Repository navigation
v0.7.170
Claude Code auto-mode server-side safety review (experiential#1252; experiential 0.7.170; exp-gateway-native 0.3.140).
/v1/messagesaccepts Claude Code's top-levelsafeguardsfield (array of objects, carried opaque) instead of failing with a 400 unknown-field error. Itsanthropic-betatokendangerous-tool-use-2026-09-03is accepted silently.- On a rung whose upstream is Anthropic's official origin (api.anthropic.com, including the subscription path), the field is forwarded byte-for-byte and the beta token is injected; on every other rung (Azure Foundry, custom Anthropic-compatible origins, every non-Anthropic dialect, and batch lines) both are dropped silently, per attempt, so failover never carries them.
- The provider's
delta.safeguard_resultsfrom the finalmessage_deltais relayed unchanged in the client's streamingmessage_delta, and as a top-level key on non-streaming Messages replies; only when the provider sent it, never on Chat or Responses surfaces. Tool-use ids are unchanged, so verdicts keyed by id match. - A request carrying
safeguardsno longer replay-matches the same request without it.
Verified live: Claude Code 2.1.294 in auto mode through a local gateway built from this commit used the server's verdicts (evaluated / not_flagged keyed by the matching tool_use id), made no separate classifier requests, and printed no "isn't eligible" notice.