The readme served two roles: design-decision source of truth and
work-status tracker. The status section grew into a plan the
orchestrator must navigate, so it moves to its own file with task
IDs, dependency edges, statuses, and per-item acceptance criteria.
readme.md remains authoritative for decisions (stack, formats,
workflow); workflow step 7 and AGENTS.md pointers now name
roadmap.md for status. Cross-references verified.
Also records: stale feature/epsilon-parser branch pointer (676e55e,
2 behind main) deleted locally; recreate from main when B1 starts.
Orientation file so agents do not have to re-derive project conventions
from the 300-line readme every session: environment paths, the honest
red/green state per crate, test-data rules (fixtures in, raw samples
out), git boundaries (feature branches freely, main only on approval),
and a definition of done demanding executed verification over
inspection. Points at readme.md sections rather than duplicating them.
Roadmap now reflects reality: config done, db tests red awaiting
implementation (with the 'do not alter test intent' guard), and the
epsilon-parser branch open-but-empty pending the step-2 spec.
Guidelines adjusted so an agent can work a feature branch
unattended: fixtures are the explicit exception to the data/ commit
ban, and per-passing-test commits need no feedback on a branch —
main still does.
Cargo needs an installed, PATH-visible Rust; record how (rustup via
brew) and where build/test commands run (Application/). Python is a
fixture-regeneration aid only, not a product dependency.
data/fixtures/: committed sanitized copies of all three input formats
(see scripts/ and prior commit), safe for any repo consumer and the
expected substrate for parser tests. Documents the sanitization
choices: values with personal content are replaced, non-PII values
(amounts, GL codes, descriptions) kept verbatim, and subfranchise
cross-value arithmetic intentionally broken.
Real samples under data/test_input/ stay local and gitignored; these
committed fixtures are format-identical copies with every identifying
value replaced one-way via salted hashes: epsilon contract cards and
customer numbers into synthetic disjoint ranges (batches renumbered
9405+, dates +2y; cumulative slice batches 5001+, dates -6y), tsdrms
R/A / DBR / location ids into ZZ-form synthetic ids (dates +2y,
filenames shifted to match cutoffs), subfranchise statements to
extracted text with sender, partner, invoice numbers and amounts
replaced (EUR x rate = SEK arithmetic deliberately NOT preserved).
Amounts, GL codes, descriptions, and station/terminal/pump/receipt
numbers carry no personal data and are kept verbatim per readme
'Data formats'.
scripts/sanitize_samples.py regenerates fixtures deterministically
from local raw samples; scripts/check_fixture_leaks.py verifies no
real value appears in fixtures (content or filenames), exit 0 =
clean. Verified passing for all sources.
hermes