v1.3.2 — PDF-by-default for Resume/Cover, fix silent Markdown fallback

[1.3.2] - 2026-07-12

Fixed

• Resume/Cover Letter silently uploaded to Drive as Markdown instead of
  PDF. The [pdf] extra (fpdf2) was a separate, easy-to-forget install
  step from [drive] — a missing fpdf2 degraded every Resume/Cover
  upload to .md with only a per-file warning, easy to miss. fpdf2 is
  now bundled into the [drive] extra, so pip install -e ".[drive]"
  alone gets PDF rendering by default; careeros doctor (with
  drive.enabled: true) now also proactively checks for it and FAILs with
  a clear fix if it's ever missing, instead of only surfacing a buried
  warning during daily.

• Only Resume and Cover Letter ever attempt PDF now. Application
  Answers, Evaluation, and Deep Report always upload as Markdown — not
  "PDF with a fallback," simply never attempted, since only Resume/Cover
  are ever attached to a job application.

• A Resume/Cover previously uploaded as .md (before PDF rendering was
  available) is no longer left orphaned in Drive once PDF rendering
  works. Drive matches existing files by exact filename, so a new
  .pdf upload didn't replace an old .md — both used to sit side by
  side in the flat Drive folder. Re-uploading now explicitly deletes the
  stale .md when a .pdf supersedes it for the same job.

• Edge case, found during this release's own verification: the stale-.md cleanup above now also handles two different jobs that happen to share the exact same Company/Role string (a genuine name collision, disambiguated with a numeric suffix like ... (2).md) — the cleanup previously only checked the un-suffixed base filename.

v1.3.1 — Applyability scoring contract, live provider quota, anti-scripting guardrail

[1.3.1] - 2026-07-12

Added

• careeros doctor --live. Opt-in flag that actually verifies
  Fantastic Jobs and every configured Apify token against their real APIs
  right now (a small, bounded amount of real quota: one 1-record fetch,
  plus a free account-usage check per Apify token, no actor run) instead of
  only reporting local/stored state — catches a bad or exhausted key before
  your first daily run, not mid-discover.
• Live provider-quota surfacing. The discovery summary table now shows
  each provider's real, provider-reported remaining quota (e.g. Fantastic
  Jobs' x-ratelimit-* headers) when available, next to its record count
  and timing.

Fixed

• Score/recommendation could disagree. evaluate's score is meant to
  mean applyability — a green (≥ threshold) score should always mean the
  candidate can actually apply — but a job with a real deal-breaker (an
  onsite location outside your accepted cities, a stated preference
  violation) could still show a high score alongside
  recommendation: "skip". careeros evaluate --finalize now
  deterministically caps a skip-recommendation eval's stored score below
  threshold, so the two signals can never disagree, regardless of which
  agent/model produced the evaluation. The 5 rubric dimensions themselves
  stay honest and untouched — only the final score is clamped — so the
  "why" behind a skip is still legible. prompts/eval_v2.md documents the
  contract explicitly, including an anchor scale for scoring logistics
  honestly as a preference-ranking signal rather than a pass/fail gate.
• Fantastic Jobs' weekly quota counter was locally calculated, not
  live-verified. The local .careeros/discovery_budget.json weekly
  record counter is independent of which API key is configured, so
  rotating in a fresh key did nothing to clear a "quota exhausted" state.
  The counter is now advisory-only; the live API call (and its
  x-ratelimit-* response headers) is authoritative, so a fresh key works
  immediately.
• An exhausted Apify token stayed "exhausted" for the rest of the billing
  cycle, even after a mid-month top-up on the same token, with zero
  re-verification. A token now only pre-emptively skips on the same day
  it was marked exhausted — any other day gets one fresh live retry before
  being trusted as exhausted again. Old on-disk state (the pre-1.3.1
  bare-list format) migrates automatically.
• AI reasoning stages could be silently replaced by scripts. A new
  standing rule in AGENT_GUIDE.md ("Reasoning stages must be reasoned,
  never scripted"), referenced from prompts/gate_v1.md,
  prompts/eval_v2.md, and skills/daily.md's Gate/Evaluate steps, makes
  explicit that Gate/Evaluate/Resume/Cover/Application-Answer output must
  always come from real per-job reasoning — never a keyword-matching or
  formulaic script standing in for it, even under batch-size pressure.

