DeepSeek Harness: First Use

┌──────────────────────────────────────────────────┐
│ [Standard ▼]  [deepseek-chat ▼]  ⚙️  📋  ❓  │
└──────────────────────────────────────────────────┘
   ↑Mode Selection   ↑Model Selection   ↑Settings ↑Logs ↑Help

┌─────────────────────────────────────────┐
│ 👤 Alice                                │
│ Help me analyze the project structure   │
│ in the current directory                │
├─────────────────────────────────────────┤
│ 🤖 Agent                                │
│ 🔍 Using tool: search                   │
│   → Searching in /home/alice/project... │
│ ✅ Found 15 files                       │
│                                         │
│ This project is an Express.js app.      │
│ Main structure:                         │
│ - src/routes/ — Route definitions       │
│ - src/models/ — Data models             │
│ - src/middleware/ — Middleware           │
└─────────────────────────────────────────┘

# The workspace determines the Agent's "field of view"
/home/alice/project     ← Agent can only see files in this directory
├── src/
├── package.json
└── README.md

📂 Select Workspace
┌──────────────────────────────────────┐
│ ○ /home/alice/project               │
│ ○ /home/alice/another-repo          │
│ ● Enter custom path...              │
└──────────────────────────────────────┘

# Pre-set workspace via environment variable
export DSH_WORKSPACE="/home/alice/my-project"

# Or in Web UI: Settings → Workspace → Change

Hello, please help me analyze the code quality of this project

🤖 Agent: I'll help you analyze this project.

🔍 Using tool: search
  → Pattern: *
  → Path: /home/alice/project
  → Max depth: 3

🔍 Using tool: file_edit (read)
  → File: package.json
  → Action: read

📊 Analysis Results:

1. Tech Stack:
   - Runtime: Node.js 20.x
   - Framework: Express 4.18
   - Database: PostgreSQL (pg 8.12)
   - Testing: Jest 29.x

2. Directory Structure:
   src/
   ├── routes/    — 8 route files
   ├── models/    — 5 data models
   ├── middleware/ — 3 middleware
   └── utils/     — Utility functions

3. Suggestions:
   - Missing TypeScript type definitions
   - Recommend adding src/services/ layer to separate business logic

│                                      │
│ 🔍 search                            │
│ ├─ Status: ✅ Complete               │
│ ├─ Duration: 0.3s                    │
│ └─ Files found: 23                   │
│                                      │
│ 📄 file_edit (read)                  │
│ ├─ Status: ✅ Complete               │
│ ├─ Duration: 0.1s                    │
│ └─ File: package.json (1.2KB)        │
│                                      │
│ 📊 Total tools: 2                    │
│ 📊 Total time: 0.4s                  │
└──────────────────────────────────────┘

// Tool execution pipeline pseudocode
async function executeTool(tool, params) {
  // 1. pre-execute: validation + approval
  await preExecute(tool, params);
  
  // 2. execute: actual execution
  const result = await tool.execute(params);
  
  // 3. post-execute: log recording
  await postExecute(tool, params, result);
  
  return result;
}

┌─ ⚠️ Approval Required ───────────────────┐
│                                            │
│ Agent wants to:                            │
│ 📝 Edit file: src/index.ts                 │
│                                            │
│ Changes:                                   │
│ - Line 12: Add import statement            │
│ - Line 45: Modify error handler            │
│                                            │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐    │
│ │  ✅ Allow │ │ 🔁 Always│ │  ❌ Deny │    │
│ └──────────┘ └──────────┘ └──────────┘    │
└────────────────────────────────────────────┘

# dsh.config.yaml
approval:
  # File reads: always allow
  file_read: always
  
  # File edits: require approval
  file_edit: ask
  
  # Shell commands: based on danger level
  shell:
    safe_commands: always     # ls, cat, grep, etc.
    unsafe_commands: ask      # rm, chmod, sudo, etc.
    dangerous_commands: deny  # rm -rf /, dd, etc.
  
  # Search operations: always allow
  search: always

# Set sandbox mode
export DSH_SANDBOX=strict    # Most secure
export DSH_SANDBOX=permissive # Default
export DSH_SANDBOX=off        # Not recommended

👤 Alice: Create health.ts under src/routes/, add a /health GET endpoint
        that returns { status: "ok", timestamp: Date.now() }

🤖 Agent: I'll create the health check endpoint for you.

📝 Using tool: file_edit
  → Action: create
  → Path: src/routes/health.ts
  → Content: 
    import { Router } from 'express';
    const router = Router();
    router.get('/health', (_req, res) => {
      res.json({ status: 'ok', timestamp: Date.now() });
    });
    export default router;

⚠️ Approval Required: Create file src/routes/health.ts
  → [Alice clicks Allow]

✅ File created: src/routes/health.ts

Now we need to register this route in app.ts:

📝 Using tool: file_edit
  → Action: edit
  → Path: src/app.ts
  → Changes: 
    + import healthRouter from './routes/health';
    + app.use('/health', healthRouter);

⚠️ Approval Required: Edit file src/app.ts
  → [Alice clicks Allow]

✅ File edited: src/app.ts

Done! Health check endpoint added. Test it at GET /health.

┌─ Tool Panel ────────────────────────┐
│ 📝 file_edit (create)                │
│ ├─ Path: src/routes/health.ts        │
│ ├─ Status: ✅ Approved & Complete    │
│ └─ Duration: 0.2s                    │
│                                      │
│ 📝 file_edit (edit)                  │
│ ├─ Path: src/app.ts                  │
│ ├─ Changes: +2 lines                 │
│ ├─ Status: ✅ Approved & Complete    │
│ └─ Duration: 0.1s                    │
└──────────────────────────────────────┘

Left session list → Click + button → New session

Sessions are automatically named (based on first conversation content), or can be manually renamed.

// Each SessionEvent is automatically persisted
interface SessionEvent {
  type: 'user_message' | 'agent_message' | 'tool_call' | 'tool_result' | 'approval';
  timestamp: number;
  data: Record<string, unknown>;
}