Security-review gate (detector fan-out + verifier + GPT-4.1 cross-family)
blocked on F6 and confirmed six lower findings; all are closed here.
- F6 (HIGH): the default run now queries a trailing 3-day window
(today-3..today-1) so the Monday run covers Friday-Sunday postings the
previous-day cadence silently skipped. A per-run staleness sweep flags
never-bank-confirmed payments (ACH > 16 days, checks > 60 days) in the
summary and via console.error.
- F2: replay identity is {event, date, amount} — bankRef excluded (feeds
omit/reformat it between runs); a paid event dated on/before the
latest return in history is a replayed original, never a redeposit;
rows without a valid valueDate are routed to unmatched instead of
being applied under a substituted date; within a day, debits sort
before credits.
- F1: a check-number match with the wrong amount no longer falls
through to the amount fallback (altered-check/collapsed-posting is a
human-review case); the amount fallback requires digit-subsequence
corroboration between posting and candidate numbers; check_return
fallback requires a bank-confirmed candidate.
- F4: the ach_return amount fallback requires cleared_date within 45
days before the credit; an unattributable PMT id is an unmatched
alert, never an amount guess.
- F5: the pmt_id rung requires amount equality; mismatch is the
partial-reversal human case.
- F3: vendor prefix matching requires >= 10 normalized chars (short
names must match exactly); candidates with a conflicting stored
pmt_id are excluded; vendor+amount matches are audit-logged.
- F7: payment writes are conditioned on the snapshot's
status/clear_status and retried once against a fresh read; residual
conflicts are counted, not silently interleaved.
- F9/summary bounds: run summary pk is append-only
(boa_recon#<from>_<to>#<runAt>); unmatched list capped at 50, unknown
codes at 20 keys/16 chars, stale list at 50 (full counts kept).
- ReDoS: description bounded to 500 chars before classification;
CHECK/embedded-number regexes use bounded quantifiers.
- FBOA2-1: both BoA res.json() parses are wrapped so a malformed 200
throws a status-only error and cannot leak a body snippet.
Handler event contract extended (optional fromDate/toDate); cross-family
review required before merge.
processPaymentCsv auto-cleared ACH the moment send_payment_on passed,
but the bank disagrees often enough to matter: settlement lags the send
date by up to 8 days, settled ACH can bounce and re-debit (Uline
$19,281.12, Heights $10,000.00 on the 5/26 overdrawn week), and voided
ACH that settled anyway returned via electronic credits invisible to a
Check-only feed. Move ACH to bank confirmation:
- processPaymentCsv: ACH rows keep their Stampli status; the send-date
auto-clear is removed. The bankConfirmed protection is unchanged, so
bank-cleared records still cannot be moved backward by the CSV.
- fetchBoaTransactions scans all payments (method filter dropped) and
processes ACH CCD lines: match by stored pmt_id first (reversals
reuse the original PMT id), then by the Stampli payment number
embedded in PMT INFO (internal spaces stripped, amount must agree),
then vendor + exact amount within send_payment_on -2..+14 days.
Settled debit -> Cleared/Cleared with dates, pmt_id persisted on the
record; credit reusing the pmt_id (or a unique same-amount return
credit against a bank-confirmed ACH) -> clear_status=Returned +
returned_date + history event. Ambiguity is unmatched, no write.
Existing DDB ACH records already Cleared by the old auto-clear are
unaffected: bank confirmation is additive and the CSV path never moves
a record backward.
Handler event contract extended (optional fromDate/toDate); cross-family
review required before merge.
The two-code 475/255 filter missed every return on the Jan-Jul statement
(29 ARP refer-to-maker credits, 24 electronic return credits, and the
redeposit cycles), leaving 5 paid-then-returned checks stuck at Cleared
($5,256.62, vendors unpaid). Rewrite fetchBoaTransactions around a pure
reconciliation module (src/boaRecon.js) covered by node:test:
- BAI transaction-code event map (only 475 = check paid is confirmed;
remaining codes pend the enumeration replay) with a description-text
fallback classifier for ARP refer-to-maker, return-of-posted-check
(check-numbered and electronic) and ACH DES:PAYMENTS formats. Unknown
check-shaped transactions are logged and counted, never dropped.
