Retrospective coverage for the October 27, 2026 issue.
A support record can lose useful context if a privacy filter mistakes an example for a secret. OpenViking’s pinned privacy documentation, available by October 27, 2026, describes a retreat from automatic discovery: new sensitive-value discovery in sessions is temporarily disabled because false positives create excessive invalid entries. It supplies no measured error rate. The product tension is concrete: discovering unknown secrets and preserving faithful diagnostic material are different objectives.
The documented session write path retains registered-value replacement before tool-output externalization and message persistence, but registration alone is insufficient. Only string values in the active Privacy catalog containing at least six Unicode characters participate in automatic scanning and substring replacement. Shorter values can still be stored, referenced and explicitly resolved. Tool inputs are traversed recursively; sensitive-looking field names no longer cause entire values to be stored and masked. Unmatched content remains unchanged, and this path creates no new Session Privacy data. The document provides no runtime switch to re-enable discovery.
Consider a hypothetical support session containing a dummy configuration example and a registered production credential represented by a string longer than six Unicode characters. Disabling discovery removes one route by which the example could be misclassified, while eligible registered-value matching can still replace the credential with a reference. An unregistered secret would not be newly discovered by this path. Our interpretation is that this shifts responsibility toward catalog maintenance—not that maintaining a catalog guarantees complete protection. The documented matcher cache also permits cross-worker staleness bounded by a 3,600-second TTL.
Preserving ordinary diagnostic text could make stored sessions more useful for later troubleshooting; that is an inference, not a demonstrated retrieval or task improvement. A reasonable alternative is to prioritize broader secret discovery despite false positives. For a support product, however, silently transforming explanatory material can undermine the record the product depends on. The restriction is session-specific: the documented add_skill path still includes LLM extraction from skill bodies.
Nor do placeholders establish redacted access. Documented session reads preserve them by default, but explicit raw=true reads through FSService.read attempt restoration using active configuration values. The feature does not encrypt storage or redact responses. The useful boundary is therefore narrower than “privacy enabled”: what gets discovered, what qualifies for replacement, and what a reader can restore are separate decisions.
