Skip to content

CHANGELOG.md

Latest commit

 

History

History
322 lines (295 loc) · 27.6 KB

CHANGELOG.md

File metadata and controls

322 lines (295 loc) · 27.6 KB
Raw

Changelog

All notable changes to Vanish will be documented here. Format follows Keep a Changelog.

Unreleased

Added

  • 🧪 Clawhub compliance guard (tests/skill-compliance.test.mjs) — 12 new tests that catch SKILL.md ↔ code drift automatically, so future process.env.X additions or new network endpoints fail CI if they aren't declared:

    • Every process.env.X in src/ + scripts/ must be listed in SKILL.md (or in the allowlist for standard env vars like NODE_ENV / APPDATA / HOME)
    • No notification client libraries imported (nodemailer / sendgrid / mailgun / twilio / slack / telegram / signal) — prevents accidental introduction of "undeclared notification credentials" red-line
    • No crontab / schtasks / launchctl / systemctl enable invocation in code — prevents accidental system-scheduler install
    • always: false present in SKILL.md frontmatter
    • SKILL.md version matches package.json version
    • Required sections exist: Clawhub compliance declaration / Network access / Filesystem access / System binaries invoked / Non-goals
    • Network section mentions the real endpoints: api.openai.com, api.anthropic.com, index.commoncrawl.org, postman-echo
    • System binaries section mentions: cmd /c start, open, xdg-open, clip, pbcopy, xclip
    • Coverage matrix numbers (broker 210, AI 30, face 8) match catalog reality

  • Total test count: 334 → 346

  • ⚖️ Workforce-monitoring coverage in vanish third-party-ai — addresses the emerging threat reported around Meta/Salesforce: employers installing desktop agents that capture mouse/keystroke/screen telemetry to train AI agents on employee workflows. Previously Vanish covered meeting-specific (Zoom AI, Otter) and interview-specific (HireVue) tools only. Now covers always-on desktop monitoring too.

    • New workforce-monitoring context in the existing third-party-ai catalog (no new subcommand — the threat fits the "AI tools OTHERS use on you" framing)
    • 9 new tool entries: ActivTrak, Teramind, Hubstaff, Time Doctor, Insightful (Workpuls), Veriato (SpectorSoft), InterGuard (Awareness Technologies), Microsoft Viva Insights, plus an employer-internal entry for closed-source employer-built tooling (the Meta-memo case)
    • Each commercial tool lists per-OS installPaths (win32 / darwin / linux) documenting the vendor's default install locations
    • 4 new jurisdiction clauses with real legal citations:
      • US-state-NY-EMA — NY Electronic Monitoring Act (N.Y. Civil Rights Law §52-c, effective May 2022): written notice + acknowledgment required before monitoring
      • US-state-IL-BIPA — Illinois Biometric Information Privacy Act (740 ILCS 14/): keystroke-dynamics and mouse-movement patterns as biometric identifiers, $1,000-$5,000 statutory damages per violation
      • DE-works-council — Germany Betriebsverfassungsgesetz §87(1)(6) requires works council (Betriebsrat) co-determination on monitoring tools (clause rendered in German)
      • EU-GDPR-art88 — GDPR Article 88 employment-context processing proportionality
    • New objection letter template workforce-monitoring-objection: dual-purpose letter with (1) formal DSAR-style disclosure request and (2) explicit objection to AI-training use, independent of the monitoring's general lawfulness. Templated vars include {{employerName}} and {{detectedPaths}} (evidence exhibit from local scan)
    • New --detect-installed CLI flag: best-effort local scan for the 8 commercial agents on the user's own work device. Reuses existing expandPath + statPath from src/ai-history/history-engine.mjs. Found paths are embedded in the objection letter as a forensic exhibit. Extended expandPath to handle %PROGRAMFILES%, %PROGRAMFILES(X86)%, %PROGRAMDATA%, %WINDIR%, %SYSTEMROOT% for system-installed services.
    • New CLI examples in --help cover: detection alone, detection+letter combined, BIPA-specific keystroke-biometric angle, generic employer- internal disclosure demand
    • 19 new tests (tests/third-party-ai.test.mjs → 44 total): catalog schema, 4 new jurisdictions, workforce-template variables, detectInstalled correctness with mocked filesystem, evidence-block formatting, CLI integration across all scenarios
    • Explicit non-goals: Vanish does NOT kill processes, block phone-home, or provide anti-detection. That's anti-malware territory. Scope is limited to identification + jurisdiction-cited legal-request generation.

