Files
jobboard/plugins/ats.py
T
WolfandClaude Opus 3abb44e10a Stream a board's body against MAX_BYTES and a 60 s deadline; read timeout 30 s
_get streams the body with a running byte count, so an oversized board is
refused at MAX_BYTES instead of after it is held whole, and bounds the whole
request, redirects included, by DEADLINE, checked between chunks. The
per-phase timeouts are httpx.Timeout(10, read=30), set from 84 live requests
against mongodb, feedzai and the five Greenhouse boards with most open roles,
with and without pay_transparency=true: the body arrives in under 1 s, the
first byte waited up to 16.6 s, and the flag made no measurable difference.
_paged works unchanged through it (tested over a mock transport, 429 included).

Correct pull.py's BATCH_SIZE comment: a batch reads its boards in sequence,
and the parallelism is the http pool's.

Co-Authored-By: Claude Opus <noreply@anthropic.com>
2026-09-29 15:12:14 +02:00

932 lines
43 KiB
Python

"""ATS boards: Greenhouse, Lever, Ashby, Workable, SmartRecruiters, Recruitee, Personio.
Every one of these hosts is a company's own careers page speaking public,
unauthenticated JSON (Personio: XML) — first-party, no scraping, and a role
that drops off the board is *closed*, which is the cheapest staleness check
there is.
The field map and the redirect allow-list are ported from
`openings.go` and `cmd/seedgen/main.go` in
[jobhunt-hq](https://github.com/rembrandtreyes/jobhunt-hq) (MIT), including
`classify`'s nine functions and their keywords. Workable, SmartRecruiters,
Recruitee and Personio are new: jobhunt-hq only read Greenhouse, Lever and
Ashby, so those four are written from one real board of each host, fetched
while this file was written (2026-09-23; Workable on 2026-09-28, when the
first board with roles was seen), rather than from the Go source.
The body cap is not a straight port — see `MAX_BYTES` below.
A target is `host:slug`, e.g. `greenhouse:databricks`, `ashby:deepl`,
`lever:acme`. Every board is fetched independently; one that fails or times
out is logged, recorded as failing under `ats:failing:<target>` (see
`ATSSource.fetch`), and skipped, and the rest continue.
Workable and SmartRecruiters page their list and state a total; every other
host answers with the whole board at once. A paged board counts as complete
only when it read as many distinct roles as the total says (see `_paged`).
"""
from __future__ import annotations
import html
import json
import logging
import re
import time
import xml.etree.ElementTree as ET
from collections.abc import Callable, Iterator
from datetime import UTC, datetime
from typing import Any
import httpx
from desk.sources import (
API_VERSION, ARTICLE_READER, Item, PluginSpec, Profile, Source,
SummaryPolicy, TaggingPolicy, TargetSpec, Tasks, USER_AGENT, strip_html,
)
log = logging.getLogger(__name__)
#: openings.go's 8 MB body cap is not kept: it was written against arbitrary URLs, and every
#: URL fetched here is `?content=true` against a fixed allow-list of seven ATS
#: API hosts. The real run of 2026-09-23 hit it on the biggest employers —
#: Anduril at 41 MB, Anthropic, OpenAI, Databricks, Carvana, Axon and
#: Airwallex between 9 and 19 MB — and skipped their boards outright. Raised
#: to 96 MB so those boards fit; a board that still exceeds it gets one retry
#: (Greenhouse only, see `_fetch_rows`) without `?content=true`, so at least
#: the titles land.
MAX_BYTES = 96 << 20
#: openings.go's 10 s did not hold: `greenhouse:mongodb` and `greenhouse:feedzai`
#: failed a pull of 2026-09-28 with a read timeout. Measured live on
#: 2026-09-29, three runs each of mongodb, feedzai and the five Greenhouse
#: boards with the most open roles (databricks, stripe, anthropic, feverup,
#: datadog), with and without `pay_transparency=true`, cached and with a
#: cache-busting parameter (84 requests; CloudFront answered every one
#: "Miss", so all reached the origin): the body arrives in under 1 s once it
#: starts (at most 0.09 s between two chunks, 10.4 MB the largest), but the
#: first byte sometimes waits: 16.6 s (feverup, no flag), 14.4 s (anthropic,
#: flag), 10.9 s (feverup, flag), 5.2 s and 3.6 s, and under 2.6 s otherwise.
