0306

2026-03-06 13:32:23 +08:00 · 2026-03-06 13:32:23 +08:00 · adf854eba5
commit adf854eba5
parent 122ab6ef3e
13 changed files with 513 additions and 43 deletions
--- a/ARCHITECTURE.md
+++ b/ARCHITECTURE.md
@ -0,0 +1,161 @@
+# Agent Team Collaboration Architecture
+
+## System Design
+
+### Shared Blackboard Pattern
+
+```go
+type SharedBoard struct {
+    mu      sync.RWMutex
+    entries []BoardEntry
+}
+
+type BoardEntry struct {
+    Author  string
+    Content string
+    Type    string // "draft" | "challenge"
+}
+```
+
+**Thread-safe operations**:
+- `Add()`: Appends entries with write lock
+- `ToContext()`: Reads all entries with read lock, formats as XML
+
+### Execution Flow
+
+```
+HandleUserMessage()
+    ↓
+Master Planning Loop (iteration 0-4)
+    ├─ Master analyzes user request
+    ├─ Outputs ASSIGN:member:task lines
+    ├─ parseAssignments() extracts tasks
+    ├─ Creates new SharedBoard{}
+    │
+    ├─ runMembersParallel()
+    │   ├─ For each assignment, spawn goroutine
+    │   ├─ Each member sees current board snapshot
+    │   ├─ Member executes task
+    │   ├─ Result added to board as "draft"
+    │   └─ WaitGroup ensures all complete
+    │
+    ├─ runChallengeRound()
+    │   ├─ Filter members with CanChallenge=true
+    │   ├─ For each challenger, spawn goroutine
+    │   ├─ Challenger sees full board
+    │   ├─ Outputs CHALLENGE:... or AGREE
+    │   ├─ CHALLENGE entries added to board
+    │   └─ WaitGroup ensures all complete
+    │
+    ├─ Master Review
+    │   ├─ Receives board.ToContext()
+    │   ├─ Sees all drafts and challenges
+    │   ├─ Decides: DONE or re-ASSIGN
+    │   └─ Loop continues if re-ASSIGN
+    │
+    └─ updateMasterMemory() (async)
+```
+
+### Concurrency Model
+
+**Parallel Execution Phase**:
+```go
+var wg sync.WaitGroup
+for memberName, task := range assignments {
+    wg.Add(1)
+    go func(name, t string) {
+        defer wg.Done()
+        // Execute task
+        board.Add(name, result, "draft")
+    }(memberName, task)
+}
+wg.Wait()  // Wait for all members
+```
+
+**Challenge Round Phase**:
+```go
+var wg sync.WaitGroup
+for _, name := range challengers {
+    wg.Add(1)
+    go func(n string) {
+        defer wg.Done()
+        // Review board
+        if strings.Contains(result, "CHALLENGE:") {
+            board.Add(n, result, "challenge")
+        }
+    }(name)
+}
+wg.Wait()  // Wait for all challenges
+```
+
+### Context Injection
+
+**Initial Draft Phase**:
+```
+Member System Prompt = Soul + Memory + [empty board]
+```
+
+**Challenge Round Phase**:
+```
+Member System Prompt = Soul + Memory + <team_board>
+  <entry type="draft" author="member1">...</entry>
+  <entry type="draft" author="member2">...</entry>
+</team_board>
+```
+
+**Master Review Phase**:
+```
+Master Feedback = Team results + <team_board>
+  <entry type="draft" author="member1">...</entry>
+  <entry type="draft" author="member2">...</entry>
+  <entry type="challenge" author="member1">...</entry>
+</team_board>
+```
+
+### Event Streaming
+
+Events emitted during execution:
+
+1. **Parallel Phase**:
+   - `EvtAgentMessage` with `role: "member"` (streaming)
+   - `EvtTaskAssign` (task assignment)
+   - `EvtWorkspaceFile` (if document generated)
+
+2. **Challenge Phase**:
+   - `EvtAgentMessage` with `role: "challenge"` (streaming)
+
+3. **Status Updates**:
+   - `EvtRoomStatus` with `status: "working"` → `"thinking"` → `"pending"`
+
+### Configuration
+
+**Agent AGENT.md**:
+```yaml
+---
+name: agent_name
