agent_debugger

Peaky Peek (agent_debugger) is a local-first, open-source audit and trust console designed specifically for debugging AI agents. Unlike traditional observability tools like LangSmith, OpenTelemetry, or Sentry, Peaky Peek focuses on answering the question operators actually care about: can I trust what this agent did? The tool captures the complete causal chain behind every agent action and reframes each run as an audit record that answers five key questions: what happened, why, with what evidence, with what result, and where it failed. Every claim is classified deterministically as verified, partially verified, contradicted, unsupported, unverified, or stale, and each session receives an explainable trust score plus a verdict card with a named posture (act, verify-first, or do-not-act). Key features include: - Agent Audit & Trust Console with deterministic verdict cards, explainable trust scores, failure narratives, verification badges, and risk signals - Decision Tree Visualization for navigating agent reasoning as an interactive tree - Checkpoint Replay for time-traveling through agent execution with checkpoint-aware playback - Trace Search for finding specific events across all sessions - Failure Clustering & Multi-Agent Coordination for adaptive analysis of similar failures - Session Comparison for side-by-side comparison of two agent runs Integration is available through three methods: a simple decorator (@trace), a context manager (trace_session), and zero-config auto-patching that works with PydanticAI, LangChain, OpenAI SDK, CrewAI, AutoGen, LlamaIndex, and Anthropic without code changes. The architecture consists of an SDK layer for instrumentation, an intelligence layer for detection and replay, a FastAPI server with SQLite storage, and a React/TypeScript frontend with nine specialized panels. The project is MIT licensed, requires Python 3.10+, and is available via pip as peaky-peek-server. Privacy features include local-first operation with no external telemetry, optional redaction pipeline, API key authentication with bcrypt hashing, and SQLite storage with no cloud dependency.