Changelog

Customer/job feedback pipeline

POST /v1/jobs/:id/reports and GET /v1/jobs/:id/reports support customer/API reporting and history with organization-scoped feed ownership validation, reason labels, and metadata.

Admin feedback triage

Observability dashboard now includes a Feedback tab with summaries, hotspots, recommendations, and one-click status/action controls for recent reports.

Merge hygiene tooling

cleanup + doctor + snapshot + triage commands are now a documented operator workflow for post-merge and post-incident review.

Observability UI refresh

Admin observability page now exposes status cards, doctor packet, attention items, pipeline cards, feedback signal, and copyable handoff context in one page.

Data cleanup controls

added legacy monitor row and stale job-state cleanup tooling (cleanup:observability) to keep signal windows clean for reliable triage.

tRPC developer ergonomics

debug logging noise in browser console reduced by gating tRPC logger noise to debug/trace-only operation.

Merged monitoring artifacts

fixed stale monitor state and legacy L1.5 blocked evidence affecting operator confidence after cleanup and real merge cycles.

Observability docs drift risk

added explicit post-merge operator checks to avoid stale assumptions after branch integration.

bun run gen:openapi:check: openapi.json is up to date.

bun run test: 821/821 passing across 72 files.

cd workers/enrichment && bun run test: 68/68 passing across 4 files.

bun run observability:doctor: completed; Apify L4 readiness is OK, snapshot remains degraded because recent job monitoring evidence still has blocked checks.

bun run observability:triage: completed; current data shows 110 failed, 114 queued, and 16 processing jobs in the 36h customer-visible window, plus 16 blocked monitor checks and 3 L4 rendered checks.

bun run observability:snapshot: completed with status=degraded for current data.

bun run cleanup:observability: dry-run completed; 16 legacy monitor logs and 16 legacy job monitor states matched. Apply was not run in this docs sync.

Feedback reports validate feed ownership but do not yet verify that the reported job belongs to that feed's filters, so feed-attributed hot spots are directional until the caller supplies the correct feed context.

Feedback admin reads sort global recent reports by created_at and status without a matching compound date/status index. Add the index before high-volume reporting.

Post-merge observability is operational but not clean: the 36h window still includes failed/queued/processing jobs and blocked L1.5 monitor checks. L4 is configured and working, but the historical blocked window needs to roll off or be cleared with an explicitly reviewed cleanup apply.

Observability triage CLI

bun run observability:triage reads the snapshot and existing webhook_log, job_source, job, and monitor_log rows to explain source ingest, customer-visible jobs, and job monitoring degradation with safe aggregate buckets.

Read-only triage builder

lib/observability/triage.ts groups recent source runs, job processing state, and monitor rows by safe source/status/error-kind dimensions while redacting emails, URLs, IDs, payloads, and raw error text.

Operator remediation commands

jobs:unlock-stuck-pending, apify:backfill-timed-out, and jobs:process-enrichment document and expose the safe remediation paths used for stranded pending jobs, timed-out Apify datasets, and local enrichment fallback.

Timed-out Apify webhooks

ACTOR.RUN.TIMED_OUT now attempts partial dataset ingest when run and dataset IDs are available instead of treating all timeouts as terminal source-ingest errors.

Worker queue dispatch safety

manual and scheduled queue dispatch batches are capped at Cloudflare's 100-message sendBatch limit, including oversized manual query overrides.

Observability evidence

customer-visible job evidence now distinguishes pending, queued, processing, failed, and completed jobs created in the 36-hour window.

Docs/runbooks

README, AGENTS, CLAUDE, and the observability runbook now make observability:snapshot then observability:triage the first diagnostic path.

Re-enrichment lock reset

existing jobs whose raw source data changes now reset both processing_status and processing_run_id, preventing pending rows from becoming unclaimable.

Retryable failed jobs

the enrichment worker can requeue retryable failed jobs with draft or active public status.

AI date normalization

invalid AI-produced date strings are written as NULL instead of failing otherwise valid enrichment jobs.

Production backlog remediation

stale queued/processing locks were reset, the Cloudflare queue was purged after DB state was restored to pending, a one-message queue canary succeeded, and 1881 jobs were redispatched through verified queue delivery.

bun run typecheck: clean

bun run lint: clean

bun run gen:openapi:check: openapi.json is up to date

bun run test -- tests/pipeline/apify-webhook.test.ts tests/pipeline/ingest.test.ts tests/pipeline/observability-snapshot.test.ts tests/pipeline/observability-triage.test.ts: 73/73 passing

bun run test: 765/765 passing

cd workers/enrichment && bun run test: 68/68 passing

bun run build: exit 0; emits the known local Google OAuth credential warning and module-type warning

Cloudflare worker deployment f40d11b2-5c80-4c10-b04c-1c5216c313ca: health ok, queue enabled

Queue canary: 1/1 queued job succeeded in production

Local enrichment fallback: 150/150 jobs succeeded, 0 failed, 1616 stale locks reset

Production triage samples after remediation: pending 0 with queue drain progressing from queued 865 / processing 38 / completed 417 / failed 18 at 2026-07-02T22:55:00Z to queued 0 / processing 0 / completed 1318 / failed 20 at 2026-07-02T23:09:00Z

Internal observability snapshot

GET /api/internal/observability/snapshot is protected by CRON_SECRET and reports API, database, release, source ingest, customer-visible jobs, AI enrichment, worker, monitoring, email, webhook, cron, and Sentry readiness signals with freshness, confidence, warnings, and safe evidence.

Snapshot CLI

bun run observability:snapshot prints the same developer/AI-agent health context locally for audits and release checks.

Release metadata helper

runtime release evidence now accepts SHA-shaped Vercel commit metadata while rejecting branch names such as main.

Observability regression tests

route, public health, release metadata, snapshot freshness, Sentry privacy, active-source ingest freshness, confirmed email delivery, disabled delivery channels, enabled-only delivery freshness, and OpenAPI fixture coverage now guard the new contracts.

Public health contract

/api/v1/health now returns structured API/database/release checks, version, nullable short commit metadata, environment, timestamp, and response time.

Observability docs

README and the observability runbook document the snapshot command, protected endpoint, component meanings, known gaps, and verification workflow.

False-green observability signals

source ingest, customer-visible jobs, AI enrichment, monitoring, email, webhook, and cron checks now degrade on stale/missing freshness evidence instead of treating unrelated historical rows as proof of current health.

Customer-visible freshness semantics

the customer-visible jobs component now uses job.processed_at for active completed freshness while keeping job.created_at counts as backlog evidence, avoiding false degradation when enrichment is processing older backlog.

Delivery signal accuracy

email delivery success is based on confirmed email_delivery.status = delivered rows, source ingest freshness is scoped to active job sources, and email/webhook success/failure evidence is filtered to currently enabled delivery configuration.

Sentry readiness baseline

fully configured Sentry reports ok while dashboard readback remains a known gap instead of permanently degrading the whole snapshot.

Protected snapshot auth

the snapshot route now fails closed when CRON_SECRET is absent, including when env validation is intentionally skipped.

Cron auth/readback hardening

CRON_SECRET-protected cron/internal routes now reuse the shared constant-time auth helper, and unhandled snapshot generation failures are logged before returning 500.

bun run gen:openapi:check: openapi.json is up to date

bun run lint: clean

bun run typecheck: clean

bun run test tests/pipeline/cron-auth.test.ts tests/pipeline/observability-snapshot.test.ts tests/pipeline/observability-snapshot-route.test.ts tests/pipeline/sentry-probe-route.test.ts tests/pipeline/scraper-slugs-route.test.ts tests/pipeline/apify-webhook.test.ts tests/pipeline/webhook-validation.test.ts tests/pipeline/push.test.ts: 93/93 passing

bun run test: 759/759 passing

bun run build: exit 0; emits the known Better Auth Google credential warning and module-type warning

bun run observability:snapshot: exit 0; local snapshot returns real degraded evidence for current ingest/job data

bun run verify:e2e-one:dry-run: stages 1-3 passed

Snapshot secret scan: snapshot_secret_leaks=0

Controlled Sentry readback route

POST /api/internal/sentry-readback is protected by CRON_SECRET, captures one synthetic exception, flushes Sentry, and returns the event ID for dashboard/API confirmation.

Readback route tests

route coverage verifies deny-by-default auth, successful capture/flush behavior, and safe marker fallback for unsafe marker input.

Observability runbook

README_OBSERVABILITY.md now documents the production readback command and the expected Sentry properties, with event ID as the canonical lookup key.

TODO hygiene

the Google OAuth credential warning stays pending until public beta auth planning; scheduled production confirmation for monitor_log writes remains tracked separately.

bun run with-dev-env -- vitest run tests/pipeline/sentry-probe-route.test.ts --reporter verbose: 3/3 passing

bun run lint: clean

bun run typecheck: clean

bun run test: 717/717 passing

bun run build: exit 0; emits the known Better Auth Google credential warning

Vercel production deployment dpl_6dWEXKJ7vR3yhC7oXu5wM8CL9q2E from commit 8587db2b7ade86a14bbee6115e9221b6ed4b6192: ready

Sentry event 70e3962abbe34f53b1cb1f9ca8cf4bef landed in craftled-se / bordfeed with environment=production, release 8587db2b7ade86a14bbee6115e9221b6ed4b6192, and probe=sentry-readback

Raw Sentry event check: probe email, authorization value, URL query token, and nested session token were absent; marker-like values were redacted by the scrubber

GitHub release-note audit: v3.0.25 through v3.0.33.0 normalized public release bodies match checked-in docs/releases files

Shared Sentry scrubbers

browser, server, and edge Sentry errors, transactions, and spans now remove request URLs/bodies, cookies, headers, query strings, email-looking values, bearer tokens, and token-like keys before ingestion.

Sentry privacy regression tests

unit/source guards cover tRPC input capture, user email/IP logging, runtime event/trace scrubber wiring, replay sampling, redaction, and noisy browser error drops.

Release observability setup

Bordfeed is connected to the craftled-se / bordfeed Sentry project via the Vercel Marketplace integration, with Sentry env vars present for Production, Preview, and Development.

Monitor audit trail writes

monitorJobBatch() now writes best-effort monitor_log rows for each checked job, so L0-L3 status decisions have a per-check audit trail.

tRPC Sentry middleware

raw RPC input capture is disabled, and Sentry user scope now uses user ID only.

Session Replay

replay capture remains disabled (0 session/error sampling) until an explicit privacy/product decision enables it.

Error logging

tRPC error logs no longer include raw user email or IP, and email send logs use a stable recipient hash instead of raw recipient addresses, reducing PII risk with the Vercel Sentry log drain.

Typecheck drift in enrichment postprocess tests

removed stale never casts so bun run typecheck is clean again.

Homepage release badge drift

changelog parsing now accepts four-part release headings such as 3.0.33.0, so the marketing hero badge follows the latest release instead of falling back to 3.0.30.

bun run lint: clean

bun run typecheck: clean

bun run test: 714/714 passing

bun run build: exit 0; emits the known Better Auth Google credential warning

Production canary for deployment dpl_94s3XDdf9gfhqa2ethjSt45J68tF: homepage 200, health API 200 with database up, no browser console errors, no Vercel error logs or 5xx logs found.

preferSourceApplyUrl()

WWR RSS external href extraction; DynamiteJobs applyLink and listing fallback when AI returns a company homepage.

scripts/cleanup-stale-jobs.ts + scripts/audit-stale-records.ts

dry-run/apply housekeeping for failed enrichments, stuck pending rows, and bare homepage apply URLs (npm run cleanup:jobs).

verify:e2e-one:dry-run npm script

safe default golden-path check without prod writes.

Unit tests

bare homepage quality gate, DynamiteJobs/WWR apply URL preference, e2e-golden ingest block (691 total).

E2E fixture leak

e2e-golden* source_job_id values blocked at ingest; verify:e2e-one requires --allow-prod (and --allow-prod-ingest for fixtures).

Bare homepage apply URLs

enrichment fails when AI returns https://company.com with no job path; postprocess prefers source listing/apply links first.

verify:e2e-one

defaults to --dry-run; full prod E2E is opt-in.

CF Worker prompts

WWR and DynamiteJobs apply URL hints (deploy worker separately for prompt changes).

npm run test:unit: 691/691 passing

Enrichment worker tests: 65/65 passing

lib/pipeline/enrichment-postprocess.ts

shared post-AI normalization: Workday apply_url resolution from scraper source_url, raw-content validation for listing-only payloads, and minimum description length gate.

scripts/verify-one-job-e2e.ts

golden-path E2E verifier with --live, --source (dynamitejobs/workday/wwr), --pipeline, and --feed flags.

scripts/adversarial-workday-test.ts

tiered adversarial tests that exercise prod CF Worker enrichment without local OpenAI bypass.

tests/pipeline/enrichment-postprocess.test.ts

unit tests for apply URL resolution and quality gates.

Workday relative apply URLs

enrichment now resolves /job/... paths to absolute URLs using the scraper’s source_url site segment; fixes broken apply links in the customer-visible pool.

Hollow enrichment completions

listing-only Workday raw and jobs with empty/thin descriptions are marked failed instead of completed.

Workday scraper time-budget fallback

when fetch_details=true, remaining listings are skipped instead of pushed as listing-only jobs.

CF Worker enrichment

process-jobs.ts applies postprocess gates before DB write-back (deploy separately via workers/enrichment wrangler deploy).

DynamiteJobs smoke config

writing-editing category filter for dev smoke runs.

verify:e2e-one npm script added to package.json.

npm run test:unit: 682/682 passing

Adversarial Workday tier 3: 15 PASS, 0 FAIL

DynamiteJobs live E2E: all stages passed (Writing Jobs feed)

AI SDK patches

ai 7.0.0 → 7.0.4, @ai-sdk/openai ^4.0.2, @ai-sdk/react 4.0.5 (lockstep).

Sentry

@sentry/nextjs 10.61.0 → 10.62.0; adds AI SDK v7 support for vercelAiIntegration.

Runtime dependencies

safe patch set: Scalar 0.9.49, TanStack Query 5.101.2, Shiki rehype 4.3.0, Fumadocs core/ui/mdx 16.10.6 / 15.0.13, motion 12.42, nanoid 5.1.16, react-email 6.6.5, Resend 6.16.0, Stripe SDK 22.3.0.

Dev tooling

autoprefixer 10.5.2, lint-staged 17.0.8, postcss 8.5.16.

Stripe API version pin

stripe@22.3.0 narrows TypeScript apiVersion to the SDK default; kept runtime pin at 2026-05-27.dahlia with an explicit @ts-expect-error so SDK bumps never silently shift live API behavior.

better-auth 1.6.22

deferred; requires twoFactor lockout column migration.

cropperjs v1, @types/node 25, testcontainers 11

unchanged from v3.0.29.

bun install --frozen-lockfile: clean

bun run typecheck, bun run lint, bun run build: clean

bun run test: 672/672 passing

bun run gen:openapi:check: up to date

AI SDK v6 → v7 (major)

ai 6.0.209 → 7.0.0, @ai-sdk/openai 3.x → 4.0.0, @ai-sdk/react 3.x → 4.0.0. Migrated toTextStreamResponse() → createTextStreamResponse({ stream: toTextStream(result) }) in the chat route handler (only affected call site). Client-side useChat + TextStreamChatTransport + sendMessage pattern was already v7-compatible.

