Crate roboticus_agent 

roboticus-agent

Agent core for the Roboticus runtime. The central module is agent_loop (mapped from loop.rs), which implements a ReAct reasoning loop as a typed state machine: Think → Act → Observe → Persist, with idle/loop detection and financial guards.

Key Types

• agent_loop – ReAct state machine with typed transitions via typestate
• tools::ToolRegistry – Trait-based tool system (10 categories)
• policy::PolicyEngine – Rule-based policy evaluation
• retrieval::MemoryRetriever – Hybrid RAG pipeline (FTS5 + vector cosine)
• analyzer::ContextAnalyzer – Topic extraction, sentiment, complexity
• recommendations::RecommendationEngine – Proactive suggestions
• obsidian::ObsidianVault – Obsidian vault integration
• skills::SkillLoader / skills::SkillRegistry – Dual-format skill system

Modules

• context – Progressive context loading (4 levels) and compression
• injection – 4-layer prompt injection defense
• prompt – System prompt builder with HMAC trust boundaries
• memory – Memory budget manager and turn ingestion
• retrieval – Hybrid search RAG pipeline with content chunking
• knowledge – KnowledgeSource trait and aggregation
• discovery – Capability discovery across tools, skills, plugins
• digest – Turn digest and history summarization
• device – Device context and hardware info
• workspace – Workspace state (file tree, git status, project metadata)
• orchestration – Multi-agent task decomposition
• governor – Rate governor and concurrency limits
• typestate – Compile-time valid state transitions
• speculative – Parallel branch evaluation with best-result selection
• manifest – Agent capability declarations
• services – Service locator and dependency wiring
• mcp – Model Context Protocol client integration
• wasm – WASM plugin runtime
• obsidian / obsidian_tools – Vault integration and read/write/search tools
• skills / script_runner – Skill loading, execution, sandboxed scripts
• analyzer / recommendations – Conversation analysis and proactive suggestions

Modules

action_planner

Deterministic action planner for task-oriented turns.

agent_loop

analyzer

approvals

capability

Unified capability registry — a catalog of invocable tools with metadata and dispatch.

compaction

Context compaction — minimize the token footprint of all injected context.

consolidation

Memory consolidation pipeline — the “dreaming” cycle.

context

device

digest

discovery

governor

ingest

Document ingestion pipeline: file -> parse -> chunk -> embed -> store.

injection

interview

knowledge

learning

Learning loop — detect successful multi-step tool sequences from completed sessions and synthesize reusable skill documents.

manifest

mcp

mcp_handler

MCP Server Handler — bridges rmcp’s ServerHandler to Roboticus’s ToolRegistry.

memory

normalization

Normalization

obsidian

obsidian_tools

orchestration

policy

prompt

ranking

Unified ranking types for memory retrieval and tool search convergence.

recommendations

retrieval

retrieval_strategy

Adaptive source discovery — policy-driven retrieval strategy selection.

script_runner

services

skills

speculative

subagents

task_state

Task Operating State — introspection-driven state synthesis.

tool_output_filter

Tool output noise filter — cleans tool output before it reaches the LLM.

tool_search

Semantic tool search — rank and prune tools before presenting to the LLM.

tools

topic

Topic segmentation for session conversation threads.

typestate

wasm

workspace