Omniscio

PM Intelligence Advisor

An advisory intelligence layer for Omni: it reasons across all your project management boards and surfaces strategic insights, completion predictions, policy recommendations, and answers to plain-English questions about your projects.

On this page

An advisory intelligence layer for Omni. It reasons across all your project management boards and surfaces strategic insights, completion predictions, policy recommendations, and answers to plain-English questions about your projects.

What it is

The PM Advisor turns the raw cross-board data (the PM Cross-Board Intelligence snapshots) into advice a human can act on. It has six parts:

  • Strategic insights -- it detects cross-board patterns (resource conflicts, cascading delays, effort drift, seasonal risk, cross-department coordination) and lists them ranked by severity, each with a plain-language description and the reasoning behind it. Seasonal risk fires only when this or next quarter is marked Busy or Peak AND the board's real work puts it at risk; its card ("Get ahead of a busy season") repeats the same reason, the same "Keep an eye on" line for late items, and the same next step Project Analytics shows. A quarter's urgency multiplier is bounded 0.5 to 3.0 by MIN_URGENCY_MULTIPLIER and MAX_URGENCY_MULTIPLIER in src/shared/pm-busy-season.ts, which both the request contract and the store read — so a season can sharpen or soften a deadline but no single quarter can dominate the urgency score, and the store refuses a value outside that range rather than clamping it.
  • Strategic coaching -- for the highest-severity insights it derives a concrete recommended action ("review and reassign overdue items across projects", "investigate the shared upstream dependency", etc.).
  • Prediction narratives -- for a board it runs a Monte Carlo simulation over the board's recent completion velocity to estimate when the remaining work will finish (p50 / p85 / p95), with a confidence label and a deadline warning when the estimate slips past a due date. A what-if mode re-runs the estimate with added/removed items or a velocity change.
  • Policy recommendations -- it watches how often you approve or deny an automation and recommends promoting a trusted action to automatic, or pulling a too-eager one back to propose-only.
  • Quick question -- ask a plain-English question about your projects and get an answer grounded in the current board data, with a confidence score and suggested follow-up questions.
  • Handoff enrichment -- enriches shift handoffs with cross-board context via GET /pm/advisor/handoff-summary (enrichHandoff()). The report it hands back covers only the boards you belong to: a shift report that spans several boards comes back with the other boards' names and actions removed. Like the rest of the advisor, it answers only while Omni is on.
  • Turning on automatic approvals asks first -- accepting a recommendation that would let an action type run without asking shows what will change and asks you to confirm. If the trust record behind it changed since it was suggested (you started refusing that action, or revoked its trust), the suggestion is set aside instead of applied.

Where to find it

The PM Advisor now lives inside Omni. It no longer has a sidebar item of its own, and its separate panel is gone.

  • What it shows now lives on the Omni page. Its insights, strategic suggestions and policy recommendations are the Suggestions section of the Projects list, a board's Project Forecast shows when you pick that board, and the busy-season check-up is a card on the welcome screen. See ai-project-manager.md.
  • Old links still work. A saved link, keyboard shortcut, bookmark, deep link or pinned toolbar button that pointed at the old PM Advisor opens the Omni page instead, and so does reopening the app when it was last closed on the PM Advisor.
  • Old sessions move. Any session that lived in the old PM Advisor hub moves into the Omni the next time the app starts. Inbox Triage sessions start there too.
  • For agents and scripts, the advisor's insights, suggestions, forecasts, policy recommendations and quick questions are served by the /pm/advisor/* CLI routes, which answer "not found" while Omni is turned off, and that answer names the switch that is off — Omni itself, Mission Control, or the approval question still to be answered — so a caller can tell a switched-off advisor from a URL that does not exist. Asking about a board that does not exist, or that the caller is not a member of, is also "not found" rather than an empty forecast or an invented what-if projection. On the insights list, ?category= accepts every category the advisor emits, including trend-risk. On the suggestions list, ?limit= (default 50, at most 200) is how many of the highest-severity insights are read. Low findings raise no suggestion, so the list can be shorter than the limit, and no other cap applies: every qualifying insight that was read comes back.

The feature's old Lab entry and its pmIntelligenceAdvisorEnabled setting are gone: once the advisor's hub folded into Omni, neither had a reader left.

How it behaves

A quick question first tries a deterministic, template-based answer grounded in the board snapshots (no AI cost). Depending on your PM AI comfort setting, a low-confidence template answer may be escalated to an LLM. A daily per-feature cap bounds the number of LLM calls, and a low-comfort setting keeps answers template-only. The call count belongs to the person signed in (two people on one computer each get their own day) and the day turns over at midnight UTC. A question the daily dollar cap stops before anything is sent does not use up a call; a call that was sent and then failed still counts.

