v0.1.0-alpha.1 — bundled SearXNG + polish

Post-alpha polish: free-by-default search, more honest UI feedback, and a handful of latent bug fixes surfaced by real use.

Added

• SearXNG is now bundled in compose. docker compose up -d now starts a self-hosted SearXNG container alongside the rest of the stack. Configure by setting SEARXNG_SECRET (recommended: openssl rand -hex 32) and picking searxng as the search provider in Settings → Sourcing — no third-party account, no card. Config lives in searxng/settings.yml.
• Per-source result limit input in the Search tab. A small numeric control (1–30, default 5) next to the source chips lets you request more results per search.
• Paginated SearXNG provider. The provider now walks pageno=1..5 until it has limit distinct results (URL-deduped). Earlier it asked only for page 1, so limit_per_source above ~10 was silently ignored.

Changed

• SMTP settings form now pre-fills from saved values. The non-secret SMTP fields round-trip through GET /api/settings. The password remains write-only; leaving it blank on save now keeps the stored password instead of overwriting it.
• "Needs paste" is now explicit. When a LinkedIn URL can't be auto-extracted, the kanban card label, the detail page stage, and an amber banner all say "needs paste" instead of the misleading "extracting".
• Gmail channel hidden from the Notify wizard. The backend route returns 501 for that channel, so the option is removed until it's wired up.

Fixed

• PATCH /api/candidates/{id} no longer crashes with MissingGreenlet when the candidate row has a non-null updated_at.
• The e2e candidate-edit spec restores the original email in test.afterEach and asserts the restore response, so a mid-test failure no longer leaks dirty state into subsequent runs.
• tests/unit/test_webpage_fetcher.py was asserting on stale error copy from a previous rewrite. Updated to assert on the HTTP status embedded in the message instead.

Upgrade

cd ~/recruiter-agent
git pull
echo "SEARXNG_SECRET=$(openssl rand -hex 32)" >> .env
docker compose up -d --build

If you were running a standalone searxng container outside compose, remove it first (docker rm -f searxng) — the bundled service replaces it.

Full commit log: v0.1.0-alpha...v0.1.0-alpha.1

v0.1.0-alpha — first public preview

Changelog

All notable user-facing changes are recorded here. The format is loosely based
on Keep a Changelog and the project
follows semantic versioning post-1.0.

v0.1.0-alpha — 2026-05-24

First public preview. Single-tenant, self-hosted recruiter assistant: add a
candidate from a URL, CV, paste, or search, and the system extracts a
structured profile, enriches it from public sources, scores it against your
job's weighted criteria, and lets you invite the candidate by email with a
calendar attachment.

Added

• Job + criteria. Create jobs with title, description, and 3–6 weighted
  criteria. The LLM proposes initial criteria from the JD; weights can be
  edited freely and are normalized to sum to 1.0.
• Candidate sources. LinkedIn URL · GitHub URL · arbitrary web page ·
  paste · PDF/DOCX upload · search-and-add across LinkedIn / GitHub / Web.
• LinkedIn extraction tiers. Apify (commercial scrape) → Playwright with
  a stored li_at cookie → GitHub-by-name fallback → manual paste prompt.
• Pipeline. Async orchestrator runs fetch → extract → enrich → score
  with SSE-pushed stage updates to the kanban.
• Scoring. LLM-driven per-criterion score (0–100) with a one-sentence
  rationale and a weighted overall.
• Validate → Notify wizard. Four-step flow (channel → slots → LLM-drafted
  email → confirm) that sends an SMTP invite with an ICS attachment and
  moves the candidate to Invited.
• Candidate editing. Inline edit form for full_name, email, phone,
  headline, location, summary (the rest stay LLM-extracted).
• Search-query suggester. "Suggest from JD" button in the Search tab
  generates a source-aware query (LinkedIn-style keywords / GitHub
  language: filters / quoted Web Boolean) from the job description.
• Conversational agent. 11 registered tools, reachable from the Chat
  panel on every candidate detail page. Reversible actions go through an
  undo store.
• Settings. LLM (provider, key, model id, monthly spend cap) · Sourcing
  (search provider + key, GitHub PAT, Apify token + actor, LinkedIn cookie
  or credentials) · Enrichment (Twitter, YouTube, Stack Exchange) ·
  Notifications (SMTP host/port/user/password/from) · Profile (recruiter
  name + email).
• Structured rejection. Reject dialog captures a reason; banner on the
  detail page until the candidate is Unrejected.
• Eager-seed bootstrap. First start with RECRUITER_DEFAULT_ACCOUNT_*
  set creates the admin row before serving requests.
• e2e tests. Playwright suite (auth · candidate edit · search-query
  suggester) under recruiter-frontend/e2e/ runnable via npm run e2e.
• CI. GitHub Actions runs backend pytest (testcontainers postgres) and
  frontend vitest + typecheck on every push and PR.

Security

• All secrets in the DB (LLM keys, search keys, GitHub PAT, Apify token,
  LinkedIn cookie + password, SMTP password, OAuth tokens) are encrypted
  with Fernet under RECRUITER_SETTINGS_KEY.
• Session cookies are HttpOnly, SameSite=Strict, SHA-256-hashed at rest;
  login is rate-limited (5/minute).
• Compose refuses to start when required secrets are missing — no weak
  fallback can ship.
• CSRF Origin allowlist gates every mutating endpoint, configurable via
  RECRUITER_ALLOWED_ORIGINS.

Known limitations

• Skills / experience / education / links are LLM-extracted only — not
  editable in the UI yet. Edit via SQL or re-paste the source if extraction
  got it wrong.
• Gmail + Google Calendar channel is not wired — only SMTP is functional;
  the Gmail option is hidden from the Notify wizard.
• SerpAPI is the default search provider but requires an API key. Brave
  and SearXNG are supported alternatives.
• Playwright e2e is not yet in CI — runs locally against docker compose up -d. CI runs backend pytest + frontend vitest + typecheck only.
• No upgrade story between releases yet — schema migrations work, but
  there is no formal pre-upgrade backup procedure documented beyond
  "back up RECRUITER_SETTINGS_KEY and the postgres volume."

Recently fixed (since the previous unstable HEAD)

• PATCH /api/candidates/{id} no longer crashes with MissingGreenlet when
  the candidate row had a non-null updated_at.
• SMTP settings form now pre-fills host/port/user/from on open and keeps the
  stored password on partial updates instead of silently overwriting it
  with an empty string.
• LinkedIn-URL applications stuck waiting on manual paste now surface an
  explicit "needs paste" banner instead of an indefinite "extracting" label.
• LLM gateway errors propagate the upstream response body (e.g. LINAGORA's
  "Model not found") instead of just the HTTP status.
• LLM tab exposes a per-provider Model input.
• README rewritten and shortened (~215 lines, down from ~666).
• First-run admin user is eager-seeded at backend startup rather than lazily
  on first successful login.