Sentry

@sentry/nextjs 10.60.0 → 10.61.0; adds streamGenAiSpans: true default for gen_ai span extraction.

Runtime dependencies

safe patch set: AWS S3 SDK 3.1075, Scalar 0.9.48, TanStack Query 5.101.1, tRPC 11.18, better-auth 1.6.20, Fumadocs core/ui 16.10.5, lucide-react 1.21, motion 12.41, nanoid 5.1.15, papaparse 5.5.4, pg 8.22, react-email 6.6.4, react-hook-form 7.80, Recharts 3.9, Resend 6.14, sharp 0.35.2, Stripe 22.2.3, and uuid 14.0.1.

Dev tooling

upgraded Biome to 2.5.1 and Playwright to 1.61.1.

Sentry config

moved reactComponentAnnotation and autoInstrumentMiddleware to the new webpack.* option paths so the production Sentry wrapper no longer emits deprecation warnings under 10.60.

Superadmin smoke reliability

the authenticated Playwright smoke now has an explicit route-suite timeout and waits for client-side network idle before checking dashboard/admin routes, so dependency regressions that surface after initial document load are less likely to slip through.

cropperjs v1, @types/node 25, testcontainers 11

intentionally unchanged. cropperjs v2 remains blocked by react-cropper; @types/node 26 is ahead of the Node 22 runtime; testcontainers 12 waits on a Node engine floor bump to 22.22+.

bun install --frozen-lockfile: clean

bun run typecheck, bun run lint, bun run build: clean

bun run test: 672/672 passing

bun run with-dev-env -- playwright test tests/e2e/superadmin.spec.ts: 1/1 passing

better-auth 1.6.18 → 1.6.19

bug-fix release, no new schema fields. Relevant fixes: cookie-cache fallback lookup (#9348), session/account cache cookies split into chunks near the browser per-cookie size limit (#10088), and single-use verification flows (magic-link) reusing active transactions on connection-limited DB adapters (#10070, useful on Neon).

@content-collections/core 0.15.1 → 0.15.2

patch release for the content/MDX build chain (blog, docs, changelog collections).

cropperjs v1, testcontainers v12

unchanged from v3.0.27 (react-cropper v1 pin; Node ≥ 22.22 floor). See v3.0.27 notes.

bun run lint, bun run typecheck, bun run build: clean

bun run test: 672/672 passing

Dependencies (Tier 1 — safe drop-ins)

Radix UI (27 packages to latest, lockstep), AI SDK (ai 6.0.206, @ai-sdk/openai 3.0.71, @ai-sdk/react 3.0.208 — security patches), Next 16.2.9, Tailwind 4.3.1, lucide-react 1.18, @scalar/api-reference-react 0.9.46, @shikijs/rehype 4.2, react-hook-form 7.79, react-email 6.6, use-stick-to-bottom 1.1.6, @aws-sdk/* 3.1069 (paired), @stripe/stripe-js 9.8, @marsidev/react-turnstile 1.5.3, @types/node + @types/react (override synced), vitest 4.1.9.

Dependencies (Tier 3 — reviewed)

zod-openapi 6.0.0 (major; spec stays OpenAPI 3.1.0 and regenerates byte-identical), better-auth 1.6.18 (session/2FA/OAuth security fixes), stripe 22.2.1, @sentry/nextjs 10.58, sharp 0.35.1, react-qr-code 2.2, fumadocs-core/fumadocs-ui 16.10.3, fumadocs-mdx 15.0.12, @playwright/test 1.61.

Biome 2.5

noShadow and noIncrementDecrement were promoted out of nursery; moved both to their stable groups (suspicious/style, still off) so bun run lint and the pre-commit hook stay green.

Stripe API version pinned

lib/billing/stripe.ts now sets apiVersion: "2026-05-27.dahlia" explicitly so SDK upgrades never silently shift the live API version (22.2.1 would otherwise have moved it from 2026-04-22.dahlia).

2FA enrollment under better-auth 1.6.x

added the two_factor.verified column (boolean, default true) that the upgraded twoFactor plugin reads/writes for the TOTP verification gate. Without it, enrollment returned a 500 (field "verified" does not exist); build, typecheck, and unit tests all passed — only live testing caught it. Migration 20260616090415.

cropperjs v1

react-cropper@2.3.3 hard-pins cropperjs ^1.5.13; v2 is a web-components rewrite that removes getCroppedCanvas(). Frozen on v1 until react-cropper ships v2 support.

testcontainers v12

requires Node >= 22.22; repo pins engines.node 22.21.1. Deferred until the Node floor is raised.

bun run lint, bun run typecheck, bun run build: clean

bun run test: 672/672 passing

Live QA against the running app: /docs (fumadocs 16.10.3 TOC), /docs/api (Scalar 0.9.46, OpenAPI 3.1.0), sign-up → onboarding → sign-in, full 2FA lifecycle (enable → QR → TOTP verify → disable), superadmin admin routes, Radix Select filters.

Closure mass-close guard

reconcileJobClosure now skips set-reconciliation when the scraped set covers less than 50% of active DB jobs, preventing smoke-test/limited runs (e.g. {seed_slugs: ["linear"], limit: 1}) from mass-closing thousands of legitimate jobs. 1,250 jobs wrongly closed on 2026-06-08 were reopened; live pool restored to 1,369.

Intra-batch duplicate externalId crash

ingestJobBatch now deduplicates jobs by externalId within a single payload before the DB lookup. Previously, RSS feeds emitting the same source_job_id twice caused the batch INSERT to hit a unique-constraint error, aborting the entire chunk silently.

Concurrent-ingest ID alignment

RETURNING now fetches (id, externalId) and uses a Map for result assignment instead of positional indexing. onConflictDoNothing() silently omits conflicting rows from RETURNING, which broke positional offset math under concurrent webhook ingests of the same jobs.

Known concurrent duplicate tradeoff

if two webhook invocations ingest the same new jobs concurrently, the loser may classify rows skipped by onConflictDoNothing() as per-job error results because those rows are absent from RETURNING. The database state is still correct and the actor-run idempotency guard makes this rare, but an all-duplicate losing invocation can surface as a false pipeline failure notification.

scripts/reopen-falsely-closed-jobs.ts

one-time remediation script that targeted and reopened the 1,250 incident-scoped closed jobs (date-fenced to 2026-06-08, source IN ashby/weworkremotely, consecutive_misses=2). Supports --dry-run.

scripts/verify-pipeline.ts

end-to-end pipeline health check script covering all 5 stages: sources, raw ingest, AI enrichment, customer-visible pool, and email delivery.

Post-merge release sync

prepared GitHub release notes for v3.0.26, refreshed README/ROADMAP/TODOS/v3 architecture docs for the closure guard and ingest dedup tradeoffs, and corrected verify-pipeline.ts wording to say email delivery only.

bun run test: 672/672 passing

Dashboard session persistence

changed proxy.ts to prefer Better Auth's signed session_data cache and validate with cookie-cache bypass plus refresh disabled only on cache miss, so an expired cache no longer logs out users while the real session_token remains valid without adding a DB-backed session lookup to every protected request.

Auth-page redirects after cache expiry

signed-in users now continue to redirect away from /auth/* pages even after the short-lived session cache expires.

Auth proxy regression coverage

added proxy tests for cached-session fast paths, valid-token/missing-cache navigation, corrupt-cache fallback, unauthenticated redirects, session lookup failures, and logged-in auth-page redirects including organization invitations.

Release state sync

aligned package/OpenAPI version metadata with VERSION (3.0.25), refreshed source docs around the current 16 DB-mapped Apify actors (2 active in production), corrected live schema table counts to 38, and refreshed the prepared GitHub release notes for v3.0.25.

bun run test: 667/667 passing

workers/company-enrichment bun run test: 25/25 passing

bun run typecheck: clean

bun run lint: clean

bun run gen:openapi:check: spec up to date

bun run scraper:check: 16 local actors mapped to job_source with v3 fields

bun run build: exit 0, 65/65 static pages

WeWorkRemotely Apify version alignment

changed the actor manifest back to Apify version 0.0, the version that owns the latest build tag used by default runs, preventing future pushes from publishing to an unused actor version.

WWR actor input/runtime contract

made the actor honor the documented rssUrl override, restrict it to HTTP(S) WeWorkRemotely hosts, guard runtime limit values against schema bypasses, declared ESM module semantics, removed the unused useProxy input, trimmed unused runtime dependencies, and replaced the template README with Bordfeed-specific v3 input/output docs.

WWR webhook ingestion

repaired the stale remote actor path that emitted legacy dataset fields and caused Apify webhook rows to complete with jobs_errored instead of ingesting v3 raw-job records.

WWR regression coverage

extended the scraper contract test to lock the actor version/tag assumption, documented input fields, RSS URL override wiring, and dependency footprint.

bun run scraper:check: 16 local actors mapped to job_source with v3 fields

bun run lint: clean

bun run typecheck: clean

bun run test: 653/653 passing

bun run build: exit 0, 65/65 static pages

Remote WWR Apify smoke: run euRWBHt4qYChkN6Y1 succeeded on build 0.0.12, validated 1 v3 dataset item, and webhook ingestion completed with jobs_errored=0

Remote Apify scraper DX

added scripts/scraper-dx.ts plus bun run scraper:* commands for listing DB-mapped actors, static v3 contract checks, smoke runs, run status, logs, dataset fetches, validation, and actor pushes. Smoke inputs cover all 16 local actors and prevent accidental full production scrapes unless explicit input is supplied.

Superadmin route smoke

added an env-driven Playwright spec that signs in as a superadmin, selects an organization, and verifies dashboard, organization, admin, and API docs routes render without auth redirects or application errors.

Protected-route auth gating

enabled Better Auth session cookie caching and changed proxy.ts to decode the signed session_data cookie directly, avoiding internal /api/auth/get-session calls that caused intermittent sign-in redirects during rapid protected-route navigation.

Scraper actor metadata and seed behavior

normalized Workable and WeWorkRemotely actor manifests, pinned Workable runtime dependencies, coerced Teleperformance job IDs to strings, and prioritized explicit Workday/Oracle seed companies without forcing full discovery first.

Docs cleanup

replaced the starter Fumadocs pages with Bordfeed quick-start and platform overview docs, and refreshed ATS.md with the 2026-06-08 Bluedoor snapshot plus Ashby coverage/status.

Superadmin E2E auth bounce

fixed the production-only route protection failure where valid Better Auth cookies could remain present while protected pages redirected to /auth/sign-in under fast navigation.

bun run check: clean

bun run typecheck: clean

bun run gen:openapi:check: spec up to date

bun run with-dev-env -- vitest run: 649/649 passing

bun run build: exit 0, 65/65 static pages

bun run test:e2e: 8/8 passing, including superadmin smoke

bun run scraper:check: 16 local actors mapped to job_source with v3 fields

Remote Ashby Apify smoke: run succeeded and validated 1 v3 dataset item

Canonical ATS registry + lifecycle schema

added ats_platform, company_alias, company_event, and company_ats_history tables; extended company with lifecycle status/evidence fields; extended company_ats with platform FK, validity windows, first/last-seen timestamps, redirects, and indexes. Drizzle relations and shared selects now expose lifecycle fields consistently.

ATS platform helpers and seeds

added canonical platform metadata/aliases for 16 ATS/source platforms, default Ashby seed slugs, board URL generation, db:seed:ats, and reset/reseed scripts for fast iteration over scraped company/job data.

DB-backed scraper slug APIs

internal scraper slug read/report/discovery endpoints now normalize platform aliases, resolve canonical platform rows, write ATS history, and upsert discovered slugs into company_ats.

Admin lifecycle tools

company detail pages now show lifecycle status, aliases, events, ATS connections, and ATS history, with admin mutations for managing aliases/events/connections and archiving companies.

Coverage for ATS and scraper route guardrails

added unit tests for platform normalization/board URLs and internal scraper slug auth/input-validation paths.

Company visibility now respects lifecycle state

public API, organization dashboard, admin listings, similar-company queries, filter options, and fuzzy fallback exclude archived companies while keeping active/acquired/merged/rebranded/unknown records available.

Greenhouse seed path upgraded

Greenhouse slug seeding now resolves the canonical platform row, writes platform FK/board URLs/status metadata, and records backfill history.

Scraper report completeness

slug report upserts now refresh board_url alongside status, scrape timestamps, and job counts.

Shared prod/testing Neon DB migrated, scraped data reset, and reseeded: 16 ATS platforms, 20 Ashby slugs, 7,101 Greenhouse slugs, 20 materialized Ashby companies, 1,241 linked Ashby jobs, and 10 WeWorkRemotely smoke jobs.

Live Ashby scrape: 20/20 boards active, 1,241 jobs, 0 errored; DB invariants verified with 0 unlinked Ashby jobs and 0 missing ATS platform links.

bun run test: 649/649 passing · bun run typecheck: clean · bun run lint: clean (3 pre-existing warnings unrelated to this diff) · env-backed bun run build: exit 0, 65/65 static pages.

Local app smoke with main .env.local: / returns 200 and /api/v1/health returns status: ok.

Hero badge polish

replaced the raw top-three changelog-title ticker with a compact announcement title derived from the latest release's first item. Technical parentheticals, extra clauses, and long words are normalized into a short public-facing label (for v3.0.20: Ashby scraper), and the badge text is capped responsively so future verbose release notes cannot blow out the hero layout.

Release announcement tests

added coverage for Ashby-style technical changelog titles, empty releases, first-clause cleanup, word-boundary truncation, hard truncation, and fallback behavior when cleanup removes the full title.

bun run test: 640/640 · bun run typecheck: clean · bun run lint: clean (3 pre-existing warnings unrelated to this diff)

Browser smoke on http://localhost:3000: mobile 390px and desktop 1440px render the badge without horizontal overflow; Next runtime diagnostics report no config or session errors.

Ashby scraper (apify-scrapers/Ashby Scraper/). Fetches api.ashbyhq.com/posting-api/job-board/{slug} per board and emits the lean v3 format ({source_job_id, source_name, url, raw}). The board slug is tagged onto raw (ashby_board_slug) because Ashby job objects carry no company-name field, so company resolution always has a signal. Listed-only by default; no date filter by default so closure reconciliation stays correct for these small boards; capped at 20 boards for now. Deployed to Apify (bordfeed-ashby-scraper, actor RgwxfmXHhi6Dm4PDI), webhook + job_source row wired, and registered in the v3 contract test (now 16 scrapers).

ATS.md rewritten as a two-source gap analysis (SourceStack market volume + Bluedoor live aggregator coverage, cross-validated). Adds a tiered build roadmap (iCIMS / UKG / SuccessFactors next), a domain-coverage filter explaining why ADP/Paycom are deprioritized despite high volume, and a "scrape-difficulty moat" read on Workday/SmartRecruiters.

bun run test: 634/634 · bun run typecheck: clean · bun run lint: clean (3 pre-existing warnings unrelated to this diff)

Ashby actor E2E in production: run #1 ingested 1,242 jobs (20/20 boards, 0 errored, 2.97s); dedup run #2 → 0 created, 1,242 updated, job count steady at 1,242 (externalId dedup confirmed); both webhook deliveries logged completed. Per-board counts match an independent API probe; live-API ID-set diff exact on linear/modal/cohere.

