The tracker that drafts itself — and never moves without your yes.

30 minutes before any recorded meeting, Slack gets a brief of attendees' open work and a suggested agenda — nobody asked for it.

A cron tick modeled on the scheduled-jobs claim-and-run clock scans calendar events starting in the next ~25–35 minutes, resolves attendees to wiki people via the calendar's identity matching, pulls each attendee's open tasks, runs a scoped agent turn to summarize what changed since the series' last occurrence, and posts the result to the meeting's Slack channel through the existing bot.

The instant a transcript finalizes, the agent reads it and proposes tasks — with a speaker, a due date, and a citation to the exact transcript page.

Hooks into the Recall webhook right after it writes raw/meetings/<month>/<date>-<botId>.md and kicks the incremental index. A scoped agent turn reads just that page, extracts explicit commitments using speaker-segment attribution, resolves names to people entities via the same identity matching calendar sync uses, and writes proposal rows — never live tasks. The identical trigger fires on the Google Meet webhook path and on ad-hoc "Record Now" calls with no calendar event behind them.

Decisions made out loud get written to a durable wiki page — the exact source the next T−30 brief reads from.

The same T+10m extraction turn, run for decisions instead of action items: outputs entities/decisions/<slug>.md (title, decided-by, date, rationale, transcript citation, linked task) via the existing wiki-write path, closing the loop the pre-meeting brief consumes as "what changed since last sync."

Every standing meeting gets a rolling scorecard — held, proposed, completed — plus a one-time flag the moment one produces literally nothing.

Groups meetings by the calendar's recurring-event ID into a series, joins against tasks and decisions sourced from each occurrence's transcript, and computes a 4-week rolling completion rate exposed as a queryable meeting_health dataset. A cheaper per-occurrence check flags "transcript finalized, extraction ran, zero outcomes" the same day with a once-only nudge to the organizer.

One inbox for every AI-drafted task action, rendered as a diff with a citation — the structural piece Linear/Jira have no concept of at all.

A /tasks/proposals view lists every pending proposal across every producing feature — meeting extraction, signal triage, stale nudges, status inference, dependency-slip flags — each rendered as current-state → proposed-state with its evidence as a clickable wiki pill, and three actions: approve, edit-then-approve, dismiss-with-reason. Approving calls the same write path a typed chat command uses. A daily digest of the same data pushes to Slack every weekday morning via the existing delivery path.

Every Slack thread, email, and doc gets scanned for an implied task before anyone files one.

A recurring per-org sweep, structurally identical to the alerts evaluation, runs right after each reindex, filters to newly-changed raw/slack, raw/gmail, and raw/googledrive pages via the wiki's change tracking (not a full rescan), and asks a scoped agent turn per active project whether the content implies an actionable task. If yes: title, description, suggested assignee, priority — citing the source page, written as a proposal.

A task stuck "blocked" gets nudged only when the wiki proves the blocker actually cleared — with the message that proves it.

A per-task condition evaluated on every reindex, the same way alert conditions run: "task X is blocked on Y; is there wiki evidence Y is resolved?" A blocked_on_ref column (ticket key or wiki path, captured when a task is marked blocked) scopes the turn to Grep only matching pages, proposing an unblock with the citation on a match. No generic time-based nag, ever.

Sees three tasks upstream that a milestone is going to slip, and says so with the full chain of evidence — not a manually-toggled red flag.

A daily scoped agent turn walks a task's declared dependency edges plus calendar proximity to a milestone date, cross-referencing wiki evidence of at-risk upstream work — a task in-progress with no recent Drive activity, a ticket that slipped estimate twice — and produces one proposal naming the root cause with every claim cited.

First-class task tools on the same agent that already reads the whole company brain — "create a task," "what's blocking the milestone," and "mark this done" just work, from chat or Slack.

New tasks/projects/task_events tables. Three agent tools registered exactly like the verified alerts pattern — create_task, update_task, query_tasks — each a zod-validated, DB-free confirmation handler (production agents run in an egress-fenced container), with the real DB write intercepted in the messages route via a new taskTool.ts matching alertTool.ts line-for-line.