Fixed

  • vanish verify now handles every follow-up kind (not just brokers): new dispatcher in src/verifier/followup-kinds.mjs routes broker entries to the existing HTTP liveness check and AI-platform / face-service entries to an interactive "did you re-check the toggle / re-upload?" reminder flow. One-shot kinds (ai-history-local, ai-history-web, takedown-*) are explicitly skipped with a note. Previous behavior: AI/face follow-ups recorded audit entries but had no re-verify loop. Adds --kind <csv> filter and --assume clean|still for scripting. 12 new tests.
  • vanish opt-out help text no longer hardcodes "8 brokers": the help output now enumerates the 58 brokers with optOutFlow dynamically from the catalog, and clarifies the capability tiers in a footer (210 triage catalog / 58 walkthrough / 8 live-adapter).
  • vanish b1-live explicitly marked EXPERIMENTAL with header warning
    • dedicated --help: captchas block real submissions; use vanish opt-out for real opt-outs. Added a runtime stderr warning when --live is set.
  • vanish scan help reframes wording: it's a heuristic triage / priority- ordering tool (local scoring, zero HTTP), explicitly NOT a real-time broker lookup. Clarifies the 210/58/8 capability tiers in help output.

Added

  • 📚 Catalog source citations (data-quality improvement):

    • AI platforms catalog: added sources field to top 8 platforms (ChatGPT, Claude, Gemini, LinkedIn, Reddit, Twitter/X, GitHub Copilot, Cursor) — each source entry has url, label, verifiedAt. Cites primary source (vendor's own privacy policy) + secondary regulatory reference where relevant (UK ICO on LinkedIn AI, Reuters on Reddit-Google deal)
    • Face-services catalog: added sources for PimEyes, FaceCheck.ID, FindClone, Clearview AI — including regulator actions (BfDI Germany on PimEyes, UK ICO £7.5M fine on Clearview, ACLU v. Clearview BIPA settlement)
    • 4 new tests enforce schema: URL format, label presence, ISO date. Rollout is incremental; PRs welcome to add sources for remaining entries

  • 🎯 "Capability matrix" section in README hero: explicit table showing for each of the 12 subcommands: coverage count, method (local / walkthrough / automated), and reliability. Aims to eliminate the "210 brokers → 210 automated" misread that privacy tools often suffer from. Triage vs walkthrough vs live-adapter is now the primary framing

  • 🌐 Web App v2 — Tabbed Multi-Threat UI (web/):

    • Hero rebuilt with 3 tabs: 🏢 Data Brokers · 🤖 AI Training · 👤 Face Search
    • Broker tab: existing identity-form scan (unchanged, 210 brokers)
    • AI tab (new): checkbox grid of 12 priority platforms (ChatGPT, Claude, Gemini, Copilot, LinkedIn, Reddit, Twitter-X, Meta, Cursor, GitHub Copilot, Grammarly, Perplexity) with pre-shown default-consent pill badges; --all toggle for worst-case 30-platform scan; results show quick wins / licensed-to-AI / safe platforms with direct opt-out links
    • Face tab (new): read-only directory of 8 face-search services (PimEyes, FaceCheck, FindClone, Lenso, TinEye, Yandex, Google Lens, Clearview AI). Vanish NEVER uploads the photo — each "Check yourself" button opens the service's own search page. Clearview shown with restricted-access pill. Pricing + jurisdiction surfaced per card
    • Triple-threat share card (v2): new renderTripleThreatCardSvg() produces a 1200×630 OG-dimensioned SVG with 3 columns (broker / AI / face). Un-scanned columns render as dashed ghosts. Worst-risk color sets the top accent line. When both broker and AI are scanned in a session, the Download PNG / SVG / Copy Text buttons auto-upgrade to the combined card
    • Copy-share-text now produces multi-line summary when multiple threats have been scanned
    • Vite aliases added: @ai-scanner, @face-scanner, @ai-catalog, @face-catalog
    • Mobile-responsive breakpoints for tabs + card grids

