Core session persistence layer using SQLite with WAL mode. Implements session lifecycle management, message storage with role/content/tool_calls, compression-driven session rotation with lineage tracking (parent_session_id), and token counting. Features atomic writes, corruption recovery, and efficient pagination for large histories.

Advanced context compression system that summarizes conversation history when token thresholds are exceeded. Implements temporal anchoring to preserve recent context, cross-session guards to prevent leakage, media stripping for vision tokens, and configurable compression triggers. Handles reasoning content preservation and tool call budget management.

Memory management abstraction supporting multiple backend providers (Hindsight, Mem0, Honcho, OpenViking, Supermemory, ByteRover, Holographic). Implements auto-recall before turns, async retention after turns, and tool-based explicit memory operations. Provider-agnostic interface enables hot-swapping memory systems.

Skill synchronization system that clones/pulls skill repositories from GitHub, GitLab, and custom sources. Handles skill versioning, dependency resolution, and local skill development workflow. Implements atomic updates with rollback capability and skill metadata extraction from SKILL.md manifests.

Central gateway process managing WebSocket/stdio transports for CLI, TUI, and messaging platforms (Telegram, Discord, Slack, etc.). Handles session lifecycle, approval routing, tool execution streaming, and cross-platform message deduplication. Implements graceful shutdown with transcript flushing.

Universal transport layer for OpenAI-compatible chat completions across 20+ providers. Handles provider-specific quirks (Qwen message normalization, Kimi reasoning config, Gemini thinking budgets), streaming delta aggregation, automatic tool call argument repair, and fallback provider chains.

Hindsight memory provider with knowledge graph, entity resolution, and multi-strategy retrieval (semantic + entity graph). Supports cloud, local embedded (with built-in PostgreSQL), and local external modes. Implements observation consolidation (deduplicated beliefs with proof counts) and cross-memory synthesis via reflection.

Terminal execution abstraction supporting 6 backends (local, Docker, SSH, Singularity, Modal, Daytona). Implements PTY management with proper signal handling, output streaming with ANSI parsing, working directory tracking, and container resource limits. Handles Windows pywinpty and Unix ptyprocess uniformly.

Main agent conversation loop orchestrating model calls, tool execution, context compression, and interrupt handling. Manages iteration budgets, streaming callbacks, reasoning content extraction, and graceful degradation on provider failures. Implements turn retry logic with state preservation.

Lazy dependency installation system that defers optional backend SDKs (Firecrawl, Exa, Browserbase, etc.) to first use. Implements supply-chain security with exact version pinning, isolated install targets on data volume, and ABI stamping for cache invalidation. Prevents compromised PyPI releases from breaking core installs.