agent-audit-logging

What it does

Implements comprehensive audit logging and reporting for multi-agent systems. Covers event capture, structured logging, traceability chains, compliance reporting, forensic analysis, and real-time monitoring dashboards for agent actions and decisions.

How it works

• Defines a structured AuditEvent schema with event types (invocation, reasoning, tool_call, tool_result, decision, llm_response, error, handoff, human_intervention, token_usage) and fields for identity, context, traceability, and metadata.
• Builds an AuditLogger that buffers events and flushes to storage in batches; provides methods to log specific event types (log_invocation, log_tool_call, log_decision, log_error).
• Establishes a TraceabilityChain to fetch and timeline events by trace_id, and to build a graph of events for visualization.
• Implements a ComplianceReporter with methods to generate reports (summary, tool_usage, error_analysis, compliance_check) from stored events, including calculations like total_events, error rates, tool usage, and suggested actions.
• Provides a real-time AuditDashboard that reports recent activity from the in-memory buffer, including counts, activity by agent, and recent errors.

When to use it

Use when you need end-to-end observability and accountability for agent actions in production multi-agent systems, including debugging, compliance auditing, billing/tracking, and forensic analysis. Trigger reporting during periodic audits or on-demand for incident reviews. Real-time dashboard is intended for live monitoring.

What it can touch

• Storage backend for batch_write and query interfaces (via storage_backend in AuditLogger and storage in TraceabilityChain, ComplianceReporter).
• Internal in-memory buffer for recent events accessed by AuditDashboard.
• No external tool calls beyond the declared tools in the skill (claude-code) are invoked by the agent through this framework.

Caveats

• The skill defines a data model and methods but relies on an external storage backend implementation for persistence and querying.
• The step-by-step code samples assume asynchronous execution (async/await) and Python types; real integration requires compatible runtime and libraries (e.g., asyncio, threading, statistics).
• Some helper imports (e.g., Counter, defaultdict, datetime, threading, asyncio) are implied but not explicitly listed in the snippet; ensure they are available in the actual implementation.