Also in this release

MemHub
Features8
  • Brain overview — a brain detail page opens on its auto-compiled Overview, the generated one-liner appears in the header, and you can ask MemHub to rebuild the overview on demand.
  • Routines — a routine's goal and schedule are editable from the detail page and from the anatomy card, without deleting the routine or losing its run history.
  • Brain sharing — share a brain with every active member of a workspace in one action, from Studio, chat, or MCP, instead of granting each teammate one by one.
  • Session detail — the Transcript tab renders a parsed session-gist preamble instead of burying the gist as a raw memory card.
  • Rulebook — a shell-command rule can block instead of advise: the agent's command is refused with the rule's reason attached, and RULEBOOK_OVERRIDE='<why>' is the deliberate way past it. Only a rulebook or org admin can arm one, because a gate stops every teammate's agent.
  • Rulebook — a rule that fires is surfaced in the session where it happened, rather than only showing up in the ledger afterwards.
  • Rulebook/memhub:rules-from-sessions proposes rules from what your team already does: it reads the repo's CLAUDE.md and your Claude Code, Codex, and Cursor transcripts, replays every candidate through the real hook, and states each proposal's origin, its cost, and what changes with it on. Survivors are filed as proposed; nothing is activated for you.
  • Capture — a markdown deliverable an agent writes is saved as a draft artifact when the session stops, instead of living only in the transcript.
Improvements16
  • Connected sources — the Sources page answers what this teamspace is actually granted — granted here, not granted here, org-wide, or unknown — with counts like repos, pages, teams, and channels.
  • Connected sources — Slack, GitHub, Notion, Linear, and HubSpot management share the Studio sources route shape and page shell, with clearer manage actions and fewer legacy organization-settings pages.
  • Brains — Shared with me is grouped by who shared the brain rather than by your own folders, so the tab loads one real section per sharer instead of many empty folder reads.
  • Brains — the Brains grid uses one shimmer loading front across folders, sections, and probes, so cards stop appearing in a waterfall as the page settles.
  • Routines — schedule prose is served in the reader's timezone, so the anatomy line, the badge, and the picker no longer describe one routine as running at two different times.
  • Sessions — long Transcript and Memory pages keep a compact sticky header with the session title, author, and back action after the original header scrolls away.
  • Usage digests — digest emails drop the streak, busiest-day, and peak-hour rhythm lines, so the email body is quieter.
  • Code execution sandbox — the Python runtime available to model-written code moves to pandas 2.3.3, scipy 1.18.1, and reportlab 5.0.1, re-verified against the same runtime probes as the previous bump. pandas keeps the 2.x semantics your code already relies on, and reportlab's canvas, image-reader, and table paths are unchanged.
  • Performance — database connections are pooled in front of the pooler and released before long streams begin, so concurrent requests no longer serialize behind fresh connection setup or chat and file streams.
  • Rulebook — an edit rule sees files a Bash command wrote, so a heredoc, a write_text(), or a sed -i is matched the same way a direct edit is.
  • Rulebook — a rule can scope itself to the repo and the person it applies to, with given conditions, path scope, and fixtures that verify it fires where you meant it to.
  • Rulebook/memhub:create-rule proves a new rule actually fires before filing it, and checks it against the rules already in the book, so a duplicate or a contradiction is caught at authoring time.
  • Rulebook — you are told when your team's rules stop reaching your agent, instead of the hook going quiet and the rules silently not applying.
  • Capture — a tool result too large for any request the server will accept is elided rather than failing the capture.
  • Cursor — captured turns carry real per-turn timestamps, pinned when the turn is first observed, so a Cursor session's timeline reflects when the work happened.
  • Install guides — the Codex setup documentation replaces the legacy MCP instructions.
Fixes9
  • MCP server — a refused tool call says what to fix, instead of reaching the caller as a bare Error executing tool.
  • Brains — a brain filed in another teamspace's folder no longer disappears from the current teamspace; it renders exactly once, in the matching folder or in Uncategorized.
  • Storage — legacy Sessions cache entries are purged on load, freeing localStorage that could otherwise make unrelated UI preferences, like collapsing Shared-with-me groups, silently stop saving.
  • Rulebook — plugin-facing rulebook routes accept personal access keys, and a JWT call without X-Org-Id resolves the caller's org instead of silently serving no rules to installed hooks.
  • Rulebook — an over-long rule pattern is refused at authoring time instead of being stored and then silently dropped by the hook, and fire history can be filtered by session id.
  • Rulebook — adding yourself to a rulebook no longer sends you a redundant "you were added" notification, while org-wide scope changes still notify the people newly bound by the policy.
  • Slack — @MemHub answers use the same run path as Studio, format for Slack's mrkdwn, and say when a scope gap or a run failure prevented an answer instead of falling back to a plausible but worse response.
  • Rulebook — a gate has no freshness timer: the cached rulebook is the book, so a blocking rule cannot lapse into advice because a refresh arrived late.
  • RulebookRULEBOOK_OVERRIDE is recognised at the start of the blocked command, so the documented escape hatch works where you actually type it.
Security4
  • Sessions — transcript access is no longer inherited from brain grants: members see their own sessions, org admins see the org's sessions, and a brain viewer no longer gets a teammate's transcript or session metrics just because derived memory lives in that brain.
  • Slack — hosted web search is an explicit surface-policy decision rather than an accidental bypass of the function-tool allowlist, so future hosted tools have to be classified at the one seam that can withhold them.
  • Brain sharing — workspace sharing creates explicit per-user grants as a snapshot, and does not reintroduce workspace membership as implicit brain access.
  • Rulebook — the rule-relevance judge no longer receives credentials from the session it is judging.