  • 6 new tests (tests/share-card.test.mjs — triple-threat tests): required-score validation, all-three rendering, ghost-column fallback, privacy invariant (no identity leaks), face-only scenario. Total: 293 → 299

  • Bundle impact: JS 145 KB → 201 KB (27 KB → 39 KB gzipped); CSS 7 KB → 12 KB (2 KB → 3 KB gzipped). Total gzipped still <45 KB

  • Deployed automatically to https://ramboxie.github.io/vanish/ on every push to main

  • ⚖️ Takedown Orchestrator (src/takedown/, new subcommand vanish takedown):

    • For anyone needing to remove non-consensual intimate imagery (NCII): OnlyFans/Patreon pirated content, revenge-posts by ex-partners, unauthorized reposts, deepfakes, content from past careers the user now regrets
    • Leak site DMCA catalog (12 sites): coomer.su, kemono.su, thothub.tv, bitchesgirls.com, leakgallery.com, erome.com, Pornhub (has faster non-DMCA form), XVideos, Telegram channels, Discord servers, Reddit subs, Twitter/X — each with abuse contact + takedownDifficulty + recommended approach
    • Hash registries (most effective free tools):
      • StopNCII.org (gold standard): Your images NEVER upload — hashes are generated locally in-browser, only the hash goes to the registry. Meta/TikTok/Bumble/Reddit/OnlyFans/Pornhub/Snap all scan against it and auto-block matches
      • Meta NCII Pilot (now integrated with StopNCII)
      • NCMEC CyberTipline (for anyone under 18 at time of creation)
    • Search engine removal: Google's dedicated intimate-imagery form (faster than general DMCA), Bing content removal, Yandex removal
    • 4 legal letter templates with jurisdiction-aware clauses:
      • DMCA §512(c) takedown notice (sworn statement + perjury attestation)
      • Cease & Desist letter (to individual distributor)
      • Police report narrative draft (cites Shield Act, state statutes)
      • Civil pre-suit demand letter (with statutory damages table)
    • Jurisdiction flags cite real law: --jurisdiction DMCA | SHIELD | TAKE-IT-DOWN | EU | UK | CA | AU
    • HMAC-signed audit trail of every drafted takedown — admissible as evidence in subsequent litigation
    • Support resources built in: CCRI hotline (1-844-878-CCRI), Revenge Porn Helpline UK, Australia eSafety Commissioner, NCMEC — surfaced via vanish takedown --support
    • Privacy-preserving by design: Vanish never stores your content, URLs you targeted, or the sites you visited. Just drafts letters + records audit
    • Unified CLI combines hash registration / search-engine removal / DMCA letters / legal templates in one tool

  • 31 new tests (tests/takedown.test.mjs) — catalog integrity, all 4 legal template rendering with substitutions, jurisdiction clause selection across 7 jurisdictions, leak site key resolution, DMCA notice planning, CLI smoke tests covering --stopncii / --dmca-letter / --cease-and-desist / --google-intimate / --support / --list flows. Total: 262 → 293

  • This addresses one of the most underserved privacy needs in the commercial tool landscape — DeleteMe/Optery/Incogni cover ZERO of this

  • 📚 Training Dataset Membership Check (src/dataset-check/, new subcommand vanish dataset-check):

