Adds a token-denominated guardrail in front of every server-side AI call
so a misconfigured port can't run up an unbounded bill. Soft caps surface
a banner; hard caps refuse new requests until the period rolls over.
Usage flows into a feature-typed ledger so future AI surfaces (summary,
embeddings, reply-draft) can drop in without schema changes.
- New table ai_usage_ledger (port, user, feature, provider, model,
input/output/total tokens, request id) with two indexes for rollup
- New service ai-budget.service.ts: getAiBudget/setAiBudget,
checkBudget (pre-flight gate), recordAiUsage, currentPeriodTokens,
periodBreakdown — all token-based, period boundaries in UTC
- runOcr now returns provider usage so the route can record the actual
spend instead of estimating
- Scan-receipt route gates on checkBudget before invoking AI; returns
source: manual / reason: budget-exceeded when blocked, surfaces
softCapWarning on the success path
- Admin UI: new AiBudgetCard on the OCR settings page — shows current
spend, per-feature breakdown, soft/hard cap inputs, period selector
- Permission: admin.manage_settings on both routes
Tests: 766/766 vitest (was 756) — +10 budget tests covering enforce/
disabled/cap-exceed/estimate-exceed/soft-warn/period boundaries/
cross-port isolation/silent ledger failure.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PR1 of Phase B per docs/superpowers/specs/2026-04-28-phase-b-insights-alerts-design.md.
Lays the foundation that PRs 2-10 will fill in with behaviour.
Schema (migration 0014):
- alerts table with rule-engine fields (rule_id, severity, link,
entity_type/id, fingerprint, fired/dismissed/acknowledged/resolved
timestamps, jsonb metadata). Partial-unique fingerprint index keeps
one open row per (port, rule, entity); separate indexes power
severity-filtered and time-ordered queries.
- analytics_snapshots (port_id, metric_id) -> jsonb cache + computedAt
for the 15-min recurring refresh.
- expenses: duplicate_of self-FK, dedup_scanned_at, ocr_status/raw/
confidence; partial index on (port, vendor, amount, date) where
duplicate_of IS NULL drives the dedup heuristic.
- audit_logs.search_text: GENERATED ALWAYS tsvector over
action+entity_type+entity_id+user_id, GIN-indexed (drizzle can't
model GENERATED ALWAYS in TS yet, so the migration appends manual
ALTER + the GIN index).
Service skeletons in src/lib/services/:
- alerts.service.ts: fingerprintFor, reconcileAlertsForPort (upsert +
auto-resolve), dismiss, acknowledge, listAlertsForPort.
- alert-rules.ts: RULE_REGISTRY of 10 rule evaluators (currently no-op);
PR2 fills in the bodies.
- analytics.service.ts: readSnapshot/writeSnapshot with 15-min TTL +
no-op compute* stubs for the four chart series; PR3 fills behavior.
- expense-dedup.service.ts: scanForDuplicates + markBestDuplicate
using the partial dedup index. PR8 wires the BullMQ trigger.
- expense-ocr.service.ts: OcrResult/OcrLineItem types + ocrReceipt
stub. PR9 wires Claude Vision (Haiku 4.5 + ephemeral system-prompt
cache).
- audit-search.service.ts: tsvector @@ plainto_tsquery + cursor
pagination on (createdAt, id). PR10 wires the admin UI.
tsc clean, lint clean, vitest 675/675 (one unrelated AES random-output
flake passes solo).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The client portal no longer uses passwordless / magic-link sign-in. Each
client now has a `portal_users` row with a scrypt-hashed password,
created by an admin from the client detail page; the admin's invite
mails an activation link that the client uses to set their own password.
Forgot-password is wired through the same token mechanism.
Schema (migration `0009_outgoing_rumiko_fujikawa.sql`):
- `portal_users` — one per client account, separate from the CRM
`users` table (better-auth) so the auth realms stay isolated. Email
is globally unique, password is null until activation.
- `portal_auth_tokens` — single-use activation / reset tokens. Stores
only the SHA-256 hash so a DB compromise never leaks live tokens.
Services:
- `src/lib/portal/passwords.ts` — scrypt hash/verify (no new deps;
uses node:crypto), token mint+hash helpers.
- `src/lib/services/portal-auth.service.ts` — createPortalUser,
resendActivation, activateAccount, signIn (timing-safe),
requestPasswordReset, resetPassword. Auth failures throw the new
UnauthorizedError (401); enumeration-safe behaviour everywhere.
Routes:
- POST /api/portal/auth/sign-in — sets the existing portal JWT cookie.
- POST /api/portal/auth/forgot-password — always 200.
- POST /api/portal/auth/reset-password — token + new password.
- POST /api/portal/auth/activate — token + initial password.
- POST /api/v1/clients/:id/portal-user — admin invite (and `?action=resend`).
- Removed: /api/portal/auth/request, /api/portal/auth/verify (magic link).
UI:
- /portal/login — replaced email-only magic-link form with email +
password + "forgot password" link.
- /portal/forgot-password, /portal/reset-password, /portal/activate — new.
- New shared `PasswordSetForm` component used by activate + reset.
- New `PortalInviteButton` rendered on the client detail header.
Email send:
- `createTransporter` now wires SMTP auth when SMTP_USER+SMTP_PASS are
set (gmail app-password or marina-server creds, configured via env).
- `SMTP_FROM` env var lets the sender address be overridden without
pinning it to `noreply@${SMTP_HOST}`.
Tests:
- Smoke spec 17 (client-portal) updated to the new flow: 7/7 green.
- Smoke specs 02-crud-spine, 05-invoices, 20-critical-path updated to
match the post-refactor client + invoice forms (drop companyName,
use OwnerPicker + billingEmail).
- Vitest 652/652 still green; type-check clean.
Drops the dead `requestMagicLink` from portal.service.ts.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
