# Pomera AI Commander - Remaining Tasks
## Overview
Focused task list for remaining features.
---
## 1. MCP Server Transport Layer
Additional transport options for the embedded MCP server.
### Tasks
- [ ] Build SSE/HTTP transport with aiohttp
- Expose MCP server over HTTP at configurable port (default: 8080)
- Handle `POST /mcp` for tool calls
- Handle `GET /mcp/sse` for server-sent events
- Add CORS headers for browser clients
- [ ] Implement WebSocket transport
- WebSocket server at configurable port (default: 8081)
- Bidirectional JSON-RPC messaging
- Connection tracking and client management
### Dependencies
- `aiohttp>=3.9.0` (already in requirements.txt)
- `websockets>=12.0` (to add)
---
## 2. Multi-File Attachment for AI Tools
Add ability to attach multiple files as context for AI prompts.
### Tasks
- [ ] Add "Attach Files" button to AI Tools widget
- Multi-select file dialog (txt, md, py, json, etc.)
- Display attached file list with remove option
- Show file size and truncation indicator
- [ ] Implement file content injection
- Read and format file contents for prompt context
- Token-aware truncation (configurable max tokens per file)
- Format: `### filename.ext\n\`\`\`\n{content}\n\`\`\``
- [ ] Add attachment persistence
- Remember last attached files per session
- Optional: save attachment presets
### UI Mockup
```
┌─────────────────────────────────────────────────────┐
│ Context Files: │
│ 📄 readme.md (2.1 KB) [×] │
│ 📄 config.json (456 B) [×] │
│ 📄 main.py (8.3 KB, truncated) [×] │
│ │
│ [+ Attach Files] [Clear All] │
└─────────────────────────────────────────────────────┘
```
---
## 3. Prompt Caching for Claude Models
Implement prompt caching to reduce costs and latency for Claude models.
### Supported Providers
| Provider | Cache Support | Implementation |
|----------|---------------|----------------|
| Anthropic (Direct) | ✅ Native | `cache_control` parameter |
| OpenRouter | ✅ Pass-through | Forward cache headers |
| Azure AI | ✅ If using Claude | Via Azure headers |
| Vertex AI | ✅ Anthropic Claude | Via Vertex API |
### Tasks
- [ ] Anthropic Direct API
- Add `cache_control: {"type": "ephemeral"}` to system prompt
- Track cache hit/miss in response metadata
- Display cache stats in UI (tokens cached, savings)
- [ ] OpenRouter
- Forward Anthropic cache headers for Claude models
- Detect Claude model and enable caching automatically
- [ ] Azure AI (Claude models)
- Implement Azure-specific cache headers
- Test with Claude models on Azure
- [ ] Vertex AI (Anthropic Claude)
- Use Vertex-specific caching mechanism
- Handle Vertex API response format for cache stats
- [ ] UI Integration
- Add "Enable Prompt Caching" toggle in AI settings
- Show cache statistics (hits, misses, tokens saved)
- Display cost savings estimate
### Implementation Notes
```python
# Anthropic cache control example
{
"model": "claude-sonnet-4-20250514",
"system": [
{
"type": "text",
"text": "You are a helpful assistant...",
"cache_control": {"type": "ephemeral"}
}
],
"messages": [...]
}
```
### Cache Strategy
1. **System prompt** → Always cache (stable across requests)
2. **Attached files** → Cache with ephemeral type (per session)
3. **User messages** → Don't cache (changes each request)
---
## Implementation Priority
1. **Multi-File Attachment** — High user value, moderate complexity
2. **Prompt Caching** — Cost savings, provider-specific work
3. **MCP Transport** — Nice-to-have, enables web clients
---
## Notes
- All new features should include logging integration
- Follow existing code patterns in `tools/ai_tools.py`
- Test with multiple file types and sizes
- Cache implementation should gracefully degrade if unsupported
