v0.1.0

Initial public release.

Added

• MCP server over stdio (dist/index.js) for Claude Desktop, Claude Code and any MCP client, with tools
  for session, jobs, applicants, local data, queue/pacing and debugging, plus a review_applicants prompt.
• Drives the user's own installed Google Chrome through patchright
  in a dedicated profile: persistent mode (launch Chrome, visible window, real fingerprint) and cdp mode
  (attach to a Chrome started with scripts/launch-chrome.sh). Never headless, never downloads a browser,
  never types credentials: the user signs in by hand (npm run login / browser_open_login).
• Posted-jobs crawl (jobs_sync, jobs_list) over every jobState= tab including Closed, with
  applicant counts, status and dates parsed from the 2026 card layout and the legacy layout.
• Applicant-list crawl (applicants_sync) supporting both LinkedIn hiring dashboards: the classic list
  (25 per page, crawled per rating bucket including the hidden Not a fit, sorted by applied date) and the
  2026 "Hiring Pro" list (rating=ALL&sort=DateApplied, numbered pages, resumable via start= offsets).
  Progress is saved after every page; rows are deduplicated by application id.
• Application details (applicants_fetch_details): contact email/phone when shared, screening-question
  answers, your rating, and the resume file (PDF/DOCX) including popup-based downloads, with text extraction
  (unpdf / mammoth) into a full-text-searchable SQLite index.
• Full profiles (applicants_fetch_profiles, or chained after details): about, experience, education,
  skills, certifications, languages, projects, honors. LINKEDIN_MCP_PROFILE_STRATEGY = auto (visit like a
  person, then one in-page Voyager request), voyager or dom (page text only).
• Raw snapshots next to every parsed record (raw-*.json, list pages, captured JSON) so parsers can be
  fixed and re-run offline after LinkedIn changes its markup.
• Human-like pacing: working hours and days (with per-day jitter, IANA time zone), daily caps for
  application pages and profile views with a warm-up ramp, hourly action cap, random long breaks, log-normal
  delays, shuffled task order, occasional feed visits, Bezier mouse paths, uneven scrolling and variable typing
  rhythm. slow / normal / brisk speed presets; runtime overrides with pacing_set (persisted).
• Resumable SQLite-backed queue with queue_status (counts, today's usage vs caps, working-hours state,
  next eligible time, per-job list progress, checkpoints), queue_start / pause / resume / cancel /
  retry_failed / tasks, and a single-owner lock so the MCP server and the standalone worker
  (npm run worker) never drive the browser at the same time.
• Checkpoint handling: verification / unusual-activity / login pages stop the worker, requeue without
  burning an attempt and report needsHuman; HTTP 429/999 back off for hours; three consecutive unexpected
  failures trigger a cool-down.
• Local data tools: applicants_list (filters + full-text search), applicant_get, resume_text,
  applicants_export (CSV / JSON / JSONL). Instant, no LinkedIn traffic.
• Local dashboard (npm run dashboard, 127.0.0.1:4173 only): jobs, progress meters, applicant table,
  detail drawer, exports.
• Debug tools for when LinkedIn changes: debug_snapshot, debug_navigate (linkedin.com only),
  debug_page_text, debug_find, debug_click, debug_captures.
• Two DOM generations (legacy Ember/artdeco and the 2026 SDUI / Hiring Pro UI) handled through a single
  selector map (src/linkedin/selectors.ts) with runtime generation detection.
• Data directory (~/.linkedin-applicants-mcp, LINKEDIN_MCP_DATA_DIR) created with owner-only permissions;
  JSON-lines logs; all configuration via LINKEDIN_MCP_* environment variables (see .env.example).
• Unit tests and a stdio smoke test that spawns the built server (no browser, no LinkedIn); research notes
  under docs/research/.