PM Cross-Board Intelligence (G2 -- cross-board reasoning infrastructure)
Shipped backend infrastructure that aggregates every Mission Control board into materialized snapshots across hot, warm, and cold computation tiers, so higher-level features read pre-computed numbers instead of querying live. It backs template-based natural-language answers, per-board seasonal urgency, lead-time tracking, and three tiers of detected patterns.
On this page
What it is
Shipped backend infrastructure for cross-board reasoning, natural-language querying, seasonal urgency, lead-time tracking, and pattern detection. It has no screen of its own: it is the shared reasoning layer that higher-level features read from.
Where to find it
Nothing here is Lab-gated -- it ships, but as backend infrastructure, so you meet it through the
features that consume it rather than by opening it directly. It is feature-gated as
pm-cross-board-intelligence (pmCrossboardIntelligenceEnabled), and all computation and IPC
handlers are no-ops when disabled.
How it behaves
It aggregates data across all PM boards into materialized snapshots that higher-level features (narrative queries, dashboards, AI recommendations) read from.
An item total counts live items only. Archiving an item takes it off the board, so it leaves the count in the same breath: an answer that still added it up would report work nobody can see (BUG-SIM-056). Note that archiving and deleting are different things and neither implies the other — a row can be live and archived at once, which is how this went wrong — so a reader has to ask for both.
It runs three computation tiers:
- Hot -- event-driven (on board sync), sub-second recompute for the synced board; it expires the cross-board snapshots rather than recomputing them, so the next reader rebuilds them once.
- Warm -- every 10 minutes. Re-computes only the boards idle past the staleness age (a deliberate cost control), then rebuilds the cross-board snapshot and runs pattern detection over every board in the active scope -- a partial sweep must never produce a partial cross-board answer (BUG-SIM-013).
- Cold -- daily with chunked processing (50 boards/batch), full cleanup of expired snapshots and patterns.
Key subsystems
- Narrative queries -- natural-language answers built from templates (deterministic, no model call for the answer itself). Seven question types are answered on the spot: status, blocking, overdue, attention, summary, dependency and risk. One step can use a model: classifying a question the templates do not recognise as one of those types. That fallback is capped at 50 classifications and $1 of model spend per day (resetting at local midnight). When it is capped, has no API key, or still cannot place the question, the question is handed to Omni (or the name you gave it; see "Asking questions" below) rather than answered with "try rephrasing" or "I can't answer that". The cap is never reported, so a question that would have been understood earlier in the day is handed over exactly like one that was never understood at all.
- Replies never dead-end. With no projects yet, the box says so and offers Create a project, Brain dump with AI (when that feature is on) and Import your work — and makes no AI call at all. When projects exist but their data is still being read, it says "I'm still reading your projects - try again in a minute." A question it cannot work out, with no project named, is handed to Omni and the reply names up to five of your projects ("and N more" past that) beside the questions it can answer, so it never leaves you with only "I didn't understand". Answers say "project", never "board".
- What's blocked by what -- ask "What's blocked?", "What's blocking what?", "What is
Launch email waiting on?" or "What slips if API schema slips?" (add "on <project>" to
scope it). The answer comes only from the Dependency column: when item B lists item A,
B waits for A. "Connect projects" links are never read as blocking, finished items never
block, and an item on another account's project is never named. Chains are followed up
to 20 steps and 200 items -- a loop is walked once -- and the answer says when it stopped
early. With no links yet, the reply explains how to add a Dependency column. These
questions never use a paid model call, and the answer is read fresh each time.
- Stuck work is never hidden by the size of your workspace. An item counts as blocked when its own status says stuck, and the answer reads that status on every project in scope, whatever the project is called and however much work the workspace holds. A project whose name sorts late is read exactly like the first one, so "what is blocked?" names its stuck items instead of reporting nothing. A status label needs only an index and its text to be read: one saved without a colour is still a label, and its items still count.
- What if I push this back a week? -- ask "If I push Add API versioning strategy on Engineering Sprint Board out a week, what else slips?" or "If we delay the Postgres upgrade by 3 days, what else moves?" The answer moves the item's own due date by the amount you gave (one week when you give none, and it says it assumed that), names the sprint it would now land past and by how many days, names anything linked as waiting on it, and quotes the project's likely finish from the same what-if calculator the predictions screens use. When nothing is linked to the item it says so, and still tells you what the slip means -- such as how far past the sprint end the new date falls, and which of the owner's other items are due inside that week. An item it cannot find is said plainly, with how to ask. Only a question that either moves work or says how far the date moves counts, and the verb carrying it can be any of the usual ones — push, move, delay, slide, or simply "slips": "What if the API schema slips two weeks?" is answered as a what-if because it states a delay, while "What slips if API schema slips?" stays a dependency answer because it states no delay and is really asking what depends on it. A question about the deadline stays a risk answer.
- A status answer says how much is left. Every status answer states how many of a project's items are finished and how many are open, alongside the total and the overdue share -- "20 items (7 finished, 13 open), 3 overdue" -- per project and again across every project the answer covers. A bare total with an overdue percentage leaves the reader unable to tell nearly-done work from barely-started work, which is the one thing a status question is asked to settle. Finished work is counted once, by the same rule the finished-in-the-last-7-days answer uses, so two answers about one project can never disagree. A project with nothing finished says "0 finished" rather than staying silent about it.
- A shortened list says how many more. When an answer names only some of the items it found, it says how many it left out -- "21 items are stuck or blocked: A, B, C, D, E and 16 more". The number stated and the names shown always add up, because the remainder comes from the full count rather than from the trimmed list. A list that happens to hold exactly as many items as it shows adds no "and more" clause.
- The attention answer leads with the on-track verdicts, never a count of cards. "What needs
attention?" opens with each project's on-track verdict (the same judgment "Is it on track?" uses:
a count line such as "Across 3 projects: 1 not on track, 1 at risk, 1 on track." and one line for
each project not on track, with its deciding reason). Then it gives the warning signs in plain
facts, then names the stuck or blocked items. It never says "N urgent issues". The stuck items are
named whether or not any warning sign exists, so blocked work is never hidden.
- "Proceeding normally" only when the verdicts agree. With no warning sign and nothing stuck, it says the work is "proceeding normally" only when every project is judged on track. If any project is at risk, not on track or can't be judged yet, it says no warning signs were found and gives the verdicts instead. When the verdicts could not be read, it states the facts and claims no verdict.
- Several projects are ranked worst first by verdict. With more than one project in scope the answer opens "Most in need of attention, worst first: …" (up to five). The order is the on-track verdict (not on track, at risk, can't tell yet, on track), then how many days the finish date slips (the likely finish past the date, or a passed date until today; no likely finish slips the most), then how much other work is held up. A project with no problem and nothing to keep an eye on is left out. Item counts never decide the order.
- The status answer states real numbers and each project's verdict. A status question gives the real counts (items, finished, open, the overdue share) and labels every project in the breakdown with its on-track verdict, for example "Backend Services (at risk): 20 items (0 finished, 20 open), 1 overdue". It never says "N urgent issues". Stuck items are named with their owners. The blocking answer works the same way: a pattern description is added to the named items, never swapped for them.
- "Is it on track?" is a judgment, not a percentage. A status question that says "on track"
is answered by judging each project against its date: the active sprint's end, or else the
latest date set on its open items. No overdue share or blended score decides it.
- What it weighs:
- the saved finish forecast, with its most likely date and its safe date, built from how fast the team really finishes work
- stuck items, and what they hold up
- late items
- work due while its owner is out, with nobody covering
- overloaded people
- pace
- work piling up in one step: items there have waited longer than the team usually takes in that step, and at least one is due within a week or holds up other work (how many items sit there never decides it; with too little history it says it can't tell yet)
- work due within a week that hasn't moved in 5 days
- the project's total budget
- How it decides:
- The most likely finish is after the date, or the date has passed with work still open: No.
