// library
Severity
Bucket
Reject PRs adding real PAN/CVV in fixtures, seeds, or mocks. Only use Luhn-valid test PANs from the gateway or opaque tokens (e.g., tok_) and never include CVV. Add a check to fail if a PAN regex is matched. (PCI DSS data minimization)
If the PR claims to fix a specific issue (e.g., 'Fixes #123' / 'Fix PAY-123'), validate it against the real production error. - If an observability MCP is available (Sentry/Datadog/Bugsnag): fetch the event/stack trace and confirm the change addresses the root cause. - Require a regression test (or a clearly documented reason why a test cannot be added). Call out fixes that only hide symptoms (catch-and-ignore, broader retries, defaulting values) without removing the underlying failure mode.
Verify that after loading a PyTorch model, either `model.eval()` or `model.train()` is called. Failing to do so can result in incorrect behavior, especially for layers like dropout and batch normalization.
Before approving/merging risky changes, check whether there is an active critical production incident. - If an incident/monitoring MCP is available (PagerDuty, Datadog Monitors): query for active Sev1/Sev2 incidents. - If there is an active incident: warn the author and suggest delaying merges/deploys or using an explicit emergency override process.
Serve text-based assets (JS, CSS, JSON, SVG) with Brotli (br) when the client sends "Accept-Encoding: br" and fallback to gzip. Always set "Vary: Accept-Encoding" and do NOT compress already-compressed formats (e.g., .png, .jpg, .woff2).
When accepting a slice or map from external snippet or returning one, consider copying it. This prevents unintentional modifications to the original data. For example, don't store a slice argument directly if the caller might modify it later; instead, make a copy for internal use.