    • Checks whether your content appears in 8 major public AI training datasets: Common Crawl, LAION-5B (via Have I Been Trained), The Pile, C4, OpenAI WebText, RedPajama, Dolma, FineWeb
    • Active query for Common Crawl — real HTTP call against the CDX Index Server across the 5 most recent monthly snapshots, returns per-snapshot hit count + digests
    • Walkthroughs for the other 7 datasets (no public query API) — each with: what's in the dataset, how to check manually (e.g., wimbd.apps.allenai.org for Dolma, c4-search.apps.allenai.org for C4, haveibeentrained.com for LAION), opt-out steps where possible (mostly robots.txt + CCBot)
    • Classification: low / moderate / high / critical based on Common Crawl hit count
    • --walkthrough-only mode skips network entirely; --json for machine-readable output
    • First open-source tool that actively queries training dataset indexes (vs only providing docs)

  • 20 new tests (tests/dataset-check.test.mjs) — catalog integrity, key resolution, mock fetch for Common Crawl CDX, fallback snapshot list, classification thresholds, CLI smoke

  • ⚖️ Third-Party AI Exposure (src/third-party-ai/, new subcommand vanish third-party-ai):

    • Catalog of 13 AI tools that OTHER people use on your data (you're the subject, not the user): Zoom AI Companion · Otter · Fireflies · Fathom · Gong · Chorus/ZoomInfo · Read.ai (facial analysis) · Teams Copilot · HireVue · Pymetrics/Harver · Abridge · Nuance DAX · Suki AI
    • Grouped by context: workplace / hr-recruiting / medical
    • Jurisdiction-aware objection letter generator — pick --jurisdiction EU|CA|IL|NY|HIPAA to get a legally-cited objection letter per context (GDPR Article 21, CCPA AB-331, Illinois AI Video Interview Act, NYC Local Law 144, HIPAA 45 CFR)
    • 4 letter templates: workplace-meeting, sales-call-recording (customer side), ai-interview (accommodation request), medical-ai (HIPAA decline)
    • --output <path> writes assembled letters to a file; --json for programmatic use
    • Designed to put vendors + deployers on notice — often enough to trigger accommodation without litigation

  • 25 new tests (tests/third-party-ai.test.mjs) — catalog + template references, jurisdiction clause selection, letter rendering with substitutions, context grouping, CLI integration

  • 🧠 LLM Memorization Check (src/llm-memory/, new subcommand vanish llm-memory-check):

    • Tests whether major LLMs (GPT-4o-mini, Claude 3.5 Haiku) output your personal identifiers verbatim when probed with 15 stalker-style prompts
    • Each probe is a realistic doxxing query ("What's X's phone number?", "Complete: X's email is..."), targeting email / phone / address / city / workplace leaks
    • Engine (memory-check-engine.mjs) ships with detectLeaks() that handles phone formatting variants, case-insensitive email matching, and skips too-short tokens that cause false positives
    • Two real providers: OpenAI + Anthropic (bring-your-own-API-key via OPENAI_API_KEY / ANTHROPIC_API_KEY env vars)
    • --dry-run mode uses mock provider so CI and curious users can run without API keys
    • Report shows per-provider leak rate (0-100%) + which leak types triggered + rating (safe/low/moderate/high)
    • --verbose flag includes per-probe response excerpts
    • --json for machine-readable output
    • Privacy-preserving by design: the report echoes your name (for identification) but NOT your email/phone/address — so the scan result itself doesn't become a leak vector
    • Cost: ~$0.01/scan in API fees (gpt-4o-mini + claude-3-5-haiku are cheap)
    • This is the FIRST open-source tool to check LLM memorization of arbitrary individuals — research labs do this but don't ship a tool

  • 30 new tests (tests/llm-memory-check.test.mjs) — probe catalog, renderProbe templating, detectLeaks fuzzy phone matching, provider creation, end-to-end mock runs, CLI integration. Total: 163 → 193

  • 🧹 AI History Cleanup Guide (src/ai-history/, new subcommand vanish clean-ai-history):