- Otherwise, the safe finish is after the date, or a real problem is found: At risk. Real problems are a stuck or late item holding up other work, work due while its owner is out with nobody covering, and spending past the total budget.
- Otherwise, the forecast lands in time on real finished work: Yes.
- Otherwise: Can't tell yet, naming what is missing.
- Smaller warnings never flip the answer. These are late items that hold nothing up, a stuck item that holds nothing up, overload, a slowing pace, a pile-up, quiet work due soon, and one item over its own budget. They are listed under "Keep an eye on".
- A rough guess never says Yes. A forecast without enough finished work behind it can only warn that the work will be late. That covers a velocity-only guess, or one built from the team's own time estimates.
- Every answer explains itself. It states the verdict and the date, then "Because:" (the deciding facts), "Keep an eye on:", "Looks fine:" and "I couldn't check:". It closes by saying it used only the project's own items, not email or meetings.
- Several projects get a count line. For example, "Across 3 projects: 1 not on track, 1 at risk, 1 on track." Then one line per project follows, worst first.
- Only saved forecasts are used. The answer never starts a new forecast. A missing one reads as "still being worked out".
- The answer carries its facts. The response includes an
onTracklist holding each project's verdict and reasons, so the command-line route returns the same judgment. - Other progress questions ("How is X going?") keep the count-first answer, never say "Things look on track" based on an overdue share, and label each project with this same verdict.
- "Check email, meetings and chat to make sure". Under the answer, a button looks for news the boards don't show yet, such as a vendor delay, someone away, or a date moved in a meeting.
- Nothing private is read until you press it.
- It reads only your email, meeting recordings and Team Chat, and only the ones you have said yes to. A source without an answer gets the same "May I check?" Yes or No as What Next, saved in the same place.
- A company admin's "outside sources off" switch blocks it, with a plain note.
- Each finding shows the words it rests on, and anything it can't back with what it read is dropped.
- The verdict and the dates never change. If a date should move, change it on the board.
- It says which sources it checked and which it couldn't read. "Nothing found" is said only about a source that answered.
- With several projects, it checks the three most at risk and says how many it left out.
- It always finishes inside 30 seconds. It is on the computer only, and it needs What Next switched on.
- An agent calling it from the command line shares What Next's limit of 2 asks an hour; after that it is told to wait, and nothing is read.
- Contract: pm-on-track-outside-check-contract.md.
- Four examples:
- Website Relaunch: At risk. "Homepage copy" is due Nov 4 while Priya is out Nov 2 to 6, and nobody else is on it, even though the work most likely finishes Nov 13 against a Nov 20 sprint end.
- Q4 Marketing Campaign: Yes. 15 of 60 items are late, but the open work most likely finishes Nov 30 (safe date Dec 6), before Dec 15. The late items are listed under "Keep an eye on".
- Warehouse Move: No. At about 4 finishes a week, the 20 open items most likely finish around Nov 21, three weeks after the Oct 31 sprint end, even though nothing is late yet.
- Mobile App Beta: Can't tell yet. Only 1 item is finished and there is no sprint or due date, so it says what to add.
- Contract: pm-on-track-judgment-contract.md.
- What it weighs:
- The Home Today strip uses the same judgment. Its "Urgent" list shows exactly the projects judged No or At risk, worst first, each with one plain reason and never a money amount. The sidebar number counts them. That count is reused until the projects, a saved forecast or the day changes, so a refresh stays cheap.
- Seasonal calendar -- per-board quarterly urgency multipliers, each held between 0.5 and 3 (one shared pair of constants for the request contract and the store, which refuses anything outside that range) so a single busy quarter can never dominate the urgency score.
- Seasonal risk (Busy seasons and risk) -- each board rates its next four quarters from its
real work: open items due by each quarter's end (overdue and Stuck items count against this
one, because they still have to be finished) against how many items the team actually finished
in the last four weeks (read from item history, never from sync timestamps). Only that pace
check sets the level: a count of Stuck or overdue items never raises it on its own. Overdue
items show as a separate "Keep an eye on" line under the reason, and Stuck items lead the next
step. Each board sets every quarter to Quiet, Normal, Busy or
Peak in Stats → Mission Control → Projects; a busier step leaves less room for planned work, but a
step alone never makes a quarter risky. Every quarter comes back with a verdict (On track ·
Tight · Likely to slip · Will miss dates without help · Not enough history yet · Nothing
scheduled yet), a plain-English reason with the real counts, and a next step. Viewers see it
read-only. Shown in Project Analytics and
GET /pm/intelligence/seasonal/risk-preview, which give the same answer. Contract: pm-seasonal-risk-contract.md. Who's out: below the quarter cards, members record time off once per person (it counts on every board that person has open work on): the dates, a plain type (Vacation, Leave, Sick, Other) and full or half time, and nothing else, no reason and no notes. Pace counts working days (Monday to Friday, and not the project's company holidays when those are on; see below) and leaves out days people were away, so a short-staffed month never reads as a slow team; each quarter's room for planned work shrinks by the share of the team's working time that is out, and the reason says so as a count only ("1 person on this board is out for part of this quarter, which takes about 20% of the team's time"). CLI:GET /pm/intelligence/time-off,GET /pm/intelligence/time-off/people,POST /pm/intelligence/time-off,DELETE /pm/intelligence/time-off/:id(viewers refused on writes, and an entry can only be added, changed or removed through a board its person is on). Each entry carriessource(manualorcalendar). Shared with the team: Who's out is one list per Mission Control account. Every save or removal is kept on the computer first and sent to the Mission Control server in the background (/time-offon the server), and every teammate's app pulls the team's list back, so an entry typed on one computer shows on everyone's within about 15 minutes. Offline changes wait and are sent later, and a change waiting to be sent is never overwritten by an older server copy. Anyone (not a guest) can record their own time off; members and admins can record a teammate's. Contract: pm-shared-whos-out-contract.md. From Google Calendar: with the switch "Add my out-of-office from Google Calendar" on (the default, settingpmTimeOffCalendarImport) and Google connected, each person's own app reads their own primary calendar every 15 minutes, from 30 days back to 120 days ahead: Google out-of-office events, and all-day events they made called PTO, vacation, holiday, leave, time off or day off. They become that person's time off, tagged "From calendar": typed Vacation when the title says vacation, PTO or holiday and Other otherwise (never Sick or Leave), full day for six or more working hours (9 AM to 5 PM), half day for three to six. Only the dates are kept; titles never leave the calendar code. Moving or deleting the event moves or removes the entry; removing an imported entry in Who's out keeps it removed. Only the person themselves or an account admin can remove a day that came from a calendar. No one's app reads anyone else's calendar: teammates' days off arrive through the shared list. Showing the type: by default a teammate sees only that someone is out and when, not whether it is Vacation, Leave, Sick or Other. A board owner or account admin can turn on "Show time-off type to the team" in the board's menu (the ⋯ next to the board name; stored on the server asshowTimeOffType, off for every new board). Then everyone who can see that board sees the type for the people on it. Each person always sees their own type, and account admins see everyone's. The server leaves the type out of each teammate's copy unless the setting allows it, so a hidden type never reaches their computer; the app keeps such an entry as just "out". Each app learns the setting, and whether its user is an admin, from the board sync. Company holidays: each project has one list of company days off, shown above Who's out (Statistics → Project Analytics → Busy seasons and risk → Company holidays). It is behind the Lab switch "Mission Control company holidays", which is off by default while in development; with it off, every count works exactly as before. Contract: pm-board-holidays-contract.md.- Who can change it: the project's owner or a company admin. Everyone who can open the project sees the list. Anyone else sees "Only project owners and company admins can change holidays."
- The starting list is the US paid holidays: New Year's Day, Memorial Day, Independence Day, Labor Day, Thanksgiving, the day after Thanksgiving and Christmas Day.