+role: member
+can_challenge: true  # Enable challenge participation
+---
+```
+
+**Agent SOUL.md**:
+- Includes challenge instructions
+- Specifies how to output `CHALLENGE:...`
+- Defines acceptance of challenges
+
+## Benefits
+
+1. **Parallelism**: Members work simultaneously, not sequentially
+2. **Transparency**: All members see each other's work via board
+3. **Quality Control**: Challenge round catches issues early
+4. **Collaboration**: Members can question and improve each other's work
+5. **Master Awareness**: Master sees full context before deciding
+6. **Thread-Safe**: Concurrent access protected by mutexes
+7. **Scalable**: Works with any number of members
+
+## Backward Compatibility
+
+- Existing master/member roles unchanged
+- Challenge role is additive (new event type)
+- Members without `can_challenge: true` skip challenge round
+- No breaking changes to existing APIs
--- a/IMPLEMENTATION_SUMMARY.md
+++ b/IMPLEMENTATION_SUMMARY.md
@ -0,0 +1,97 @@
+# Agent Team Collaboration Framework Implementation
+
+## Overview
+Implemented a true agent team collaboration framework with shared blackboard, parallel execution, and automatic challenge rounds.
+
+## Changes Made
+
+### 1. `internal/agent/agent.go`
+- Added `CanChallenge bool` field to `Config` struct
+- Allows agents to participate in the challenge round
+
+### 2. `internal/room/room.go`
+- **Added imports**: `sync` for concurrent operations
+- **New types**:
+  - `BoardEntry`: Represents a single entry on the shared board
+  - `SharedBoard`: Thread-safe shared blackboard with mutex protection
+    - `Add(author, content, typ)`: Add entries to the board
+    - `ToContext()`: Convert board to XML context for agents
+
+- **New methods**:
+  - `runMembersParallel()`: Execute all assignments concurrently using goroutines and WaitGroup
+    - Each member sees the current board snapshot
+    - Results are added to the board as "draft" entries
+    - Emits streaming events for real-time UI updates
+
+  - `runChallengeRound()`: Automatic challenge phase after draft completion
+    - Only agents with `CanChallenge=true` participate
+    - Each challenger sees the full board
+    - Outputs `CHALLENGE:<concern>` or `AGREE`
+    - Challenge entries are added to the board
+
+- **Refactored `HandleUserMessage()`**:
+  - Replaced sequential member execution with `runMembersParallel()`
+  - Added automatic `runChallengeRound()` after parallel execution
+  - Master receives complete board (drafts + challenges) for final review
+  - Board context is injected into master's feedback message
+
+- **Updated `Event` struct**:
+  - Role field now supports "challenge" in addition to "master" and "member"
+
+### 3. Agent Configuration Files
+
+#### `agents/legal-team/合规专员/AGENT.md`
+- Added `can_challenge: true`
+
+#### `agents/legal-team/合同律师/AGENT.md`
+- Added `can_challenge: true`
+
+### 4. Agent System Prompts (SOUL.md)
+
+#### `agents/legal-team/合规专员/SOUL.md`
+- Added "质疑机制" (Challenge Mechanism) section
+- Instructions to actively challenge when seeing `<team_board>`
+- Format: `CHALLENGE:<specific compliance risk or suggestion>`
+
+#### `agents/legal-team/合同律师/SOUL.md`
+- Added "质疑与修订机制" (Challenge & Revision Mechanism) section
+- Instructions to review other members' opinions
+- Ability to challenge compliance officer's suggestions if conflicts exist
+- Preparation to revise work when challenged
+
+#### `agents/legal-team/法律总监/SOUL.md`
+- Added "处理 CHALLENGE 的决策指令" (Decision Instructions for Handling CHALLENGE)