- Matching on check number AND amount over all candidates; wrong-amount
or collapsed-number postings fall back to a unique exact-amount match
within checks issued in the last 120 days; ambiguity means unmatched
with no write.
- Transitions always set BOTH status and clear_status (the missed
returns slipped through the divergence between them). clear_status is
bank truth: returns apply even to Cleared records, a second paid debit
on a Returned check is a redeposit back to Cleared, and a return on a
voided check is terminal voided-and-bounced. Every applied event is
appended to a history list; identical replayed events are noops.
- Optional {fromDate, toDate} replay payload (strictly validated) for
gap replays and the BAI-code enumeration runs; default stays
yesterday.
- Each run writes a boa_recon#<runDate> summary item (90-day TTL) with
per-event counts, unmatched check numbers/amounts, and unknown codes;
unmatched/unknown also console.error (Slack alerting is #71).
ACH transactions are classified but not yet acted on; bank-confirming
ACH lands with #69.
* fix(csv): parse M/D/YY dates with loud rejects; stop canceled rows zeroing stored fields (#65, #68)
Stampli exports switched from MM/DD/YYYY to M/D/YY, which toISODate
mangled into garbage ISO strings and slackAppHome's parseMDYLocal turned
into year-1926 dates. Canceled rows also blank 'Amount in USD' and can
blank dates, which the unconditional upsert wrote over previously stored
real values (116 records at amount_usd=0 as of the 2026-07-21
reconciliation).
- src/dates.js: shared strict parser for M/D/YY + MM/DD/YYYY (real-date
validation, 2-digit years 2000-based, plausibility window helper)
- processPaymentCsv: canonicalize send_payment_on to MM/DD/YYYY; reject
rows with unparseable/implausible dates and fail the invocation after
valid rows are processed (surfaces via errors alarm + async DLQ;
retry-safe since upserts are idempotent and existing checks are not
re-offered to BoA); only overwrite amount_usd/send_payment_on with
real values, falling back to the 'Amount' column on cancels
- BoA add_Issue/cancel_Issue now use stored record values when the CSV
row is blank (cancel_Issue previously sent amount 0.00 for voids) and
skip registration on missing date/amount instead of sending garbage;
same guard on the backfill path
- slackAppHome: use the shared parser (fixes 1926 aging)
* fix(csv): validate BoA registration data before writes; harden logging (security review)
Hardening from the /sh-security-review pass on this branch:
- BoA eligibility (resolvable amount + issue date) is now decided and
validated BEFORE the DDB upsert: a cancel-transition row we cannot act
on is rejected with the record untouched, so the transition gate stays
open for a later clean file instead of silently losing the cancel
forever (the skip guard previously ran after the status write)
- First-seen rows that are already canceled are stored but never
offered to add_Issue — registering a voided check as an active issue
created a live issue with no cancel to follow (worsened by the Amount
fallback making the previously-failing call succeed)
- isPlausibleSendYear window is now relative (year-7..year+2) instead
of hardcoded 2020-2035
- Untrusted CSV values are JSON-encoded and truncated before log
interpolation (quoted CSV cells carry newlines -> CloudWatch log-line
forgery, CWE-117)
- OAuth token-exchange failures no longer throw the raw BoA authn
response body (aligns with the PR #63 log-scrub posture)
Verified: unit smoke on the date module; full replay of the real
2026-07-21 export (1,197 rows) against live table state produces 1,197
upserts, 0 rejects, 0 spurious BoA submissions.
* fix(csv): comma-tolerant amount parsing in backfill (PR #72 review)
Number() on a hand-inserted comma-formatted string amount would NaN and
skip the check during backfill; hoist the CSV path's parseAmount to
module scope and share it. No live records are affected (all amount_usd
values are DDB Number type) — defensive consistency.