- An admin can remove any day, including a starting one, and it stays removed in later years. To add a day, pick from the US list (MLK Day, Presidents' Day, Good Friday, Juneteenth, Columbus/Indigenous Peoples' Day, Veterans Day, Christmas Eve, New Year's Eve, or a starting holiday that was removed), enter a day of your own that repeats every year, or enter a one-time range of up to 31 days (the last day is optional). A list holds at most 60 days, and a name is 1 to 60 characters.
- Weekend rule: a yearly holiday on a Saturday is taken the Friday before, and on a Sunday the Monday after. That can reach into the previous year. A day off is counted once, even when two holidays land on it.
- Shared with the team: the list is kept on the Mission Control team server, and each computer keeps a copy that updates with the normal project sync. A project that is only on your computer keeps its list there, and its owner or admin on that computer edits it.
- What it changes: a holiday is not a working day in the Risk Heatmap's business-day windows and pace, quiet item working days, seasonal risk (workdays left and pace), and Who's out capacity. Time off that falls on a holiday removes no extra capacity.
- Two notes: until the team server has been updated, the editor shows "Showing the US holidays. Changes will be possible after the next Mission Control server update." and the counts still use the US list. Offline, it shows the last saved list.
- CLI:
GET /pm/boards/:boardId/holidays,POST /pm/boards/:boardId/holidays,DELETE /pm/boards/:boardId/holidays/:key(a key isus:<id>for a US-list day orrow:<id>for a day added by an admin). The phone does not reach these, the same as Who's out. Learn our busy season: a button above the quarter cards opens a window that suggests each quarter's level from history. - Sources, up to 6 at once: this project's past work; another project you can open, including ones brought in with Import; a CSV or Excel file (read once in memory, 10 MB limit, never saved); or a Google Sheets link (needs Google connected; first visible tab, up to 50,000 rows, read for at most 20 seconds).
- For a spreadsheet you pick the date column, and either count each row or add up a number column such as orders or revenue.
- Each source shows its share of the year per calendar quarter, with a plain reason ("About 63% of this work lands in Oct to Dec, about 2.5 times a typical quarter. Based on 80 dated entries since Feb 2024."). Several sources are averaged, each counted once.
- A quarter's share over an even 25% maps to a level: up to 0.875 times is Quiet, up to 1.25 times Normal, up to 1.75 times Busy, above that Peak.
- It needs at least 30 dated entries across 3 different quarters from the last three years; otherwise it says "Not enough history to suggest a season yet".
- Nothing changes until Use these, which saves all four levels through the normal save and keeps each quarter's note. Viewers can look but not apply.
- CLI:
POST /pm/intelligence/seasonal/learn/columns,POST /pm/intelligence/seasonal/learn. Over the CLI one learn request carries up to 6 files of 10 MB each. - Documents can teach it too: a Mission Control document from the project's workspace, a Google Doc link (needs Google connected), a .md or text file (2 MB limit), a Share (read from its KMS page or original file), a KMS note, or pasted text.
- Before any document is sent, a warning titled "Documents are read by generative AI" explains that its text goes to an AI service (Claude, by Anthropic) to be read once, that Omniscio keeps no copy, and asks people not to send anything proprietary or confidential. Nothing is sent until Send to AI; Cancel sends nothing. Projects and spreadsheets never show the warning.
- The AI returns a split per quarter and one sentence quoted from the document. The source counts only when the split adds up and the sentence really appears in the document. The reason then reads '"Sales plan" says: "<that sentence>"'. A document that says nothing about timing is listed but not counted. Text is capped at 100,000 characters, and nothing is stored.
- With Mission Control's AI helpers set to low, or no API key, document sources say so and read
nothing; projects and spreadsheets still work. When a request leaves nothing readable, the CLI
route refuses by cause: an AI-off source among them answers 422
UNPROCESSABLEwith the sentence naming where to turn AI on, and a purely unreadable request answers 502BAD_GATEWAY. CLI:GET /pm/intelligence/seasonal/learn/documentslists what can be picked, andPOST .../seasonal/learnneedsconsentToAi: truefor any document source. Linked teams: a "Linked teams" line above the quarter cards shows each linked project as a chip ("Warehouse, 2 quarters ahead"), with "Link a team" for an editor. - Pick a project you can open in the same workspace and how far ahead this team works (same quarter up to 3 quarters ahead). Up to 3 links per board.
- A link only SUGGESTS. When the linked project is busier in the matching quarter, that quarter card shows "Warehouse is at Peak in Oct to Dec, and this team works 2 quarters ahead. Make this quarter Peak too?" with Use Peak and Keep Normal. Nothing changes until one is pressed; Keep hides it until the linked project's season changes.
- One step only: a linked project's own links are never followed.
- Someone who cannot open the linked project sees "A linked team".
- CLI:
POST /pm/intelligence/seasonal/links,DELETE /pm/intelligence/seasonal/links/:id,POST /pm/intelligence/seasonal/suggestions/decline; the preview carrieslinksandsuggestions. Busy-season check-up: the built-in "Busy Season and Risk Check-Up" super prompt ships with every install. Its Start the check-up button lived on the PM Advisor panel, which was removed when the PM Advisor folded into Omni. The whole section uses the Mission Control teal.
- Lead-time tracker -- SOP-derived lead-time configs and status-transition observation tracking.
- Pattern detector -- three tiers of patterns (operational, strategic,
predictive) with TTL expiry. Its agent-stall pattern reads an item's status the
same way the board summary does, below.
A pattern's identity is its tier, type and scope -- never its numbers. So a
condition that keeps being detected stays ONE card that refreshes, rather than a
new card per detection with a superseded figure on the old one. A pattern you
dismiss stays dismissed; the detection does not bring it back inside its TTL.
An urgent card about one item -- an item nearing its worst-case lead time -- is one card
per item (and per lead-time setup), so a board with three such items shows three cards,
each refreshing on its own. The usual cap of three new urgent cards per board per hour still
applies.
The cross-board overdue trend is scoped to the whole workspace, not to the
boards a particular sweep happened to cover, so it no longer shows a
"projects affected" count and appears once rather than once per sweep. On a shared
computer each signed-in person gets their own copy of it, and dismissing a card
hides it only for the person who dismissed it.
A project with at least one item waiting on unfinished work gets a "work is waiting"
card (a link icon, warning level) naming the biggest hold-up when that item sits on the
card's own project, and saying instead that the hold-up is on another project when it does
not, so a card never names work on a project it does not cover; while it shows, the question
box offers "What's blocked by what?". Once nothing is waiting the card is no longer produced
and fades within 4 hours.
Every Home card that judges appears only on a real signal, never on one count or one percentage
(Nov 2026). The Overdue card is a plain fact: one overdue item shows it at High.
Each card's High / Medium / Low level comes from the judgment behind it, never from a number:
the project's on-track answer (No is High, At risk is Medium, otherwise Low), the worst answer
across projects for the "overdue rising" card, and the workload answer for a person card
(Overloaded is High). Omni suggestions follow the same rule, and only High and Medium insights
become a suggestion.
- Blocks may be cascading -- a late or blocked item holds up other unfinished work through a real "waits on" link. Many blocked and overdue items with no links between them show nothing.
- Agent stalled -- the agent session on an unfinished item stopped (stalled or errored), no other session is working on it, and nothing has moved for 4 hours.
- Stuck or blocked share -- shows the correct share, and appears only when a blocked item has waited 2 or more days or is due within 7 days.
- Bottleneck -- items in one step have sat longer than 85 of every 100 past items took there (and at least 2 days). Backlog, To do and Not started are never called a bottleneck, and a step needs 10 past stays before it is judged.
- Velocity may be low -- the finish forecast from the team's real pace lands after the project's date.
