feat: add Memory & Skill Injection from Proxy (Network Level)
Implements transparent memory and skill injection at the proxy layer,
enabling AI agents connecting through OmniRoute to automatically inherit
memory and tool capabilities without client-side code changes.
Memory System:
- SQLite-backed memory store with 4 types: factual, episodic, procedural, semantic
- Token budget retrieval with configurable max tokens (default 2000)
- Memory injection into chat requests (system message + message prefix)
- Memory fact extraction from responses
- Dashboard page: /dashboard/memory
- MCP tools: omniroute_memory_search, omniroute_memory_add, omniroute_memory_clear
Skills System:
- Skill registry with semver resolution (^, ~, >=, etc.)
- Docker sandbox runner with resource constraints (CPU 100ms, RAM 256MB)
- Skill executor with timeout handling
- Built-in skills: file_read, file_write, http_request, web_search, eval_code, execute_command
- Skill schema injection for OpenAI, Claude, Gemini formats
- Tool call interception and execution
- Dashboard page: /dashboard/skills
- MCP tools: omniroute_skills_list, omniroute_skills_enable, omniroute_skills_execute
Advanced Features:
- Browser automation skill (Playwright-based)
- A2A memory-aware routing skill
- Hybrid execution mode (direct/sandbox/auto-upgrade)
- Custom skill registration API
- Integration tests
- Memory caching layer
- Memory summarization
- Performance benchmarks
Resolves: GitHub Issue #812
oyi77
e6e54822f5