IMPLEMENTATION_PLAN.md

# Implementation Plan - Vast.ai SSH Tunnel

## Primary Objective
Replace the complex MCP/Cloudflare auto-deploy system with simple SSH tunnel for remote GPU access.

## Current Status
- ✅ Requirements documented
- ✅ Design documented  
- ✅ Tasks broken down
- ✅ Bugs documented
- ⏳ Ready to implement

## Implementation Order

### PHASE 1: Fix Critical Bugs (1-2 hours)
These bugs block testing and need to be fixed first:

1. **Fix `deliberate` float parsing bug**
   - Location: agent_compiler.py council deliberation code generation
   - Issue: Float values being concatenated without separators
   - Impact: Blocks council deliberation feature

2. **Implement `to` command**
   - Location: agent_compiler.py command parser
   - Issue: Command documented but not implemented
   - Impact: Can't test individual council members

### PHASE 2: SSH Infrastructure (2-3 hours)
Core remote execution system:

1. **Add paramiko to requirements.txt**
2. **Create RemoteGPUExecutor class** (new file: remote_gpu.py)
   - SSH connection management
   - Remote Python REPL
   - Code execution and output capture
3. **Test SSH connection** with real Vast.ai instance

### PHASE 3: Simplify vast rent (3-4 hours)
**CRITICAL**: This is quine code - must be extremely careful

1. **Remove complexity from vast rent**:
   - Delete agent upload code
   - Delete MCP server setup
   - Delete Cloudflare tunnel code
   - Keep only: rent instance + SSH setup

2. **Add SSH tunnel setup**:
   - Wait for "running" status
   - Get SSH details
   - Create RemoteGPUExecutor
   - Store in TUI state

3. **Test**: Rent GPU, verify SSH connection works

### PHASE 4: Remote Model Loading (2-3 hours)

1. **Modify plug command**:
   - Check if remote executor exists
   - If yes: load on remote GPU
   - If no: load locally

2. **Implement remote loading**:
   - Generate model loading code
   - Execute via SSH
   - Verify success

3. **Test**: Plug BGE-small on remote GPU

### PHASE 5: Remote Inference (2-3 hours)

1. **Modify infer/embed commands**:
   - Check if model is remote
   - Execute remotely if yes
   - Return results to local TUI

2. **Test**: Run embeddings on remote GPU

### PHASE 6: Polish & Documentation (1-2 hours)

1. **Add status display**: Show local vs remote GPU
2. **Add vast stop command**: Destroy instance, close SSH
3. **Update documentation**: relay_center.cmd, FelixBag guides

## Total Estimated Time: 11-17 hours

## Risk Mitigation

### Quine Structure Risk
- **Risk**: Breaking quine structure when editing agent_compiler.py
- **Mitigation**: 
  - Make small, incremental changes
  - Test after each change
  - Keep old code commented out initially
  - Verify quine hash after compilation

### SSH Compatibility Risk
- **Risk**: paramiko might not work on Windows
- **Mitigation**:
  - Test early (Phase 2)
  - Have fallback plan (use subprocess + ssh.exe)

### Remote Execution Risk
- **Risk**: Code serialization/deserialization might fail
- **Mitigation**:
  - Start with simple examples
  - Test incrementally
  - Add robust error handling

## Success Criteria

1. ✅ User can rent GPU with `vast rent 1`
2. ✅ SSH connection establishes automatically
3. ✅ User can `plug` model on remote GPU
4. ✅ User can run `embed` using remote GPU
5. ✅ Results return to local TUI correctly
6. ✅ Quine structure preserved (hash matches)
7. ✅ Works on Windows

## Next Steps

1. Start with Phase 1: Fix critical bugs
2. Test fixes with Beholder Council
3. Move to Phase 2: SSH infrastructure
4. Continue through phases sequentially

## Notes

- Primary focus: SSH tunnel implementation
- Secondary: Fix bugs that block testing
- Tertiary: Polish and additional features
- Always test after each change
- Document any new bugs discovered