Company active_job_count no longer freezes after job deletions. workers/company-enrichment/src/stats.ts recomputed counts only for companies present in the job table (FROM job ... GROUP BY company_id), leaving companies that dropped to zero jobs at stale values. Added a second pass that zeroes job_count/active_job_count for companies with no jobs (NOT EXISTS + change guard). Production verified: sum(active_job_count) equals active jobs with a company (37 == 37), 0 phantom rows. (+4 tests in stats.test.ts)

.env.example

documented CRON_SECRET and WORKER_ENRICHMENT_URL. The admin Enrichment "Worker Health" badge proxies the company worker's /stats via WORKER_ENRICHMENT_URL; it was undocumented and unset in prod, so the badge always read "Worker unavailable" despite the worker's hourly cron running fine.

bun run test: 617/617 (main) · workers/company-enrichment: 25/25 (incl. 4 new) · typecheck clean

Company worker deployed (bordfeed-company-enrichment); worker CRON_SECRET synced with Vercel prod (/stats → 200); WORKER_ENRICHMENT_URL wired in production

React core

react + react-dom 19.2.6 → 19.2.7 (FormData regression fix), @types/react 19.2.14 → 19.2.16 (bumped in both devDependencies and the overrides block), @types/node 25.6.2 → 25.9.1, react-day-picker 10.0.0 → 10.0.1.

Next.js

next + @next/bundle-analyzer 16.2.6 → 16.2.7 (bug-fix backport, not a security release), @sentry/nextjs 10.52.0 → 10.56.0 (no Next-facing API or withSentryConfig changes).

AI SDK + email

ai 6.0.176 → 6.0.195 + @ai-sdk/react 3.0.178 → 3.0.197 (lockstep), @ai-sdk/openai ^3.0.63 → ^3.0.67, react-email 6.1.1 → 6.5.0 (dev-CLI only), resend 6.12.3 → 6.12.4.

Docs

fumadocs-core + fumadocs-ui 16.8.8 → 16.9.3 (exact-peer lockstep), fumadocs-mdx 15.0.0 → 15.0.10 (Vite/Rolldown work only — no loader-path or next.config.mts change), @content-collections/core 0.15.0 → 0.15.1, @shikijs/rehype 4.0.2 → 4.1.0.

Styling/UI

tailwindcss + @tailwindcss/postcss 4.2.4 → 4.3.0 + tailwind-merge 3.5.0 → 3.6.0 (3.6.0 adds Tailwind 4.3 class-set support), postcss 8.5.14 → 8.5.15, lucide-react ^1.14.0 → ^1.17.0, motion 12.38.0 → 12.40.0 (React 19 AnimatePresence fixes), geist 1.7.0 → 1.7.2.

Storage

@aws-sdk/client-s3 + @aws-sdk/s3-request-presigner 3.1045.0 → 3.1059.0 (lockstep; "version bump only" across the range).

Forms/data

react-hook-form 7.75.0 → 7.77.0 (get() prototype-pollution hardening), @hookform/resolvers 5.2.2 → 5.4.0 (Zod 4 ./zod path unchanged), @tanstack/react-query 5.100.9 → 5.101.0, pg 8.20.0 → 8.21.0, date-fns 4.1.0 → 4.4.0, @stripe/stripe-js 9.4.0 → 9.7.0.

Tooling

vitest 4.1.5 → 4.1.8, @playwright/test 1.59.1 → 1.60.0 (new Chromium 148 / Firefox 150 binaries), @biomejs/biome ^2.4.14 → ^2.4.16 (nursery-only rule additions, no default changes), lint-staged 17.0.2 → 17.0.7.

cropperjs 1.6.2 → 2.1.1

blocked by react-cropper 2.3.3, which pins cropperjs ^1.5.13; no react-cropper release supports the v2 Web Components rewrite, and v2 drops the bundled dist/cropper.css that app/layout.tsx imports. Already tracked in TODOS.md ("Replace react-cropper with react-easy-crop").

testcontainers + @testcontainers/postgresql 11.14.0 → 12.0.1

v12 requires Node ≥ 22.22; engines.node is 22.21.1. Test-only (gated behind RUN_DB_TESTS), deferred until the Node floor is raised.

stripe 22.1.1 → 22.2.0, better-auth 1.6.9 → 1.6.14, @scalar/api-reference-react 0.9.34 → 0.9.42

held for a separate PR: Stripe silently shifts its pinned API version, Better Auth carries auth-behavior changes (invitation-verification default) worth testing in isolation, and Scalar is a fast-moving client SPA needing a browser smoke.

bun run lint: clean · bun run typecheck: clean · bun run gen:openapi:check: spec unchanged

bun run test: 617/617 passing · bun run build: exit 0, 65/65 static pages

Browser smoke against the production build: all 8 public routes HTTP 200; homepage FAQ accordion (AnimatePresence) and /docs (fumadocs 16.9.3) render with zero console errors

Enrichment over-tagging (workers/enrichment/src/prompt.ts)

removed the two "prefer writing-family categories" nudges and the writing-dominated example list that biased GPT-4o-mini toward "Technical Writing". The prompt now classifies by the role's primary function and returns null when unsure. Verified on 25 freshly-scraped jobs: 0 mis-tagged (was the #1 category), with a labeled eval covering both directions including the adversarial "mentions writing docs" case.

Enriched-only customer surfaces

a new activeEnrichedJob() helper (lib/jobs/conditions.ts) enforces status = active AND processing_status = completed, replacing 17 call sites that previously filtered only status = active: the public Feeds API (/v1/feeds/:id/jobs + facets) and export, company detail + company jobs, similar jobs, template preview, the org dashboard (feed preview/export/digest/send, jobs browser + fuzzy fallback + filter options, job detail), and the email-export and push crons — so the scheduled newsletter never delivers an unenriched job.

lib/jobs/conditions.ts

shared activeEnrichedJob() visibility condition (single source of truth for "what end users may see").

Tests

prompt.test.ts regression guard (fails if the writing bias returns); tests/jobs/active-enriched.test.ts (helper compiles to status + processing_status); tests/feeds/writing-feed-flow.db.test.ts (DB integration: writing feed returns only enriched writing jobs, excludes a pending job).

Dev/QA scripts

scripts/eval-occupational-category.ts (classification eval), verify-feed-flow.ts, run-ingest-enriched.ts, test-enrich-sample.ts, seed-test-accounts.ts; TEST_ACCOUNT_PASSWORD placeholder in .env.example.

bun run typecheck: clean · bun run lint: clean · bun run test: 617/617 passing · bun run gen:openapi:check: spec unchanged

Prod behavioral check: flipped a live job to processing_status = pending → it dropped from the engineering feed (15 → 14) → restored → 15.

Full customer flow in-browser: created a custom "Remote Engineering" feed → digest rendered 12 correctly-categorized engineering jobs with a real GPT-4o-mini intro → sent via Resend, confirmed delivered in email_delivery.

Self-hosted Umami analytics (components/analytics/umami.tsx)

a cookie-free next/script tag gated on NEXT_PUBLIC_UMAMI_SCRIPT_URL + NEXT_PUBLIC_UMAMI_WEBSITE_ID. Renders nothing until both are set, so the app builds and runs before analytics is configured. Both values are public (they appear in page HTML), not secrets.

Max and Ultra pricing plans (config/billing.config.ts)

Max at $299/mo · $2,870/yr, Ultra at $999/mo · $9,590/yr, both with a 14-day trial and ~20% annual discount (matching Pro). Checkout for these tiers goes live once their NEXT_PUBLIC_STRIPE_PRICE_MAX_* / _ULTRA_* price IDs are set; until then the existing min(1) guard blocks any malformed checkout.

Env vars (lib/env.ts, .env.example)

NEXT_PUBLIC_UMAMI_SCRIPT_URL, NEXT_PUBLIC_UMAMI_WEBSITE_ID, and four NEXT_PUBLIC_STRIPE_PRICE_MAX_* / _ULTRA_* price IDs (all optional).

Legal pages rewritten (content/legal/{privacy,terms,cookies}.mdx)

full GDPR/CPRA policies for an API-first SaaS: third-party processor list (Vercel, Neon, Cloudflare R2 + Turnstile, Resend, Stripe, OpenAI, Anthropic, Sentry, Apify, Logo.dev, People Data Labs), legitimate-interest basis for public job-data aggregation, Lithuania governing law, support@bordfeed.com as the single contact. Tables avoided in favor of CommonMark lists since the legalPages content-collection doesn't load remark-gfm.

Analytics provider (app/layout.tsx)

<Analytics /> + <SpeedInsights /> replaced with <UmamiAnalytics />.

@vercel/analytics + @vercel/speed-insights (package.json)

no longer imported anywhere; Umami replaces them.

Enterprise plan (config/billing.config.ts)

superseded by the priced Max/Ultra tiers. The isEnterprise handling in the pricing components is now inert (left in place for a possible future custom tier).

bun install: clean (2 packages removed)

bunx biome check on changed files: clean

bun run build: exit 0

content-collections generated 19 docs (incl. the 3 legal pages), TypeScript check passed, all routes built

bun run test: 616/616 passing

bun run gen:openapi: spec unchanged

Live browser fetch of the source policies (marketful.com/privacy, bestwriting.com/terms) + standalone compileMDX of all three legal pages

fumadocs-mdx 14.3.2 → 15.0.0

v15 drops the CommonJS loader-mdx/loader-meta shims; webpack loader paths are now fumadocs-mdx/webpack/{mdx,meta} (ESM-only). load-from-file-*.mjs uses import('file://…?hash=…') semantics that don't work when Next.js transpiles next.config.ts to a CommonJS next.config.compiled.js. Renaming to next.config.mts makes Next preserve the ESM loader context so v15's loader resolution succeeds.

lint-staged 16.4.0 → 17.0.2

drops Node 20 (floor is 22.22.1, publish-time-only — engines.node left at 22.21.1 since lint-staged runs in pre-commit hooks, not on Vercel). Optional yaml dep removed (we use .lintstagedrc.js). gitWorkflow.js refactored around Git 2.32+ git update-index --again. End-to-end tested via this PR's own commit hook.

next.config.ts → next.config.mts

required by fumadocs-mdx 15. Three knock-on fixes:

bun install: 24 packages installed clean; postinstall regenerated .source/* against fumadocs-mdx@15

bun run lint: clean

bun run typecheck: exit 0

bun run test: 616/616 passing

bun run gen:openapi:check: spec up to date

bun run build: 65/65 static pages, no warnings (one Node MODULE_TYPELESS_PACKAGE_JSON reparse note for lib/env.ts — performance-only, not an error)

Live browser smoke against /docs (fumadocs MDX rendering with Quick Start content, sidebar nav, TOC, search, themed callouts) and /docs/api (Scalar with v3.0.14 spec, sidebar nav, auth panel, client libraries)

both clean, zero console errors

react-day-picker 9.14.0 → 10.0.0

v10 removes the DeprecatedUI<T> typing fallback and a batch of v9-deprecated public APIs. Ships smaller install (transitive date-fns-jalali and @tabby_ai/hijri-converter extracted to optional @daypicker/* packages we don't use). Peer ranges and date-fns ^4.1.0 dep unchanged. CSS hooks (.rdp-button_next, data-selected, data-range-*, data-focused) and getDefaultClassNames() key set unchanged.

components/ui/custom/date-picker.tsx:130

initialFocus → autoFocus on the DateRangePicker <Calendar> (the only one of the v10-removed prop names in use across the codebase).

components/ui/calendar.tsx:93

renamed table className key to month_grid. The table key was a v9 no-op aliased through DeprecatedUI (its JSDoc said "renamed to UI.MonthGrid"); v10's strict Partial<ClassNames> typing surfaces it as a TS error. Now actually styles the slot. Pre-existing latent bug — v9 was silently dropping the w-full border-collapse styles on the table wrapper.

bun run lint clean, bun run typecheck exit 0

616/616 unit tests passing

Production build: 12.2s compile, 65/65 static pages, no warnings

Live browser smoke against all three <Calendar> shapes used in the codebase (single mode like DatePicker, range mode like DateRangePicker with autoFocus, single mode with disabled predicate like ban-user-modal): chevron nav, today highlight, range start/middle/end ARIA + visual states, and disabled-predicate enforcement (past dates correctly blocked, future dates selectable) all behave identically to v9. Zero console errors.

next 16.2.4 → 16.2.6

closes a batch of advisories that map directly to surfaces this app uses: GHSA-26hh-7cqf-hhc6 (segment-prefetch middleware bypass — 16.2.5's fix was incomplete, 16.2.6 finishes it), GHSA-h64f-5h5j-jqjh (Image Optimization DoS — relevant to next/image with dangerouslyAllowSVG: true in next.config.ts), GHSA-wfc6-r584-vfw7 (RSC cache poisoning), GHSA-mg66-mrh9-m8jx (Cache Components DoS), plus GHSA-267c-6grr-h53f, GHSA-492v-c6pp-mqqv, GHSA-c4j6-fc7j-m34r, GHSA-8h8q-6873-q5fj and four moderate-severity advisories. Repo uses proxy.ts (Next 16's renamed middleware) and App Router throughout, so directly affected.

resend 6.12.2 → 6.12.3

bumps transitive svix 1.90.0 → 1.92.2 to silence GHSA-w5hq-g745-h8pq. Relevant to app/api/webhooks/resend/route.ts which calls Resend.webhooks.verify(). svix also drops its uuid dep (small bundle win).

react + react-dom 19.2.5 → 19.2.6

RSC type hardening + perf (#36425). No API changes.

AI SDK trio

ai 6.0.174 → 6.0.176, @ai-sdk/openai ^3.0.58 → ^3.0.63, @ai-sdk/react 3.0.176 → 3.0.178. ai@6.0.176 adds opt-in toolMetadata field; existing streamText/generateText/generateObject callers in app/api/ai/chat/route.ts, lib/feeds/digest-intro.ts, lib/pipeline/job-status.ts, and workers/enrichment/src/enrichment.ts unaffected. Remaining patches cascade through @ai-sdk/provider-utils@4.0.27 and @ai-sdk/gateway@3.0.111.

@sentry/nextjs 10.51.0 → 10.52.0

browser log envelopes now gate IP/UA inference behind sendDefaultPii (already false at instrumentation-client.ts:23, so this is a privacy fix in our favor). Internal Vercel-AI event processor migrated to span streaming. Transitive @sentry/node drops @opentelemetry/instrumentation-{ioredis,redis} (-3 packages; Postgres-only project, no tracing regression).

AWS SDK

@aws-sdk/client-s3 and @aws-sdk/s3-request-presigner 3.1041.0 → 3.1045.0. Daily snapshots (VPC Lattice DNS, Bedrock AgentCore, MediaTailor, GuardDuty); only S3-specific change is outpost validation in 3.1044, irrelevant to R2/MinIO endpoint in lib/storage/s3.ts.

@scalar/api-reference-react 0.9.32 → 0.9.34

bundled @scalar/api-reference engine ships tag-label overflow fix and command-palette inline errors. No API surface change for /docs/api.