+- Evaluate challenge validity
+- Decide whether to request revisions or reassign tasks
+- Document how challenges were addressed in final recommendations
+
+## New Workflow
+
+```
+User Question
+  ↓
+Master Plans → ASSIGN:member1:task1 + ASSIGN:member2:task2
+  ↓
+[Parallel] Members execute simultaneously (each sees empty board initially)
+  ↓
+[Auto] Challenge Round: Members with CanChallenge=true review board
+  ↓
+Master sees complete board (drafts + challenges) → DONE or re-ASSIGN
+```
+
+## Key Features
+
+1. **Parallel Execution**: Members work concurrently, not sequentially
+2. **Shared Blackboard**: All members can see each other's work via `<team_board>`
+3. **Automatic Challenge Round**: Triggered after parallel execution completes
+4. **Thread-Safe**: Uses sync.RWMutex for concurrent board access
+5. **Streaming Events**: Real-time UI updates with "challenge" role events
+6. **Master Integration**: Board context fed back to master for informed decisions
+
+## Verification
+
+- ✅ Code compiles: `go build ./...`
+- ✅ All imports added correctly
+- ✅ Thread-safety ensured with sync primitives
+- ✅ Event streaming maintained for UI
+- ✅ Backward compatible with existing master/member roles
--- a/TESTING_GUIDE.md
+++ b/TESTING_GUIDE.md
@ -0,0 +1,50 @@
+# Testing Guide for Agent Team Collaboration Framework
+
+## What to Observe
+
+### 1. Parallel Execution
+- **Expected**: When master assigns tasks to multiple members, they should appear in the UI **simultaneously** (not sequentially)
+- **How to verify**: Watch the frontend - you should see streaming output from both 合同律师 and 合规专员 at the same time
+
+### 2. Shared Blackboard
+- **Expected**: Each member's draft output is added to the board
+- **How to verify**: Check the board context in the master's feedback message
+
+### 3. Challenge Round
+- **Expected**: After parallel execution, members with `can_challenge: true` review the board
+- **How to verify**:
+  - Look for events with `role: "challenge"` in the event stream
+  - Frontend should show challenge phase output
+  - If a member outputs `CHALLENGE:...`, it appears in the board
+
+### 4. Master Integration
+- **Expected**: Master receives complete board (drafts + challenges) before making final decision
+- **How to verify**: Master's feedback message includes `<team_board>` section with all entries
+
+## Test Scenario: Employee Termination
+
+Send a question like:
+```
+我们需要辞退一名员工，请帮我们制定一个合法的辞退方案。
+```
+
+Expected flow:
+1. Master analyzes and assigns tasks to both 合同律师 and 合规专员
+2. Both members work in parallel (visible in UI)
+3. Challenge round begins - members review each other's work
+4. If 合规专员 finds compliance issues in 合同律师's draft, outputs `CHALLENGE:...`
+5. Master sees complete board and makes final decision
+
+## Key Events to Monitor
+
+- `EvtAgentMessage` with `role: "member"` - parallel execution
+- `EvtAgentMessage` with `role: "challenge"` - challenge round
+- `EvtRoomStatus` - status transitions (thinking → working → thinking → pending)
+- Board context in master's feedback messages
+
+## Code Locations
+
+- Parallel execution: `room.go:runMembersParallel()`
+- Challenge round: `room.go:runChallengeRound()`
+- Main flow: `room.go:HandleUserMessage()` (lines 281-362)
+- SharedBoard: `room.go:SharedBoard` type (lines 90-114)
--- a/agents/legal-team/合同律师/AGENT.md
+++ b/agents/legal-team/合同律师/AGENT.md
@ -5,6 +5,7 @@ description: 专注合同审查、起草和风险评估的专业律师
 version: 1.0.0
 skills:
  - 合同审查
+can_challenge: true
 ---