    • Catalog of 9 AI tools with documented cache paths per OS (Windows/macOS/Linux) and web-UI walkthroughs
    • Local-app targets: Cursor, VS Code + GitHub Copilot cache, Claude Desktop, ChatGPT Desktop
    • Web-UI targets: ChatGPT web, Claude web, Google Gemini/Activity, Perplexity, Grammarly
    • Does NOT auto-delete (by design) — instead, for each tool:
      • Discovers which cache paths actually exist on your OS
      • Reports approximate size per path
      • Prints the exact PowerShell/bash command you can copy-paste to delete
      • For web-UI tools: opens the settings page + prints 3-5 step walkthrough
    • Records HMAC-signed audit trail on user confirmation of deletion
    • --local-only / --web-only filters; --all for full wipe audit
    • Philosophy: destructive shell commands belong in your shell under your control, not hidden behind a tool. Vanish shows you where to look and exactly what to type.

  • 24 new tests (tests/clean-ai-history.test.mjs) — catalog integrity, cross-platform path expansion (%APPDATA%, ~/), byte formatting, filter logic, CLI integration on Windows/macOS/Linux paths. Total: 193 → 217

  • 👤 Face-Search Scanner + Opt-Out (src/face-scanner/, new subcommands vanish face-scan + vanish face-opt-out):

    • New catalog: 8 face-recognition services (PimEyes, FaceCheck.ID, FindClone, Lenso, TinEye, Yandex Images, Google Lens, Clearview AI)
    • Each service has: category (face-search / reverse-image / face-database), accessModel (free / freemium / paid / restricted), jurisdiction, pricing, knownFor description
    • vanish face-scan — opens each service's search page in your browser and prints a step-by-step walkthrough of how to upload your selfie + what free vs paid tiers show. Vanish never handles your photo — you upload it yourself on each service's own page
    • vanish face-opt-out — browser-assisted opt-out request for all 8 services including Clearview AI (where individuals can't even search themselves but have CCPA/GDPR deletion rights)
    • Per-service opt-out walkthroughs encode the real form flow (PimEyes: upload 1-3 photos + identity verification + email confirmation; Clearview: CCPA/GDPR request with government ID; Yandex: per-URL removal via search console)
    • Clearview AI gets 60-day reverify (slower processing); others 30-day reverify
    • Follow-up entries use kind: 'face-service' to distinguish from broker/AI opt-outs — shared queue
    • HMAC-signed audit trail with jurisdiction field (important for GDPR/CCPA evidence)
    • --free-only flag filters to services without paid tier; --all covers every service
    • Privacy notes built into scan walkthroughs — e.g., "PimEyes retains uploads 48h — create an account and delete after"

  • 21 new tests (tests/face-scan.test.mjs). Total test count: 142 → 163

  • Catalog at src/face-scanner/face-services-catalog.json. All 8 services verified April 2026

  • 🎯 AI Training Opt-Out (scripts/ai-opt-out.mjs, new subcommand vanish ai-opt-out):

    • Browser-assisted opt-out for 26 of the 30 AI platforms (4 are already safe by default)
    • Each platform has a walkthrough entry in ai-platforms-catalog.json (upgraded to schema v2):
      • targetSetting — exact UI string users search for (e.g., "Improve the model for everyone")
      • steps[] — step-by-step instructions (e.g., "Click profile → Settings → Data controls → toggle OFF")
      • verification — what success looks like ("toggle shows grey/off")
      • tierOverrides — when the opt-out isn't needed (e.g., "ChatGPT Team/Enterprise already opted-out")
    • Opens each platform's settings URL in your browser, prints the walkthrough, waits for your confirmation
    • --clipboard flag copies the exact toggle name to clipboard so you can Ctrl/Cmd+F it on the page
    • Records HMAC-signed audit trail + 60-day re-verify followUp (AI platforms silently reset settings after policy updates)
    • Follow-up entries use kind: 'ai-platform' to distinguish from broker opt-outs — shared queue, different reverify cadence
    • Three input modes: explicit flags (--chatgpt --linkedin), CSV (--use chatgpt,linkedin), or --all (every non-safe platform)
    • --no-open test mode skips browser open and auto-confirms (used in CI)