v1.1.0

Added

• Automatic Application Answers. For every Apply-tier job, daily now
  drafts real answers to that job's actual application questions — no
  waiting until you've manually opened the form. A new background form-reader
  (careeros/apply/browser.py) fetches the form's visible text: a
  zero-dependency HTTP tier first, an optional headless-Playwright fallback
  only when the form genuinely needs JavaScript to render. Neither tier uses
  per-site scraping selectors — the agent identifies the real questions from
  plain text and drafts grounded answers the same "selector, not writer" way
  as the resume.
• careeros publish <job-id> — upload one job's current artifacts
  (resume, cover, evaluation, deep report, application answers — whichever
  exist) to Drive and patch just that Sheet row, without a full daily run.
  The prep and on-demand apply <job-id> skills now run this
  automatically as their last step.
• A specific status for every unreadable form, replacing one generic
  fallback: ✅ Generated, 🔒 Login Required, ❌ Closed, ⚙️ Playwright
  Missing, 📄 No Essay Questions, 🌐 Network Error, 🛡️ Bot-Blocked, or the
  generic Manual review required only when none of those match. Each is
  detected generically (login walls, closed-posting notices, Cloudflare-style
  bot challenges, and pages that render real text but never reach an actual
  form) — no per-ATS logic.
• A Status column in the Sheet (dropdown: Not Applied by default,
  Applied, Received Call, Interview, After Interview,
  Ongoing / In Process, Offer, Rejected) for tracking your own
  application progress by hand. The pipeline sets the default on a new row
  and never touches it again, exactly like Notes.
• New Sheet columns: Evaluation (Drive), Deep Report (Drive),
  Application Answers (Drive) — direct, per-job links to artifacts that
  were already being generated (and, for Evaluation, already uploaded) but
  had nowhere to show up.
• Newest-on-top Sheet ordering. New rows are now inserted directly below
  the header instead of appended at the bottom, so each day's run reads at
  the top without scrolling past a growing history.
• careeros sheets sync-status — patch the Application Answers status
  of existing Sheet rows after re-running apply --prepare/--finalize
  (e.g. reclassifying old jobs into the newer, more specific statuses)
  without appending a duplicate row.
• Personal/logistics questions, asked once. Notice period, work
  authorization/visa status, salary expectations, earliest start date, and
  employment type are the same answer on every application. The first time
  one is genuinely needed and missing, it's asked once and saved to
  profile.yaml's new logistics: section — every later application,
  batch or on-demand, reuses it automatically. Voluntary EEO/demographic
  self-identification questions are deliberately excluded and always left
  for you to answer yourself.
• careeros --version — prints the installed version and exits.
• careeros doctor now checks Playwright independently at two levels: the
  [apply] extra's Python package, and the chromium browser binary
  (playwright install chromium) — "package installed but browser binary
  missing" and "package not installed at all" are reported as distinct,
  actionable messages instead of one opaque failure.

Changed

• Sheet columns removed: Resume Path, Cover Letter Path,
  Report Path (local filesystem paths, useless outside your own machine)
  and Drive Folder (redundant once every artifact has its own direct
  link). careeros sheets migrate / sheets append remove these
  automatically from an existing Sheet — see Migration notes below.
• CI now installs the optional [apply] extra and runs
  playwright install chromium --with-deps, so the real headless-browser
  fetch path runs in automated tests instead of being skipped.
• The package version now has a single source of truth
  (pyproject.toml); careeros.__version__ reads it back via
  importlib.metadata instead of duplicating the string.

Fixed

• The Playwright fallback used wait_until="networkidle", which never
  fires on pages with persistent background network activity (analytics
  beacons, a bot-check's own verification polling) — it could hard-timeout
  with zero text captured even though the real content rendered within a
  second or two. Switched to wait_until="load" plus a short fixed buffer.
• PDF rendering crashed instead of falling back to Markdown when resume,
  cover, or answers content contained a non-Latin currency symbol (₹, €, £,
  ¥) — blocking the entire Drive upload for that job. Now sanitized to a
  latin-1-safe equivalent, and any other still-unmappable character falls
  back to Markdown instead of failing the upload.
• careeros sheets sync-status could overwrite a just-published
  Application Answers link with a blank cell, because the local
  drive_links.json cache isn't refreshed by careeros publish. It now
  never touches a generated-status job's cell — only publish does.
• A login-wall, closed-posting, or bot-blocked page returns real,
  substantial (non-empty) text — it's just the wrong page. The batch apply
  stage now checks for these BEFORE treating any non-empty fetch as a
  readable form, so a LinkedIn login-wall page can no longer be silently
  sent to the drafting step as if it were the real application form.