Fumadocs

fumadocs-core + fumadocs-ui 16.8.7 → 16.8.8 — adds copy-ID button on headings and Linux ctrl shortcut hint in search. fumadocs-mdx held at 14.3.2 (v15 requires renaming next.config.ts → next.config.mts).

stripe 22.1.0 → 22.1.1

restores multipart for /v1/files (fixes a v22 regression we don't hit) and exports Stripe.Decimal type.

react-email 6.0.5 → 6.1.1

CLI dev tool only (bun email:dev); production rendering uses @react-email/render@2.0.8 pinned separately. Tailwind compilation tweaks (media-query un-nesting, non-inline config vars, clearer <Head> outside <Tailwind> error).

@marsidev/react-turnstile 1.5.1 → 1.5.2

release notes literally say "no significant changes".

@types/node 25.6.0 → 25.6.2

types-only refinements (http2 sessionError, webstreams TransformStream cancel callback). engines.node: 22.21.1 runtime unchanged.

@next/bundle-analyzer 16.2.4 → 16.2.6

locked to next version per convention; build-time-only wrapper gated behind ANALYZE=true.

postcss 8.5.13 → 8.5.14

custom-syntax regression fix. We use @tailwindcss/postcss, so no impact.

Homepage hero pill leaks bold markdown

lib/changelog.ts parser rewritten around a single splitOnSeparator walker that finds the first em-dash (—), en-dash (–), or - *outside* parens, and cleanMarkdown now strips bold and ` code consistently for both titles and descriptions. Fixes three real-world regressions on v3.0.11 cards: zod 4.3.6 → 4.4.3 — … (now title zod 4.3.6 → 4.4.3), @sentry/nextjs 10.50.0 → 10.51.0 — … (now @sentry/nextjs 10.50.0 → 10.51.0), and react-email 6.0.4 → 6.0.5 (CLI dev tool only — production rendering uses …)` (no longer mis-split inside the parenthetical aside; rendered as a single title-only card).

6 regression tests added in tests/changelog.test.ts covering bold-prefix-with-trailing-text, bold+code prefix, bold inside descriptions, em-dash inside parens, multi-bold lines without a top-level dash, and hyphenated package names like lucide-react. Suite is 23/23 green; full project test run 616/616.

AI SDK trio

ai 6.0.168 → 6.0.174, @ai-sdk/openai 3.0.53 → 3.0.58, @ai-sdk/react 3.0.170 → 3.0.176. ai@6.0.171 ships an MCP prototype-pollution fix; ai@6.0.170 adds allowSystemInMessages (we use the dedicated system: parameter in lib/feeds/digest-intro.ts, so no warning fires). No call-site changes required.

zod 4.3.6 → 4.4.3

only minor in window is 4.4.0 (.merge() with refinements throws, z.url() rejects malformed forms, base64 rejects whitespace, JSON Schema strips redundant id). No call sites trigger the new tightenings; OpenAPI spec output is byte-identical (gen:openapi --check passes).

AWS SDK

@aws-sdk/client-s3 and @aws-sdk/s3-request-presigner 3.1038.0 → 3.1041.0. Daily releases targeting unrelated services; no S3 client or presigner changes.

@sentry/nextjs 10.50.0 → 10.51.0

internal "span-first APIs" migration for edge/server processors, span streaming filter, Turbopack detection attribute. No withSentryConfig or instrumentation hook changes.

@tanstack/react-query 5.100.6 → 5.100.9

devtools-only patches (Angular theme option, onClose type fix, NoInfer restore on persister TQueryKey). No runtime semantics changed.

@scalar/api-reference-react 0.9.29 → 0.9.32

bundled engine moves to @scalar/api-reference@1.55.1, deepObject query-param display fix, internal slugger refactor.

Fumadocs

fumadocs-core/fumadocs-ui 16.8.5 → 16.8.7. TOC hotfix in fumadocs-ui and Clerk-style RTL layout fix in 16.8.6.

@stripe/stripe-js 9.3.1 → 9.4.0

adds ReleaseTrain type and hashedValue overload to handleNextAction. (No source imports — server-side stripe@22.1.0 handles all flows; checkout uses Stripe-hosted redirect.)

react-hook-form 7.74.0 → 7.75.0

dirtyFields now prunes empty-false nodes, useWatch no longer re-renders on unrelated field validation, TypeScript 6.0 support added. We only read formState.errors and formState.isValid, so the dirtyFields shape change is inert.

lucide-react ^1.12.0 → ^1.14.0

additive icons only (waves-vertical, repeat-off). No renames or removals across our 90 import sites.

react-email 6.0.4 → 6.0.5 (CLI dev tool only — production rendering uses @react-email/render@2.0.8 pinned separately).

nanoid 5.1.9 → 5.1.11

fixes a regression where requesting a very large ID broke generation.

react-qr-code 2.0.18 → 2.0.21, ufo 1.6.3 → 1.6.4 (withoutBase() slash-collapse fix), use-stick-to-bottom 1.1.3 → 1.1.4.

biome ^2.4.13 → ^2.4.14

adds three new lint rules (useTestHooksOnTop, noReactStringRefs, useMathMinMax); none auto-fire on this codebase.

postcss 8.5.12 → 8.5.13

postcss-scss comment regression fix (we don't use SCSS, no impact).

AWS SDK @aws-sdk/client-s3 and @aws-sdk/s3-request-presigner 3.1033.0 → 3.1038.0 (additive client updates, retry-on-200 fix).

Stripe Node 22.0.2 → 22.1.0

pinned API version moves to 2026-04-22.dahlia. New enum values across BalanceTransaction.type, PaymentMethod.type, tax-id types, and invoice/subscription payment_method_types (now includes pix/upi). Adds parseEventNotificationAsync and async stack-trace preservation.

@stripe/stripe-js 9.2.0 → 9.3.1

adds contactDetails element and PMC update support.

@sentry/nextjs 10.49.0 → 10.50.0

additive integrations (Effect v4 beta, Hono Bun) and LCP/console fixes.

tRPC 11.16.0 → 11.17.0 across @trpc/{client,server,react-query}

fixes React 19 proxy coercion in createInnerProxy and adds subscription inference helpers.

@tanstack/react-query 5.99.2 → 5.100.6

adds retryOnMount callback, fixes hydration for resolved promises and infinite-query AbortSignal propagation.

better-auth 1.6.5 → 1.6.9

preserves response headers when APIError is thrown, fixes Partitioned cookie forwarding, organization team field inference, and OAuth profile email fallback.

Fumadocs fumadocs-core/fumadocs-ui 16.8.1 → 16.8.5, fumadocs-mdx 14.3.1 → 14.3.2

Shiki lazy-mode language fix, locale page-tree fix, schema type narrowing.

@scalar/api-reference-react 0.9.24 → 0.9.29

bundled @scalar/api-reference engine moves 1.49 → 1.53 (kepler theme position:fixed flare fix, several SSR/hydration fixes, lazy rendering, pre/post request scripts).

react-hook-form 7.73.1 → 7.74.0

adds setValues API; fixes nested-field unregister and valueAsNumber NaN coercion.

react-email 5.2.10 → 6.0.4 (CLI dev tool only — @react-email/components and @react-email/render imports continue to work).

@react-email/render 2.0.7 → 2.0.8

strips nul bytes from renderToPipeableStream output to prevent multi-byte char truncation in feed-digest emails.

lucide-react ^1.8.0 → ^1.12.0

new icons, ESM bundles switch to .mjs.

@marsidev/react-turnstile 1.5.0 → 1.5.1

fixes params being passed to turnstile.execute().

biome ^2.4.12 → ^2.4.13

noUselessEscapeInRegex v-flag fix and organizeImports declare-module sort fix.

postcss 8.5.10 → 8.5.12

fixes "reading any file via user-generated CSS" and nested-bracket parsing performance.

Lower-frequency worker schedules

job enrichment and company resolution now run hourly in low-cost MVP mode, and embedding generation runs every 6 hours instead of every 5 minutes.

Lower-frequency Vercel cron schedules

monitor checks now run twice daily, push/email stay hourly, and TTL expiry moves to a midnight UTC daily slot.

Worker documentation synced

README and agent docs now reflect the low-cost schedules for the active pipelines.

Vercel Cron route compatibility

internal cron routes now expose GET handlers in addition to POST so Vercel Cron can invoke them directly while keeping CRON_SECRET auth.

Nested source company names

company resolution now safely extracts names from raw company objects such as { company: { name: "Atari" } }, including slug-conflict recovery.

Greenhouse ATS domain detection

job-boards.greenhouse.io is now treated as an ATS host instead of a company domain in both app and worker normalizers.

Company enrichment test fixtures

stale enrichment tests now account for the current daily-limit query shape.

Zod-driven OpenAPI generator

scripts/generate-openapi.ts builds public/api/openapi.json from Zod schemas via zod-openapi@5.4.6. Wired into prebuild (Vercel regenerates on every deploy) and lint-staged (auto-regen + restage on schema/route edits).

Drift-detection tests (tests/openapi/)

13 Vitest unit tests cover: spec is valid OpenAPI 3.1, all 18 operationId values present, route↔spec parity (walks app/api/v1/, excluding cron), deterministic output, magic-string + limit validation regression coverage.

Documented query params

GET /search/suggest, /feeds/{id}/jobs, /feeds/{id}/export, /companies/export, /companies/{id}/jobs had inline private query parsing. Now centralised in schemas/api-v1-queries.ts and surfaced in the spec.

Documented filter fields

salary_unit, source_name, status, processing_status, tech_stack were already validated by feedFiltersSchema but absent from the published spec. The generated spec now reflects what the API actually accepts.

?since=last_poll regression caught pre-ship

earlier in this branch, switching since to z.string().datetime() would have broken the magic string used by /feeds/{id}/jobs and /feeds/{id}/export to fetch jobs since the last poll. Schema is now z.union([z.literal("last_poll"), z.string().datetime()]). Regression tests in tests/openapi/queries.test.ts.

/api/openapi.json 500 in dev mode

Next.js was complaining about a conflicting public file and route handler at the same path. Removed the redundant CORS-header route handler; Scalar fetches same-origin so CORS was never needed. The static file at public/api/openapi.json now serves directly with sane default cache headers.

/api/v1/templates envelope consistency

list response now includes request_id like every other v1 endpoint.

Spec version drift

info.version is now read from package.json, killing the manual sync step that produced the v2.4.0/v3.0.7 mismatch.

**Feed response shape

snake_case + secret stripped — POST /feeds, GET /feeds, GET /feeds/{id}, PATCH /feeds/{id} previously returned raw Drizzle rows: camelCase fields (deliveryMode, deliveryFrequency, createdAt, updatedAt, lastDeliveryAt, lastPollAt) and the webhook signing secret. Now they return snake_case via the new formatFeedForApi() shaper, consistent with jobs/companies. Breaking** for any consumer reading data.deliveryMode etc.; switch to data.delivery_mode. Public response fields are now: id, name, filters, delivery_mode, delivery_frequency, url, enabled, template_id, last_delivery_at, last_poll_at, created_at, updated_at. Fields explicitly stripped from responses: secret (security — write-only), organizationId (internal multi-tenant key), events, failureCount, emailSchedule, emailFormat, emailRecipient, lastEmailAt (internal state and undocumented features).

GET /feeds/{id} stats.last_delivery_at removed

same value is now at the top-level data.last_delivery_at. The stats object only carries deliveries_sent.

?limit= validation

was permissive (clamped out-of-range to bounds, fell back to default on NaN, truncated decimals). Now strict: out-of-range, non-integer, or empty limit returns 422 with a clear error instead of silently coercing.

?since= validation

was permissive (any string new Date() could parse). Now strict ISO 8601 datetime, with the literal "last_poll" accepted as a magic value. Date-only strings like 2024-01-01 return 422.

Error envelope schema

published spec now correctly nests errors and retry_after inside error.{} (matches lib/api/v1/errors.ts runtime shape).

Status code for invalid format on /feeds/{id}/export

was 400 bad_request, now 422 validation_error. Consistent with every other v1 validation error.

Spec serving

public/api/openapi.json is now a generated artifact, not hand-edited. Editing it directly fails CI via the gen:openapi:check lint hook. Regenerate with bun run gen:openapi.

Dependency updates (#171, #172)

Biome lint warnings (#171)

low-risk refactors (unused vars/params, optional chaining, safe type casts) to bring the repo back to a clean lint state.

Repository doc drift (#173)

command docs now distinguish headless Playwright runs from the Playwright UI, cron route docs match the actual /api/v1/... POST endpoints, the embedding worker schedule notes the real 5-minute cron + daily token budget, and the granular email scheduling design doc is linked from the main repo entry points.

22-package dependency patch sweep (#168)

better-auth 1.6.1→1.6.3, @tanstack/react-query ^5.96.2→^5.99.0, @sentry/nextjs ^10.47.0→^10.48.0, ai ^6.0.154→^6.0.161, @ai-sdk/openai ^3.0.52→^3.0.53, @ai-sdk/react ^3.0.156→^3.0.163, @stripe/stripe-js 9.1.0→9.2.0, resend 6.10.0→6.11.0, @react-email/components 1.0.11→1.0.12, @react-email/render 2.0.5→2.0.6, @scalar/api-reference-react ^0.9.20→^0.9.22, @aws-sdk/client-s3 + @aws-sdk/s3-request-presigner ^3.1027.0→^3.1030.0, fumadocs-core / fumadocs-ui 16.7.11→16.7.16, fumadocs-mdx 14.2.11→14.2.14, @fumadocs/content-collections 1.2.7→1.2.9, @biomejs/biome ^2.4.10→^2.4.12, @types/node 25.5.2→25.6.0, autoprefixer 10.4.27→10.5.0, dotenv 17.4.1→17.4.2. Every bump audited against our call sites before merging (see PR #168 body). 554 unit tests still pass.

16-package pre-release dep bump (unreleased, commit d15b038)

next 16.2.2→16.2.3 (security fix), react / react-dom 19.2.4→19.2.5, better-auth 1.6.0→1.6.1, stripe 22.0.0→22.0.1, ai ^6.0.149→^6.0.154, @ai-sdk/openai ^3.0.51→^3.0.52, @ai-sdk/react ^3.0.151→^3.0.156, @aws-sdk/client-s3 + presigner ^3.1025.0→^3.1027.0, fumadocs-core / fumadocs-ui 16.7.10→16.7.11, @next/bundle-analyzer 16.2.2→16.2.3, @testcontainers/postgresql + testcontainers 11.13.0→11.14.0, postcss 8.5.8→8.5.9.

Lucide + Vitest (unreleased, commit a82a434)

lucide-react 1.7.0→^1.8.0, vitest 4.1.3→^4.1.4.

packageManager

bun@1.3.11 → bun@1.3.12.

Documentation drift

README.md now reflects 34 DB tables (was 33), stale LICENSE reference removed (file deleted in 36d4003). CLAUDE.md corrects API-key attribution — we use a custom resolver in lib/feeds/api-keys.ts, NOT the Better Auth apiKey plugin (which is not registered in lib/auth/index.ts) (#169). CLAUDE.md, ROADMAP.md, and the historical v2.7.5 CHANGELOG.md entry corrected from "19 retryable patterns" to "18" (matches code: 18 regex patterns in lib/pipeline/enrichment-retry.ts). ROADMAP.md test count updated to 554 across 40 files (was 508 across 36).

