AI Cost Firewall v0.4.0
AI Cost Firewall v0.4.0 adds structured evidence lifecycle events and finalizes guard orchestration behavior for request and response processing.
Highlights
vcal.evidence.eventschema v1.1 structured application-log events;- stable
trace_idcorrelation; - exactly one
request.completedorrequest.failedterminal event for every trace that emitsrequest.received; - global rejection of
stream=truewith HTTP 422 before cache, guard, or upstream processing; - request-side and response-side Security Guard block handling;
- Security Guard
rule_idpropagation into responses, metrics, and evidence; - Privacy Guard restore skipping after response-side Security Guard blocks;
- guard orchestration metrics and latency histograms;
- safer OpenAI-compatible metadata preservation;
- sanitized upstream errors without upstream response-body leakage;
- updated deployment examples, configuration examples, dashboards, and documentation.
Dashboard boundaries
AI Firewall dashboards show cost, cache, runtime, and high-level guard orchestration health.
VCAL Security Guard and VCAL Privacy Guard keep their own module-specific dashboards.
VCAL Audit is planned as a separate evidence consumer and dashboard for retained trace search, lifecycle analysis, retention, export, and SIEM/SOAR integration.
Streaming
Streaming is not supported in v0.4.0. Requests with "stream": true return HTTP 422 before cache, guard, or upstream processing.
Evidence lifecycle
A successful request emits request.received followed by request.completed.
A failed request emits request.received followed by request.failed.
Each trace has exactly one terminal event.