Migration notes: upgrading an existing Sheet

If you were running CareerOS before this release, your Google Sheet needs
one one-time cleanup pass:

careeros sheets migrate

This removes the four deprecated columns, adds the four new ones, applies
header/Score/Status formatting, backfills Status to Not Applied on
existing rows, and sorts your Sheet's existing rows by Date descending (a
one-time fix for history that was written bottom-up before this release).
It's safe to re-run — every step is idempotent, and a Sheet already on the
current schema is a no-op. After this one pass, every future daily run
keeps the Sheet current automatically; you never need to run it again
unless you skip several releases at once.

CareerOS v1.0.0

CareerOS v1.0.0

An AI-powered, deterministic job discovery and recommendation engine.
Deterministic where possible, AI only where reasoning adds value.

This is the definitive v1.0.0 — hardened through a full production audit
(P2.8) after the initial public commit, with discovery now permanently
frozen on evidence.

Highlights

• End-to-end daily pipeline: discover → normalize → dedupe → constraints →
  AI gate → evaluate → two-tier threshold → resume/cover artifacts → daily
  report → Google Sheets — driven from your host coding CLI (Claude Code,
  Codex, Gemini CLI, etc.), zero server, zero database.
• Discovery (frozen default): queries both Fantastic Jobs endpoints —
  active-ats (career sites/ATS) and active-jb (+LinkedIn/YC/Wellfound) —
  every run, splitting the record budget 50/50 rather than doubling it.
  Backed by a full 107-job production acceptance audit: the two sources
  score a statistically equal ~8% ≥4.0 rate but are 92% disjoint, so
  querying both roughly doubles interview-worthy jobs found at the same
  quota cost. A single endpoint remains selectable via config.
• Quota guard: recommends a daily discovery limit from your plan/quota/
  goal, warns before exhaustion, and never silently overrides your
  configured limit — fully config-driven, no hardcoded numbers.
• Two-tier selection: jobs scoring ≥ threshold (default 4.0) get the
  full pipeline (resume, cover, report, Drive, Sheet row); jobs in the
  Consider band (default 3.5–3.9) get a Sheet row with score + a concise
  reason only — no AI artifacts, no Drive cost — so near-misses stay
  visible for free.
• Profile-driven segmented discovery — one query per work-mode tier instead
  of a single broad fetch.
• Cache-everything AI stages, fingerprinted on job content + profile version
  • prompt version — an unchanged re-run costs zero tokens.
• Optional Google Drive backup of shortlisted artifacts, additive and off
  by default. New step-by-step setup guide: docs/google-setup.md.
• 204 tests, including a parity test proving the REST and legacy-actor
  providers map identical raw data to an identical Job shape, and a
  regression suite for the two-tier selection, quota guard, and Sheets
  schema-migration logic.

What changed in the P2.8 hardening pass

• Fixed a real Sheets bug: dedup and row-writes now key columns by header
  name, not position, with additive self-healing migration for sheets
  created under an older column layout (dedup was silently broken on drift).
• Fixed the day-level summary re-deriving Apply/Consider from raw evals
  independently of the actual selection — it now renders the same
  partitioned lists the pipeline already computed, so the report can never
  disagree with what's in your Sheet.
• Fixed the quota guard crashing discover on a non-ISO --date label.
• Removed dead code found during the audit (an unused selection function,
  unused helper constants) and corrected several stale docs.

Installation

pip install -e .

See the README for Quickstart and provider setup.

Advanced configuration

The default REST provider supports two interchangeable transports for the
same Fantastic Jobs dataset: the direct API
(developer.fantastic.jobs) or RapidAPI's
"Active Jobs DB" listing — pick whichever fits your pricing/free-tier needs
via config.api.transport. Discovery endpoint (both/active-ats/
active-jb) and the quota guard's plan/limit are also configurable — see
careeros/providers/README.md and templates/config.example.yaml.

Discovery is now frozen

No further discovery changes (endpoint, query strategy, provider set) unless
a production issue is found. Future work moves to P3 (outcome tracking) and
the v1.1 onboarding/Drive+PDF overhaul.