CI reproducibility (#168)

.github/workflows/deploy-enrichment-worker.yml now pins bun-version: 1.3.12 on oven-sh/setup-bun to match packageManager, preventing drift between local and CI Bun versions.

cropperjs 1.6.2

v2 is a Web Components rewrite. react-cropper@2.3.3 is unmaintained (last commit Sep 2023) and still pins cropperjs ^1.5.13. The CSS path, getCroppedCanvas(), and ref pattern we use are gone in v2. Upstream issue requesting v2 support has no activity. Tracked in TODOS as P2.

drizzle-orm 0.45.2 / drizzle-kit 0.31.10

these ARE the latest stable on the 0.x line. 1.0.0-beta introduces a migration-system overhaul; plan a dedicated PR when 1.0 GAs.

stripe 22.0.1

22.1.0-beta is public-preview of an unreleased Stripe API version. We just migrated to v22 in v3.0.2.

ai / @ai-sdk/* majors

v7 beta / v4 beta pre-stable, ESM-only. Wait for GA.

react-email (CLI) 5.2.10

v6 is a major restructuring (consolidates per-component packages) and needs a dedicated migration PR.

next-themes 0.4.6

Bun patch retained until upstream pacocoursey/next-themes#386 merges.

554 tests across 40 files

34 user tables (lib/db/schema/tables.ts + __drizzle_migrations)

15 active scrapers (+1 archived: JobDataAPI)

18 OpenAPI operations across 15 paths

30 feed filter fields (Zod feedFiltersSchema)

18 retryable enrichment error patterns

5-level job monitoring pipeline

L0 (time expiry), L1 (HEAD status), L1.5 (bot/CAPTCHA detection), L2 (keyword scan), L3 (AI classifier). Easy cases caught for free, AI only for ambiguous pages (~10% of checks)

L1.5 bot detector

detects Cloudflare challenges, reCAPTCHA/hCaptcha, 403 forbidden, JS-only shells, empty pages. Flags blocked jobs for future L4 (Apify rendering)

L2 keyword scanner

30+ regex patterns for closed/filled/expired job pages plus active confirmation patterns. Catches ~55% of status changes at zero cost

Fixed L3 AI classifier

replaced broken dynamic import("zod") with static Zod schema, limited input to 2K chars (was 50K), added structured output with reason field

Adaptive backoff

6h for fresh jobs (0-3 days), 12h for settled (3-14 days), 24h for stale (14-30 days), 72h for very old (30+ days). Max 8 attempts

Structured monitoring detail

monitorLevel and monitorDetail columns on jobTable track which level caught each job and what keyword matched

monitor_log table

schema created for per-check audit trail (writes not yet wired in the pipeline)

Monitor API richer stats

blocked count and levelBreakdown in response

Monitor pipeline rewrite

monitor-jobs.ts now runs L0→L1→L1.5→L2→L3 sequentially, short-circuiting at the first conclusive level. ~90% of checks resolved without AI

Monitor route

logs and returns blocked + levelBreakdown for observability

Broken Zod import in AI classifier

import("zod") dynamic import never worked; replaced with static z.object() schema

L3 AI classifier

now strips HTML tags before sending to AI (was sending raw HTML, would mostly be <head> metadata)

L2 keyword patterns

bounded greedy .* to .{0,80} to prevent false positives matching across long pages

L1.5 bot detector

cached regex matches to avoid double scans; gated empty-page check behind content-type to avoid false positives on JSON/binary responses

Monitor pipeline

handle 405/501 HEAD rejection by proceeding to GET; wrap URL parsing in try/catch; map private_host/invalid_url to monitorStatus: "no_url"; use inArray instead of raw SQL

Apify schedule two-way sync

toggling a source on/off in the admin UI now enables/disables the corresponding Apify schedule via their API. Previously the toggle only updated the database, leaving actors running on Apify's cron

Webhook rejection for inactive sources

webhooks from paused or archived sources are now rejected at the handler level (HTTP 200 with source_inactive status). No more silent data ingestion from disabled sources

Cache invalidation for non-active sources

the webhook handler's source cache no longer stores paused/archived sources, so pause takes effect immediately instead of after 60s TTL

Source resolution

resolveSource() now returns both name and status (was name-only). Queries all sources regardless of status (was filtered to active-only)

Admin jobs: 4 new filter dropdowns

Salary Unit, Source, Status, Processing. Source/Status/Processing are admin-only. Full DB/API/UI filter parity (30 feed filter fields)

Clickable stat cards

Total/Active/Closed/Draft cards on admin jobs page now filter the table by status. Active state with highlighted border

Job detail: 8 new fields

status badge, salary unit, first scraped date, last updated date, source type, external job ID, processing status, processed date, tags (jsonb). Full DB/UI parity (46 fields in detail select)

Admin jobs default view

now shows only enriched jobs (processing_status=completed) instead of all 52K jobs. Hides "Untitled" draft/failed jobs from default browse. Explicit filter dropdowns or stat card clicks override the default

Feed filter schema

status and processing_status use Zod enums matching Postgres enums (type-safe at API boundary)

Fuzzy search fallback

now respects the same status/enrichment filter as the main query (was previously unfiltered)

Hydration error

CompanyLogo fallback changed from <div> to <span> to fix React hydration warning when rendered inside SheetDescription's <p> tag

Stripe SDK v22

upgraded from v21.0.1. Migrated type annotations in checkout and subscription billing code (SessionCreateParams → Parameters utility type, UpcomingInvoice → Invoice, StripeError → InstanceType). Pinned to exact 22.0.0 to prevent silent API version drift

16 dependency updates

AI SDK (ai 6.0.145, @ai-sdk/openai 3.0.50, @ai-sdk/react 3.0.147), AWS SDK 3.1023.0, Scalar API docs 0.9.19 (fixes production crash), react-hook-form 7.72.1 (fixes dirty state bugs), fumadocs 16.7.10, TanStack React Query 5.96.2, slugify 1.6.9, Playwright 1.59.1, TypeScript types, dotenv 17.4.0

company_ats table (#154)

DB-driven ATS platform connections. Links companies to ATS slugs (1:many, nullable companyId). Unique constraint on (platform, slug). Status tracking (active/empty/dead), discoveredVia provenance, job count + last scraped timestamps

Internal scraper API (#154)

3 CRON_SECRET-protected endpoints replacing per-scraper Apify KV stores: GET /api/internal/scraper/slugs (read by platform/status), POST /api/internal/scraper/slugs/report (upsert scrape results), POST /api/internal/scraper/slugs/discover (submit new slugs)

Admin ATS Connections card (#154)

inline add/delete UI on company detail page with platform-specific board URL generation (8 ATS platforms), status badges, external links, job count display

ATS connections in public API (#154)

ats_connections array in GET /v1/companies/:id response

Greenhouse slug seed script (#154)

scripts/seed-greenhouse-slugs.ts imports 7,101 pre-discovered slugs from Apify KV store into company_ats

Greenhouse scraper (#154)

migrated from Apify KV store to DB-driven slug management via internal API. Reads slugs from API, reports results back, discovers new slugs via Common Crawl. Seed slugs merged into current run's poll list for immediate scraping

Greenhouse actor input schema (#156)

added api_base and cron_secret fields to Apify input schema. cron_secret is required and marked as secret. Apify schedule updated with prod credentials

Admin company router (#154)

createAtsConnection uses upsert (onConflictDoUpdate on platform+slug) to handle linking existing seed slugs to companies

Admin sources page (#157)

replaced static status Badge with interactive Switch toggle (optimistic UI). Removes redundant pause/play action button. Both desktop and mobile layouts updated

DB schema

33 user tables (up from 32). New company_ats table with indexes on (platform, status) and companyId

v3 pipeline architecture

scrapers do transport only, AI extracts all fields. Design doc: docs/designs/v3-architecture.md

externalId-based dedup

{source_name}_{source_job_id} replaces content hash + fuzzy title/company matching. Exists → UPDATE (reset processingStatus if raw data changed), new → INSERT

Source-aware enrichment prompts

buildExtractionPrompt(rawJson, sourceName) tells GPT-4o-mini what ATS format it's looking at. Raw JSON >50K chars gets truncated with title+description fallback

Set reconciliation closure

lib/pipeline/closure.ts diffs source_job_id sets between scrapes. 2-miss grace period via consecutiveMisses column. Gated on webhook_log.status = 'completed'

90-day TTL expiry cron

app/api/v1/cron/expire-jobs/ marks jobs with no scrape activity >90 days as expired

Job status module

lib/pipeline/job-status.ts extracted from deleted lib/enrichment/

Scraper v3 contract tests

106 static analysis tests validating all 15 scrapers output {source_job_id, source_name, url, raw}

New tests

ingest dedup (21), webhook validation (20), enrichment prompts (13), enrichment pipeline (15), process-jobs (37), closure (12), scraper contract (106)

All 15 scrapers migrated to v3 format

output {source_job_id, source_name, url, raw} only. ~3,000 lines of convertToRawJob() field mapping deleted

Enrichment worker

v3 path reads rawSourcedJobData + sourceName from claimed jobs. Legacy fallback for pre-v3 jobs using description/title string

Webhook handler

validates source_job_id presence, rejects items without it, tracks rejected count in webhook log. Runs closure reconciliation on completed scrape runs

Jobvite re-enabled

now v3 format with all other scrapers (was disabled at 24 jobs)

Test count

508 tests across 36 files (up from 379 across 34)

DB schema

consecutiveMisses integer column on job table, closed + expired added to JobStatus enum

Safe dependency refresh

updated Next.js to 16.2.2, @next/bundle-analyzer to 16.2.2, Playwright to 1.59.0, Fumadocs to 16.7.9, TanStack React Query to 5.96.1, and the React Email patch trio (@react-email/components, @react-email/render, @react-email/tailwind)

Documentation sync

aligned README, quickstart, AI/chat docs, source inventory, and enrichment docs with the current v2.7.5 codebase and routing structure

lib/enrichment/

entire directory deleted (extract.ts, prompt.ts, schema.ts, index.ts). CF Worker is sole enrichment path

/api/internal/ingest

deleted, webhook handler is sole ingest entry point

/api/internal/process

deleted, CF Worker is sole enrichment path (no Vercel fallback)

lib/pipeline/process-jobs.ts

Vercel-side enrichment fallback deleted

Content hash dedup

replaced by externalId-based upsert

Fuzzy title+company matching

replaced by externalId-based upsert

Per-scraper field mapping

~3,000 lines of convertToRawJob(), convertListingToRawJob(), convertDetailToRawJob() functions across all 15 scrapers

Auth E2E smoke spec

updated Playwright auth checks to match current auth metadata (Create an account) and the actual card-title DOM structure used by the auth UI

Oracle HCM Cloud job source (#149)

reverse-engineered Oracle Recruiting Cloud REST API (/hcmRestApi/resources/latest/recruitingCEJobRequisitions). No auth required. Multi-tenant ATS architecture with KV state persistence across 11 FA pod variants. Dual discovery: Common Crawl (20 indexes × 11 pods) + Wayback Machine CDX fallback. DNS wildcard pollution filtering for Wayback results (4-char slug validation). 535 active companies discovered, 32K+ jobs in pool. Input-sanitized siteNumber via VALID_SITE_RE. Actor 0oejH2uT5fnHgeUtc, 12h schedule (6am/6pm UTC), webhook 1Xwc4iQIAO3yDTuCd

Enrichment retry system (#150)

new lib/pipeline/enrichment-retry.ts with classifyEnrichmentFailure() (18 retryable error patterns: 500, 429, quota, timeout, network errors), buildFailureMetadata() (exponential backoff: 15min → 1h → 6h → 24h, max 7 retries), getRetryCountFromMetadata(). Both lib/pipeline/process-jobs.ts and workers/enrichment/src/process-jobs.ts now requeue retryable failed jobs before each batch, prioritize retry jobs in claim queries

Writing category expansion (#150)

lib/feeds/filter-builder.ts now expands writing/content category filters across 4 fields (title, occupational_category, department, skills) using word-boundary regex matching. New constants: WRITING_CATEGORY_TOKENS, WRITING_TITLE_SIGNALS (13 terms), WRITING_OCCUPATIONAL_CATEGORY_SIGNALS (16 terms), WRITING_DEPARTMENT_SIGNALS, WRITING_SKILL_SIGNALS

Writing-aware enrichment prompts (#150)

both lib/enrichment/prompt.ts and workers/enrichment/src/prompt.ts updated to prefer specific writing-family categories (Technical Writing, Copywriting, Editorial, Communications, SEO, Medical Writing, Journalism) over generic "Marketing" or null

Writing re-enrichment script (#150)

scripts/requeue-writing-reenrichment.ts targets active writing jobs with blank/generic occupational_category and requeues them for enrichment with updated prompts. Supports --dry-run and --limit flags

New tests (#150)

tests/feeds/category-matching.test.ts (5 tests for writing category expansion), tests/pipeline/enrichment-retry.test.ts (2 tests for failure classification + retry metadata), updated tests/pipeline/process-jobs.test.ts (retry requeue + priority ordering), updated workers/enrichment/src/process-jobs.test.ts (152 additions for worker retry logic)

Multi-Tenant ATS Pattern in skill docs

148-line section added to .claude/skills/new-job-source/SKILL.md covering architecture, KV state, 3-source discovery (CC + Wayback + robots.txt), time budgets, unknown classification caps, scrape modes, 15K ingest cap

Scraper count

14 active sources (up from 13), 15 total actors in apify-scrapers/

Active sources in DB

oracle added with actor ID 0oejH2uT5fnHgeUtc

Test count

379 tests across 34 files (up from 370 across 32)

Enrichment error handling

both pipelines now use buildFailureMetadata() + classifyEnrichmentFailure() instead of hardcoded { lastError, failedAt } objects. Failed jobs get retryable, retryCount, nextRetryAt, lastErrorType metadata

Job claim priority

retry-due jobs are prioritized ahead of content-role priority and chronological ordering in both lib and worker pipelines

Biome lint cleanup

removed 3 biome-ignore comments from filter-builder.ts

JazzHR job source

HTML scraping of {slug}.applytojob.com career pages + JSON-LD JobPosting extraction from detail pages. 2,696 seed company slugs discovered via crt.sh, Brave, Google, Wayback Machine, and Common Crawl. Self-healing pattern: classifies boards as active/empty/dead, rotates dead sample daily. Seed-only fast path for quick test runs (3s vs 10+ min for full store). Actor L8IXtCnjtPNUzKy7w, 12h schedule, webhook wired to ingest pipeline. JazzHR (Employ Inc) powers SMB hiring — complements Lever (mid-market) and Jobvite (enterprise, disabled)

Scraper count

13 active sources (up from 12), 14 total actors in apify-scrapers/. 109K+ active jobs

Active sources in DB

jazzhr added with actor ID L8IXtCnjtPNUzKy7w

Concentrix job source (#133)

reverse-engineered WordPress REST API (Talkpush JDQ plugin at jdq/v1/search). ~1,933 jobs, no auth, paginated. Apply URLs → Workday. Type normalization (full_time → "Full Time", permanent → "Full Time"). Actor MAfceoYD9uN48sR0P, 12h schedule

