PolarisBuddy

PolarisBuddy is the global assistant that rides along on every page of Polaris. Press ⌘J (Ctrl+J on Windows/Linux) anywhere — or click the floating bubble in the corner — and a chat panel opens beside whatever you were doing. Unlike the six purpose-built chat surfaces (paper chat, library chat, and so on), Buddy is not tied to one page: it can call the platform's whole read-only tool registry, so it answers questions that cut across papers, libraries, ideas, experiments, and manuscripts.

Buddy runs the same Navigator / Helm / Sextant split as a Voyage, scaled down to a conversation: a Navigator decides how many tool rounds are left, a Helm executes tool calls in parallel, and a Sextant sanity-checks the finished answer. Every step of that loop is visible in the panel as it happens.

Opening it

• ⌘J / Ctrl+J toggles the panel from any page.
• The floating bubble does the same on click. It is draggable, remembers its position, and is also a drop target (see below). Its tooltip reads "PolarisBuddy (⌘J) · drag me · drop a paper or selected text on me".
• On wide screens the panel docks as a column next to your work; on narrow screens it becomes an overlay drawer. Same content either way.

When the panel opens on an empty conversation, Buddy shows a time-of-day greeting, an opening question, and four suggestion cards. The question and cards are chosen from what is true right now — first from the page you are on, otherwise from your own recent activity (running experiments, this week's ideas and saved papers, papers you left half-read, today's feed). None of this goes through a model: the signals come straight from SQL and the sentences are picked, not generated, so the opening is instant, free, and never wrong. When there is genuinely something new — an experiment running, fresh papers in today's feed — the closed bubble can also show a one-line nudge; no news, no nudge.

Page awareness

Buddy knows what you are looking at. The frontend derives a page context from the current route — reading a paper, viewing an idea, watching an experiment, browsing a library, writing a manuscript, sitting in a topic workspace, or scanning the daily feed — and sends it along with each question. Two things follow:

• The opening cards change with the page. On a paper you get "Explain this paper", "Compare with prior work", "Is the evidence strong enough?"; on an experiment you get "Where is it up to?", "Can the results be trusted?"; and so on.
• When you say "this paper" or "this experiment" mid-conversation, Buddy resolves it to the thing on your screen.

Context is a hint, not a permission

The page context only tells the model what you are looking at. When Buddy actually reads content, it goes through the same tools and the same permission checks as everything else — a forged id can only make Buddy look up something you were already allowed to see.

Drag a paper in

Drag a paper card from anywhere in the platform onto the floating bubble. The bubble highlights, and on drop Buddy immediately asks itself to walk you through that paper — the problem, how the method works, how strong the evidence is — using the platform's stored full text and wiki notes.

You can also drag selected text onto the bubble: Buddy explains the passage, checking the platform's corpus where needed.

How a turn runs

Each question starts a tool loop of up to 8 model⇄tool rounds:

1. The model streams its thinking and answer text. A collapsed "Thinking" row shows the latest line while it reasons.
2. When it decides to call tools, the panel shows a tool card per call — name, a one-line summary of what came back, and duration; click to expand a result preview. Read-only calls run in parallel (up to 4 at a time, 60 s each).
3. Results are fed back and the loop continues until the model answers, the round limit is reached (the loop then forces a wrap-up with the material already gathered), or you press stop.

Alongside the answer you get:

• Papers looked at — every paper any tool returned this turn (up to 20), each linked to its reading page. This is deliberately not labelled "citations": the platform cannot verify which paper each [n] marker in the prose refers to, so it shows you what was actually retrieved and lets you judge.
• Inline figures and citation chips — the model can embed platform figures and numbered references directly in the answer; figures render inline, and [n] chips link to papers.
• A verification note when something does not line up — for example the answer cites [5] but only 3 papers were retrieved, or it claims "I found…" without a single successful search. These checks are deterministic and heuristic; they flag, they never rewrite or block.
• Follow-up chips derived from what actually happened this turn.

Which tools it has

The platform keeps a single registry of 47 tools (the same registry the MCP server exposes). Buddy gets every read-only tool by default — around 43 once a couple of exceptions are removed — spanning literature search, full-text grep, wiki notes, figures, concepts, the knowledge graph, ideas, experiments and their logs, tasks, gates, manuscripts, and external lookups (arXiv, Semantic Scholar, OpenAlex). Enabling memory adds remember/recall; plan mode adds submit_plan. Buddy cannot write to anything except its own opt-in memory.

