Vinktar

Create an empty board in the workspace, to leave the user a view of what you found; fill it with add_panel, where each panel says which projects it reads. First decide what kind of board the request is, from the decision the user wants to make with it: "launch" when one feature or release just shipped: is it used, by whom, does it convert, did it break anything; "conversion" when a funnel someone cares about: where people drop and which segment converts; "growth" when the business at a glance: new people, active people, conversions and money over time; "engagement" when how much and how often people use the product, and whether they come back; "revenue" when money: how much, from whom, and the path to paying; "acquisition" when where people come from and which channels bring the ones who sign up; "reliability" when what is breaking and whether it hurts the flows that matter; "custom" when none of the above: a board shaped around one specific question. The result is a plan for that kind, filled in with the project's own signup, activation, conversion and revenue events: follow it, dropping what the user did not ask about. If the request fits none, ask the user what the board is for before building. Pass `expiresInDays` for a board built for one launch, so it deletes itself afterwards instead of staying forever. Only do this when the user asked for a board or agreed to one.

Define an event in the project's tracking plan: what it means, its role, the properties it carries (type, unit, whether every send must include it, an example) and where the code sends it. Call it in the same step as adding or changing a track() call, before any data arrives; and for existing events once you have read the code that sends them. It takes effect at once, labelled as written by an agent, and the team can correct it in Vinktar. get_setup_status then reports defined events that never arrive and sends missing a required property. Conventions (get_install_guide with section "conventions" has them all): new event names and property keys are snake_case, object then past-tense action (signed_up, report_exported); a description is one sentence of at most 100 characters and a longer one is refused; an event already shipping under another spelling is defined as it is sent, never renamed unasked. Do not overwrite a description the team wrote unless the user asks. Setup call: never counted.

Delete a board with its panels. Only when the user asked for it, or agreed when you proposed it; name the board before you do. A person can restore it in Vinktar from the boards page for 30 days.

Remove a panel from its board. Only when the user asked for it, or agreed when you proposed it; say which panel before you do. The panel id is in get_board. A person can restore it in Vinktar from the boards page for 30 days.

Write the project's notes for AI agents: what the product does and for whom, the flows that matter (signup, activation, the paid action), what counts as an active user, and anything that makes the data easy to misread. Write them from the codebase during setup, in plain sentences, up to 2,000 characters; they replace the current notes, so read them in get_schema first and keep what the team wrote. Every later agent reads them before the data. Setup call: never counted.

Change an alert from get_alert: its lines, window, timing, severity, recipients (workspace members) and channels (saved ids or "#name" in the connected Slack). Or hold its news with snoozeMinutes (0 ends a snooze; a week at most). What it reads and what it alerts on are fixed; switching an alert off or deleting it is a person's, in Vinktar. Only change an alert when the user asked.

Rename a board, rewrite the line under its title, change when it deletes itself, or lay its panels out. A board is a grid: every row spans the full width, its panels share it (one to four) at one height, equally unless you give widths. Pass `rows` to set the layout outright, top to bottom, as rows of panel ids from get_board; panels you leave out keep their order after them. Or pass `arrange` to straighten the rows as they read now. Pass only what changes. Panels are changed with update_panel and delete_panel.

Change an error issue's status or assignee: resolve it (optionally in the release that ships the fix, or in the next release when it has not shipped yet, so a later occurrence reopens it as a regression), ignore it, reopen it, or assign it to a teammate. Only act on an issue when the user asked you to.

Change a panel on a board: its title, how it draws, its number format or what it shows. Pass only what changes; the rest stays as it is. The panel id is in get_board. To turn it into another type, pass `type` with the full spec for it, the same fields as add_panel. A changed panel is run and held to the same rules as add_panel before it is saved (structured types before SQL, one question per panel, the board's time range), and one that breaks them is refused with nothing changed.

