Expand description
rx4 — the agent harness engine.
Models write. rx4 gives them tools, memory, loops, permissions, sessions, and control planes. Hosts (CLIs, TUIs, IDEs) embed rx4.
§Safety
This crate is #![forbid(unsafe_code)] — no unsafe code is allowed anywhere.
use rx4::Agent;
let mut agent = Agent::new();
agent.set_scope(rx4::Scope::Coding);
agent.prompt("fix the failing test").await?;Re-exports§
pub use agent::normalize_tool_name;pub use agent::Agent;pub use agent::Event;pub use agent::ToolCall;pub use agent::ToolContext;pub use agent::ToolDefinition;pub use agent::ToolEffect;pub use agent::ToolExecuteBox;pub use agent::ToolExecuteFn;pub use agent::ToolExecutor;pub use agent::ToolFuture;pub use agent::ToolRegistry;pub use agent::ToolResult;pub use compaction::apply_compaction;pub use compaction::compact_messages;pub use compaction::CompactionConfig;pub use compaction::CompactionMarker;pub use compaction::CompactionResult;pub use context::compose_system_prompt;pub use context::load_project_instructions;pub use context::ProjectInstructions;pub use cost::CostEntry;pub use cost::ModelPricing;pub use cost::PricingRegistry;pub use cost::SessionCost;pub use cost::TokenUsage;pub use extract::extract_knowledge_loose;pub use extract::extract_proactive_loose;pub use extract::parse_knowledge;pub use extract::parse_proactive;pub use extract::ExtractedKnowledge;pub use extract::ProactiveItem;pub use guardrails::classify_tool;pub use guardrails::GuardrailConfig;pub use guardrails::GuardrailDecision;pub use guardrails::SelfHealingRetry;pub use guardrails::ToolClass;pub use guardrails::ToolGuardrails;pub use hooks::HookDecision;pub use hooks::HookEvent;pub use hooks::HookFn;pub use hooks::HookRegistry;pub use mode::Profile;pub use mode::Scope;pub use model_router::ModelRouter;pub use model_router::ModelRouterError;pub use model_router::ModelTier;pub use model_router::ProactiveMonitor;pub use model_router::RouterConfig;pub use model_router::SkillSuggestion;pub use model_router::SubagentModelSelector;pub use model_router::TaskTier;pub use model_router::TaskType;pub use models::CompatConfig;pub use models::ModelInfo;pub use models::ModelRegistry;pub use multiagent::CoordinatorEvent;pub use multiagent::AgentProfile;pub use multiagent::AgentRole;pub use multiagent::MultiAgentCoordinator;pub use multiagent::MultiAgentError;pub use multiagent::TeamResult;pub use multiagent::TeamTask;pub use permissions::command_from_args;pub use permissions::is_dangerous_shell_command;pub use permissions::is_process_tool;pub use permissions::is_read_only_tool;pub use permissions::is_write_tool;pub use permissions::path_outside_workspace;pub use permissions::shell_argv;pub use permissions::shell_ast;pub use permissions::shell_command_allowed;pub use permissions::shell_command_matches_all;pub use permissions::shell_command_matches_any;pub use permissions::shell_rule_matches;pub use permissions::shell_segments;pub use permissions::shell_simples;pub use permissions::AlwaysAllow;pub use permissions::AlwaysDeny;pub use permissions::ApprovalRequest;pub use permissions::Approver;pub use permissions::AsyncApprover;pub use permissions::Authorizer;pub use permissions::ChannelApprover;pub use permissions::ChannelAsyncApprover;pub use permissions::Decision;pub use permissions::PermissionMode;pub use permissions::Policy;pub use permissions::PolicyAuthorizer;pub use permissions::ShellNode;pub use permissions::ShellSimple;pub use prompt_cache::apply_cache_control;pub use prompt_cache::CachePoint;pub use prompt_cache::CachePosition;pub use prompt_cache::CacheProvider;pub use prompt_cache::CacheStats;pub use prompt_cache::CacheStatsTracker;pub use prompt_cache::CacheTtl;pub use prompt_cache::PromptCacheConfig;pub use provider::Message;pub use provider::Provider;pub use provider::ProviderRegistry;pub use provider::Role;pub use provider::StreamEvent;pub use ranking::rank;pub use ranking::rank_with_query;pub use ranking::top_n;pub use repomap::RepoMap;pub use repomap::RepoMapError;pub use rollout::RolloutEntry;pub use rollout::RolloutManager;pub use rollout::TraceWriter;pub use routing::AgentRoute;pub use routing::AgentRouter;pub use routing::RoutingConfig;pub use routing::RoutingStats;pub use routing::SmartRouter;pub use routing::TurnComplexity;pub use sandbox::detect_sandbox;pub use sandbox::OsSandbox;pub use sandbox::OsSandboxConfig;pub use sandbox::OsSandboxRunner;pub use sandbox::SandboxConfig;pub use sandbox::SandboxError;pub use sandbox::SandboxManager;pub