You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Constituent of #769. Reviewed current main at 1de4135 on 2026-10-08; this issue proposes work, no implementation is included.
CWEs: CWE-778.
Finding and evidence
The documented strict/strict_mutations audit contract says failed audit writes become client-visible errors (pkg/mcp_server/README.md). In pkg/mcp_server/gate.go, recordAudit discards the error after handleAuditFailure logs it; finalizeAudit unconditionally returns nil.
An in-memory MCP test registered the real tool wrapper with a sink whose Record returns probe audit failure, ran a mutation in full_access with failure_mode=strict, and received a successful tool result. The sink was called once and its failure was printed only to stderr.
Required outcome
Propagate sink failures according to the documented strict, strict_mutations and best_effort contracts.
Preserve the fact that execution may already have succeeded. Return an explicit audit failure that does not encourage blindly retrying a potentially completed mutation.
Decide separately whether a pre-execution audit-health gate is needed; do not imply the existing post-execution write can roll back a cloud mutation.
Validation and acceptance
Test real tool middleware with a deterministic failing sink: strict fails reads and writes, strict_mutations surfaces mutation/lifecycle failures while reads continue, and best_effort continues. Include a black-box injectable I/O-failure scenario with a mock mutation counter to prove exactly one execution and clear failure reporting.
Scope and related work
This additional finding arose while validating #769. It is distinct from secret redaction and SQL permission policy; the defect is an advertised audit guarantee that is not enforced.
Activity