Sparebank1 context

sparebank1-context

Turn sb1 from "shows transactions" into "reasons over your money". This skill runs a brief interview over the user's real data, then stores the answers as a personal finance context that future finance sessions load up front. Read sparebank1-shared first for auth and flags.

The CLI already surfaces the data. This skill captures the personal meaning layer the bank cannot infer, and nothing more.

The privacy contract (read this first)

This skill's mechanism is public (it ships in a public repo and installs for anyone), but the answers are private personal data. Therefore:

• Write the context file to a user-owned path outside this repo. Default: ~/.config/sparebank1-cli/context.md (alongside where the file secret store writes). Override with the SB1_CONTEXT_FILE environment variable.
• Never write answers, balances, account numbers, or counterparties into this repository, a commit, a PR, or any shared/cloud-synced location. Treat the file like the plaintext file secret store: keep it off backups and out of anything that syncs.
• Prefer account names + last 4 digits over full account numbers. Don't record figures the user didn't ask you to keep.
• The file is for this user's agent sessions. Don't paste its contents into chats, issues, or external services.

Resolve the path once at the start:

ctx="${SB1_CONTEXT_FILE:-$HOME/.config/sparebank1-cli/context.md}"
mkdir -p "$(dirname "$ctx")"

When to run

• First time: no context file exists → run the full interview below.
• Refresh: the file exists → read it, show the user what's captured, and only ask about what's new or changed (new counterparties, new Uncategorized rows, subscriptions to revisit). Don't re-ask settled questions.

Confirm login before pulling data:

sb1 status   # if not logged in, ask the user to run `sb1 login` (see sparebank1-shared)

Step 1 — gather the data (no code change needed)

Drive the interview from existing commands. Start with a short window:

sb1 --json summary --months 3

From the JSON, the rows worth asking about are:

• categories[] — especially the "Uncategorized" bucket (the bank couldn't classify these; you can, with the user's help).
• topOutgoing[] — the biggest counterparties by spend.
• subscriptions[] — what the bank already flagged as recurring.

To see the actual Uncategorized rows behind that bucket, drill in per account:

sb1 --json transactions -a <account> --from <start> --to <end> --classified

Each row has category, recurring, subscription, counterpartyName, amount (negative = outgoing), and date. Filter to rows with a missing/"Uncategorized" category or to the top counterparties — those are your interview targets.

Step 2 — the interview (keep it short)

Prioritize the Uncategorized bucket and the top counterparties. Do not walk every transaction. Group similar rows and ask once per group. Aim for a handful of focused questions, not a questionnaire. Capture these dimensions:

Dimension	The question to resolve
Income classification	For each inbound counterparty: salary, reimbursement, transfer from a partner, or a refund? (So "income" isn't overcounted.)
Work-expensed / reimbursed	Which outgoing rows does work pay back (travel, equipment)? These should net OUT of "real" spending.
Account semantics	What is each account for — daily, buffer, a savings goal, BSU, joint?
Shared / split	Anything split with a partner where only the user's share should count?
Counterparty aliases	Cryptic strings (e.g. KLARNA*…, acquirer codes) → a human label.
Subscriptions	For each flagged (or user-known) subscription: keep, cancel, or annual?
Fixed vs discretionary	Which spend is fixed (rent, loan, insurance) vs discretionary?

Ask in plain language, tied to what you actually saw: "You sent ~kr 4 000/month to BERG EIENDOM — is that rent (fixed)?" beats an abstract checklist. Skip dimensions that don't apply to this user.

Step 3 — write the context file

Save the answers to the resolved path in the format below. On a refresh, merge into the existing file rather than overwriting blind, and bump the date. After writing, tell the user the exact path and restate that it stays out of git and backups.

# SpareBank 1 — personal finance context

<!-- Private. Maintained by the sparebank1-context skill. Never commit to git;
     keep off cloud-synced folders and backups. Last updated: YYYY-MM-DD -->

## Accounts — what each is for
- <name> (····<last 4>): <daily | buffer | savings goal | BSU | joint | ...>

## Income — how to classify inbound money
- <counterparty / pattern>: <salary | reimbursement | partner transfer | refund | other>

## Work-expensed / reimbursed  (net these OUT of real spending)
- <counterparty / pattern>: <what it is; reimbursed in full / partially>

## Shared / split expenses  (only my share counts)
- <counterparty / pattern>: <split, e.g. 50/50 with partner>

## Counterparty aliases  (cryptic string → human label)
- <raw string>: <label>

## Subscriptions
- <name>: <keep | cancel | annual> — <note>

## Fixed vs discretionary
- Fixed: <rent, loan, insurance, ...>
- Discretionary: <eating out, shopping, ...>

## Notes / preferences
- <anything else to know when reasoning about this money>

Leave out any section the user had nothing for. Don't pad it with invented rows.

How future finance sessions use this

At the start of any session that reasons about the user's SpareBank 1 money (budgeting, "how much did I really spend", savings rate, "what can I cancel"), read the context file first:

cat "${SB1_CONTEXT_FILE:-$HOME/.config/sparebank1-cli/context.md}" 2>/dev/null

Then apply it on top of the raw sb1 output: net out work-reimbursed and internal/partner inflows before calling something "income", count only the user's share of split rows, resolve aliases to readable labels, and respect the keep/cancel notes on subscriptions. If the file is missing, offer to run this skill's interview. If it looks stale (new big counterparties show up that aren't in it), offer a quick refresh.

Future idea (not built)

A sb1 summary --uncategorized helper could surface exactly the rows worth asking about. It isn't needed: summary plus transactions --classified already provide everything, and the conversational interview belongs to the agent, not the binary.