A quick question's text is free-form user input. When a low-confidence template answer escalates, the question, that template answer and the board data it is grounded in are scrubbed of recognizable secrets and contact details — keys, tokens, email addresses, phone numbers, file paths and IP addresses — before they are sent. Names and other free text are left as they are, because no filter can safely tell a person's name from ordinary words; the filter and that residual are recorded in PII_INVENTORY.md under "AI / LLM prompt egress". The question is still not recorded to feature_events: telemetry for this surface is deferred (telemetry-surface-baseline.json), and feature_events metadata is never aggregated anyway (getFeatureUsageSummarySince groups by feature and action), so a PII tag there would be a row nothing queries.

The advisor only sees projects you are a member of: its answers, the handoff summary, and the insights and suggestions lists leave out every board you have not been added to, and an insight that also involves such a board is left out entirely.

A column kept private from you stays private in AI answers. When a project's people, status or due column is hidden from you, the advisor leaves out that column and everything worked out from it: who is assigned, workload and capacity, the finish date it would aim for, items due soon, and steps where work is piling up. A pattern card about a project you are not a member of is not shown, and while nobody is signed in no pattern cards are shown at all. A pattern card that names no particular project is shown only to someone who is a member of every project, because it is built from all of them. A project you have no membership on shows none of its people, status or due details, and if the privacy rules cannot be read at all, nothing private is shown.

An item's history follows the item. Once an item moves to another project, what happened to it while it was on the old project no longer counts there: answers, insights, effort drift, pace and velocity, behavior patterns, automation suggestions, the board history export and the board timeline for the old project leave it out, so someone who cannot open the item's new project does not see its history through the old one. A project's pace and velocity therefore drop the finishes of items that later moved away.

With no projects yet, or while your projects are still being read, the advisor answers from its built-in reply and never calls the AI. Its answers say "project", never "board" (for example "Project Forecast").

The AI Project Manager's list of projects that need attention keeps private columns private: it does not call a project late from a due date you cannot see, and it does not count stuck work from a status you cannot see, including a second status column that is hidden. On a project whose status is hidden from you it reports no work as held up, and it never says late work is holding other work up. It also stops reusing an earlier answer once a column becomes private or someone else signs in.

The advisor's own wording follows the app's language: insight titles, descriptions and reasons, strategic suggestions and their actions, and the evidence and effect lines of a policy recommendation. Suggestions are worded each time the list loads. Insights and policy recommendations are worded once, when they are created, and saved that way, so after you change the language an insight keeps its old wording until it expires (within a day), and a policy recommendation keeps it for as long as it waits for your answer. The seasonal-risk insight reuses the busy-season wording from Project Analytics.

Project Forecast

Each project gets a forecast. A project with enough history (at least 7 items finished in the last 90 days, or at least 10 timing measurements) shows its pace and its Monte Carlo dates. A project below that bar shows a Ballpark instead: "Ballpark: about 2 to 4 weeks (Oct 30 – Nov 13)" with the note "An early guess from 3 finished items. Mission Control learns your team's pace and tightens this over time.", and no items/day pill. With nothing finished yet it reads "Ballpark: once a few items are finished" and "No finished items yet." A finished project still reads "All items are completed." The forecast's sentences are translated. The rule is described in pm-predictive-analytics.md.

Questions about what is blocked by what ("What's blocked?", "What slips if API schema slips?") are answered from the projects' Dependency links and always count as a confident template answer, so they are never escalated to an LLM and cost nothing.

The quick-question answers also stop at a daily dollar cap, $5/day unless you set your own (the pmAdvisorDailySpendCapUsd setting, read through getAdvisorDailySpendCapUsd()). Every other Mission Control AI feature has its own fixed daily ceiling, and all of them sit in one table, PM_LLM_DAILY_SPEND_CAP_USD_BY_COST_TAG in src/main/services/pm/shared/pm-defaults.ts. The sprint retrospective's ceiling is also $5; features whose calls are cheaper or run unattended have lower ones (the import column resolver's is $2), so $5 is not a promise that every PM AI feature may spend that much. The caps reset each day.

Settings

  • Turn on Omni -- the one switch for the advisor's background reasoning, its AI answers and its CLI routes. Off until you turn it on. See ai-project-manager.md.

The advisor also reads the PM AI comfort setting to decide when a question may be escalated to an LLM.

The advisor is part of Omni, so ai-project-manager.md is the parent page, and mission-control.md covers the boards it reads. Its insights and prediction narratives are built on the snapshot layer described in pm-cross-board-intelligence.md, and the underlying Monte Carlo completion forecasting is covered in detail by pm-predictive-analytics.md. The other panel that reasons over the same boards operationally is pm-operations-engine.md, and the full list of library pages is in INDEX.md.

Last verified 2026-11-10