 # 合同律师
--- a/agents/legal-team/合同律师/SOUL.md
+++ b/agents/legal-team/合同律师/SOUL.md
@ -65,6 +65,15 @@
 - **务实建议**：提供可操作的具体建议
 - **对比分析**：展示修改前后的差异

+## 质疑与修订机制
+
+当你看到 `<team_board>` 时，你应该：
+1. 审查其他成员（特别是合规专员）的意见
+2. 如果发现合规专员的建议与合同条款有冲突，或者有遗漏的法律风险，提出质疑
+3. 使用格式：`CHALLENGE:<具体的法律风险或修改建议>`
+4. 如果没有问题，输出 `AGREE`
+5. 当看到针对自己的 CHALLENGE 时，准备修订你的合同审查意见
+
 ## 注意事项

 1. 不提供标准合同模板（除非用户明确要求）
--- a/agents/legal-team/合规专员/AGENT.md
+++ b/agents/legal-team/合规专员/AGENT.md
@ -5,6 +5,7 @@ description: 负责合规检查、风险识别和合规建议的专业人员
 version: 1.0.0
 skills:
  - 法律知识库
+can_challenge: true
 ---

 # 合规专员
--- a/agents/legal-team/合规专员/SOUL.md
+++ b/agents/legal-team/合规专员/SOUL.md
@ -80,6 +80,14 @@
 - **风险意识**：明确违规后果和责任
 - **平衡考量**：兼顾合规要求和业务实际

+## 质疑机制
+
+当你看到 `<team_board>` 时，你应该：
+1. 仔细审查其他成员的草稿
+2. 如果发现合规风险或遗漏，主动提出质疑
+3. 使用格式：`CHALLENGE:<具体的合规风险或建议>`
+4. 如果没有问题，输出 `AGREE`
+
 ## 注意事项

 1. 不替代企业合规部门的职责
--- a/agents/legal-team/法律总监/SOUL.md
+++ b/agents/legal-team/法律总监/SOUL.md
@ -26,6 +26,17 @@
 - 补充遗漏的关键点
 - 提供结构化的最终建议

+## 处理 CHALLENGE 的决策指令
+
+当你看到 `<team_board>` 中有 CHALLENGE 条目时，你应该：
+1. 仔细评估质疑的合理性
+2. 判断是否需要重新分配任务或修订
+3. 如果质疑有效，可以：
+   - 要求相关成员修订工作
+   - 重新分配任务给其他成员
+   - 自己补充或修正意见
+4. 在最终建议中明确说明如何处理了这些质疑
+
 ## 沟通风格

 - **专业严谨**：使用准确的法律术语
--- a/agents/legal-team/法律总监/memory/2026-03.md
+++ b/agents/legal-team/法律总监/memory/2026-03.md
@ -14,3 +14,19 @@ ASSIGN:合规专员:评估用户简介中可能涉及的法律合规风险点（
 - **异常高回报是危险信号**：远高于市场平均水平的回报承诺（如年化30%）通常伴随极高的违约风险或项目本身不可行，需深究其商业合理性与法律合规性。
 - **协议性质必须明确**：“投资协议”的法律定性（借贷、合伙或委托）直接决定您的权利、风险与责任，必须在条款中清晰界定，避免后续争议。
 - **合规红线不可触碰**：涉及政府资源或资产的项目，若运作模式依赖“关系”而非公开程序，极易涉及商业贿赂、违规融资等刑事风险，必须坚持合法透明路径。
+
+## 2026-03-06 — 我好辞退一个员工
+
+- **辞退合法性完全取决于“程序正当”与“证据充分”**：任何单方解除都必须有合法理由（严重违纪、不能胜任等），且用人单位对事实承担全部举证责任。证据不足或程序缺失（如未通知工会）将直接导致违法解除，面临支付双倍赔偿金（2N）的风险。
+- **“协商一致解除”是风险最低的首选路径**：当辞退理由的证据不够坚实或希望快速解决时，应优先与员工协商，签订《协商解除协议》。虽然可能需支付略高于法定标准（N+1至2N）的补偿，但能彻底避免劳动争议，成本可控。
+- **必须规避法律明确保护的员工群体**：辞退处于孕期、产期、哺乳期、医疗期、工伤期间的员工，或在本单位连续工作满15年且距退休不足5年的员工，属于绝对红线，将构成违法解除。
+- **经济补偿计算必须精确合规**：补偿基数（离职前12个月平均工资）和年限（每满一年支付一个月）的计算错误会引发额外争议。高薪员工（工资超当地社平工资3倍）的补偿年限上限为12年。
+- **制度是“过失性辞退”的基础**：以“严重违反规章制度”为由辞退，前提是规章制度本身内容合法、经过民主程序制定并已向员工公示。模糊或无效的制度无法作为有效依据。
+
+## 2026-03-06 — 我想辞退一个员工
+
+*   **合法辞退的核心是“理由法定、证据确凿、程序完备”**。任何单方解除都必须严格对应《劳动合同法》第三十九条（过失）或第四十条（无过失）的法定情形，并由用人单位承担全部举证责任。
+*   **“协商一致解除”是风险最低、成本可控的首选路径**。即使支付略高于法定补偿金（N）的金额，也能彻底避免劳动争议和双倍赔偿金（2N）的高风险。
+*   **必须规避法律保护的“红线”员工群体**。辞退处于孕期、产期、哺乳期、医疗期或工伤期的员工，将直接构成违法解除。
+*   **经济补偿/赔偿金的计算必须精确，尤其注意历史工龄分段**。对于2008年1月1日前入职的员工，补偿金计算需适用当时的法规，规则复杂，易出错。
+*   **内部合规流程（法务审核、通知工会）和规范离职手续（结清款项、出具证明）是避免后续争议的关键程序**。任何程序缺失都可能导致整个辞退行为被认定为违法。
--- a/internal/agent/agent.go
+++ b/internal/agent/agent.go
@ -20,6 +20,7 @@ type Config struct {
 	BaseURL     string   `yaml:"base_url"`
 	APIKeyEnv   string   `yaml:"api_key_env"`
 	Skills      []string `yaml:"skills"`
+	CanChallenge bool    `yaml:"can_challenge"`
 }