  • 13 new tests (tests/ai-opt-out.test.mjs) — walkthrough integrity, safe-platform handling, CLI batching, state persistence. Total test count: 129 → 142

  • Catalog version bumped 1 → 2 (walkthrough schema)

  • 🤖 AI Training Exposure Scanner (src/ai-scanner/, new subcommand vanish ai-scan):

    • New catalog: 30 major LLM platforms across 6 categories (chat / content / productivity / creative / email / dev)
    • Classifies each as exposed (opted-in by default), licensed (data sold to AI companies), safe (opted-out by default), action-needed (policy unclear), or not-applicable (you don't use it)
    • Per-platform metadata: default consent state, opt-out URL, opt-out method + difficulty, estimated time, data types used, AI models trained
    • Quick wins list (easy + medium difficulty opt-outs with URLs)
    • Licensed-content list (platforms that have sold your data — opt-out only affects future training)
    • Hard opt-outs list (GDPR/CCPA email-only paths)
    • Exposure score 0-100 with colored bar + risk level (critical/high/moderate/low)
    • Accepts explicit flags (--linkedin --twitter) or CSV (--use linkedin,twitter) or --all (worst-case)
    • Outputs: terminal banner + Markdown report + JSON (--json)
    • Zero data transmission — scan takes no personal information, just platform names
    • Covered platforms include ChatGPT, Claude, Gemini, Copilot, Meta AI, Perplexity, LinkedIn, Reddit, Twitter/X (Grok), Stack Overflow, Tumblr, Medium, Quora, Facebook, Pinterest, Grammarly, Notion AI, Otter, Zoom, Slack, Gmail, Outlook, GitHub Copilot, Cursor, Adobe, Canva, DeviantArt, Shutterstock, Figma, ArtStation
    • All data verified April 2026. Notes call out recent policy changes (LinkedIn 2024, Reddit deals, Meta EU objection form)

  • 20 new tests (tests/ai-scan.test.mjs): catalog integrity, classification logic, score boundaries, CLI integration. Total test count: 109 → 129

  • Static web app at web/:

    • Zero-install browser experience for non-developers
    • Dark-themed single-page app (Vite + vanilla JS, no framework)
    • Uses the same src/scanner/ modules as the CLI (single source of truth)
    • Share card preview + PNG/SVG download directly from browser
    • 58 broker opt-out cards with direct links, captcha warnings, processing time
    • 100% client-side — no server, no tracking, no cookies
    • Deployed via GitHub Actions to GitHub Pages on every push
    • URL: https://ramboxie.github.io/vanish/
    • Bundle size: ~145 KB (27 KB gzipped) including the 210-broker catalog

  • scan-engine.mjs refactored to be isomorphic (Node + browser):

    • Dropped node:module.createRequire and node:crypto
    • Uses globalThis.crypto.getRandomValues() (Node 20+ and all modern browsers)
    • options.catalog is now required — callers explicitly pass it
    • Call sites updated: scripts/scan-demo.mjs, src/wizard/engine.mjs, and tests

  • CI: New .github/workflows/deploy-web.yml — builds and deploys web app automatically when web/, src/scanner/, or the catalog changes

  • Share Card feature (src/scanner/share-card.mjs):

    • Privacy-preserving SVG card generator (1200×630, OG-image standard)
    • Shareable terminal banner printed at top of scan output by default
    • Contains ONLY aggregate score + category stats — no name, email, or phone
    • Color-coded by risk level (red/orange/yellow/green)
    • New --share-card <path> flag on scan command
    • New --no-banner + --no-color flags for CI / quiet modes

  • .github/workflows/test.yml — CI running tests on Ubuntu / macOS / Windows × Node 20 / 22

  • SECURITY.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md

  • CHANGELOG.md (this file)

  • Issue templates: broker-broken, bug report, feature request

  • .gitattributes normalizing line endings across platforms

Tests

  • 13 new tests in tests/share-card.test.mjs — banner consistency, SVG validity, privacy-preserving content, color-by-risk-level, CLI integration
  • Total test count: 95 → 108

0.2.0 — 2026-04-17

Major release: scan → opt-out → verify loop complete. 210 brokers cataloged, 58 with browser-assisted opt-out.