use sandbox::SandboxProfile;pub use sandbox::SandboxViolation;pub use secrets::filter_env_vars;pub use secrets::is_sensitive_env_var;pub use secrets::RedactionConfig;pub use secrets::Redactor;pub use secrets::SecretMatch;pub use secrets::SecretPattern;pub use session::Session;pub use slash::help_text as slash_help_text;pub use slash::parse as parse_slash;pub use slash::Command as SlashCommand;pub use sse::SseError;pub use sse::SseEvent;pub use sse::SseParser;pub use subagent::SubagentConfig;pub use subagent::SubagentError;pub use subagent::SubagentHandle;pub use subagent::SubagentManager;pub use subagent::SubagentResult;pub use subagent::SubagentStatus;pub use tools::register_builtin_tools;pub use tools::register_spawn_agent_tool;pub use work_pack::WorkPack;pub use work_pack::WorkPackError;pub use acp::AcpHost;pub use acp::AcpSession;pub use marketplace::verify_plugin_integrity;pub use marketplace::InstalledPlugin;pub use marketplace::MarketplaceError;pub use marketplace::MarketplaceIndex;pub use marketplace::McpServerConfig;pub use marketplace::McpTransportKind;pub use marketplace::PluginBlocklist;pub use marketplace::PluginInstaller;pub use marketplace::PluginManifest;pub use lsp::Diagnostic;pub use lsp::DiagnosticSeverity;pub use lsp::Location;pub use lsp::LspManager;pub use lsp::LspServer;
Modules§
- acp
- ACP (Agent Client Protocol) host — JSON-RPC session surface over an embedded agent.
- agent
- Agent loop: event-driven turn cycling with tool execution, permissions, scopes, cancellation, caching, and parallel tool dispatch.
- compaction
- Semantic context compaction: token estimation, compaction markers, and oldest-first message removal that preserves the system prompt and a recent token window.
- config
- Configuration: load config from file + env.
- context
- Context: project instruction files (AGENTS.md), system prompt composition, compaction.
- cost
- Cost tracking per-model and per-session (Crush pattern).
- extract
- Structured extraction: JSON contracts for knowledge/proactive extraction.
- guardrails
- Guardrails: empty-turn detection, repeated tool failure, loop hygiene, bounded tool recursion, abort signaling, and tool effect batching.
- hooks
- Lifecycle hooks: real async dispatch with parking_lot (grok pattern).
- ipc
- JSON-RPC IPC server over Unix socket.
- lsp
- LSP client manager — JSON-RPC 2.0 over stdio with Content-Length framing.
- marketplace
- Marketplace: plugin discovery, installation, and lifecycle management.
- mode
- Work scopes (not named agents). Profiles shape tools, policy, and prompt tone.
- model_
router - Intelligent model routing: map task tiers to model preferences, with prompt heuristics, fallback resolution, and low-power proactive monitoring inspired by Hermes/omi.
- models
- Model registry: built-in model metadata, user overrides, capability detection, and per-provider compatibility configuration.
- multiagent
- Multi-agent coordination: role-based agent profiles, team spawning, and an optional in-memory event bus.
- permissions
- Permissions: policy modes, allow/deny lists, host Approver / Authorizer (pi beforeToolCall).
- plugin
- Plugin registry: host extension points loaded from marketplace installs.
- prompt_
cache - Prompt caching support for Anthropic and OpenAI providers.
- provider
- Provider abstraction with async streaming (pi_agent_rust pattern). Real SSE streaming via reqwest + eventsource-stream.
- ranking
- Proactive ranking: score and rank items by relevance.
- repomap
- Repository map: structural summary of a codebase ranked by importance.
- rollout
- Rollout persistence: durable JSONL event logging with buffered I/O.
- routing
- Smart routing: classify each user turn as simple or strong and route to different models (OpenClaude pattern), plus per-agent model routing.
- sandbox
- Sandbox: userspace filesystem, command, and environment validation layer.
- secrets
- Secrets: detection and redaction of common credential patterns in text and environment variables.
- session
- Session: conversation tree with fork/merge/persist (JSONL).
- slash
- Slash command parser.
- sse
- Optimized SSE (Server-Sent Events) parser.
- subagent
- Subagent: isolated agent execution contexts with optional git worktree isolation.
- tools
- Built-in coding tools: read, write, edit, bash, grep, find, ls + extended tools. Uses rayon for parallel search (grok pattern).
- work_
pack - Specialist work packs as markdown data (not hard-coded agent names).