Teleperformance job source (#134)

reverse-engineered Umbraco CMS REST API (/Umbraco/Api/Careers/GetCareersBase). ~2,192 jobs, no auth, paginated. Apply URLs → iCIMS via short.sg. Actor xnUX4LT7uyMJgyhgi, 12h schedule

Scraper count

12 active sources (up from 10), 13 total actors in apify-scrapers/

Feed filter security hardening

escapeRegex() prevents regex injection in category/title filters, escapeIlike() prevents wildcard injection in company/location ILIKE filters. Zod schema enforces ^[a-zA-Z0-9 /&,._-]+$ regex allowlist on category values. Client-side validation on filter editor with inline error messages

Feed filter UX improvements

lazy export (tab renders on first visit), notification form field persistence, deduplication of filter arrays, filter preview count before save, create-feed page validates filters with same Zod schema

Stored XSS prevention

company_size and company_funding_stage feed filters validated at save time (not just query time)

Scraper limits removed

all 10 active scrapers now default to Infinity (no artificial cap). Input schemas have no maximum constraint on limit field. Apify schedule inputs no longer pass a limit. The only real constraints are: date freshness (max_age_days/max_age_hours), content-hash dedup, and 15K webhook ingestion cap. Jobvite schedule disabled (dead source)

DynamiteJobs max_age_days

schedule input corrected from 1 to 7 days

FTS category matching

feed category filter changed from broken content_search @@ ... full-text approach to occupational_category ~* ... regex matching (10x more accurate for writing/content jobs)

Email digest cron

category filter in cron route updated to match tRPC regex approach (was still using old FTS path)

Invalid regex crash

invalid regex patterns in feed filters now caught with try/catch instead of crashing the query

CC+PDL+browser enrichment pipeline

replaces Apollo as primary company enrichment source. 4-layer approach: PDL free dataset (35M companies, industry/size/location/LinkedIn), tech detection from Common Crawl WARC records (4,036 Wappalyzer + CSP patterns), AI extraction via GPT-4o-mini (description/keywords/socials), agent-browser fallback for SPAs

Company discovery from Common Crawl

1,282 companies extracted from Ashby ATS pages in CC March 2026 crawl, 1,186 seeded to DB with 16 domain backfills

Tech stack detector (scripts/cc-extract/lib/tech-detector.ts)

7-layer detection: Wappalyzer rules (2,809 technologies) + CSP header parsing + cookies + preconnect/DNS-prefetch + inline script globals + external host detection + data attributes. Relevance filter drops noise (HTTP/3, HSTS, jQuery, embedded video)

10 enrichment scripts in scripts/cc-extract/

discover, seed, enrich-from-pdl, enrich-from-cc, backfill-pdl-state, fix-pdl-casing, tech detection tests, AI enrichment tests

Wappalyzer detection rules

3,931 technology definitions in data/wappalyzer/ (MIT license, Lissy93 fork)

Dependencies

updated 4 packages: AWS SDK 3.1020, @stripe/stripe-js 9.0.1, resend 6.10

CF Worker Phase 2 disabled

Apollo enrichment removed from cron pipeline. Worker still handles Phase 1 (resolve), Phase 3 (dedup), Phase 4 (stats). Enrichment now runs locally via scripts/cc-extract/

Admin enrichment dashboard

Apollo budget card replaced with CC+PDL provider info card. Removed unused Progress import

Admin enrichment router

updated worker health type to match new stats endpoint (removed dailyApiCalls/dailyLimit)

Shell injection in fetchLive

replaced execSync with execFileSync + domain regex validation (CC data is untrusted)

--limit arg parsing

was returning NaN when flag omitted, silently processing 0 companies

3,692 companies enriched via PDL (industry, size, founded, city, state, country, LinkedIn)

3,475 states backfilled, 5,943 values title-cased

209 companies enriched via CC+AI (tech stack, description, keywords, socials)

3,545 stale Apollo "failed" statuses corrected to "enriched"

Total cost: ~$0.05 in OpenAI credits

Dependencies

updated 8 packages: AWS SDK 3.1019, tRPC 11.16, Turnstile 1.5, Scalar 0.9.17, TypeScript 6.0.2

TypeScript 6.0 migration

removed deprecated baseUrl, added explicit types field, simplified lib (dom now includes dom.iterable)

Company logo component (components/companies/company-logo.tsx)

reusable CompanyLogo with Logo.dev CDN (img.logo.dev), retina support (2x), graceful fallback to building icon. Replaces inline logos across 5 views

Admin company CRUD

create/edit/delete companies via CompanyFormDialog (identity, location, financials, social, tech stack). Soft-delete via mergedIntoId pattern, duplicate detection on DB constraint

Company detail enhancements

enrichment snapshot diffs (before/after with field completeness), funding events table, identity/financials/meta admin cards, Facebook social link, multi-industry badges

ATS landscape research (ATS.md)

survey of 318 ATS platforms by job volume with P0/P1 gap analysis

Common Crawl extraction plan (CC-EXTRACT.md)

Web Data Commons extraction strategy (4.3M jobs across 63K domains)

Company discovery scripts (scripts/)

SourceStack extraction, domain backfill, discovered company seeding utilities

Workday parallel discovery

24 Common Crawl indexes queried in parallel (batches of 10) + robots.txt site enumeration (1,214 tenants probed in 37s, found 2,043 hidden sites). From ~1.7K to 4,337 known companies, 1,122 unique active boards, 254K total jobs

Workday time budget

graceful shutdown before actor timeout with periodic state saves, unknown company cap (200/run), priority-ordered poll list

SmartRecruiters concurrent fetching

batches of 5 parallel detail fetches with AbortSignal.timeout(15s), pagination guard (MAX_PAGES=50), named KV store for company persistence

PW concurrent fetching

3 parallel detail fetches with 500ms batch delay, AbortSignal.timeout, error-safe discovery via Promise.allSettled

Scraper KV persistence

all 5 stateful scrapers (SmartRecruiters, Workday, PW, Greenhouse, Lever) now use named Apify KV stores. Previously used per-run default stores, losing all classification state between runs

SmartRecruiters timeout

was timing out at 30min with sequential fetching, now completes in ~53s

Workday timeout

was rediscovering from scratch every run (broken KV), timing out at 5min. Now persists 4,337 companies and completes scrape runs in ~90s

PW timeout

was timing out at 10min with sequential 2s-delay fetching, now completes in ~201s

PW crash

unhandled AbortError from listing page fetch crashed the entire actor

Workday retry with backoff

transient errors (timeout, 429, 502-504) get 1 retry with exponential backoff. 5 consecutive failures auto-downgrade company to dead

PostgreSQL array filter

technology_names filter used invalid SQL, now uses proper ARRAY[]::text[] cast

Apollo enrichment

skip early if no domain (was returning 422 from name-only fallback)

CompanyLogo stale state

reset errored state when domain prop changes

Admin company domain

preprocess empty string to null (unique index conflict prevention)

Form NaN validation

annualRevenue/totalFunding validated with Number.isFinite before submit

Migration breakpoint

added missing statement-breakpoint marker on second ALTER

Database schema

annual_revenue and total_funding upgraded from integer to bigint (supports billion-scale values)

Job detail queries

include companyDomain via company table join for logo rendering

API response

adds enrichment_error, apollo_id, normalizedName to company responses

Removed .cursorrules (project uses CLAUDE.md for AI guidance)

Admin Enrichment Command Center

dedicated /dashboard/admin/enrichment page with stats cards (5 status counts), Apollo daily budget bar with color thresholds (sage/amber/red), worker health card, pipeline run history table, and merge audit log

Enrichment card on admin company detail

status badge, data completeness bar (22 fields), error display, Apollo ID, re-enrich button with 1-hour cooldown

Enrichment status column + filter on admin companies DataTable with DESIGN.md-aligned badge colors

Re-enrich All Non-Enriched bulk action

resets all failed+skipped companies to pending with confirmation dialog

Worker /stats endpoint

returns last run, daily API calls, daily limit, environment (CRON_SECRET auth)

Pipeline run log

enrichment_run_log table tracks every worker execution with phase-by-phase metrics, auto-prunes after 7 days

Enrichment snapshots

enrichment_snapshot table stores pre-re-enrich field values for diff view

enrichment_error column on company table

worker writes failure reason on error, clears on success and skip

EnrichmentStatusBadge component

shared badge with semantic colors per DESIGN.md (sage for enriched, warning for pending, info for processing, error for failed, neutral for skipped)

ENRICHMENT_FIELDS constant

22-field list used for completeness scoring and snapshot building

Domain/website_url preserved during enrichment

COALESCE prevents Apollo from NULLing Phase 1 resolve data when it doesn't return primary_domain, avoiding duplicate companies on next resolve cycle

Apollo budget metric

daily limit now counts all companies that entered processing today (not just 'enriched'), preventing over-provisioning

Apollo provider failure model

API errors now throw (map to 'failed' status) instead of returning null; only missing org data maps to 'skipped'

processing status added to EnrichmentStatus enum

was used by worker but missing from Drizzle schema

enrichmentStatus filter validated

constrained to z.enum of 5 valid statuses instead of accepting arbitrary strings

Progress bar fallback color

fixed for Tailwind v4 (var(--primary) not hsl(var(--primary)))

Re-enrich/diff guard merged companies

added mergedIntoId IS NULL filter to prevent operating on soft-deleted duplicates

Worker run logging

both cron and manual /process triggers now write pipeline results to enrichment_run_log

Worker health proxy

implemented as tRPC workerHealth procedure (admin-only, 5s timeout) instead of separate API route

18 dependencies updated

security patches (drizzle-orm SQL injection fix, vitest CVE), major bumps (stripe v21, @stripe/stripe-js v9), and routine patches across AI SDK, AWS SDK, tRPC, Sentry, lucide-react, recharts, fumadocs, Biome, Scalar

drizzle-orm 0.45.2

patches SQL injection in sql.identifier() / sql.as() (CWE-89), used in lib/feeds/facets.ts

vitest 4.1.2

resolves CVE in flatted dependency

CRON_SECRET required

changed from optional to required in env schema, preventing "Bearer undefined" auth bypass on cron/webhook endpoints when secret is unset

API key revocation org-scoped

revokeApiKey WHERE clause now includes organizationId, closing cross-tenant IDOR that could disable another org's keys

Worker auth deny-by-default

all 3 CF workers (enrichment, company-enrichment, embedding-generation) now return 401 when CRON_SECRET is missing instead of skipping auth entirely

GitHub Actions SHA-pinned

actions/checkout@v4, oven-sh/setup-bun@v2, cloudflare/wrangler-action@v3 pinned to full commit SHAs for supply chain hardening

Embedding worker lockfile

committed bun.lock for deterministic builds

Rate limiting TODO

added P1 security TODO for real API rate limiting implementation (stub headers only, no enforcement)

Faceted search counts

GET /feeds/:id/jobs?facets=workplace_type,industry returns value/count distributions per field. Typesense-style: each facet excludes its own filter condition so users see "what else is available"

Search highlighting

?include=highlights adds <mark>-tagged snippets in title and description via PostgreSQL ts_headline(). HTML stripped before highlighting for XSS safety

Search synonym expansion

common job terms automatically expand with OR groups (e.g. "frontend" matches "front-end", "ui"; "engineer" matches "developer", "programmer"). 30 synonym groups seeded, admin CRUD via tRPC

Admin synonym management

admin.synonyms tRPC router for CRUD operations on synonym groups with cache invalidation

OpenAPI spec

documented facets param, highlights in include, relevance_score, highlights response fields, exclude_domains filter, synonym behavior

Facet counts respect since filter

temporal conditions now passed to facet queries so incremental polls return consistent facet counts

Fuzzy fallback uses raw search

synonym-expanded query no longer breaks pg_trgm similarity matching in dashboard browsers

Synonym cache TTL reset on failure

prevents thundering herd of DB retries when the database is temporarily unavailable

Feed search relevance scoring

REST feed jobs endpoint returns relevance_score and orders by ts_rank_cd when feed has a search filter

Companies similar endpoint

fixed Drizzle array serialization for technology_names && overlap operator (was passing raw array, now uses sql.join with explicit ARRAY[]::text[] cast)

Feed search pagination safety

cursor-based pagination and lastPollAt checkpoint are now disabled when search is active, preventing silent data loss from mixing relevance ordering with time-based cursors

Full-text search

replaced ILIKE with Postgres tsvector + GIN indexes for ranked, stemmed search across jobs (title/company/description weighted A/B/C) and companies (name/industry)

Fuzzy search fallback

pg_trgm trigram similarity for typo-tolerant matching in dashboard browsers when FTS returns zero results

Relevance ordering

search results in admin and org dashboards now rank by ts_rank_cd relevance instead of chronological

Similar jobs API

GET /api/v1/jobs/{id}/similar returns top 10 semantically similar jobs via pgvector cosine distance

Search suggestions API

GET /api/v1/search/suggest?q= returns up to 5 fuzzy-matched job title suggestions

Embedding generation worker

Cloudflare Worker (5min cron) generates OpenAI text-embedding-3-small vectors for active jobs with daily token budget

DRY company filters

extracted buildRestCompanyConditions() to deduplicate REST API company filter logic

Search field max length increased to 200 chars (feeds + companies schemas)

REST companies API and export routes refactored to use shared filter builder

API response includes optional relevance_score when search is active

Companies browser rewrite

replaced custom grid/table view with shared DataTable component, nuqs URL state persistence, faceted multi-select filters (industry, country, funding stage), secondary toolbar (employee range, has active jobs checkbox), sortable columns, and offset pagination with total count

Admin companies dashboard

full admin companies list + detail pages reusing shared CompaniesPageClient and CompanyDetailClient with context="admin". Admin sidebar nav updated with Companies link

Two-way job↔company linking

company detail page shows Jobs tab with clickable job cards linking to internal job pages. Job detail pages link company name back to company detail

Shared filter builder

extracted lib/companies/filter-builder.ts with companyListInput Zod schema and buildCompanyConditions() to DRY admin + org company routers

Total funding formatting

company detail page now shows raw dollar amounts ($484.4M) instead of treating values as millions ($484.4T)

Company jobs link

job cards in company detail Jobs tab now link to internal job page (/dashboard/.../jobs/{id}) instead of external apply URL

Companies page crash

Radix Select crashed on load when a filter value resolved to an empty string. Fixed by passing undefined instead of "" and filtering empty strings from filter option queries

Apollo enrichment throughput

raised batch size (10 → 40) and daily limit (50 → 5000), unlocking ~480 company enrichments/hour — approaching Apollo's 600/hour rate limit

Companies API

6 REST endpoints (/v1/companies, /v1/companies/:id, /v1/companies/:id/jobs, /v1/companies/:id/similar, /v1/companies/trending, /v1/companies/export) with cursor pagination, search, and filtering by industry, size, funding stage, tech stack, and country

Company table

new company entity with 30+ columns for identity, enrichment data, and derived stats. company_merge_audit table for dedup audit trail. company_id FK on job table

3-tier entity resolution

deterministic name/domain matching (Tier 1), Apollo enrichment signal merging (Tier 2), and LLM dedup placeholder (Tier 3). Handles ATS domain filtering (57+ known platforms), normalized name matching, and domain-first enrichment ordering

Apollo Organization Enrichment

CF Worker enriches pending companies via Apollo.io API with daily credit limits, FOR UPDATE SKIP LOCKED concurrency control, and automatic Tier 2 dedup on apolloId/domain conflicts

CF Worker pipeline

4-phase cron every 5 min: Resolve (500 jobs/tick) → Enrich (10/tick via Apollo) → LLM Dedup (future) → Stats. Deployed at workers/company-enrichment/

Companies dashboard

list page with DataTable, search, faceted filters (industry, country, funding stage), sortable columns, and offset pagination. Detail page with tabs (Overview, Jobs, Similar). Admin companies pages with full parity. Trending widget on org dashboard home

Company feed filters

4 new feed filter fields: company_size, company_funding_stage, company_tech_stack, company_id. Only matches jobs at enriched companies (documented semantic)

Hiring velocity

on-the-fly computation of 7d/30d/90d job counts with trend detection (accelerating/stable/decelerating)

Company normalization

pure functions for name normalization (suffix stripping, word-boundary safe), domain extraction (ATS filtering), and slug generation. Shared between Next.js app and CF Worker

OpenAPI spec

6 new company endpoints added to openapi.json with Company, CompanyBrief, and HiringVelocity schemas

42 unit tests

normalization (10), domain extraction (8), slug generation (5), and velocity computation (6) plus API and filter tests

ATS domain list expanded

from 28 to 57+ known ATS/HRIS platforms to prevent wrong company enrichments (added Taleo, ADP, Workday, Oracle, SAP, Personio, etc.)