Create an alert: what to read, when it is news, and who hears it. Subjects: issues (the whole error stream: new_issue, regression, spike, occurrences, users_affected), issue (one issue: regression, spike, occurrences, users_affected), panel (a board panel from get_board: threshold, change, anomaly), board (every panel on a board: change), event (one event's count or unique users, no panel needed: threshold, change, anomaly; "below 1 per hour" means it stopped arriving), query (a VinktarQL value; use {from} and {to} for the window: threshold, change, anomaly). Timing: everyMinutes and pendingFor (checks in breach before it fires); once it fires it stays quiet for a cooldown (one window for a number, an hour per issue for errors) rather than repeating. Who hears it: recipients (workspace members by address), channels (saved channel ids from get_alert, or "#name" in the connected Slack), notifyTeam; with none, the owners and admins. You cannot name a webhook, a URL or an outside address: a person adds those in Vinktar. For a fix, use check_fix. Only create an alert when the user asked for one.

Add a panel to a board. Each panel answers one question and reads at a glance: a number, a line or a bar. Build it from the structured types; `sql` is the last resort. Choose in this order: `funnel` for conversion through ordered events (viewed pricing → started checkout → paid), with a breakdown for conversion per segment (per shop, per plan): always a funnel, never a query or trend lines per step. `retention` for whether people come back. `trends` for everything that counts or sums events: `chart` "scalar" for one number (revenue, installs, failed charges; several numbers are several scalar panels, never one wide table), "line" for a metric over time, "bar" to compare segments over time, "hbar" to rank them, "share" for parts of a whole, "heatmap" for when in the week; set `display.format` to "currency" for money and "percent" for a rate. A series can carry its own `breakdown`, so one panel shows signups by plan beside checkouts by country; reach for that before SQL. A table only when the rows themselves are the point (the operations or shops to look at), with at most four columns. `issues` for the top errors, `text` for a heading over a row of panels: a single "## Heading" line renders as a slim section title. `sql` only when none of those can show it (a ratio between two measures), with at most four columns, the board's range as `{from}`/`{to}` rather than now(), and `sqlReason` saying why. A funnel and retention count the project's identified users: if the people in the question are not them, say so in a text panel instead. The spec is exactly what query_trends, query_funnel, query_retention or run_sql just ran (same field names), so check the numbers with that tool first. The panel is run once before it is saved: one that fails or breaks these rules is refused and nothing is saved, so follow the message and call again. The board is a grid: a panel joins the last row when it is the same kind and there is room (numbers up to four to a row, charts two, a table or text on its own), and every row shares its width equally at one height. Add panels in the order they should read; to change the rows afterwards, use update_board with `rows`.

After a fix for an issue ships, start a server-side check of whether it comes back: occurrences on the fix release (or the next release seen) and later, against how many sessions reached that release, over a horizon of up to 14 days. It keeps running after this session. Read the result later with get_alert. The result is never "fixed": it is recurred, no recurrence in N sessions, or not enough evidence.

Run a read-only VinktarQL SELECT over the project's `events` table, for anything query_trends, query_funnel and query_retention cannot express; start with those, since what they run is what a board panel should be. A ClickHouse-flavoured subset: WHERE, GROUP BY, HAVING, ORDER BY, LIMIT, DISTINCT, IN, LIKE and allowlisted functions (count, uniq, sum, avg, quantile, countIf, toStartOfDay…); no joins, subqueries, CTEs or INTERVAL, so filter time as `timestamp > now() - 86400 * 7`. Payload properties are `properties.<key>`, user traits `user.<key>`. Tenant scoping and plan retention are applied automatically. Call get_schema first for columns, functions and property names, and aggregate in SQL rather than paging through raw rows.

Before the person has signed in to Vinktar: start a project for their app now, so setup can run and data can arrive before anyone makes an account. Returns the project's write key (public, safe in browser code) and a claim link that only the person should open: it signs them in, or makes them an account, and keeps the project with its data. Unclaimed, the project takes up to 10,000 events and is deleted with its data after 7 days. Only call it when the person has not signed in and the other tools ask for sign-in; once signed in, use list_projects and get_project_keys instead.

Start here. One project's window, or several pooled with `projects` (every row then names its project), compared with the one before it, across analytics and errors: traffic (events, active users, sessions), the events that rose, fell, appeared or stopped, error volume, new, regressed and fastest-growing issues, and the releases that shipped, and what the project's alerts said. Each row names the tool to go deeper.

Raw events, newest first, under simple filters: event names, one user (distinctId), page search, property filters or a VinktarQL WHERE expression. Use it to see what one user did, or to inspect real payloads. Pass `eventId` instead to get that one event in full, every column and the whole payload. Windows longer than 30 days are narrowed to their last 30. For counts, use query_trends or run_sql instead of paging.

