checkout-analysis

Noibu Checkout Performance & Health Analysis

Surfaces where shoppers drop off in checkout and what to do about it, as a ranked triage board built from Noibu session, value, and error data.

How it works

Run Setup, then pick one of two behaviors from the user's prompt:

• Quick answer (one focused question) → run 1–2 queries, answer directly, offer to go deeper.
• Full analysis (broad request, bare invocation, or "yes" to the offer) → the four-step workflow below.

Full analysis flow: (1) broad overview (Q1–Q7, current + prior window) → (2) cross-reference + period-over-period → (3) pick the top 3 regressions → (4) run targeted follow-ups → render the board (one widget). Signals are flagged on change vs the prior window, not absolute level — see references/queries.md "Signal model".

Setup — before any query

Work quietly. Don't narrate plumbing — resolving the domain, loading reference files, and reading the board format all happen silently, with no "let me…" commentary. The first thing the user sees is the Step 1 overview line ("Starting with a broad look…"); after that, narrate only real analytical progress (what the data shows), never file reads or tool setup.

• Resolve the domain first; keep the company id it returns. If the user gave a domain (name or UUID), use it. If not, ask which one via AskUserQuestion populated from the user's domains — don't interrogate for anything else; take the default window and proceed. If the account has exactly one domain, skip the question and use it. Domain resolution returns a company id alongside the domain UUID; some tools (priority errors, data-connection checks) need that company id too — carry both.
• Load the Noibu context reference (the querying-noibu-data skill/reference). It maps the role-based names used here ("session query tool", "funnel depth field", etc.) to the real Noibu tools/columns and documents query constraints. If it isn't available, discover tools and field names from the live Noibu API / tool schema instead — don't stop or guess.
• Confirm every field name by role before using it (from the context reference, else the live schema). Steps below name fields by role, never by hard-coded column, so when the API changes only the lookup moves.
• Default analysis window: the one in the context reference; if none, last 30 days.
• Call list_scheduled_tasks now — note whether any task's prompt references this domain; this sets the action-bar Schedule button label later ("Schedule insights" vs "Edit scheduled insights") without blocking rendering.