Scope

The bar above the input box shows what this question can reach. By default it follows the topic you are working in; click it to pick another topic, or choose All assets to search everything you can see. Lab-level assets (the daily feed, public libraries) are always included — they belong to no topic. Once you pick a scope by hand, it stays put instead of following your navigation. Each conversation remembers the scope it was asked in.

Modes

Open the + menu next to the scope picker to choose one of three modes:

Mode	UI label	Behavior
Normal	"Normal · just answer"	The default. Ask, get an answer.
Plan	"Plan · propose first, act on approval"	This turn does research only — no acting. Buddy investigates with tools, then submits a step-by-step plan and stops.
Goal	"Goal mode · keep pursuing one goal"	Hand it an objective and it drives toward a result across turns.

Plan mode ends its turn with a plan card titled "Shall I go ahead?" and two buttons: Go ahead and Revise it. Approving is one click — it switches back to normal mode and tells the model to execute the plan it just proposed, updating a live progress card (n/m steps, running step highlighted) as it goes. "Revise it" hands the floor back to you to say what is wrong. The point of the mode is to move the "I thought you wanted A" conversation to before any time or tokens are spent.

Goal mode takes your first message as the standing goal — no separate field to fill in. The goal is stored on the conversation and re-sent every round, so the model cannot forget it by turn three. In this mode Buddy is instructed to report progress against the goal each round, keep moving on its own (look things up rather than asking "which should I check first?"), stop only for genuine decisions — direction, spend, anything that would change data — and to say plainly when the goal is reached. Click the mode label to drop back to plain chat.

Conversations

• Buddy keeps named conversations. A title is generated after the first turn; the conversation rail (dock layout) or the history popover (overlay layout) lists them, newest activity first.
• Conversations run in parallel: you can start a turn, switch to another conversation, and come back — a live dot marks every conversation that is still running, and stop only stops the one you press it in.
• Stopping or losing the connection mid-answer keeps everything generated so far, marked as interrupted.
• Replayed history is budgeted to roughly half the model's context window, and older tool results are compacted to one-line stubs after a couple of rounds — long conversations stay cheap, but very old retrieval detail is summarized away. Start a new conversation for a new subject.

Memory

Buddy can keep long-term memory about you — research direction, preferences, standing agreements — but it is off by default: an assistant that records things should be switched on by its user. Enable it in Settings → PolarisBuddy → Memory. Once on:

• Buddy gets two extra tools, remember and recall — the only writing tools it ever has, and they touch nothing but its own notes about you.
• Stable facts are injected into every turn (capped at ~1 200 characters, newest first); timestamped notes surface only when recalled.
• Every memory is visible and deletable in the same settings card.

Key settings

Setting	Where	What it does
Assistant enabled	admin settings	Deployment-wide switch. When off, the panel explains that an admin must enable it.
Memory	Settings → PolarisBuddy	Opt-in long-term memory; list and delete individual memories.
Skills	Settings → PolarisBuddy	The on-demand skill playbooks Buddy can load; built-ins plus your own. See Skills.
MCP	Settings → PolarisBuddy	The same tool registry exposed to external clients. See MCP.
Scope / mode	in the panel	Per-conversation topic scope and chat/plan/goal mode.

Tips and limits

• Buddy is read-only. It can look at everything you can see, but it cannot edit papers, approve gates, start experiments, or change manuscripts. Long, stateful work belongs to tasks; Buddy can inspect their status and logs.
• 8 tool rounds per turn. Sprawling questions ("survey this whole field") are better asked in goal mode, or phrased so Buddy can delegate to its research sub-agent, which keeps intermediate retrieval out of your conversation.
• Tool results are truncated (about 4 000 characters each fed back to the model), so "read me the entire paper" works better as targeted questions.
• The "Papers looked at" list is retrieval, not citation. If a claim matters, click through and check — the verification note will warn you about the most obvious mismatches, but it is a heuristic, not a guarantee.
• If your model provider does not support function calling, Buddy degrades to answering from context alone and says so, rather than failing the turn.