Enrichment ordering

domain-first priority (ORDER BY (domain IS NOT NULL) DESC) improves Apollo match rate from ~0% to ~90% for name-only lookups

Granular email scheduling

new feed_notification_config table replaces the rigid daily/weekly email schedule. Supports per-recipient configs with independent frequencies (every 6h, every 12h, daily, weekly), preferred delivery hour + timezone, and minimum job threshold

Smart auto-frequency

adaptive "auto" mode sends based on job volume: 50+ jobs at 6h intervals, 20+ at 6h, 5+ at 12h, 1+ daily fallback. 6-hour floor caps sends at 4/day

Preferred hour + timezone

schedule emails for a specific hour in the recipient's timezone using native Intl.DateTimeFormat (no external libs). Handles DST transitions gracefully

Minimum job threshold

minJobCount column skips digest sends when fewer than N new jobs match, preventing low-value emails

Channel-aware table design

feed_notification_config supports future Slack/webhook channels via channel column (only email implemented now)

Multi-recipient data model

one feed can have N notification configs with different recipients, frequencies, and formats

Parallel cron processing

email-export cron rewritten with Promise.allSettled batches of 5 concurrent sends, maxDuration=300

Per-config idempotency

keys include configId + hour (notif-{configId}-{YYYY-MM-DDTHH}) enabling sub-daily sends without duplicates

Neon driver unique constraint detection

Postgres error code 23505 is embedded in the error message string by the Neon serverless driver, not as a .code property. Duplicate notification config attempts now correctly show "This email is already configured" instead of a 500 error

Cron schedule

email-export changed from daily (0 8 * * *) to hourly (0 * * * *) to support sub-daily frequencies

Email delivery tracking

extracted trackEmailDelivery() helper from 3 duplicated call sites (cron markdown, cron HTML, sendDigest mutation). Added configId FK on email_delivery table

tRPC mutations

added createNotificationConfig, updateNotificationConfig, deleteNotificationConfig with org-scope auth. Legacy updateEmailSchedule retained as wrapper

Feed detail UI

email schedule card replaced with notification config manager (config list, frequency dropdown with 5 options, add/remove)

Vercel enrichment cron

Removed /api/internal/process from vercel.json crons. CF Worker (workers/enrichment/) is the sole enrichment path — 50 jobs/5 min with atomic claiming. Vercel route kept as manual fallback trigger

pg SSL deprecation warning suppressed

Filtered pg-connection-string v2.12.0 SECURITY WARNING about sslmode=require being treated as verify-full. Vercel's Neon integration auto-sets this value and it cannot be changed

content-collections config migrated

Renamed collections to content in defineConfig per content-collections v0.14.0 breaking change

next-themes React 19 script warning patched

Bun patch on next-themes@0.4.6 skips rendering ThemeScript on the client (script only needs SSR for FOUC prevention). Matches upstream PR pacocoursey/next-themes#386

next-themes patch

patches/next-themes@0.4.6.patch should be removed when next-themes releases a fix (watch for v0.4.7+ or PR #386 merge)

Dependencies updated

Patch bumps for 8 packages: @ai-sdk/openai 3.0.48 (Bun fetch retry fix), @ai-sdk/react 3.0.139, ai 6.0.137 (gateway model catalog), @aws-sdk/client-s3 3.1015.0 (error deserialization fix), @aws-sdk/s3-request-presigner 3.1015.0, @tanstack/react-query 5.95.2 (NodeJS.Timeout type leak fix), fumadocs-core 16.7.5 (TOC observation rewrite), fumadocs-ui 16.7.5

Careers page

Removed /careers route, section components, navigation links (header + footer), and sitemap entry. Page contained hardcoded placeholder positions, not real listings

Component render contract tests

9 new tests verify React.memo wrappers on Button and all Card subcomponents

Feed detail actions

Dropdown menu on feed detail page with Rename, Pause/Resume, and Delete actions

SessionProvider context memoized

Context value wrapped in useMemo to prevent tree-wide re-renders on every parent update

JobsBrowser memoization overhaul

browseInput, columns, facetedFilterEntries, and columnFilters wrapped in useMemo; event handlers (commitSalary*, commitCompany, handleFiltersChange, handleSearchQueryChange, handleSaveAsFeed) wrapped in useCallback

Button wrapped in React.memo

Skips CVA variant recomputation when props unchanged

Card subcomponents wrapped in React.memo

All 7 exports (Card, CardHeader, CardTitle, CardDescription, CardAction, CardContent, CardFooter) memoized

Feed detail handlers use useCallback

handleCopy, handleDownload stabilized

OrganizationSwitcher Icon extracted to useCallback

Prevents inline component recreation on every render

NiceModal pattern for feed actions

Rename and delete dialogs use NiceModal.create() / ConfirmationModal instead of manual useState, consistent with codebase conventions

Sidebar context missing dependencies

Added setOpenMobile, setWidth, setIsDraggingRail to useMemo dependency array, fixing potential stale closure bug

Column filters dependency chain

Replaced fragile spread-as-deps pattern with explicit useMemo chain, removed eslint-disable comment

Delete clears detail cache

Removes cached feed data on delete so browser back button shows "Feed not found" instead of a stale ghost page

Rename trim comparison

Compares trimmed input against current name so whitespace-only changes are correctly rejected

Delete double error toast

Removed duplicate onError handler that caused two overlapping toasts on delete failure (ConfirmationModal already handles the error path)

Email delivery tracking

New email_delivery table tracks every email sent via Resend with status updates from webhooks (delivered, bounced, complained, failed). Unique index on resend_message_id prevents duplicates

Resend webhook endpoint

POST /api/webhooks/resend receives Svix-signed delivery status events, verifies signatures, and updates email delivery records

Email stats in feed overview

4-column stat grid now shows Emails Sent, Last Email, Webhook Deliveries, and Failures with tabular-nums and border-destructive on failure state

RESEND_WEBHOOK_SECRET env var

Supports webhook signature verification

sendFeedDigestEmail returns messageId

Enables delivery tracking in both cron and manual send paths

Cron email export tracks deliveries

Both markdown and HTML format paths insert into email_delivery table (best-effort with try/catch)

Dashboard sendDigest tracks deliveries

Manual sends record email delivery with error-resilient DB insert

Feeds list badge accounts for email deliveries

Badge now queries email_delivery table for counts and last-sent dates per feed. Cascades: webhook date → email date → email count → "No deliveries yet" (PR #98)

"Last Email" stat derives from actual sends

Feed detail stat card now shows max(email_delivery.createdAt) instead of the cron cursor (lastEmailAt), so manual sends are reflected immediately (PR #98)

Jobvite Apify scraper

Reverse-engineered hidden XML feed at app.jobvite.com/CompanyJobs/Xml.aspx?c={companyId}. One request returns all jobs per company with full HTML descriptions, category, jobtype, region, location, and dates. 548 company slugs from Common Crawl, self-healing KV store, CompanyId extraction from page JS. Total sources: 12

Dependency patches

Bumped content-collections, t3-env, react-query to latest patch versions

Removed newsletter/ folder

Marked TODO done, cleaned stale deps

Digest email template polish

neutral title (was sage), removed horizontal dividers, tighter workplace tags (inline with meta row), lighter card borders (neutral-100), smaller footer

Job limit picker

configurable 1-200 (default 20) for generateDigest and sendDigest, number input in Digest tab

Dark mode audit

all elements have semantic CSS classes with explicit prefers-color-scheme: dark overrides; added job-divider class to card borders

sendDigest no longer advances cron cursor

manual "Send Now" was updating lastEmailAt, causing scheduled cron to skip unsent jobs. Manual sends are now one-off with no side effects on cron scheduling

workplaceLabel test synced

test had stale copy of old workplaceTag function (returned { label, color } instead of string)

Footer link contrast

neutral-300 (1.6:1) → neutral-400 (2.7:1) for readability

Unused FileText import removed from feed-detail-client

CLAUDE.md

updated template description from "sage accent" to "neutral palette"

Feed digest email template

Rich React Email template for scheduled feed digests. Job cards with salary emphasis, workplace tags (sage pill for Remote), deadline countdown, skills, and empty state handling. Uses pixelBasedPreset for email client compatibility, dark mode via prefers-color-scheme, sage accent from DESIGN.md

AI-generated digest intro

GPT-4o-mini generates a 1-sentence contextual summary for each digest email. 10s timeout, maxOutputTokens: 100, static fallback on error. Via AI SDK v6 generateText()

exclude_domains feed filter

Block jobs from specific domains in any feed. Domain-boundary-safe ILIKE matching (won't block smart.com when blocking art.com). Capped at 20 domains per feed. Uses isNull() for proper NULL handling

Idempotency keys for digest emails

feed-digest-{feedId}-{date} prevents duplicate sends on cron retries via Resend SDK native idempotencyKey parameter

33 new unit tests

digest-intro.test.ts (4 tests: AI fallback, empty jobs, error handling), exclude-domains.test.ts (11 tests: SQL generation, domain format validation, schema constraints), digest-email-helpers.test.ts (18 tests: deadline labels, workplace tags, meta content)

Digest preview tab

new dashboard tab in feed detail page. Generate → preview (HTML iframe or markdown source) → copy to clipboard → download as file → send now to any email. Two new tRPC mutations: generateDigest (render-only) and sendDigest (with 1-min per-feed rate limit)

Email export cron upgraded

HTML format now uses React Email template with AI intro instead of basic formatJobsHtml(). Markdown format preserved

toDigestJob() extracted into shared lib/feeds/digest-job.ts

used by both cron and dashboard tRPC (DRY)

formatSalary() and formatLocation() exported from job-formatter.ts for reuse across cron and future consumers

newsletter/ standalone app

All functionality (Kadoa scrapers, OpenAI inline processing, React Email template, domain blacklist, streaming UI) replaced by native Bordfeed infrastructure. Can be safely deleted

Workday Apify scraper

Reverse-engineered Workday's CXS API (POST /wday/cxs/{tenant}/{site}/jobs). 1,765 company sites discovered via Common Crawl across 22 wd variants. Self-healing KV store, per-job detail fetch for full descriptions, 500 job cap per company. Total sources: 11

YCombinator Apify scraper

Scrapes YC's Work at a Startup job listings. 217 hiring companies with full job descriptions

Poets & Writers (PW) Apify scraper

Sitemap discovery + RDFa extraction for literary job postings

Lever Apify scraper

Self-healing KV store pattern (like Greenhouse). 2,714 company slugs from Common Crawl, daily scrape + monthly discovery mode. Seed slugs in Apify KV store mnmz5EL4jZcUGE6pC. Total sources: 8

RemoteOK Apify scraper

Dynamic tag discovery from sitemap (~250 tags), per-tag API fan-out, ID-based dedup. Respects 1s crawl-delay. Employment type inference from tags (Contract, Part Time, Freelance, Internship). Locale-safe salary formatting via Intl.NumberFormat

Intra-batch dedup

Identical payloads within the same batch now deduplicated via in-memory hash set (previously all inserted)

Null byte sanitization

Strip \x00 and control chars before DB insert (PostgreSQL rejects them, causing silent INSERT failures)

15K ingestion cap

Datasets exceeding 15K items auto-capped and marked partial instead of silently timing out

Stuck webhook recovery

Stale received entries (>5min) auto-cleaned to error on each webhook invocation

Greenhouse actor limit

Default capped at 10K (safe margin under 15K tested ceiling)

DynamiteJobs, SmartRecruiters, Greenhouse Apify scrapers

3 new job source actors in apify-scrapers/, deployed to Apify with webhooks + schedules. Total sources: 6 (was 3). Greenhouse alone covers 19K+ jobs across 7,101 company slugs

/new-job-source skill

Codified workflow for building future Apify scrapers

Chunked batch INSERT

Ingest pipeline chunks inserts (500 rows/chunk) to handle 15K+ job ingestion without hitting Postgres query size limits

Webhook processing for large batches

Switched from after() to inline processing to prevent premature function termination on large datasets. Increased maxDuration to 300s (Vercel Pro ceiling)

Ingest input sanitization

Strip HTML tags from titles/companies, cap title length at 500 chars, handle empty strings as "Untitled"

Apify scrapers excluded from TypeScript build

apify-scrapers/ added to tsconfig exclude list

Dependency updates

@tanstack/react-query 5.91.3→5.94.5, fumadocs-core/ui 16.7.3→16.7.4 (TOC CSS fix), react-hook-form 7.71.2→7.72.0

cropperjs 1→2 (react-cropper incompatible — need to swap to react-easy-crop)

Deduplicated jobs pages

Admin and org jobs now share the same rich browser component (components/jobs/), detail sheet, and detail page via a context prop pattern

Admin jobs page upgraded

Now uses the full-featured filterable DataTable browser (was a basic table with pagination), plus stats cards

Admin job detail route

Added /dashboard/admin/jobs/[id] detail page (was org-only)

Shared select objects

Extracted jobBrowseSelect and jobDetailSelect to lib/jobs/selects.ts, reused by both admin and org routers

Admin jobs router

Added browse, filterOptions, and get endpoints matching org shape (admin sees all statuses, not just active)

fumadocs-core 16.7.2 → 16.7.3

Sidebar scroll fix

fumadocs-ui 16.7.2 → 16.7.3

Home layout navigation menu redesign

recharts 2.15.4 → 3.8.0

Major version upgrade. Fixed chart.tsx tooltip/legend types for v3 compatibility

Removed react-resizable-panels

Unused dependency and wrapper component deleted. Re-add via bunx shadcn@latest add resizable if needed

better-auth 1.4.17 → 1.5.5

Auth framework major upgrade with breaking changes: renamed email confirmation callback, updated error code types, fixed session/user type inference for plugin-added fields (activeOrganizationId, twoFactorEnabled)

Worker Zod 3 → 4

Upgraded enrichment worker's Zod dependency from v3.23 to v4.3.6, aligning with root app. Import path changed to zod/v4 for explicit version pinning.