#: The flag made no difference that three runs could tell apart from that
#: noise. So a read may wait 30 s, about twice the worst first byte seen,
#: and a whole request, redirects and body included, 60 s (see `_get`),
#: about four times the slowest seen, which still leaves a 96 MB body room
#: on a link far slower than the one measured.
TIMEOUTS = httpx.Timeout(10.0, read=30.0)
DEADLINE = 60.0
#: host -> (feed URL template, careers URL template). `{slug}` is filled in;
#: recruitee and personio put the slug in the hostname instead of the path.
_BOARDS: dict[str, tuple[str, str]] = {
"greenhouse": ("https://boards-api.greenhouse.io/v1/boards/{slug}/jobs"
"?content=true&pay_transparency=true",
"https://job-boards.greenhouse.io/{slug}"),
"lever": ("https://api.lever.co/v0/postings/{slug}?mode=json",
"https://jobs.lever.co/{slug}"),
"ashby": ("https://api.ashbyhq.com/posting-api/job-board/{slug}"
"?includeCompensation=true",
"https://jobs.ashbyhq.com/{slug}"),
"workable": ("https://apply.workable.com/api/v3/accounts/{slug}/jobs",
"https://apply.workable.com/{slug}"),
"smartrecruiters": ("https://api.smartrecruiters.com/v1/companies/{slug}/postings",
"https://jobs.smartrecruiters.com/{slug}"),
"recruitee": ("https://{slug}.recruitee.com/api/offers/",
"https://{slug}.recruitee.com"),
"personio": ("https://{slug}.jobs.personio.de/xml",
"https://{slug}.jobs.personio.de"),
}
#: host -> the name the host gives itself: the hosts this plugin reads, in
#: the order a select offers them. What the Sources page's host select is
#: rendered from, so it cannot offer a host `is_known_target` refuses.
HOSTS: dict[str, str] = {
"greenhouse": "Greenhouse", "lever": "Lever", "ashby": "Ashby",
"workable": "Workable", "smartrecruiters": "SmartRecruiters",
"recruitee": "Recruitee", "personio": "Personio",
}
assert set(HOSTS) == set(_BOARDS)
#: The hosts a request or a redirect may land on for a given board host — a
#: fixed API host for five of them, and the company's own subdomain for the
#: two that put the slug in the hostname. Anything else is refused, at the
#: first hop and at every redirect: a feed is external content, and a board
#: row must never be able to point this fetch at an internal address.
def _allowed_host(host: str, slug: str) -> Callable[[str], bool]:
fixed = {
"greenhouse": "boards-api.greenhouse.io",
"lever": "api.lever.co",
"ashby": "api.ashbyhq.com",
"workable": "apply.workable.com",
"smartrecruiters": "api.smartrecruiters.com",
}
if host in fixed:
want = fixed[host]
return lambda h: h == want
if host == "recruitee":
want = f"{slug}.recruitee.com"
return lambda h: h == want
if host == "personio":
want = f"{slug}.jobs.personio.de"
return lambda h: h == want
return lambda h: False
def _is_web_url(url: str) -> bool:
"""A trimmed http(s) URL, or refuse — a feed is external content and a
`javascript:` or `data:` link must never reach an href. Port of
openings.go's `webURL`."""
url = (url or "").strip()
if not url:
return False
parsed = httpx.URL(url)
return parsed.scheme in ("http", "https") and bool(parsed.host)
class _Refused(Exception):
"""A board answered somewhere its allow-list does not cover."""
class _TooLarge(ValueError):
"""A response exceeded `MAX_BYTES`. A `ValueError` so a caller that only
knows the old shape still catches it; `_fetch_rows` catches it by type to
retry Greenhouse without `?content=true`."""
def _get(client: httpx.Client, url: str, host: str, slug: str, *,
method: str = "GET", payload: dict[str, Any] | None = None) -> bytes:
"""Fetch one board feed: https only, the host's own allow-list, at most
three redirect hops each checked the same way, and the body cut off at
`MAX_BYTES`. Port of openings.go's `fetchOne`/`feedCache.allow`. A POST
sends `payload` as JSON, `{}` when there is none.
