server@0.75.0

Minor Changes

• 0cd8e96: Add an agent type filter to the Agent Sessions page, populated from the agent sources actually present in each project's chats via a new chat.listSources endpoint.
• 7763a1b: Tool-call blocks are now durable, first-class entities with a stable /blocks/<id> URL and 👍/👎 feedback. When the risk engine blocks a tool call, the block is persisted and its reason is injected into the agent-facing response (Claude PermissionDecisionReason, Cursor AgentMessage, Codex reason) along with a link to the block page, so the agent can reason about the denial instead of hallucinating one. New session-scoped, org-admin-gated getRiskBlock and submitRiskBlockFeedback endpoints back an in-app BlockDetailPage (under AppLayout) and a slug-free redirect resolver for the agent's external link, with a "More Info" link from the Risk Events modal.

Patch Changes

• 3464cb8: Show the assistant's creator as its owner. Assistants already recorded who created them; that attribution is now surfaced as a profile avatar (reusing the org-home member avatar treatment) on both the assistant card and the assistant setup page's overview panel. The owner resolves to one of three states: the creating member (avatar + name, full name on hover), "No owner" when the assistant was never attributed, or "Orphaned, no owner" when the creator is no longer a member of the organization. Backed by a new optional created_by_user_id field on the Assistant API type.
• a5d57cb: Fix the chat detail "Risky only" filter and rework search-within-thread. The filter previously showed nothing on threads whose findings sat on other transcript pages, and only worked for org admins via the separate risk-results endpoint. chat.load (risk_only) now returns risk_seqs — the seqs of the flagged messages — so the panel windows the full thread and filters on the authorized load (the toggle is shown only to org admins). Search now steps through every occurrence in document order — within a message's text and inside a tool call's arguments and output — with the active occurrence highlighted distinctly, instead of stepping per message and washing every hit the same colour.
• e13497f: Claude Code prompt correlation no longer stalls on high-volume sessions. Previously a chat with a large backlog of unlinked prompts could exceed the correlation time budget and fail entirely, leaving prompts unlinked from their telemetry; correlation now bounds its work and drains the backlog incrementally so prompts stay reliably linked.
• d3bad97: Shorten risk policy bypass ("Request access") links. The blocked-tool-call message now embeds a short cache-backed rpbr2.<id> token instead of a 1000+ char encrypted blob in the URL fragment. Links already issued in the legacy rpbr1 format keep working until they expire.