- Cross-project overdue -- overdue work across projects is rising and the typical late item is 5 or more days late; the overdue share is shown alongside. Omni's insights follow the same rule: "shared team members are overcommitted" needs the workload check to find someone on 2 or more projects Overloaded or Tight, "cascading delays" needs a real link between projects, and "effort drift" needs work past its estimate while a due date moved later. Detected patterns belong to the person signed in to Omniscio, the same as their boards, never to whichever Claude login is active. Switching or adding a Claude login does not hide them from the Advisor or Intelligence answers. (Until 2026-10-16 the detector filed them under the Claude login, so no reader could see them.)
- Feature gate -- all IPC handlers check
requirePmEnabled()(no multi-tier permission model is implemented).
Asking questions
The question box on Mission Control's Home and My Work pages, and the "Ask the advisor" box in the Project Manager Advisor panel, never answer with a dead end such as "try rephrasing" or "I can't answer that".
The question box reads the question before it is answered. One model step turns what you typed into a structured reading: what kind of question it is, which board, item and person it means, and whether it asks more than one thing. The reading is checked against the boards, items and people you can see, so it can only point at something that exists — a name you cannot see is never answered from. The reading decides what is being asked; the numbers in the answer still come from the same fact builders that always computed them, so nothing is invented. While it reads, the box shows Thinking…, and if the reading has not come back within about eight seconds the answer is built from the built-in word matcher instead so a question never waits.
When two projects share the name you used, the question is answerable. If the name you typed could mean more than one board, the box asks which one — "I found ADV-Team Week 2026-11-01 and ADV-Team Week 2026-11-02. Which one did you mean?" — and remembers what it offered. Your next message picks one: the part of the name that tells them apart ("the 11-02 one", or just "11-02", and "11-2" counts as the same thing) or the position it was listed in ("the second one"). The original question is then answered in full for that one project, so the clarification finishes rather than becoming a dead end (BUG-SIM-167). The offer lasts for one following message and is then gone, so it can never hijack a later question; a message that names its own project, or that is itself a question, is answered as its own question and the offer is dropped. Nothing is kept across an app restart — ask again and you get the same choice question.
A count of a project's items means the items the project shows. Subitems are work nested under a parent item and the board lists them under that parent, so they are not counted as project items. "How many open items" therefore reports the same number the board's own item count shows: a board with 40 items and 6 subitems answers 40, not 46 (BUG-SIM-171). The same applies to the finished and open split, so open can never exceed what the board holds and overdue can never exceed open. The same rule holds for the counts a group or a status question reports.
The project you name outranks a word that only appears inside its name. A group or status whose title happens to sit inside the project's name can no longer take the question: "How many open items are on ADV-Sourcing & MFG?" is answered about that project even though it contains a group titled
ADV-Sourcing(BUG-SIM-171). A question that really does name a group still gets the group's own items.A name that several projects carry is stated once, not once per project. An overdue answer names the item a single time and says which projects carry it — "across 3 projects: A, B, C" — so one item never reads as three (SIM-172).
A project that keeps subitems says so beside its item count. Subitems are work nested under a parent, and the answer names them separately — "8 items, plus 2 subitems" — so the number the board shows and the number its raw item list holds are both visible rather than one being silently chosen (SIM-171).
Two items can share a name, and the answer says which project each is on. When the same item name appears more than once in one answer — two projects each holding an item called "Retest the export flow on a sample board", say — each entry names its project, in the same
"Name" (Project, detail)shape the owner answers already use (BUG-SIM-172). A name that appears only once reads exactly as it always has, so ordinary answers do not grow.The advisor's fast answer deliberately does not. The Project Manager Advisor panel has its own model step, grounded on its own board context, so its first built-in answer runs no reading — the reading there would only add up to eight seconds before the advisor even decides whether to think harder, and nothing in the advisor's own step ever uses its result. The question box is where a reading buys an answer; the advisor is where its own model does.
When no AI is available the answer says so. With no key, AI comfort turned off, or no connection, questions are answered by the built-in word matcher, and the answer ends with a plain line saying it used the built-in matcher. A question asked twice in the same session costs nothing, because the answer is cached per person.
Answers only cover projects you are a member of. Every answer, pattern card, Intelligence snapshot and project count leaves out boards you have not been added to, on the desktop and over the CLI alike. A card that spans one of your boards and one you are not on is left out too. Someone who is a member of no project gets the "no projects yet" reply. Being an admin does not widen this.
Risk questions ("What is the riskiest item on Power Board Alpha?", "What is most at risk of missing its deadline?") are answered on the spot. The answer names the riskiest item and why: overdue (most days late first), then running past the time that kind of work usually takes, then stuck. It names up to two more items to watch and the project's overall risk level from the risk heatmap. With no project named, it picks the riskiest project: by the heatmap when the heatmap scores every project in scope, otherwise by the same worst-first on-track ranking the attention answer uses, and then it states that project's verdict. It never picks by counting risky items. When no project can be judged it says "I can't tell yet which project is most at risk", and a quiet project that is still at risk gets its verdict, never "nothing looks at risk".
Overdue questions ("What's overdue?", "What is overdue on Backend Services?") give the count in plain English ("1 item is overdue", "3 items are overdue") and name the overdue items with their project, most days late first. The answer shows up to five and says how many more there are. An older saved summary that carries no item names gives each project's count instead.
- "…and who owns each one?" is answered too. When an overdue question also asks who owns the items ("who owns", "owner", "owned by", "who is responsible"), each listed item carries its owner, and when none of them has one the answer says "None of them has an owner." A project that keeps no owner column at all never gets that sentence.
- Owners read the same for everyone who can see the project. An owner comes from the project's People column, so the same item names the same owner whether the asker is the project owner, an editor or a viewer; the only thing that hides an owner is that person's own column visibility hiding the People column (BUG-SIM-168). So when two people get different owners for the same item, the cause is the DATA each computer holds, not the asker's role: check whether the copy that says "no owner" actually has rows in that project's People column. An empty People column on a project that keeps one is what reads as "no owner", and an absent column value is a sync gap, not an answer bug.
- A first-person overdue question answers about the person asking. "Do I have anything overdue?", "Anything overdue for me?" and "Am I behind on anything?" give the count of the signed-in person's own overdue items and name them, ending "Showing only what is assigned to you." so the narrowed answer can never read like the whole organization's. When that person has nothing overdue the answer says "Nothing is overdue for you right now." and then still gives the team picture. This joins the other first-person question families — what is assigned to me, what is on my plate, what is stuck on my work, what is due this week for me. A teammate named in the question always wins over the asker ("Is Priya behind on anything?" answers about Priya), a question about the team stays team-wide ("What is overdue across the team?"), and when the People columns are not visible the answer falls back to the team-wide report rather than claiming nothing is overdue.
Sprint questions ("Sprint 42 (restart) ends today. What is not going to make it, and who owns those items?", "What is left in this sprint?") are answered about that sprint: when it ends and how many days are left, how many of its items are not done, and up to five of them with their owners. "This sprint" means the one active sprint. A question that names a sprint the box cannot find says so plainly and lists the sprints it knows (or says there are none); it never answers about some other project. A project question that only mentions a sprint as a deadline ("Is Engineering Sprint Board on track for Sprint 42?") keeps its project answer, and a project whose name contains "Sprint" is never read as a sprint.
Long answers stay short. An answer over many projects names the five most urgent (most overdue, then most stuck) and says how many more there are, so a workspace with hundreds of projects never gets a list of all of them. More than five warning signs lists five and says how many more are not listed, and a long stuck-item or "who to ask" list is cut the same way. A bare "pace" in a question never outranks a clearer subject: "At this pace, what is overdue?" is an overdue question.