The actual Linear-parity UI — but every task's citations open the same reader panel chat answers already use, scrolled to the exact cited line.

src/app/tasks/page.tsx copies the verified alerts full-canvas template; a new tasks TabId joins the rail. The view renders a keyboard-navigable list (j/k, x to select, e to edit status) and a drag-between-columns board. A task's citations render through the existing wiki-pill system; the ReaderPanel gains a #L<n> fragment handler so a transcript citation opens scrolled and highlighted to the exact segment.

Before a task is created, the agent checks whether it — or its outcome — already exists, and says so instead of filing a duplicate.

The create_task flow runs a pre-check — Grep/Glob over entities/ and raw/ plus a query_tasks call — for near-duplicate titles and topics before confirming; on a match, it surfaces the existing task or the wiki page proving it's already resolved.

Tasks inherit the same allow/deny rules as wiki pages and alerts — a contractor sees only their own work, an HR-sensitive task never leaks, zero new access-control code.

Every task gets selector labels stamped at creation. Both the API and query_tasks call the same verified subset check alerts use: every selector on the task must be within the viewer's allowed set from the org's one access policy, else owner+admin-only.

Every task is born with a citation trail — the Slack thread, doc, or meeting that created it — permanently attached. No "why does this exist" archaeology.

When the agent creates a task, it captures the wiki paths it was reading that turn (the same touched-selectors mechanism alerts already derive from Read-tool activity) as evidence_paths, rendered as clickable pills. The same field seeds an on-demand Context Pack — a one-page synthesized brief generated by the existing artifact pipeline the moment someone opens a task cold.

Any chat question — in the Tasks tab or the main chat — can answer from live task state composed with the wiki, cited the same way.

Registers a tasks dataset in the analytical lane, following the calendar dataset's verified privacy template: status, priority, project, cycle timing, assignee department — deliberately omitting titles, free text, and named individuals, mirroring the calendar's confirmed stance (no title, no organizer, no attendee list). The agent may query it before answering blocking/status/velocity questions, then compose the aggregate with a qualitative wiki pass for the "why."

One glance at who's overloaded and which roadmap items have gone quiet — composed from tasks, calendar, and the strategy docs already in the wiki.

A pre-built artifact visualizes per-person task load against real calendar availability and milestone dates, plus a priority-alignment view bucketing work into "tied to a live milestone," "orphaned," and "milestone with zero open tasks" by cross-referencing evidence paths against milestone/OKR/roadmap wiki pages. Auto-refreshed weekly and pinned to Dashboards, or generated on demand from a chat question.

Turns the already-existing read-only Jira wiki ingestion into a live overlay — every open issue appears instantly, and Jira stays the boss until a team asks it to stop.

Imports every open issue (already ingested as raw/jira/<project>/<issue>.md) into tasks as a shadow row keyed by external_id. On every LiveCEO edit, pushes the delta back via the Composio connector. On every reindex, diffs the freshly-pulled page against the shadow row and reconciles drift using the durable-comparison pattern that already guards meeting transcripts. A per-project toggle moves overlay → mirror → master.

When the PR merges, when the build ships, or when Jira and the meeting transcript disagree — the task proposes fixing itself instead of quietly going stale.

One evaluator, three triggers: (1) watches raw/github pages for merge-state changes and correlates them to in-progress tasks, proposing done the moment a reindex captures the merge; (2) for overlay tasks, this is the sync bridge's drift check; (3) for any task linked to an entity, compares status against the freshest mention in recent Slack or transcripts and flags a conflict on divergence — rather than trusting the system-of-record blindly.

Approve a proposal, close a task, or reassign it with a button click inside the Slack channel where the work actually happened — never opening the app.

Extends the bot's existing Block Kit posting to interactive messages for proposals and task mentions. The Slack events webhook gains a block_actions case calling the exact same write path the chat agent tool uses — a Slack click and a chat command are two entry points into one handler.