Read alerts and what they said. Pass alertId for one alert: whether it is on or snoozed, its timing, who hears it and what it said recently; a fix check reports awaiting_release, watching, recurred, no_recurrence, not_enough_evidence or stopped with exposure and recurrences on the fix release. issueId gives that issue's latest fix check; neither lists every alert in the project, what fired in the last 24 hours, and the saved channels an alert can post to. An alert is any standing question: the error stream, an issue, a panel, a board, an event or a query.

Read a board: every panel computed over a window with the change against the previous period, as headline numbers. The quickest way to answer "how are we doing" in the terms the team already tracks. Each panel shows its id, for update_panel and delete_panel.

The step-by-step guide for setting Vinktar up in the app you are working in: which SDK fits the stack (with current versions), the recommended file layout and environment variables, defining the tracking plan as you add events, proving it with get_setup_status, and the handover telling the person where each variable goes. Call this first when asked to install, set up or audit Vinktar, and follow it. It ends with Vinktar's conventions (event and property naming, descriptions, how an event changes, releases and environments, identifying users and connecting the browser visit to the server); pass section "conventions" for those alone before adding or changing a track() call in an app that is already set up. It answers before sign-in too, and then starts with where the person signs in from each app. Setup call: never counted.

Everything needed to debug one error issue: what it is, its status and history, how often and for whom it happens in the window (by release, environment, browser, URL, tag), the latest occurrence's stack trace (source-mapped when available, each in-app frame linking to its file in the repository once the project's repository URL is set), breadcrumbs and request, and the product events that user sent in the 30 minutes before it.

The keys installing Vinktar needs: the project's public write key (VINKTAR_KEY) with the ingest host, the Sentry-compatible DSN and the Bugsnag-compatible endpoints, and whether a source-map upload key (VINKTAR_CLI_KEY, a secret) exists, with where a person creates one. The secret is never returned: tell the person to add it as a CI secret. Needs a read and write connection. Setup call: never counted.

What a project tracks (or several, with `projects`, each under its own heading): event names with their volume over the last 30 days, the payload properties each event carries, user trait keys, and the columns and functions VinktarQL (run_sql) accepts. Call before any query so names are real. Pass `event` to see every property of one event, or `propertyKey` to see that property's most common values, which is how you find the exact spelling of one before filtering on it.

Check a project's Vinktar install from the data that actually arrived: events and errors received, users identified, releases and environments tagged, source maps for the latest release, items the SDK dropped, and key events described. Each gap comes with the next step. Call it after installing, and again after each fix, until nothing is left to do. Setup call: never counted against the agent allowance.

Whether events and errors have arrived for a project started with start_project, by its write key, and whether the person has claimed it yet. Call it after installing the SDK and sending a first event. Projects someone already claimed answer only after sign-in, through get_setup_status.

The boards whose panels read from a project, or from any of several with `projects`. A board shows what the team already decided matters; read one with get_board.

Recent individual occurrences of one issue, newest first: when, which user and session, release, environment, page, browser and OS. Use it to spot a pattern (one browser, one release, one customer) or to pick a user or session to follow in browse_events.

Error-tracking issues (errors grouped by fingerprint) for a project, or across several with `projects`, with occurrences and affected users inside the window. Default: open issues (unresolved + regressed) sorted by occurrences over the last 14 days. Follow up with get_issue for the stack trace.

List the Vinktar projects this connection can read (a project is one app or site sending events and errors), with the workspace plan and what this connection is allowed to do. Call this first when you do not know the project handle.

Conversion through an ordered sequence of events (e.g. view_pricing → start_checkout → purchase): how many entered, how many reached each step, where they dropped, how long converting took, optionally by segment.

Cohort retention: users who did targetEvent in a period form a cohort, and the table shows what share of each cohort did returningEvent in each later period. Use it for "do people come back" and "did retention improve after the change".

Metrics over time: event counts, unique users or sessions, the sum, average or p95 of a numeric property, optionally split by a property (one breakdown for every series, or each series by its own) and compared to the previous period. The workhorse for "how many / how much / is it growing" questions. interval "total" returns one de-duplicated number per series for the whole window.