Questions about time ("What changed on Power Board Alpha since yesterday?", "How many items did we finish on Q4 Design Sprint in the last 7 days?", "Which items are due this week?", "Is anything due today or tomorrow?") are answered from the window's own records. The window comes from the words of the question: "since yesterday", "today", "tomorrow", "today or tomorrow", "this week", "last week", "next week", "this month", "last month", "the last 7 days" and "in the next 3 days" each mean a specific range of days. A question with no time phrase gets the last seven days, or the next seven for a due question.
- Three shapes. What changed names what moved and when, from the project's change history. What got finished counts the items that finished in the window, through the same finished rule every other screen uses. What's due names the open items due in the window with their dates. Each list shows up to five and says how many more there are.
- A due window that includes today also names already-late work. "Due by today", "due this week" and "due in the next N days" are about work that must be done, so an open item whose due date has already passed is included and marked "(due DATE, overdue)" — never hidden. A purely past or purely future window ("due last week", "due next month") stays exactly as it was: no backlog fills an answer the user did not ask about.
- Older late work is listed apart, never counted as due in the window. An open item due before the window starts gets its own sentence after the answer: "Nothing is due on Q4 Design Sprint between 2026-10-26 to 2026-11-01. Still open from before then: Accessibility audit of the new color palette (due 2026-10-16, overdue)." An item due earlier in the same week that is already late still counts as due this week and is marked overdue.
- An empty window is answered honestly, not skipped. With nothing in the window the answer says so, and a change question also names the date of the last real change: "Nothing has changed on Power Board Alpha since 2026-10-15." A project with no recorded changes at all says that instead.
- Finished work is never listed as due, whatever its date cell says, and archived items are left out.
- These answers report facts only. They never rank items, never pick one as the next thing to do, and never advise. Asking which item to work on next is a choice question (below).
- "Overdue" and "blocked" keep their own answers even with a time phrase attached: "what's overdue this week" is an overdue report, and "what's blocked this week" is a blocking answer.
- Names stay on your computer. A window answer is built from your own data with fixed wording and is never sent to an outside AI to reword.
Budget questions ("Is anything on Operations & Finance over budget?", "Are we overspending?", "Any cost overruns?", "How much have we spent on HRIS?") are answered from the project's own number columns. Contract: pm-narrative-budget-contract.md.
- Two columns are compared. A number column whose title says what was planned (Budget, Planned, Allocated, Estimate, Forecast, Approved, Target, Cap, Limit) is the budget side; one whose title says what was really spent (Actual, Spent, Spend, Cost, Used, Invoiced, Billed) is the spend side. "Budget Spent" counts as spend. Titles that name a worked-out figure (Remaining, Left, Variance, Difference) are ignored. Title matching reads English words only.
- Overruns are named with both amounts. "On Operations & Finance, 2 items are over budget: SOC 2 audit prep (Actual Spend $61,500 vs Budget $50,000, over by $11,500); …". Biggest overrun first, up to five per project, then "and N more". Amounts use each column's own number format (currency symbol, before or after, decimals).
- Nothing over is said with the count checked: "Nothing on Operations & Finance is over budget (12 items checked against Budget)."
- No budget columns is said plainly, never guessed. A project with no number columns, with number columns that are not a budget and a spend, or with both columns but no item that has both filled in, gets one sentence saying which, and what to add. A whole-workspace question answers for every project that has a pair and says how many other projects have none.
- Only live items count. Archived and deleted items are left out. Finished items still count, because money spent on finished work is still spent.
- Names and amounts stay on your computer. The answer is built from your own data with fixed wording and is never sent to an outside AI.
Who owns the work. The box knows who each item belongs to, from the project's People column — or, on a project with no People column, from a text column titled Owner, Owners, Assignee, Assignees or Assigned to, whose names are shown as typed. A project with neither is never said to have unowned items. It answers "Who has the most work on Engineering Sprint Board?", "Is anyone overloaded?" and "What is Priya working on?" (also "What does Priya have?" or "What's on Priya's plate?"). A workload answer lists each person's open items, busiest first, with how many are overdue, and how many open items have no owner. A person answer lists that person's open items with the project and whether each is overdue, stuck or due soon. Naming a project answers for that project only.
- Only open work counts. Finished, archived and deleted items are left out. An item with two people counts for both.
- Overloaded means someone owns at least 5 open items AND at least twice the average of everyone else on the same projects (people named only on finished work count as zero). With nobody else on those projects there is nothing to compare, so the answer says all the open work is with that person instead. "Nobody looks overloaded" also names the busiest person's count. A person can have their own limit, set by an admin on the People page (below). With one, they are overloaded at that many open items or more, with no comparison to anyone else, and the answer says "at or over the limit of N set for them on the People page".
- Someone is also overloaded when their work can't fit before their deadlines. Mission
Control looks ahead to each person's furthest due date (up to 3 months) and asks, deadline by
deadline: is the work due by then more than the workdays they're in before then?
- Time off moves deadlines earlier. A due date that lands while the person is out moves to their last day in, and a part-time day counts as half a day.
- Big work counts as big. An item's size comes from the project's Estimate or Story Points column, compared with a typical finished item there; with no estimate it counts as one typical item. A shared item is split between its owners.
- Pace is real. It's how much work the person has actually finished per workday in the last 8 weeks, or the team's pace if that is faster, so a board where people rarely mark work done can't flag everyone.
- The answer explains itself. It says "can't finish everything in time", gives the tightest date, the days of work needed against the days in, and the biggest item. With too little finished work to measure a pace, it says it can't tell yet.
- Mission Control flags it without being asked. Each overloaded person gets a card on the Mission Control home, refreshed every 10 minutes and gone within about half an hour once the work fits. The card shows the person's current name, never why they're out, and moving work stays your decision. Dismissing a card hides it until the next day's clean-up. To get an inbox alert too, schedule the "Alert When Team Is Overloaded" recipe.
- Names are read when you ask, from Mission Control's member list, so a name change shows on the next question. An invited member who has not set a name yet shows the name Mission Control holds for them (the start of their email address). If the member list cannot be reached, the owner answer says so plainly and other answers leave names out; an id is never shown.
- Time off counts. Using the time off recorded in Busy seasons and risk, a workload or person answer adds who is out in the next 7 days and when ("part-time" for a half-time entry), and which open items are due on a day their owner is fully out with no co-owner in: "May need cover". It only ever says someone is out, never why. Adding, changing or removing time off shows on the very next question. Time off also counts in the overload check (above).
- Away questions ("Who is out of office this week?", "Is anyone out while they have work due?") are answered from the same time-off data, showing who is out, when, and items that may need cover. They respect the same privacy rule (never the type or reason for an absence).
- Who is out answers the scope it was asked about. Time off is shared across the whole account, so a bare "Who is out?" lists everyone on the account who is away, whatever boards they happen to be on. Naming a project scopes the answer to that project's people and the answer says so: "Nobody on ADV-Personal is out in the next 7 days." A project with nobody assigned never answers "Nobody is out" on behalf of the whole account. The away words themselves are never read as a project name, whatever their capitalization, so "Who is on time off?" and "Who is on Time Off?" both answer account-wide even when a project on the account is called ADV-Time Off, and "Who is on vacation for Mobile?" still asks which of two Mobile projects was meant rather than widening to the whole account. The one exception is a project named the way this page names one: when the away words are themselves followed directly by "board" or "project" — "What is overdue on the Time Off board?", "What is overdue on the Time Off project?" — they are that project's name and the question is answered about it. Away words followed by anything else ("Who is on time off on the Costco board?", "Who is out of office across all boards?") are still just away words.
- Lead-time questions ("How long does design work usually take on Q4 Design Sprint?", "What's the turnaround on design work?", "How quickly do support tickets get closed?", "What does the sprint board typically take to move?") are answered from the board's lead-time analytics. The answer shows observed times (average, median, worst) per transition. A board with no lead-time setup gets a prompt explaining how to set it up.