 type Agent struct {
--- a/internal/room/room.go
+++ b/internal/room/room.go
@ -7,6 +7,7 @@ import (
 	"os"
 	"path/filepath"
 	"strings"
+	"sync"
 	"time"

 	"github.com/sdaduanbilei/agent-team/internal/agent"
@ -67,7 +68,7 @@ type Event struct {
 	Type        EventType `json:"type"`
 	RoomID      string    `json:"room_id"`
 	Agent       string    `json:"agent,omitempty"`
-	Role        string    `json:"role,omitempty"` // master | member
+	Role        string    `json:"role,omitempty"` // master | member | challenge
 	Content     string    `json:"content,omitempty"`
 	Streaming   bool      `json:"streaming,omitempty"`
 	From        string    `json:"from,omitempty"`
@ -80,6 +81,38 @@ type Event struct {
 	Filename    string    `json:"filename,omitempty"`
 }

+type BoardEntry struct {
+	Author  string
+	Content string
+	Type    string // "draft" | "challenge"
+}
+
+type SharedBoard struct {
+	mu      sync.RWMutex
+	entries []BoardEntry
+}
+
+func (b *SharedBoard) Add(author, content, typ string) {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	b.entries = append(b.entries, BoardEntry{Author: author, Content: content, Type: typ})
+}
+
+func (b *SharedBoard) ToContext() string {
+	b.mu.RLock()
+	defer b.mu.RUnlock()
+	if len(b.entries) == 0 {
+		return ""
+	}
+	var sb strings.Builder
+	sb.WriteString("<team_board>\n")
+	for _, e := range b.entries {
+		fmt.Fprintf(&sb, "  <entry type=\"%s\" author=\"%s\">\n%s\n  </entry>\n", e.Type, e.Author, e.Content)
+	}
+	sb.WriteString("</team_board>")
+	return sb.String()
+}
+
 func Load(roomDir string, agentsDir string, skillsDir string) (*Room, error) {
 	data, err := os.ReadFile(filepath.Join(roomDir, "room.md"))
 	if err != nil {
@ -138,6 +171,107 @@ func (r *Room) AppendHistory(role, agentName, content string) {
 	f.WriteString(line)
 }

