A single AuditLog is shared across all middleware layers.

S13.11 -- AuditLog.verify_integrity must return (True, None) for valid log.

Concurrent emit() calls from multiple threads must not interleave. Spawn N threads writing M entries each; assert every output line is valid

Test audit logging functionality.

Test audit logging captures tool calls.

Test that actions are audited

Test audit with agents.md but no security.md fails.

Test audit without .agents/ directory returns 1.

Test audit passes with valid configuration.

Test audit reports missing files.

Test retrieving audit entries by agent and type.

An audit rule allows the action (for logging purposes).

S17.2 -- AuditSource must be a runtime_checkable Protocol.

Should calculate summary statistics.

Test the structure of audit results (without API calls).

Test audit text output uses ✗ for failures.

Test audit text output uses ✓ and ✗ symbols.

Test audit trail recording

Detection populates the audit trail.

After invocation, node_execution_log contains entries for each node.

Test audit with fully initialized project passes.

Test audit with empty .agents/ directory reports missing files.

S10.11 -- AuthCheckResult has allowed, server_name, auth_method, reason.

Agent ID is auto-generated when not provided.

S5.1 -- total_penalty is 0 when no violations.

S5.3 -- total_penalty sums multiple violation penalties.

S5.2 -- total_penalty sums single violation penalty.

S15.5 -- quarantine auto-releases after configured duration.

Session is created automatically on first validation.

AutoGen kernel with deep_hooks_enabled=False should skip hooks.

AutoGen adapter enforces tool call limits.

When governance check itself fails, on_error is called for generate_reply.

S17.3 -- AutoGenKernel has health_check.

S2.4 -- AutoGenKernel is in __all__.

AutoGen signals don't affect other adapters.