- Goals questions ("What are the current goals?", "What goals do we have?", "Please list the objectives", "What are the key results?") are answered from the goals themselves, not from board status: each goal is named with its status and how far along it is. A workspace with no goals says so plainly instead of returning a board summary.
- A project can be named loosely. "The sprint board", "the marketing team" and "the hiring
pipeline" all reach the board the user meant. The match ignores capitals, a leading short
prefix such as
ADV-, a leading "the", and a trailing "board", "project", "team", "workspace" or "space", and it compares whole words only. The words are taken from a board-reference position — after "in", "within", "on", "for", "about", "of", "from" or similar, all treated the same way — never from anywhere in the sentence. Two projects matching equally well is never guessed and never answered workspace-wide: the answer names them and asks which one. A single-word phrase matches only when exactly one live project contains that word. A deleted project never wins a work match: when a live project and a deleted one share a name, the live one is used. - A named subject never gets an answer for the whole account. When the name matches no project, Omni checks your goals: a goal with that name is answered from the projects linked to it, or Omni says plainly that the goal has no linked projects yet. A name that matches nothing gets "I could not find a project named ..." with your projects listed.
- Workload questions ("Who has the most work on Data Migration?", "How is work distributed across the team on Backend Services?", "How much open work does each person carry?") are answered from the same owner workload figures, not from a general board summary.
- Who to ask. The on-track, needs-attention, blocking, waiting-on and risk answers name the owner beside each item they list (or say "no owner"), and end with a short "Who to ask" line.
- Names stay on your computer. These answers are built from your own data and are never sent to an outside AI to reword. A question that names a teammate is recognised before the paid question classifier, so the name is not sent there either.
- No People column yet? The answer says so and explains how to add one.
- Time off will count too: a person away during the coming week who has work due then will be called out as needing cover. That arrives in a follow-up change built on "Who's out".
A question that asks two things gets both answered. "What is overdue on My First Board, and what is due next week?" is read as two questions, and each half is answered by the same answer types a single question would get, in the order asked — one sentence each. A second half that names no project answers about the project the whole question named, never across every project. The split only happens when the question has a real second question in it AND the two halves need two different kinds of answer: "What is stuck and overdue?" is one answer, not two, and a trailing fragment such as "…and why?" never starts one. A question is capped at two parts, and if either half would have to be handed to the project manager session, the whole question is answered as one exactly as before.
- "And who owns it?" is answered from the half before it. "What is overdue on Software Sprint and who owns it?" lists the overdue items and then names their owners, because the pronoun points back at the items the first half just named. This works the same way after a stuck or blocked half. A second half that names its own subject ("who owns Wendy Lin") is not read this way, and a question about one item still answers about that item.
- An owner the answer cannot name is said, never dropped. When the items carry no owner the answer can read, it says it has no owner information for those projects rather than printing items with no names beside them. It never says why.
A project's own name is never read as a keyword. A project called "OKRs & Goals" does not turn every question asked on it into a question about goals: the words of the project's own name are set aside before the question is classified, so "current status of OKRs & Goals" is a status question and "what is overdue on OKRs & Goals" is an overdue question. A question that really is about goals still is, because that word sits outside the project's name — "what are the goals on OKRs & Goals" answers the goals.
A question comparing two time windows gets both windows and the comparison. "Did we finish more items on Power Board Alpha this week than last week?" answers this week, answers last week, and then says which is higher ("That is fewer than last week (2)"), or fewer, or the same. It compares this week with last week, this month with last month, and today with yesterday; two windows of different kinds are not a comparison. Each window is read from its own words, so the two answers really do cover two different ranges of days.
- "The week before" means the week before the one you named. "Did we finish more last week than the week before?" compares last week with the week before it and says "That is fewer than the week before (1)". "The previous week", "the prior month", "the month before" and "the day before" work the same way; "the week before last", "the month before last" and "the day before yesterday" are fixed periods.
A question about a named subject is answered from the items about it. "Are we on track for SOC 2 Type II by Nov 30?" names no project, so the box looks for open items whose names carry the subject ("SOC 2") across your projects and answers from those alone: "At risk. 2 open items match "SOC 2": Backfill missing audit-log fields for SOC 2 (Engineering Sprint Board, due 2026-10-28, owner Priya, who is overloaded with 8 open items), SOC 2 Type II audit prep (Operations & Finance, due 2026-11-20, over budget by $11,500). Every open one is due by 2026-11-30."
- The subject is the name after "for", "on", "in", "within", "about" or "of", up to two words, and it counts only when one of those words starts with a capital letter or holds a number ("SOC 2", "Q4 Launch", "soc 2"). So "on schedule" or "for this quarter" is never read as a subject, and month and weekday names never are.
- Each item says its project, its due date, and whether it is overdue, stuck, over budget (by how much, from the same Budget and Actual Spend columns a budget question reads) and who owns it, with "who is overloaded" when that person is. A "by <date>" deadline adds whether every open item is due by it, or names the ones due after it.
- The verdict comes from those items only: No when one is overdue or due after the deadline, At risk when one is stuck, over budget or owned by someone overloaded, otherwise Yes. The answer states the facts and never picks the one thing to do.
- Only your projects, only what you can see. Items on projects you are not a member of are never named, and a project whose date column is hidden from you shows no due dates.
- No match, no change. A subject that matches no item gets the same answer the question got before. When every matching item is finished, the answer says so.
Every item a question names is mentioned in the answer. When a question names a project and names two of its items ("the accessibility audit and the payment flow prototype both need me this week"), any named item the answer does not already cover is added as a short closing clause. It never reorders or re-ranks the answer, and it never picks between the items itself; a question that asks which to do first is a choice question (next).
A question that asks the box to choose gets a choice, or an honest "can't call it". "What should I work on first today?", "Which should I do first?", "Where do I start?", "What should I focus on?" and "What do we cut?" are handed to the project manager's What Next judgment (the same one behind the "What should I do next?" card). Contract: pm-narrative-choice-answer-contract.md.
- A pick it is sure or fairly sure of reads
Start with "Write the escalation notes" (My First Board): it is 3 days overdue and Aisha is waiting on it. Then "Security training".A cut question readsCut "Dashboard polish" first (Q4 Design Sprint): …andNext to cut: …. The item and project names come from the board, never from the AI's wording. - Anything less is said plainly. No pick, or a pick it is not sure of, reads "I can't call this one from what's on the board.", then the one question that would settle it, then the needs-attention answer the box gave before. A not-sure pick is never shown as the answer.
- Whose items it weighs. "I" questions weigh your own open items; "we", "us", "our" or "team" questions, and any question that names a project, weigh every open item on the projects in scope; "Which item should Jordan do first?" weighs Jordan's. Only projects you are a member of are ever read.
- "First" alone never counts. "What is due on My First Board?" or "the first week" are not choice questions; the phrase needs a doing word right before "first", or "which should I/we", "where do I start", "what should I focus on/prioritize" or "what do we cut/drop/push back".
- No AI, no change. With the AI helpers off, the AI failing or out of its daily allowance, nobody signed in, or nothing open to weigh, the answer is exactly the needs-attention answer.
- What leaves your computer. Unlike the fixed-wording answers above, a choice question sends the open items in scope, their recent written updates and your question to the AI, the same way the "What should I do next?" card does, under the same daily cost cap. It does not look in your email or calendar.