Added

  • Privacy Scanner (src/scanner/): heuristic 0-100 privacy score across 210 brokers, 5-factor confidence algorithm (dataTypeCoverage + categoryRisk + jurisdictionMatch + brokerReach + optOutComplexity), no external API calls
  • Browser-assisted opt-out (scripts/opt-out.mjs): opens real opt-out URLs in user's browser, pre-fills form data, user solves captcha + submits; records HMAC-signed audit trail + 30-day follow-up queue entry
  • Verify command (scripts/verify.mjs): HTTP liveness check on follow-up profile URLs. Classifies each as removed (404/410/redirect-to-root) / still-present (200 same URL) / unknown (403/429/timeout). Writes results back to queue state with verification fields
  • Catalog-driven architecture (broker-catalog.json): 210 brokers in single JSON file; registry auto-generates adapters. Replaces 23 individual broker .mjs files + template JSONs + official-endpoints.json
  • 200 → 210 broker expansion across 12 categories: 70 people-search, 21 public-records, 20 marketing-data, 18 background-check, 15 email-data, 14 phone-lookup, 12 financial, 8 social-media, 8 location-data, 7 reputation, 7 identity-resolution, 1 property
  • 58 brokers with opt-out support, including:
    • All 3 US credit bureaus (Equifax, Experian, TransUnion) + ChexSystems + LexisNexis + CoreLogic
    • Top people-search: Spokeo, Whitepages, BeenVerified, Intelius, Radaris, TruePeopleSearch, PeekYou, etc.
    • B2B marketing: Acxiom, LiveRamp, Oracle BlueKai, Epsilon, ZoomInfo, Clearbit, Neustar
    • Phone lookup: Truecaller, Hiya, USPhoneBook, SpyDialer, RoboKiller, etc.
  • 18-state wizard (was 13) with scan phase prepended: SCAN_WELCOME → SCAN_INPUT → SCAN_RUNNING → SCAN_REPORT → SCAN_HANDOFF → [existing 13-state cleanup]
  • Live broker factory (_live-broker.mjs): reusable factory for brokers that support real HTTP submission (currently 8: spokeo, thatsthem, peekyou, addresses, cocofinder, checkpeople, familytreenow, usphonebook)
  • Unified CLI router (scripts/index.mjs): vanish scan|opt-out|verify|wizard|cleanup|queue|... subcommands
  • Zero-install via npx: npx github:RAMBOXIE/vanish scan --name "..."
  • Clawhub publishing metadata in SKILL.md frontmatter
  • Follow-up queue (followUp[] in data/queue-state.json) for 30-day re-verification

Changed

  • Audit HMAC key now required in production (warns in dev, silent in test) — VANISH_AUDIT_HMAC_KEY env var
  • Secret store upgraded to scrypt KDF + per-secret salt (backward-compatible with legacy SHA-256)
  • Queue state-store added stale-lock detection (30s timeout, PID liveness check)
  • --simulate transient-error now works in live mode (was dry-run only)
  • b1-live.mjs --brokers argument now accepts comma-separated list for multi-broker live submission
  • README rewritten: value prop first, competitor comparison table, 3-step "how it works"
  • Version bumped 0.1.0 → 0.2.0

Fixed

  • Hardcoded D:/Projects/vanish paths replaced with import.meta.url-relative paths (wizard engine, 2 test files)
  • Queue state lock no longer leaks on process crash (stale detection)
  • Audit canonical JSON handles Date/undefined/circular values via pre-sanitization

Security

  • scrypt KDF with per-secret salt for secret store (replaces weaker SHA-256 derivation)
  • HMAC-SHA256 signing with canonical JSON serialization + timing-safe comparison
  • Fail-loud audit key requirement in production

[0.1.0] — earlier

Initial dry-run MVP with 23 brokers and P0 safety gates (manual trigger, triple confirm, export decision). See git history for commit-level detail prior to v0.2.