v1.95.0-rc.1

🐛 Bug Fixes

• Keep cache leakage time range picker inline at narrow widths.
• Handle None metadata in post_call _process_response for guardrails/model_armor.
• Stop scheduling sync failure_handler concurrently with async_failure_handler.
• Incrementally parse accumulated Gemini stream JSON to prevent multi-value wedge for Vertex AI.
• Install bedrock-realtime extra for Nova Sonic realtime in gateway.
• Handle explicit outputInfo: null in Vertex AI batch response.
• Keep an MCP tool call in one trace, anchored to its own request for OpenTelemetry.
• Keep a key's MCP toolsets when saving an edit.
• Bind key duration input to one Form.Item so pre-filled expiry submits.
• Derive the dashboard object_permission type from the generated schema.
• Keep entity usage tabs aligned with their panels.
• Stop reporting a no-op guardrail as applied on passthrough.
• Return 400 instead of 500 for chat completions without messages.
• Keep guardrail information in spend logs when the caller sends its own metadata.
• Strip include_usage from stream_options instead of dropping the param.
• Add /v1/messages support for Straiker plugin in guardrails.
• Resolve managed unified input_file_id to storage_url with ownership check before dispatch for proxy/batches.
• Derive tokens_saved when Headroom compression service omits it for guardrails.
• Attribute org spend for team-linked credentials minted without org_id for proxy.
• Make managed-file resolution additive, restore fall-back for missing-row and lookup errors for proxy/batches.
• Backfill usage on non-streaming /v1/messages responses for bedrock-mantle.
• Cap /v1/tool/spend window at 30 days and bound every SpendLogs read for proxy.
• Scope and bound the End User filter on the logs page.
• Merge model-level guardrails before pre_call_hook.
• Don't cool down parent deployment on advisor sub-call failure for router.