The body is streamed and counted as it arrives, so an oversized one is
refused at `MAX_BYTES` rather than after it has been read whole, and
the whole request, redirects included, must end within `DEADLINE`: the
per-read `TIMEOUTS` alone let a board that trickles bytes run on for
good. The deadline is checked between chunks, so it can be overrun by
at most one read timeout."""
allow = _allowed_host(host, slug)
deadline = time.monotonic() + DEADLINE
current = url
for _hop in range(4):
parsed = httpx.URL(current)
if parsed.scheme != "https" or not allow(parsed.host):
raise _Refused(f"{current} is not an allowed host for {host}")
with client.stream(
method, current, timeout=TIMEOUTS, follow_redirects=False,
headers={"User-Agent": USER_AGENT, "Accept": "application/json, text/xml"},
json=(payload or {}) if method == "POST" else None,
) as response:
if response.is_redirect:
location = response.headers.get("location") or ""
if not location:
raise ValueError("redirect with no location")
current = str(httpx.URL(current).join(location))
continue
response.raise_for_status()
chunks: list[bytes] = []
size = 0
for chunk in response.iter_bytes():
size += len(chunk)
if size > MAX_BYTES:
raise _TooLarge(f"response too large (over {MAX_BYTES} bytes)")
if time.monotonic() > deadline:
raise httpx.ReadTimeout(f"no whole answer within {DEADLINE:.0f} s",
request=response.request)
chunks.append(chunk)
return b"".join(chunks)
raise ValueError("too many redirects")
def _parse_dt(value: Any) -> datetime | None:
"""An aware datetime from whatever a board sends: ISO-8601 with an
offset, `Z`, a bare date, epoch milliseconds, or "YYYY-MM-DD HH:MM:SS
UTC". None for anything else, which `Item` reads as "now"."""
if value is None:
return None
if isinstance(value, int | float):
if value <= 0:
return None
seconds = value / 1000 if value > 10**12 else value
return datetime.fromtimestamp(seconds, tz=UTC)
text = str(value).strip()
if not text:
return None
try:
parsed = datetime.fromisoformat(text.replace("Z", "+00:00"))
except ValueError:
match = re.match(r"^(\d{4}-\d{2}-\d{2})[ T](\d{2}:\d{2}:\d{2})", text)
if not match:
return None
parsed = datetime.fromisoformat(f"{match.group(1)}T{match.group(2)}+00:00")
return parsed if parsed.utcoffset() is not None else parsed.replace(tzinfo=UTC)
# ------------------------------------------------------------- the parsers
#: What every parser turns a feed into: one dict per row, with these keys.
Row = dict[str, Any]
# ---------------------------------------------------------- structured pay
#
# A board's own salary range, mapped to yearly `salary_min`/`salary_max`
# (integers) and `currency` (ISO 4217) — the columns `upsert_role` unpacks
# `raw` into, per docs/decisions.md's "raw is transport, columns are
# storage". Only Ashby, Greenhouse, Lever, Recruitee and Personio give
# structured pay on their list endpoint; Workable and SmartRecruiters do not
# (checked live, 2026-09-29), so their parsers are untouched and `row.get`
# below reads a plain '' or None for them.
def _num(value: Any) -> float | None:
"""A positive number from whatever a board sends (int, float, or a
numeral string), or None for anything else, including zero and a
negative — a board's placeholder 0 (an unfilled field, not a real
figure) must not become a salary floor of zero."""
try:
number = float(value)
except (TypeError, ValueError):
return None
return number if number > 0 else None
def _yearly_amount(min_value: Any, max_value: Any, currency: Any, interval: Any,
factors: dict[str, int]) -> tuple[int | None, int | None, str]:
"""`(salary_min, salary_max, currency)` as yearly integers, or `(None,
None, "")` when there is no currency, no bound, or `interval` is not one
of `factors`' own keys — which is how an hourly or a daily rate is left
out rather than annualised into a number nobody stated. `factors` maps
the board's own interval spelling to the multiplier that makes it
yearly: 1 for a figure already yearly, 12 for a monthly one."""
code = str(currency or "").strip().upper()
factor = factors.get(str(interval or "").strip())
if not code or not factor:
return None, None, ""
lo, hi = _num(min_value), _num(max_value)
if lo is None and hi is None:
return None, None, ""
return (int(lo * factor) if lo is not None else None,
int(hi * factor) if hi is not None else None, code)
def _greenhouse_salary(ranges: list[dict[str, Any]]) -> tuple[int | None, int | None, str]:
"""The yearly range from `pay_input_ranges` (needs `pay_transparency=true`
on the feed URL, see `_BOARDS`), which the list endpoint already carries
in the same request — no per-role fetch. Only when there is exactly one
range: a board with several ("Zone 1 Pay Range", "US Tier 2 Locations",
one per office) gives no way to tell which office the role in hand is
at, and picking one would misrepresent the columns. Greenhouse's own
ranges are cents, and yearly by convention except an hourly or a daily
one, named as such in `title` with no separate field to check — those
are skipped rather than annualised into a number nobody stated."""
