feat(hub): add Agent Hub — cross-agent context sharing via /save, /hu…

docs/agent-hub-design.md

@@ -0,0 +1,112 @@
+# Agent Hub Design — 跨 Agent 上下文共享
+
+## Problem
+
+WeClaw supports multiple AI agents (Claude, Codex, Gemini, etc.), each with isolated conversation sessions. Users cannot easily share context between agents for collaborative workflows like:
+- Multi-agent debate
+- Chain-of-thought analysis (Agent A outputs → Agent B reviews)
+- Collaborative content creation
+
+## Solution: Agent Hub
+
+A shared context layer built directly into WeClaw's Go codebase, using the filesystem as a persistent message board.
+
+### Architecture
+
+```
+WeChat ←→ WeClaw (handler.go)
+              │
+              ├── Agent A (isolated session)
+              ├── Agent B (isolated session)  
+              └── Agent Hub (shared context)
+                    │
+                    ├── ~/.weclaw/hub/shared/    # shared context files
+                    ├── ~/.weclaw/hub/templates/  # prompt templates
+                    └── /hub, /save commands
+```
+
+### New Commands
+
+| Command | Description | Example |
+|---------|-------------|---------|
+| `/hub` | Read all shared files and inject as context | `/hub 基于以上分析，给出你的反驳` |
+| `/hub {filename}` | Read specific file from shared | `/hub round1_claude.md 基于此反驳` |
+| `/save {filename} {message}` | Send message and save reply to shared | `/save round1.md 分析AI未来` |
+| `/hub ls` | List files in shared directory | `/hub ls` |
+| `/hub clear` | Clear all shared files | `/hub clear` |
+| `/hub pipe {target}` | Send to agent, save result, auto-chain | `/hub pipe gemini` |
+
+### Workflow Examples
+
+#### Multi-Agent Debate
+```
+1. /save round1_claude.md 从哲学角度分析AI代理是否会替代人类决策
+   → Claude replies, result saved to shared/round1_claude.md
+
+2. @gemini /hub round1_claude.md 从技术可行性角度反驳以上观点
+   → Gemini reads round1_claude.md, replies with rebuttal
+
+3. /save round2_gemini.md @gemini 从技术可行性角度反驳
+   → Gemini's rebuttal saved to shared/round2_gemini.md
+
+4. @claude /hub round2_gemini.md 作为哲学派，回应技术派的反驳
+   → Claude reads the rebuttal, responds
+
+5. /hub 综合两方观点，给出最终结论
+   → Default agent sees all shared files, synthesizes
+```
+
+#### Chain Collaboration
+```
+1. /save draft.md 写一个关于量子计算的技术博客大纲
+2. @gemini /hub draft.md 基于大纲扩写完整文章
+3. /save article.md @gemini 基于大纲扩写完整文章
+4. @claude /hub article.md 审查文章质量并优化
+```
+
+### Implementation Plan
+
+#### Phase 1: File-based shared context (current)
+- `~/.weclaw/hub/shared/` — shared context files
+- `~/.weclaw/hub/templates/` — prompt templates
+- New commands in `handler.go`: `/hub`, `/save`
+
+#### Phase 2: Auto-save with context injection
+- Agent replies auto-saved when `/save` is used
+- `/hub` auto-injects shared files as system prompt prefix
+- File naming with timestamp and agent name
+
+#### Phase 3: Chain mode (future)
+- `/hub pipe {agent}` — automatic chain: send → save → next
+- Template system for structured workflows
+- History tracking per collaboration session
+
+### File Format
+
+Shared files use Markdown with YAML frontmatter:
+
+```markdown
+---
+agent: claude
+timestamp: 2026-04-02T01:30:00+08:00
+session: debate-001
+round: 1
+---
+
+# Claude's Analysis
+
+[content here]
+```
+
+### Integration Points in weclaw
+
+1. **`messaging/handler.go`** — Add command parsing for `/hub` and `/save`
+2. **`hub/hub.go`** (new package) — Hub logic: read/write shared files, inject context
+3. **`cmd/start.go`** — Initialize Hub with default directory
+
+### Key Design Decisions
+
+1. **Filesystem over database** — Simple, inspectable, no extra dependencies
+2. **Markdown with frontmatter** — Human-readable, agent-friendly, extensible
+3. **Opt-in via commands** — No automatic cross-contamination of agent sessions
+4. **Go-native** — No Python dependencies, fits weclaw's architecture