Please open a GitHub Security Advisory on this repository, or email the maintainer privately if you cannot use advisories.

• Do not store real production secrets in cache without understanding local disk retention (TTL_SECONDS, class TTLs under freshness mode, MAX_ENTRIES).
• Semantic tier sends truncated, redacted prompt text to the configured adjudicator: OpenRouter Decisions (default Jev) or a local/self-hosted System One URL (ADJUDICATOR=kev|laya|…). Redaction is best-effort — treat the cache DB as sensitive. Local backends still receive prompt text on loopback; trust the host.
• Never expose /stats or /admin/* on a public bind without JEVCACHE_ADMIN_TOKEN.
• Do not use jevcache to cache tool-calling / side-effecting agent traffic (bypassed by design when tools present).
• Single-node SQLite — one process, one volume. Multi-instance against the same file is unsupported.
• Multi-tenant: set X-Jevcache-Tenant (or distinct Authorization material). Without tenants, all clients share one namespace. This is not a hardened multi-tenant cloud security boundary.

Personalized balances, secrets, or anything that must never be reused across users/sessions. Prefer bypass (X-Jevcache-Bypass: 1) or don’t run the proxy for those calls.

v0 bypasses streaming for cache population; false intent admits are possible — tune INTENT_THRESHOLD and use JEVCACHE_SHADOW=1 when validating. Freshness cues are English-biased; world changes with identical wording inside a stable TTL can still HIT; multi-turn referent shifts are not validated (last user text only). Rollback: FRESHNESS_MODE=off.