- A pick it is sure or fairly sure of reads
A question about one named item is answered about that item, never about the whole project. You can name the item however you would say it out loud —
when is the Accessibility audit due,who owns Accessibility audit,what is the due date of Accessibility audit on Marketing— and it finds that item by matching your wording against the real item titles in the projects you can see, then answers from that item alone: its status, its due date, its owner, whether it is finished, and which project it is on. It never falls back to the whole-project summary, so a question about one item can no longer come back as "88% are overdue across 50 items". Thethe item named "…"form still works and is the surest way to name an item whose title is also an ordinary word. Why a board question cannot be hijacked by an item's name. The name only counts when it is what you are asking about, not when it is part of what you are asking for. So on a project that happens to have an item calledReview,what is the review status on Marketingstays a question about the project, whilewhen is Review dueanswers the item. Two items whose names overlap resolve to the longer, more specific one, and the same name on two different projects asks you which one you mean rather than picking for you. A name that matches nothing is reported as not found, naming the wording you used, so a typo reads as a typo rather than as a different answer. When the item has an owner, it names them; if the member list cannot be read at that moment it says an owner is assigned without inventing a name. Column values are not read this way yet. Asking for one item's value in an arbitrary column (a contact, a budget line, a custom field) says plainly that single-item column values are not supported yet and names the column you asked for — it does not substitute a project summary and it does not invent a value. Reading arbitrary per-cell values is a separate feature.A question the box cannot answer shows what it can see right away (a short status of the project, or of the whole workspace) and a Keep going with Omni button (the name is whatever you gave Omni). A Project Manager session is a paid AI session, so nothing starts until you tap it. Tapping opens the session in the sessions sidebar with the question, the project and that partial answer already loaded; you stay on the dashboard and the box says "Omni is on it in the sidebar" with an Open it link. Asking the same question again within 30 minutes reuses that session instead of opening a new one. The reply is labeled by what it says, not by how it was made. A reply that carries real project information — a scoped status summary, a summary of a named dashboard or goal — is an answer, even when it also offers to hand off to Omni. Only a reply whose whole content is "I did not understand that" is a cannot-answer, and a "still reading your projects" reply is a not-ready. Screens that hide or downplay a cannot-answer reply therefore never hide a real answer.
Open it automatically instead. The Mission Control setting "Ask before opening Omni" (it shows the name you chose) is on by default. Turn it off and the session opens on its own instead of waiting for the tap. If your projects cannot be reached, the box always offers the button.
A project name that matches nothing says so, lists your real projects, and offers up to five of them as one-tap re-asks of the same question. It does not open a session.
A project you name is the only one answered about. Say "the Software Sprint board" or "the finance close" and every kind of question — status, overdue, what is due this week, what changed, whether it is on track — is answered for that board alone. The everyday name works even when the board is stored with a prefix on it, and an exact name wins outright over boards that merely share a word with it. When two boards genuinely fit what you said, the box asks which one you meant, in one short question, listing the real names rather than guessing. Project names are matched before anything else, so a name that only partly matches still narrows the answer to those boards instead of the whole workspace.
"My" means you. "What are my open tickets?" answers with the work assigned to the account asking, not someone else's. "Our" does not: it names a group, so "our backlog" is answered about the boards, not turned into one person's list.
A teammate you name by name is who the answer is about. "What does Priya have?" and "How is Priya doing?" answer about Priya rather than about the boards — no particular wording is needed to make that happen. A question that is really about something else still answers about that thing: a count of unassigned work stays a count.
An everyday word on its own does not pick a project. When the only thing that could match is a whole phrase of ordinary English ("in review", "in progress", "at risk"), the box never answers silently about a project whose name merely contains those words. What it does instead depends on the phrase: "What is in progress?" asks which project you meant, in one short question that names ADV-In Progress, while "What is at risk?" is an ordinary risk question and answers account-wide without naming ADV-At Risk at all. A longer name behind the same word still resolves as always ("on the finance close"), and a project named exactly
In Reviewis still taken as named.A project that carries a date answers to both of its names. A project is known by its whole name, date and number included, so ADV-Budget 2025 and ADV-Budget 2026 stay two different projects and "the spend on the Budget 2025 project" reaches the one it names. The same project is also known by its name with the trailing date dropped, so "team week" reaches ADV-Team Week 2026-11-01. When a question matches both forms, the longer one wins, so naming the date always picks that project.
Words from a column you cannot see never reach the answer. If a column is hidden from you, nothing printed on it — its heading and any label it holds — is used, whatever kind of column it is. If your access cannot be read at all, the box uses none of the boards' own words rather than using all of them.
A column headed Owner is read by an exact heading. A text column counts as the owner column when its heading is exactly Owner, Owners, Assignee, Assignees or Assigned to — the same list the answer reads and the visibility rules withhold by, so a hidden owner column never leaks. A heading like "Salesperson", "Personal notes" or "Co-owner" is not one of them, and a dropdown headed Owner never becomes an owner column.
It reads the words your boards actually use. Because the answer is grounded on the column titles, status labels and group names a board has — "Escalated", an SLA column, a group called "Busy" — questions phrased the way a real teammate talks land on the right thing, instead of only the phrasings written into the product.
A question that names several projects ("What needs my attention across Operations & Finance and Customer Success?") is answered about all of them together: their numbers are added up, and the answer ends by naming the projects it covered, so a project with nothing wrong is still shown as checked. When one project's name contains another's ("Marketing" and "Marketing Q4 Campaign"), naming the longer one means that project alone. A risk question picks the riskiest of the projects you named. If the box hands such a question to your Project Manager, the session gets the question and every project in it rather than just one.
Goal "Suggest with AI" never pastes a non-answer into the description; when the box cannot draft one it offers "Draft it with Omni" (the name you chose).
Where it runs. Only a question a person typed or tapped on screen can open a session. The command-line route and other agents asking the same question get the answer and a hand-off flag, never a session. On a phone the question box itself stays on the computer, together with the pattern cards under it: a paired phone or web access cannot reach either (decided 2026-10-17). With Omni on, asking from the phone still works, because that question goes to Omni and not to the box.
The rules for all of this are in pm-project-manager-handoff-contract.md.
The People page
- Where: Mission Control's left menu → People. One row per person in the Mission Control account: name and email, Working hours, Overloaded at, and Team Chat.
- Working hours: the days a person works and how many hours a day, set by an admin (see "Working hours" in pm-team-workload.md). It shows "Standard week (not set)" until an admin sets it, and the question box and the workload panel judge each person in their own hours.
- Overloaded at: "Team rule" until an admin presses Set a limit and enters a whole number from 1 to 99 (Enter saves, Escape cancels). Use team rule clears it. A part-timer can carry less, a lead more. The next question in the question box uses the new number.
- Team Chat match: when someone in your Team Chat workspaces has the same email, or the same full name of two or more words, the row asks "Is this person also Priya Nair in your Acme Team Chat workspace?" with Yes and No. Nothing links without a Yes, a No is never asked again, and Unlink removes a match (and counts as a No). Each person links to at most one Team Chat person. Only the first 5 Team Chat workspaces are compared; the company workspace is named "your company Team Chat".
- Who can change it: only a Mission Control company admin. Everyone else sees the page read-only. The app checks the role with the Mission Control server before every change.
- Where it is saved: working hours and limits are saved on the Mission Control server, so every admin sees the same ones. Until your team's server is updated, limits keep saving on this computer as before and working hours show "Not available yet". Team Chat matches are kept for the signed-in Omniscio user on this computer and are not shared with teammates.
- Private: Team Chat emails never reach the page; only the matched name and workspace show.
- Time off still lives in Who's Out under Busy seasons and risk.
What should I do next? (in development)
A card at the top of the Mission Control home page answers "what should I do next?" for the signed-in person. It is on by default for anyone with Mission Control, has its own off switch in Settings → Lab (Mission Control: What should I do next?), and works on the desktop app, the paired phone and the web.
- It asks only when you press the button, unless you choose otherwise. "What's next for me?" asks once; "Ask again" asks again.
- Ask automatically. Settings → Mission Control → How Omni answers → Ask automatically when I open the home page (off by default) gets a fresh answer every time you open the home page. It reaches the home page only: the What should I work on next? widget on an Omni control center page asks only when you press its button, whatever this is set to. Turning it on first warns you that each ask uses AI and costs money.
