feat: Add unified agent configuration system

[marklubin]

Summary

Implements a single agents table as the source of truth for all agent configuration, replacing scattered config across agent_definitions, agent_voice_settings, and hardcoded values in code.

This is Phase 1 of the Unified Agent Configuration design.

Key Changes

• New agents table with comprehensive config:

  • Primary LLM config (inference_model, inference_provider_url, context_window, etc.)
  • Background LLM config (for summarizer, insights agents)
  • Block schema (JSONB defining all memory blocks with KP3 ref suffixes)
  • Voice (FK to voices table)
  • KP3 integration (ref prefix, hooks enabled)
  • Letta sync state (agent_id, sync_hash for change detection)

• SQLAlchemy model + Pydantic schemas + CRUD service

  • Full type safety with pyright strict mode
  • Hash-based sync detection to avoid unnecessary Letta updates

• agent-admin CLI commands:

  • list - List all configured agents
  • show <name> - Show detailed configuration
  • create - Interactive agent creation
  • delete <name> - Delete from config DB
  • export <name> - Export config as JSON
  • migrate - Migrate existing Letta agents to unified config

• Voice service updated:

  • Falls back to agents table if not found in legacy agent_voice_settings
  • TTS provider selection now uses voice.provider from database instead of env var

Migration Path

1. Run ./kx migrate to create the new table
2. Run agent-admin migrate --dry-run to preview what will be migrated
3. Run agent-admin migrate to create unified config from existing Letta agents
4. Voice lookup automatically checks both tables

Deferred to Follow-up PR

• Phase 2: BlockManagerAgent → KP3 refs integration (requires KP3 ref hooks)

Test plan

• Run migrations locally: ./kx migrate
• Test agent-admin CLI commands
• Verify voice lookup works from both legacy and new tables
• Test TTS provider selection uses voice.provider from DB

🤖 Generated with Claude Code