53 dependency upgrades

Next.js 16.1.6→16.2.1, tRPC 11.9→11.14, AI SDK 6.0.57→6.0.134, Sentry 10.37→10.45, motion 12.29→12.38, Tailwind CSS 4.1→4.2, Vitest 4.0→4.1, Biome 2.3→2.4, fumadocs 16.4→16.7, shikijs/rehype 3.21→4.0, react-dropzone 14→15, AWS SDK 3.975→3.1014, plus ~30 patch bumps

Worker AI SDK sync

Enrichment worker @ai-sdk/openai and ai packages synced to match root versions

Vercel Analytics/Speed Insights v2

Upgraded from v1 to v2 (React API unchanged)

Dropzone rejection handling TODO

Tracked react-dropzone v15 follow-up for upload error feedback in TODOS.md

Upload dropzone rejection handling

Both avatar and logo upload flows now validate file type (images only) and size (5MB max), show toast errors for rejected files, and guard against empty drops opening the crop modal

Worker CI uses bun

deploy-enrichment-worker.yml now uses oven-sh/setup-bun and bun install instead of npm, matching the project's package manager

Removed stale package-lock.json

Deleted npm lockfile (58 mismatches with package.json) and added packageManager: bun@1.3.11 to package.json for explicit bun selection

Apify Console link

External link button on each Apify source row opens the actor's Apify Console page directly. Desktop icon button + mobile labeled button. Only visible for sources with an actor ID

Run Now button

Trigger Apify actor runs directly from the admin Sources page. Rocket icon button for Apify-type sources in both desktop table and mobile card layouts. Backend triggerApifyRun function with full error handling

6 new triggerApifyRun tests

Full coverage: not found, non-Apify rejection, missing token, success with runId, API error response, network failure

Bulletproof Apify webhook handler

Refactored to respond 200 immediately and process via Next.js after(), eliminating Apify's 30-second timeout risk. Includes paginated dataset fetch (1000 items/page), timeout tracking, and auto-retry once on failure

Webhook idempotency

New webhook_log table with unique apifyRunId index prevents duplicate processing. Race conditions handled via insert constraint catch

Pipeline failure notifications

Dual-channel alerts (Resend email + Slack webhook) fire on FAILED/ABORTED/TIMED_OUT events and all-error ingests. Fire-and-forget with graceful degradation

Admin Sources page

Dashboard at /dashboard/admin/sources with source CRUD (create, edit, archive, toggle active/paused), 4 stats cards, webhook log table with status/source filters, retry button for failed runs. Includes toast notifications, AlertDialog confirmation for destructive actions, inline form validation, accessible icon buttons, and mobile card layout

DB-driven source resolution

Webhook handler resolves Apify actor IDs to source names from job_source table (with 60s in-memory TTL cache) instead of hardcoded map. Sources managed via admin UI

Apify run traceability

apifyRunId column on jobTable links every ingested job back to its exact Apify actor run

26 new pipeline tests

15 webhook handler tests (auth, validation, idempotency, event routing, source mapping) + 11 notification tests (channel selection, failure resilience, content formatting)

DESIGN.md

Formalized design system: Industrial/Utilitarian aesthetic, Geist Sans typography, sage accent, compact admin density, motion and component conventions

Webhook handler now handles ACTOR.RUN.FAILED, ACTOR.RUN.ABORTED, and ACTOR.RUN.TIMED_OUT events (previously ignored all non-success events)

ingestJobBatch accepts optional IngestOptions with apifyRunId passthrough for traceability

Enrichment pipeline now writes clean descriptions

AI extraction already produced a clean description field but the UPDATE query never wrote it back to the DB. Raw RSS/HTML from ingest stayed in the description column forever. Fixed in both CF Worker and lib pipeline

stripHtml hardened for double-encoded HTML

RSS feeds often contain &lt;strong&gt; (HTML entities inside HTML). Now runs two-pass strip: decode entities, then strip revealed tags. Also strips RSS metadata preambles (Title/Category/Posted headers)

Enrichment prompt updated

AI instructed to return description as clean plain text, preserving original wording verbatim — no summarizing or rephrasing, only stripping markup and encoding artifacts

Job detail views

Side panel sheet on row click in jobs browser + full-page detail view at /dashboard/organization/jobs/[id] with 100% API field parity (salary, location, career, skills, languages, description, dates & source)

Shared DetailRow component

Reusable label-value display component at components/ui/custom/detail-row.tsx, used across sheet and full page views

stripHtml utility

Converts HTML job descriptions to clean text with newline preservation, entity decoding, and whitespace normalization

formatSalaryFull utility

Formats salary ranges with currency/unit defaults (USD, /year)

18 new unit tests

Full coverage for stripHtml (11 tests) and formatSalaryFull (7 tests)

Feed cards

Added Clock icon next to delivery timestamps, skeleton loading states

Feed detail

Stat card icons (Send, Clock, AlertTriangle), skeleton loading, copy-to-clipboard button for filter JSON

API keys

Click-to-copy key prefix, standardized empty state icon size (h-12 w-12)

Jobs router

get procedure now filters by active status, consistent with browse

Cloudflare Workers enrichment pipeline

Replaces Vercel cron-based enrichment with a dedicated CF Worker at workers/enrichment/. Processes 50 jobs every 5 minutes (vs 10 jobs/15 min on Vercel). No 60s timeout limit

Atomic job claiming

FOR UPDATE SKIP LOCKED prevents race conditions between concurrent worker runs. processingRunId guard prevents stale overwrites

Configurable concurrency

Worker enriches jobs in parallel chunks (default 5 concurrent). Batch size and concurrency configurable via env vars

Stale lock auto-cleanup

Crashed workers' locks are automatically released after a configurable timeout (default 10 min)

GitHub Action auto-deploy

Worker auto-deploys on push to main when workers/enrichment/** changes via cloudflare/wrangler-action@v3

AI SDK v6 token cleanup

Removed as any casts from token usage. Uses inputTokens/outputTokens directly. Captures new v6 fields: reasoningTokens, cachedInputTokens

Enrichment throughput

50 jobs/5 min (600/hour) vs previous 40 jobs/hour on Vercel cron. 15x improvement

Neon HTTP driver

Worker uses @neondatabase/serverless HTTP driver (not WebSocket) for Cloudflare Workers compatibility

Root tsconfig exclusion

workers/ excluded from root TypeScript compilation so Vercel builds don't fail on worker-only dependencies

Live changelog

Homepage hero pill and /changelog page now show real release history from CHANGELOG.md instead of boilerplate placeholder content

Full date display

Changelog page now shows "March 19, 2026" instead of abbreviated "March 2026"

Changelog parser

Build-time parser at lib/changelog.ts handles 5 markdown bullet formats with server-only guard

Total tests

149 passing (13 suites), including 17 new changelog parser tests

CSV formula injection (CWE-1236)

CSV export now prefixes cells starting with =, +, @, - with a tab character to prevent Excel/Sheets formula injection

URL scheme validation

HTML and Markdown formatters now block javascript: and data: URL schemes in job apply_url fields, preventing XSS in exports and emails

Webhook secret leak

tRPC feed endpoints no longer return the HMAC webhook secret to the frontend on any operation (list, get, create, update, email schedule)

Export credit timing

Export API now checks credit balance before fetching data, preventing charges on failed queries

Dashboard export cursor

Dashboard export no longer incorrectly filters by email delivery cursor, which caused it to return incomplete results after cron runs

Type-safe status enum

Replaced "active" as any casts with JobStatus.active enum constant across 3 files

Ingest dedup performance

Secondary title+company dedup now uses batched OR queries (chunks of 50) instead of N+1 per-job queries

Total tests

132 passing (12 suites)

Dashboard export + scheduled email delivery

Feed detail page with Export tab. Download matching jobs as HTML, Markdown, CSV, or JSON. Configure daily/weekly email delivery via Resend — Bordfeed emails formatted newsletter content to your inbox

Feed detail page

Click any feed card to see overview stats, filters, and the new Export tab. Breadcrumb navigation back to feeds list

Export API endpoint

GET /v1/feeds/:id/export?format=json|csv|html|markdown returns up to 1000 jobs in a single response without pagination. Supports since=last_poll for incremental exports

since=last_poll auto-cursor

Server-side cursor tracks your last poll/export. GET /v1/feeds/:id/jobs?since=last_poll returns only jobs added since your last API call, eliminating client-side state tracking

include=html_snippet

GET /v1/feeds/:id/jobs?include=html_snippet adds a pre-formatted HTML block per job for newsletter use, with XSS sanitization

Job formatter utility

lib/feeds/job-formatter.ts with HTML, Markdown, CSV formatters. All user-sourced fields sanitized to prevent XSS

Secondary dedup

Ingest pipeline now catches semantic duplicates via case-insensitive title + company match, in addition to existing content hash dedup

Email export cron

POST /api/v1/cron/email-export runs daily at 08:00 UTC via Vercel Cron. Sends formatted jobs to configured recipients

17 new tests

Job formatter test suite covering HTML, Markdown, CSV output, XSS sanitization, empty states, and edge cases

ROADMAP.md

Rewritten to reflect 4-phase plan: Data Flow (done) → Dogfood+DX (current) → Monetization → Scale. SDK removed from roadmap per CEO plan decision

OpenAPI spec

Updated to v1.3.0 with export endpoint, since and include parameters documented

Feed cards

Now clickable, navigate to feed detail page

Total tests

125 passing (12 suites)

OpenAPI 3.1 spec + Scalar interactive docs

Full API specification at public/api/openapi.json covering all 11 public endpoints. Interactive docs at /docs/api via @scalar/api-reference-react with kepler theme. API route serves spec with CORS headers

Dashboard metrics

Replaced achromatic boilerplate email stats with real Bordfeed metrics: Active Feeds, Credits Remaining, API Keys, Webhook Deliveries. New organization.stats.dashboard tRPC procedure with org-scoped parallel queries

OpenAPI route

GET /api/openapi.json serves the spec with caching and CORS for Scalar client-side rendering

Enrichment cron frequency

Now runs every 15 minutes instead of every 2 hours. Clears the enrichment backlog of 1,253 jobs in ~31 hours instead of ~10 days — jobs reach your feed much faster

Org selection cards

Cards now have a visible shadow and a cleaner hover effect so it's obvious which org you're selecting

Boilerplate dashboard charts

Removed dashboard-demo-charts.tsx (640 lines of fake email metrics data) replaced by real org-scoped stats

Avatar and org logo upload failing

S3 signed upload URL was generated with ContentType: "image/jpeg" hardcoded, but the image cropper outputs PNG. S3 rejected the PUT due to content type mismatch. Made contentType a required parameter throughout the upload chain (schema → tRPC → S3) so the signed URL always matches the actual upload

**Dashboard empty

tRPC 503 errors** — getBaseUrl() used stale per-deployment Vercel URL for tRPC batch requests. Added NEXT_PUBLIC_VERCEL_PROJECT_PRODUCTION_URL as stable fallback before the ephemeral deployment URL. Also set NEXT_PUBLIC_SITE_URL env var in Vercel production (#60)

**Enrichment schema

OpenAI 100% failure** — status: z.string().default("active") produced a JSON schema without status in the required array, causing OpenAI structured output to reject every request. Changed to .nullable() with default handled in process-jobs.ts (#56)

Feed templates returning 0 results

AI enrichment produces free-form categories ("Product Design", "Full-Stack Programming") but filter builder used exact inArray match against slug-style template values ("design", "full-stack"). Changed to case-insensitive ilike partial matching. Same fix for workplace_type (#57)

Homepage hero copy

replaced achromatic boilerplate with Bordfeed messaging: "The API for job data. Fresh, structured, delivered." CTAs: "Start Building" / "View Pricing". Pill links to /changelog (#58)

Features section copy

"Everything your feed needs" with 20+ Filter Fields and Webhook Delivery cards (#59)

Stats section copy

real product metrics: 11 endpoints, 20+ filters, 6 templates, 50+ fields (#59)

FAQ section copy

5 Bordfeed-specific questions replacing generic SaaS boilerplate (#59)

Apify webhook adapter

POST /api/internal/apify-webhook receives Apify run metadata, fetches dataset items from Apify API, passes raw data to shared ingest logic. Actor→source mapping for Workable, WWR, JobDataAPI

Shared ingest module

extracted lib/pipeline/ingest-jobs.ts used by both /api/internal/ingest and /api/internal/apify-webhook

Pipeline tests

60 new tests across 4 suites (ingest, process, monitor, push). Total: 108 tests / 11 suites

SSRF protection in monitor

blocks private/internal IPs before fetching source URLs

AI job classification in monitor

wires up classifyJobStatus() to detect filled/closed jobs returning HTTP 200

Batch ingest

single SELECT + INSERT replaces N+1 per-job queries

Max batch size guard

rejects payloads >200 jobs with 413

Pipeline reliability TODOs

concurrent push protection (P2), batch push at scale (P3)

skills column

migrated from text to text[] array with arrayOverlaps filter (migration: 20260319082018_damp_vulcan.sql)

Content hash

JSON keys sorted before hashing for deterministic dedup across actor runs

Processing lock

now filters by specific job IDs (inArray) instead of locking all pending jobs

Race condition in job processing lock

concurrent cron runs could process same jobs twice

isPrivateHost() exported from webhook-validation for reuse in monitor SSRF checks

Pre-push git hook (tsc --noEmit)

failed on generated content-collections types; Vercel build already gates type errors

Feed-Centric API

complete rewrite on achromatic SaaS boilerplate

Feeds API (public REST)

POST/GET/PATCH/DELETE /v1/feeds, GET /v1/feeds/:id/jobs (poll with cursor pagination), GET /v1/health

Feed filter builder

20+ strict filter fields (category, workplace_type, country_code, career_level, salary range, languages, skills, industry, visa_sponsorship, and more)

Feed templates

6 pre-built templates (Remote Writing, AI/ML Engineering, Growth Marketing EU, DevRel, Remote Design, Full-Stack Remote US)

Webhook delivery

HMAC-SHA256 signing, SSRF protection, retry with backoff

Feed tRPC router

dashboard management (list, create, update, delete, templates)

Job admin tRPC router

pipeline management (list, stats, status updates, bulk actions)

Dashboard pages

feed list with cards, admin jobs table with stats

Database schema

job table (50+ fields), feed table, webhook delivery table, feed template table

Zod schemas

strict validation with unknown field rejection

Unit tests

48 tests across 7 suites (content hash, salary, pagination, webhook validation, webhook signing, Zod schemas, utils)

Credit-based billing

feed poll and webhook delivery consume credits (Starter $9.99/1K, Growth $49/10K, Scale $299/100K)

Design doc

docs/designs/feed-centric-api.md

Foundation

replaced v0.5.4 codebase with achromatic SaaS boilerplate (Better Auth + orgs + Stripe + Resend + Sentry + marketing pages)

Product model

enrichment is now internal-only, feeds are THE product

API architecture: hybrid

tRPC for dashboard, REST for public /v1/* API

Billing

API keys and feeds scoped to organizations (multi-tenant)

Public enrich endpoints (enrichment is internal infrastructure)

Old subscription/webhook system (replaced by feeds)

Custom auth, dashboard, admin (replaced by achromatic)