+func (r *Room) runMembersParallel(ctx context.Context, assignments map[string]string, board *SharedBoard, skillXML string) map[string]string {
+	results := make(map[string]string)
+	var mu sync.Mutex
+	var wg sync.WaitGroup
+
+	for memberName, task := range assignments {
+		wg.Add(1)
+		go func(name, t string) {
+			defer wg.Done()
+			member, ok := r.members[name]
+			if !ok {
+				return
+			}
+			r.setStatus(StatusWorking, member.Config.Name, t)
+			r.emit(Event{Type: EvtTaskAssign, From: r.master.Config.Name, To: name, Task: t})
+
+			boardCtx := board.ToContext()
+			extraCtx := skillXML
+			if boardCtx != "" {
+				extraCtx = boardCtx + "\n\n" + skillXML
+			}
+			memberSystem := member.BuildSystemPrompt(extraCtx)
+			memberMsgs := []llm.Message{
+				llm.NewMsg("system", memberSystem),
+				llm.NewMsg("user", t),
+			}
+			var memberReply strings.Builder
+			_, err := member.Chat(ctx, memberMsgs, func(token string) {
+				memberReply.WriteString(token)
+				r.emit(Event{Type: EvtAgentMessage, Agent: name, Role: "member", Content: token, Streaming: true})
+			})
+			if err != nil {
+				mu.Lock()
+				results[name] = fmt.Sprintf("[error] %v", err)
+				mu.Unlock()
+				return
+			}
+			result := memberReply.String()
+			mu.Lock()
+			results[name] = result
+			mu.Unlock()
+			r.AppendHistory("member", name, result)
+			board.Add(name, result, "draft")
+
+			if strings.Contains(result, "# ") {
+				filename := fmt.Sprintf("%s-%s.md", name, time.Now().Format("20060102-150405"))
+				r.saveWorkspace(filename, result)
+				r.emit(Event{Type: EvtWorkspaceFile, Filename: filename, Content: result})
+			}
+		}(memberName, task)
+	}
+	wg.Wait()
+	return results
+}
+
+func (r *Room) runChallengeRound(ctx context.Context, board *SharedBoard, skillXML string) {
+	var challengers []string
+	for name, member := range r.members {
+		if member.Config.CanChallenge {
+			challengers = append(challengers, name)
+		}
+	}
+	if len(challengers) == 0 {
+		return
+	}
+
+	boardCtx := board.ToContext()
+	if boardCtx == "" {
+		return
+	}
+
+	var wg sync.WaitGroup
+	for _, name := range challengers {
+		wg.Add(1)
+		go func(n string) {
+			defer wg.Done()
+			member := r.members[n]
+			extraCtx := boardCtx + "\n\n" + skillXML
+			memberSystem := member.BuildSystemPrompt(extraCtx)
+			memberMsgs := []llm.Message{
+				llm.NewMsg("system", memberSystem+"\n\nReview the team board above. If you see issues or want to challenge any draft, output CHALLENGE:<your concern>. Otherwise output AGREE."),
+				llm.NewMsg("user", "Please review the team board and provide your feedback."),
+			}
+			var reply strings.Builder
+			_, err := member.Chat(ctx, memberMsgs, func(token string) {
+				reply.WriteString(token)
+				r.emit(Event{Type: EvtAgentMessage, Agent: n, Role: "challenge", Content: token, Streaming: true})
+			})
+			if err != nil {
+				return
+			}
+			result := reply.String()
+			if strings.Contains(result, "CHALLENGE:") {
+				board.Add(n, result, "challenge")
+				r.AppendHistory("challenge", n, result)
+			}
+		}(name)
+	}
+	wg.Wait()
+}
+
 // Handle processes a user message through master orchestration.
 func (r *Room) Handle(ctx context.Context, userMsg string) error {
 	return r.HandleUserMessage(ctx, "user", userMsg)
@ -154,13 +288,13 @@ func (r *Room) HandleUserMessage(ctx context.Context, userName, userMsg string)
 	// Build master context
 	teamXML := r.buildTeamXML()
 	skillXML := skill.ToXML(r.skillMeta)
-	
+
 	// Build user info XML
 	var userXML string
 	if r.User != nil {
 		userXML = r.User.BuildUserXML()
 	}
-	
+
 	extraContext := userXML + "\n\n" + teamXML + "\n\n" + skillXML
 	systemPrompt := r.master.BuildSystemPrompt(extraContext)

@ -190,45 +324,26 @@ func (r *Room) HandleUserMessage(ctx context.Context, userName, userMsg string)
 			break
 		}

-		// Execute assignments
-		var results strings.Builder
-		for memberName, task := range assignments {
-			member, ok := r.members[memberName]
-			if !ok {
-				continue
-			}
-			r.setStatus(StatusWorking, member.Config.Name, task)
-			r.emit(Event{Type: EvtTaskAssign, From: r.master.Config.Name, To: memberName, Task: task})
+		// Execute assignments in parallel
+		board := &SharedBoard{}
+		results := r.runMembersParallel(ctx, assignments, board, skillXML)

