Field discovery now samples an account's actual transactions instead of
assuming a shape. flatten() passes through every scalar the bridge sends
rather than whitelisting eleven keys, so institution-specific fields turn
up on their own, and inferFields() — extracted from the CSV suggest route
so both paths share it — unions keys across the sample because API feeds
omit optional fields entirely.
Three bugs the live bridge exposed:
- posted=0 on pending transactions became 1970-01-01; falsy epochs are
now "no date", with date falling back to transacted_at and posted_date
kept separate.
- days=0 omitted start-date, which returns only the few most recent
transactions rather than everything — 4 instead of 89. A start-date is
always sent now, clamped to 89 days (the bridge hard-caps at 90).
- Sampling asked for more than 45 days, and the bridge's advisory notice
about that surfaced in the UI as an error. Samples use 44 days; the
threshold is exclusive.
The Sources page can now link an account: a picker in both the create
dialog and the detail panel, populated on demand, which fills the field
table from the sample and defaults the constraint field to the
transaction id with an explanation of why.
manage.py option 10 claims a setup token and writes the access URL to
.env, replacing the throwaway script.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01G2HFeU5neCKagTnmA6o9Tu
Sources with a `simplefin` block in their config can pull transactions
straight from the bridge instead of taking a CSV upload. Only the fetch
differs — dedupe, logging, and transformation reuse the import path.
The access URL is the whole credential, so it lives in .env rather than
the database that manage.py offers to reset. Claiming a setup token is
exposed as an endpoint because the token is single-use and easy to burn.
The bridge answers 200 with a populated `errors` array when a bank is
failing, which would otherwise read as a successful empty pull — those
errors ride along in the sync response and show on the Import page.
Pending transactions are skipped by default: they get a new id once they
post, which would import the same charge twice under two keys. Sources
should use ['id'] as constraint_fields — the transaction id makes
overlapping pulls free while keeping genuinely repeated charges distinct.
Verified against a stubbed bridge response, not a live account.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01G2HFeU5neCKagTnmA6o9Tu
- lit() was calling Math.trunc() on numbers, dropping decimals from balance_offset and any other numeric SQL params
- Stacks page now saves last selected stack to localStorage and restores it on load
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add database/queries/{sources,rules,mappings,records}.sql — one file per
route, all business logic in PostgreSQL functions
- Replace parameterized queries in all four route files with lit()/jsonLit()
literal interpolation for debuggability
- Add api/lib/sql.js with lit(), jsonLit(), arr() helpers
- Fix get_view_data to use json_agg (preserves column order) with subquery
(guarantees sort order is respected before aggregation)
- Fix jsonLit() for JSONB params so plain strings become valid JSON
- Update manage.py option 3 to deploy database/queries/ instead of functions.sql
- Add SPEC.md covering architecture, philosophy, and manage.py
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
