What’s new
- Agents Seller Risk now supports flexible metric weights with normalized score shares.
- Agents Data Analysis now accepts images, PDFs, CSV, JSON, Markdown, and text files.
- Agents Browser agent outputs now preserve task-level final results alongside page text.
- Agents PDF page selection now uses cached page images and windowed processing for large documents.
- Agents PDF source tracking now keeps heartbeats running during heavy citation matching.
- Agents KYC now adds China national ID checksum validation and visa MRZ checks.
- Agents Face match now uses ArcFace-based recognition with a gender mismatch review signal.
- Agents Knowledge base investigation agents now ship with built-in base skills and richer findings.
- QA Agent tools now include stricter output schema guidance for structured results.
- QA Web, search, scraping, and PDF child workflows now use larger timeout buffers.
- QA Bright Data Reddit and social scraping now wait longer for status polling.
- Dashboard Workflow results now support dashboard-style output blocks and a dedicated dashboard tab.
- Dashboard Risk results now use a shared score ruler and clearer score composition.
- Dashboard Connector tables now remember filters, search, page size, and sorting per organization.
- API Trace responses now use a shared typed contract with payload status metadata.
- Models Claude Opus 5 and Claude Sonnet 5 are now available in Roe.
- Models GPT-5.6 Sol is now the default model for agents.
- Models Job name generation now preserves non-Latin characters and uses GPT-5.4 Mini.
- Security Job failures now show one consistent customer-safe error while private diagnostics stay hidden.
- Security Destructive MCP tools now require explicit confirmation before deletion runs.
- Security MCP responses now redact stored connector credentials and default finalized knowledge bases to private.
- Workflows Reruns now work for completed jobs that were created without per-job inputs.