What's New
RLM (Recursive Language Models)
- Phase 1: Recursive sub-cognition — agent spawns deeper thinking loops for complex problems
- Phase 2: ContextStore for persistent sub-cognition state across turns
Session Hardening
- Context compression — LLM-based mid-conversation summarization prevents context overflow
- Memory injection scanning — blocks prompt injection via poisoned memories/knowledge
- Skill capture nudge — wires up existing nudge infrastructure for skill creation prompts
- User modeling — builds evolving user profiles (role, expertise, preferences) from conversation observation, injected into system prompt as
<user_context>
Security Hardening
- RBAC on gateway commands (owner-only for config, vault, identity)
- Session LRU eviction when max_sessions exceeded
- HMAC fingerprint verification on session resume
Settings UI
- Full editable Settings panel: API keys, identity toggles, provider config
- Model selection dropdown from config.yaml routing
- Provider toggle with save feedback and error handling
Payments & Prospecting
- Payment request tool for Solana wallet
- Prospecting tools for outreach automation
- Planner BUILD enforcement + swarm skill injection
Other
- Search.sh web search integration + vault CLI key-value mode
- Dashboard: Ctrl+Y copy-to-clipboard, event feed compacted to 5 rows
- Planner spawn discipline + smarter build-vs-swarm decision
Fixes
- Browser window size now matches viewport config
- Chat hanging resolved (timeouts on auto-retrieve, embedder re-init)
- Cloud deployment fixes (WebSocket URL, health check, host binding)
- Provider routing fallback to default model
- Settings responses no longer leak into chat