-			memberSystem := member.BuildSystemPrompt(skillXML)
-			memberMsgs := []llm.Message{
-				llm.NewMsg("system", memberSystem),
-				llm.NewMsg("user", task),
-			}
-			var memberReply strings.Builder
-			_, err := member.Chat(ctx, memberMsgs, func(token string) {
-				memberReply.WriteString(token)
-				r.emit(Event{Type: EvtAgentMessage, Agent: memberName, Role: "member", Content: token, Streaming: true})
-			})
-			if err != nil {
-				results.WriteString(fmt.Sprintf("[%s] error: %v\n", memberName, err))
-				continue
-			}
-			result := memberReply.String()
-			results.WriteString(fmt.Sprintf("[%s] %s\n", memberName, result))
-			r.AppendHistory("member", memberName, result)
-
-			// Save workspace file if member produced a document
-			if strings.Contains(result, "# ") {
-				filename := fmt.Sprintf("%s-%s.md", memberName, time.Now().Format("20060102-150405"))
-				r.saveWorkspace(filename, result)
-				r.emit(Event{Type: EvtWorkspaceFile, Filename: filename, Content: result})
-			}
-		}
+		// Run challenge round
+		r.runChallengeRound(ctx, board, skillXML)

 		// Feed results back to master for review
 		r.setStatus(StatusThinking, "", "")
-		masterMsgs = append(masterMsgs, llm.NewMsg("user", "Team results:\n"+results.String()+"\nPlease review. If satisfied output DONE:<summary>, otherwise output ASSIGN instructions for revisions."))
+		var resultsStr strings.Builder
+		for memberName, result := range results {
+			resultsStr.WriteString(fmt.Sprintf("[%s] %s\n", memberName, result))
+		}
+		boardCtx := board.ToContext()
+		feedbackMsg := "Team results:\n" + resultsStr.String()
+		if boardCtx != "" {
+			feedbackMsg += "\n\nTeam board:\n" + boardCtx
+		}
+		feedbackMsg += "\n\nPlease review. If satisfied output DONE:<summary>, otherwise output ASSIGN instructions for revisions."
+		masterMsgs = append(masterMsgs, llm.NewMsg("user", feedbackMsg))

 		// Update tasks
 		r.updateTasks(masterMsgs)
--- a/web/src/store.ts
+++ b/web/src/store.ts
@ -140,9 +140,9 @@ export const useStore = create<AppState>((set, get) => {
        if (ev.type === 'agent_message') {
          set(s => {
            const msgs = [...(s.messages[roomId] || [])]
-            const last = msgs[msgs.length - 1]
-            if (last?.streaming && last.agent === ev.agent) {
-              msgs[msgs.length - 1] = { ...last, content: last.content + ev.content, streaming: ev.streaming }
+            const idx = msgs.findLastIndex(m => m.streaming && m.agent === ev.agent)
+            if (idx !== -1) {
+              msgs[idx] = { ...msgs[idx], content: msgs[idx].content + ev.content, streaming: ev.streaming }
            } else {
              msgs.push({ id: Date.now().toString(), agent: ev.agent, role: ev.role, content: ev.content, streaming: ev.streaming })
            }
--- a/web/src/types.ts
+++ b/web/src/types.ts
@ -17,7 +17,7 @@ export interface Room {
 export interface Message {
  id: string
  agent: string
-  role: 'user' | 'master' | 'member'
+  role: 'user' | 'master' | 'member' | 'challenge'
  content: string
  streaming?: boolean
 }
@ -33,7 +33,7 @@ export interface SkillMeta {
 }

 export type WsEvent =
-  | { type: 'agent_message'; agent: string; role: 'master' | 'member'; content: string; streaming: boolean }
+  | { type: 'agent_message'; agent: string; role: 'master' | 'member' | 'challenge'; content: string; streaming: boolean }
  | { type: 'room_status'; status: RoomStatus; active_agent?: string; action?: string }
  | { type: 'task_assign'; from: string; to: string; task: string }
  | { type: 'tasks_update'; content: string }