Crate nexo_driver_loop 

Driver agent loop. See README.md and docs/src/architecture/driver-subsystem.md for the architecture.

Re-exports

pub use acceptance::AcceptanceEvaluator;

pub use acceptance::CustomVerifier;

pub use acceptance::CustomVerifierRegistry;

pub use acceptance::DefaultAcceptanceEvaluator;

pub use acceptance::GitClean;

pub use acceptance::NoPathsTouched;

pub use acceptance::NoopAcceptanceEvaluator;

pub use acceptance::ShellResult;

pub use acceptance::ShellRunner;

pub use config::AcceptanceConfig;

pub use config::BindingStoreConfig;

pub use config::BindingStoreKind;

pub use config::CompactPolicyConfig;

pub use config::DeciderConfig;

pub use config::DeciderMemoryConfig;

pub use config::DenyShortcutConfig;

pub use config::DriverBinConfig;

pub use config::DriverConfig;

pub use config::EmbeddingProviderConfig;

pub use config::NamespaceConfig;

pub use config::PermissionConfig;

pub use config::ReplayPolicyConfig;

pub use config::WorkspaceConfig;

pub use config::WorkspaceGitConfig;

pub use error::DriverError;

pub use events::NatsEventSink;

pub use events::DriverEvent;

pub use events::DriverEventSink;

pub use events::NoopEventSink;

pub use harness::ClaudeHarness;

pub use llm_decider::LlmDecider;

pub use mcp_config::write_mcp_config;

pub use memory::DecisionMemory;

pub use memory::Namespace;

pub use memory::NoopDecisionMemory;

pub use memory::SqliteVecDecisionMemory;

pub use orchestrator::DriverOrchestrator;

pub use orchestrator::DriverOrchestratorBuilder;

pub use orchestrator::GoalOutcome;

pub use prompt::compose_turn_prompt;

pub use replay::DefaultReplayPolicy;

pub use replay::ReplayContext;

pub use replay::ReplayDecision;

pub use replay::ReplayOutcomeHint;

pub use replay::ReplayPolicy;

pub use socket::DriverSocketServer;

pub use socket::SocketMessage;

pub use workspace::GitWorktreeMode;

pub use workspace::WorkspaceManager;

Modules

acceptance

AcceptanceEvaluator and friends.

attempt

Per-turn loop. Owns the spawned claude subprocess for one turn, projects events back to the orchestrator, and synthesises an AttemptResult at the end.

compact_store

CompactSummaryStore implementations.

config

Driver runtime configuration. The full YAML at config/driver/claude.yaml deserialises into DriverConfig.

error

evaluators

LLM-as-judge AcceptanceEvaluator implementations.

events

DriverEvent enum + sink trait. Subjects when wired to NATS: agent.driver.{goal,attempt}.{started,completed}, agent.driver.{decision,acceptance,budget.exhausted,escalate}.

extract_memories

Post-turn LLM memory extraction.

extract_memories_prompt

Extraction prompt constants.

harness

ClaudeHarness — closes the AgentHarness contract from 67.0 by delegating to attempt::run_attempt.

llm_decider

LlmDecider — production PermissionDecider that consults a LlmClient (typically MiniMax). 67.4 wires this; 67.7 will plug a real DecisionMemory for semantic recall.

mcp_config

Generate the .nexo-mcp.json Claude reads via --mcp-config.

memory

DecisionMemory trait and impls.

orchestrator

Goal-level loop. Takes a Goal, drives it to completion through the per-turn attempt loop. Emits NATS events at every major transition.

post_compact_cleanup

PostCompactCleanup.

proactive

prompt

Compose the prompt fed to Claude per turn.

replay

Replay-policy. Classifies mid-turn errors observed by attempt::run_attempt into a ReplayDecision the orchestrator acts on (mark_invalid, rollback, bump consecutive_errors).

socket

Daemon-side Unix socket server. Bins spawned by Claude (SocketDecider in nexo-driver-permission) connect here and forward each permission_prompt request to the in-process decider.

workspace

Goal workspace management. Supports a git-worktree mode on top of the plain mkdir + traversal-guard surface.