Core Orchestration and Features

Core Orchestration and Features

This page summarizes the orchestration loop, core components, and operational features for bearlike/Assistant.

Execution flow

• Input arrives from a client (CLI, API, chat, or Home Assistant integration).
• The orchestrator builds a context snapshot (summary, recent events, and selected history).
• A planner produces a plan; tool selection narrows allowed tools when needed.
• Steps run through the action runner with permission checks and tool schemas applied.
• Results are synthesized into a response and written to the session transcript.

Core components

• Orchestrator (meeseeks_core.orchestrator.Orchestrator): plan/act loop and session-scoped execution.
• Planner / ToolSelector / StepExecutor / PlanUpdater / ResponseSynthesizer (meeseeks_core.planning): plan generation, tool choice, step decisions, plan updates, and response output.
• ActionPlanRunner (meeseeks_core.action_runner.ActionPlanRunner): executes action steps and applies permission policies.
• SessionRuntime (meeseeks_core.session_runtime.SessionRuntime): shared facade for CLI and API.
• SessionStore (meeseeks_core.session_store.SessionStore): transcript + summary storage.
• ToolRegistry (meeseeks_core.tool_registry.ToolRegistry): local tools and external MCP tools.

Feature highlights

• Auto-compact runs when token budget or event thresholds are reached; /compact forces a summary pass. Token thresholds are configured with token_budget.auto_compact_threshold.
• Step reflection validates tool outcomes before results are finalized.
• Langfuse tracing is session-scoped when enabled, keeping multi-turn work in one trace context.
• External MCP servers are supported via configs/mcp.json and auto-discovered at startup.
• LiteLLM-backed chat models support multiple providers and model aliases; different models can be used for planning and tool execution.
• Permission policies gate tool execution; approvals can be automatic, denied, or prompted.

Extensibility points

• Add tools by implementing AbstractTool or by registering MCP servers with schemas.
• Add hooks through HookManager for pre/post events and compaction transforms.
• Add new interfaces by reusing SessionRuntime and the event transcript model.