- Only your projects. Every ask — on any device, and for an agent — reads only the projects you are a member of. Asking about a teammate narrows the answer to that person's items inside those projects; it never reaches a project you are not on, and being an admin does not widen it. With nobody signed in the card reads no project and answers the way it does when there is no work.
- One pick, a couple of backups, in short bullets. The pick sits on its own line with its project, then Why now, How sure (Sure, Fairly sure or Not sure, with the reason — never a number) and the Evidence: the quoted words and whose update they came from, with its date. Backups follow the same shape.
- Every claim was checked first. The answer passes through the claim checker before it is shown, so only quotes, dates and names it found on the board appear. When the pick does not hold up, the card says "I couldn't work this out confidently" and, at most, asks one question that passed every check — it never fills the gap with a guess.
- A plain word that opens a sentence is not a name. A capitalized word anywhere else must be
found on the board, but the first word of a sentence ("Upcoming", "Stalled", "Latest") passes
when it is an ordinary English word. The app checks that against an English word list it ships
(SCOWL, level 50, about 61,500 words, no names; its copyright notice ships beside it in
resources/what-next/), read from disk once, the first time What Next checks an answer. A made-up person such as Dana or Ernest is not in that list, so it still fails wherever it sits. A name that is also a word (Will, Grace, Long) passes only as a sentence's first word. If the list cannot be read, every opener is checked as a name, so the card trims more rather than letting a name through. The list is rebuilt by hand, with no download at install, build or start, bynode scripts/build-what-next-english-words.mjs <extracted SCOWL folder>. - Look deeper. Under that honest fallback the card asks Would you like me to look deeper? and says first, in plain words, that a deeper look uses a more capable AI and costs more than the quick look. Only when you press Yes, look deeper does it open a session with Omni (or the name you gave it) in the sidebar (you stay on the home page) that already holds your question, the one settling question if there was one, and the facts Mission Control gathered: your open items, who you are waiting on, the days you promised and any date clashes. That session is told never to guess, to quote what it read, and to say plainly if it still cannot tell. Pressing again while it is open opens nothing new. It has no keyboard shortcut on purpose, because each press starts a paid session. The deeper session's replies are an ordinary project manager conversation and are not re-checked by the claim checker.
- Check with you. Where the board and your own note disagree (the board says the test is due today, your note says Friday), the card asks which is right. Yes, move it to <day> moves that one date on the board to the day your note states; No, keep the board changes nothing. Nothing moves before you press Yes. The app re-reads the board first and refuses — saying so — if the board or the note changed since, or if the item holds that day in two date columns. Moving a date needs edit rights on that project.
- Waiting on. Each person you are waiting on is listed with the item and how long ago you asked. Draft a nudge shows a short message ("Hi Alex, checking in on …") built only from those facts, with a Copy button. You send it yourself; nothing is sent from the card.
- Since you last asked. When something important changed since your last ask on any of your computers — the pick is different, someone you waited on is no longer listed or a new one is, or a new clash appeared — a short box at the top says what. Otherwise it does not appear. Your last answer is kept on the Mission Control server for you alone (teammates and admins cannot see it), so asking on a second computer, or on the phone, compares with the ask on the first. Each device also keeps its own copy, so the card still works when the server cannot be reached; signing out clears that local copy only.
- No pick, said plainly. Nothing assigned to you, the AI helpers switched off, the AI not reachable, or an answer that did not make sense each get their own one-line reason, and the card still shows who you are waiting on and anything to confirm. Before it answers, it reads the latest updates on your boards from the server, your own items first, for up to 10 seconds, and the card shows "Getting your team's latest updates…" while it does. When a board could not be read in full and this computer's last complete copy of it is more than 20 minutes old, the card says the copy may be missing recent updates.
- Which tasks it weighs. Your open tasks, plus two kinds you do not own: a task where you promised a day in your own written update, and a teammate's task where they wrote that they are waiting on you, since unblocking someone can be your best next move. On a board with too little to go on, it gives its best guess marked "not sure" with the one question that would settle it.
- The rest of your plate. It also weighs your own Tasks due today or overdue, agent sessions waiting on you, and open inbox alerts. Any of them can be the pick or a backup when it is the right next move: the card then shows "Your tasks", "Session waiting on you" or "Inbox alert" where a project name would go, and the claim checker checks its name and dates against that task, session or alert. Everyday words that describe it, such as "Session" or "Longest", don't count as unknown names, while a made-up person's name still sends the answer to the honest fallback. They are never listed as someone you are waiting on or as a date to confirm. Someone with nothing assigned on any board but a task due today still gets a pick. A question about a named project or about the team leaves your personal plate out, so its answer is always from that project.
- Which AI answers. The answer runs on the smart AI tier, on your own AI key, because it was the only tier that stated no false facts when tested on planted trap boards; an answer costs about 4 cents and can take up to 45 seconds. Settings → Mission Control → How Omni answers → "Use a faster, cheaper AI for What Next" moves it to the worker tier: faster and about half the cost, but less careful. "Look deeper" always uses the most capable (genius) tier, so it stays a step above either choice.
- For agents. The same answer is
POST /pm/intelligence/what-nextwith{ asker: { name, memberId } }, and the date fix after a yes isPOST /pm/intelligence/what-next/move-datewith{ asker, itemId, updateId }. - It can look outside the board. When the board cannot explain something, the card checks the person's own approved email, calendar, Team Chat and other boards. While it looks it lists "Checking your email…"-style lines; before reading a new source it asks "May I check your email?" with the reason and a Yes / No that is saved for that person and source; and it shows what it found, with the snippets behind it, under What I found and Checked and ruled out. Everything there passed the claim checker first. Details: pm-what-next-investigator.md.
- Agents save the same yes or no with
POST /pm/intelligence/what-next/source-consent({ sourceId, decision }). - On the phone and the web. The same card shows on Mission Control Home on the paired phone and in a browser, and the computer answers it. A phone ask uses only projects you are a member of, in workspaces that have an owner — being an admin does not widen this — and if the computer cannot tell which projects you are on, it uses none. Moving a date from the phone works only on those projects. Looking further uses the same sources and the same saved yes or no as on the computer. The Google Docs sign-in opens a window on the computer, so on the phone the card says to do it on the computer and then ask again. The saved-answers list stays in the computer's Settings. The "keep going?" question and its Stop work on the phone the same way.
- When an answer takes a while. At about 30 seconds without a checked answer, the card asks
whether to keep going; Stop shows the honest answer so far. Details:
pm-what-next-investigator.md. Agents stop a running ask with
POST /pm/intelligence/what-next/stop({ askId }).
Contracts: pm-what-next-answer-view-contract.md · pm-what-next-deeper-look-contract.md · pm-what-next-since-sync-contract.md ("since you last asked" across devices) · pm-what-next-time-check-contract.md (the 30-second question and Stop). For the investigator on the card: pm-what-next-investigator-on-card-contract.md. On the phone and the web: pm-what-next-remote-access-contract.md.
Reading a status
A status cell is stored as an index into its column's own label list
({"index":N}); the label text is not in the cell. Every reader that reports a
status as words — the board summary's status tally and the agent-stall pattern
here, and the risk heatmap's bottleneck factor in
pm-predictive-analytics.md — resolves that index
against the column it came from. Resolution is per column, not per board: a
board may carry several status columns (a Status and a Priority, say) whose
indexes mean different things, so resolving a board-wide map would report a
confident wrong label. An index with no matching label falls back rather than
inventing one.
Related
The boards whose data is aggregated here are described in mission-control.md. The gate that decides whether a board has accumulated enough usage to be reasoned about at all is pm-data-density.md. The forecasts and risk views built on this snapshot layer are in pm-predictive-analytics.md. When a what-next answer finds something the board cannot explain, the pm-what-next-investigator.md looks in the person's approved outside sources.
Last verified 2026-11-03