right-rudder

Catch the step where an AI agent contradicts its own earlier decision. CLI and adapters for LangGraph, CrewAI, Letta, OpenInference, DBOS and multi-agent logs, plus an expiry read for long-running agents.

Reproducible coherence failures in real AI agents. One row per framework (LangGraph, CrewAI, Letta, OpenHands, Phoenix, DBOS), each with a trace that replays in one command.

Coherence read for Foreman, the Codex supervisor. Flags duplicate commits and stale references on the same run Jev verifies.

Coherence read for agents in NVIDIA OpenShell sandboxes. Captures model traffic beside OpenShell's delivery record and flags writes the agent believed landed.

LangChain agent middleware that flags the step where an agent contradicts its own earlier action. Deterministic, no model calls, fits tests and CI.

Coherence read for Prime Agent (Prime Intellect) sessions. Flags the step where the orchestrator acts on something outside its committed state, such as an undelivered child message.

Benchmark of a committed-state coherence read against Jev, the typed verifier, on agent runs with injected defects. Standard library only.