式（50%）：对话碎片聚类 → LLM 行为洞察 → 部分验证
-              延伸模式（10%）：用户已知领域 → LLM 深度建议
-            → 对比反重复 + 安全网去重
-              → 投递到你的飞书会话
-                → 收集反馈 → Thompson Sampling 更新偏好模型
-```
-
-### 自我进化流程
-
-```
-Agent 完成任务（≥3 次工具调用）
-→ 代码检测 ≥3 次工具调用（噪音过滤，不调 LLM）
-→ 直接注入系统事件 → 触发 Agent turn
-      → Agent 看到完整对话上下文 + 近期建议历史
-        → 值得做技能？→ 生成技能草稿 → 问用户或静默创建
-        → 不值得？→ 忽略
-        → 太频繁但值得？→ 默默创建，稍后告知
-```
-
-### 纠错自进化流程
-
-```
-双路径检测
-  Path A: Agent 自报（record_correction 工具）    Path B: 会话后 LLM 提取（/new 或 /reset）
-              ↓                                              ↓
-        CorrectionStore.addOrReinforce（Jaccard 去重 + 强化）
-              ↓
-        ~/.kaijibot/cognitive/corrections/{userId}.json
-              ↓
-        下次对话 → context-writer 注入系统提示
-              ↓
-        Agent 看到历史纠错 → 避免重复
-```
-
-### 记忆整合流程
-
-```
-每日凌晨 3 点（cron 调度）
-  → 扫描最近 7 天会话记录（按 userId 分组）
-    → LLM 批量提取结构化知识（TypedInsights + 行为模式 + 纠错）
-      → Jaccard 去重 + 冲突解决
-        → 路由到认知存储：
-            PersonaStore（领域知识、偏好、目标）
-            FragmentStore（行为模式）
-            CorrectionStore（高置信度纠错）
-          → 高置信度内容（≥ 0.7）写入 MEMORY.md inline sections
-              domain_knowledge / stated_preference / behavioral_pattern → ⚡ Core Memory
-              goal_or_aspiration → 🔥 Active Context
-          → 写入每日记忆文件（memory/YYYY-MM-DD.md）
-```
-
-### 技术架构
-
-Gateway 提供 WebSocket + HTTP 双协议，100+ RPC 方法，兼容 OpenAI API（`/v1/chat/completions`）和 MCP 协议。插件 SDK 支持 20+ 生命周期钩子，扩展可按 npm 包、Git 仓库或内置方式加载。会话按渠道 + 对话方隔离。
-
-Agent 系统实现完整的推理循环：系统提示组装（上下文文件 + 认知模式 + 工具描述 + 记忆搜索）→ LLM 推理 → 工具调用 → 观察 → 继续推理 → 流式输出。支持上下文压缩、子 agent 并行派生、模型故障转移和 API Key 轮换。
-
-项目规模：`src/agents/`（762 文件）、`src/infra/`（484）、`src/gateway/`（356）、`src/plugin-sdk/`（341）、`src/plugins/`（256）、`src/cognitive/`（10+ 模块）。
-
-## 📦 扩展与技能
-
-**62 个扩展**覆盖全部能力层：
-
-| 类别              | 扩展                                                                                                                                       |
-| ----------------- | ------------------------------------------------------------------------------------------------------------------------------------------ |
-| **消息渠道**      | feishu（飞书）                                                                                                                             |
-| **国内 LLM**      | 智谱 GLM · DeepSeek · 通义千问 · Kimi · MiniMax · 百度千帆 · 阶跃星辰 · 火山引擎 · BytePlus · Kimi Coding · 小米 · Alibaba                 |
-| **国际 LLM**      | Claude · Gemini · Grok · Mistral · Perplexity · Groq · Nvidia · HuggingFace · OpenAI                                                       |
-| **聚合 / 网关**   | OpenRouter · LiteLLM · Together · Fireworks · Cloudflare AI · Vercel AI · Copilot Proxy · Microsoft · Microsoft Foundry · Anthropic Vertex |
-| **自部署**        | Ollama · LMStudio · SGLang · vLLM                                                                                                          |
-| **开发工具**      | OpenCode · OpenCode-Go · Open-Prose · OpenShell · Kilocode · Arcee · Chutes · Venice · Vydra · Runway                                      |
-| **搜索 / 浏览器** | Exa · Tavily · Browser（Playwright）                                                                                                       |
-| **记忆**          | Memory-Core · Memory-LanceDB · Memory-Wiki                                                                                                 |
-| **语音 / 媒体**   | Speech-Core · Talk-Voice · Media-Understanding · Image-Generation                                                                          |
-| **工具类**        | Diffs · LLM-Task · Device-Pair · Webhooks · Shared · GitHub-Copilot                                                                        |
-
-**21 个内置技能**：github、gh-issues、weather、summarize、coding-agent、mcporter、skill-creator、session-logs、healthcheck、notion、obsidian、canvas、nano-pdf、taskflow、taskflow-inbox-triage、clawhub、video-frames、gifgrep、node-connect、blogwatcher、sherpa-onnx-tts。
-
-需要更多？`kaijibot skills install <name>` 从 ClawHub 安装。
+配置文件位于 `~/.kaijibot/kaijibot.json`，支持热重载。认知系统可通过 `cognitive.enabled: false` 关闭。详细配置参考 `AGENTS.md`。
 
 ## 致谢
 

package/dist/.buildstamp

@@ -1 +1 @@
-{"builtAt":1780413984397,"head":"39fe3c172ca524e2c84547527c4f177ae3649b5b"}
+{"builtAt":1780451334406,"head":"953f11e8c195a331263a91ecf54adb17a936ec23"}