NEXT STEPS:

• add a dependency view to tasks

• consider introducing document snapshots

• mentions in docs comments and relative push notificaitons and notification subscription settings

keyboard accessiblity - when pressing tab in a chat, it must focus the message composer - esc closes the comments sideba in docs - on login, pressing tab from email field should focus pasword field instead of password reset btn - keyboard shortcut for focus mode enter and exit - pressing tab while unfocused in a spreadsheet page should focus a cell - add keyboard shortcut to enter cell formula editor in the spreasheet toolbar

• internationalization (i18n and localization)

• add timezone setting to workspace, this should be inherited by all time-related values in the app. Find efficient way for this, evaluate join vs denormalization. Consider that most likely update is seldom.

• AI integrations

  • videocall transcript
  • AI bot in chat (called on mention)
  • AI dictate content
  • AI agent in videocall (unclear how to handle diarization)
  • AI document agent
  • AI tasks agent
  • agent is able to read docs
  • agent is able to read spreadsheets
  • agent is able to read diagrams
  • agent is able to read tasks

• avatars:

  • user custom avatars: look at dicebear
  • facepile not using user avatar (investigate convex cost)

• tag system follow-ups

  • cycle-scoped tag filter — taskTags already has room for a [cycleId, tagId] index when needed

• task query scaling

  • kanban active-backlog overflow when a project's uncompleted set grows past the read cap. Decision deferred. Findings so far: a board-wide limit on the active query is wrong — it truncates in index order, so one triage/backlog column (fed by inbound integration sync, so not self-limiting) eats the budget and starves "In Progress". The cap has to be per column. Preferred shape: one tasks.listBoard(projectId, perColumnLimit) query that fans out .take(cap+1) per status over by_project_status_position (single subscription; the DnD optimistic update stays a one-query patch), the existing overflow pill on any truncated column, and a [projectId, statusId] aggregate so column counts stay honest once windowed. Product alternative to weigh first: cycle-scope the board (Linear's shape) and leave the backlog to the paginated list view — smallest code, shrinks the whole problem.
  • auto-archive (Linear's structural bound): archivedAt on tasks + a workspace setting, excluded from every default read. The only thing that bounds every completed axis at once — done column, project lists, cycles, My Tasks — instead of capping each surface. Until then My Tasks caps both axes at the newest 200 and links into the project list view.

• github issues milestone ↔ cycle sync (Linear skips it too; needs a new inbound kind + outbound op in both adapters)

• self-hosted GitLab - GITLAB_BASE in gitlab/oauthClient.ts (already parameterized on cfg.base), plus literals in gitlab/outboundGateway.ts, gitlab/webhook.ts, gitlab/forceResyncAction.ts, core/syncOutMutations.ts (issue URL) and web TaskGithubBranchActions.tsx (MR URL). A per-integration base URL covers the PAT path; OAuth would also need per-workspace app credentials. Linear does PAT-only.

• I don't really know how to handle unread messages, they kind of work right now but god save me

• identity model refactor

  • split users (auth-only) from a new profiles table that can host synthetic / bot identities (integration bots, system actors); creatorId/assigneeId/userId refs point at profiles. Lets integrations create non-auth identities without polluting the auth surface.

• explore pdf reading, annotation and embedding

• evaluate cross-workspace aggregates to re-introduce admin overview

• link previews in chat

• external ticketing system??:

  • connected to project
  • connected somehow to tasks
  • external ticketer can be given access to specific resources in workspaces through links (attach to ticket??)