if len(ranges) != 1:
return None, None, ""
entry = ranges[0]
title = str(entry.get("title") or "").lower()
if any(word in title for word in ("hour", "/hr", "daily", "per day", "day rate")):
return None, None, ""
code = str(entry.get("currency_type") or "").strip().upper()
if not code:
return None, None, ""
lo, hi = _num(entry.get("min_cents")), _num(entry.get("max_cents"))
if lo is None and hi is None:
return None, None, ""
return (int(lo / 100) if lo is not None else None,
int(hi / 100) if hi is not None else None, code)
def _ashby_salary(components: list[dict[str, Any]]) -> tuple[int | None, int | None, str]:
"""The yearly range from `compensation.summaryComponents` (needs
`?includeCompensation=true` on the feed URL, see `_BOARDS`): the first
`Salary` component, yearly or monthly (a monthly one multiplied by 12);
an hourly one is left out rather than annualised."""
for component in components:
if component.get("compensationType") != "Salary":
continue
return _yearly_amount(component.get("minValue"), component.get("maxValue"),
component.get("currencyCode"), component.get("interval"),
{"1 YEAR": 1, "1 MONTH": 12})
return None, None, ""
def _parse_greenhouse(body: bytes) -> list[Row]:
data = json.loads(body)
rows = []
for j in data.get("jobs") or []:
title = (j.get("title") or "").strip()
url = (j.get("absolute_url") or "").strip()
if not title or not _is_web_url(url):
continue
departments = j.get("departments") or []
# Greenhouse ships the description HTML-entity-escaped inside the
# JSON string (`&lt;p&gt;`), so it needs unescaping before the tags
# themselves can be stripped.
content = html.unescape(j.get("content") or "")
salary_min, salary_max, currency = _greenhouse_salary(j.get("pay_input_ranges") or [])
rows.append({
"source_id": str(j.get("id") or ""),
"title": title,
"location": ((j.get("location") or {}).get("name") or "").strip(),
"url": url,
"posted": _parse_dt(j.get("first_published") or j.get("updated_at")),
"department": (departments[0].get("name") if departments else "") or "",
"description": strip_html(content),
"remote": None, # Greenhouse's feed does not say
"salary": "",
"salary_min": salary_min,
"salary_max": salary_max,
"currency": currency,
})
return rows
def _parse_lever(body: bytes) -> list[Row]:
data = json.loads(body)
if not isinstance(data, list):
return []
rows = []
for j in data:
title = (j.get("text") or "").strip()
url = (j.get("hostedUrl") or "").strip()
if not title or not _is_web_url(url):
continue
categories = j.get("categories") or {}
description = j.get("description") or j.get("descriptionPlain") or ""
salary_range = j.get("salaryRange") or {}
salary_min, salary_max, currency = _yearly_amount(
salary_range.get("min"), salary_range.get("max"), salary_range.get("currency"),
salary_range.get("interval"), {"per-year-salary": 1, "per-month-salary": 12})
rows.append({
"source_id": str(j.get("id") or ""),
"title": title,
"location": (categories.get("location") or "").strip(),
"url": url,
"posted": _parse_dt(j.get("createdAt")),
"department": (categories.get("department") or categories.get("team") or "").strip(),
"description": strip_html(description),
"remote": (j.get("workplaceType") or "").strip() or None,
"salary": "",
"salary_min": salary_min,
"salary_max": salary_max,
"currency": currency,
})
return rows
def _parse_ashby(body: bytes) -> list[Row]:
data = json.loads(body)
rows = []
for j in data.get("jobs") or []:
title = (j.get("title") or "").strip()
url = (j.get("jobUrl") or j.get("applyUrl") or "").strip()
if not title or not _is_web_url(url):
continue
description = j.get("descriptionHtml") or j.get("descriptionPlain") or ""
compensation = j.get("compensation") or {}
salary = (compensation.get("compensationTierSummary")
or compensation.get("scrapeableCompensationSalarySummary") or "")
salary_min, salary_max, currency = _ashby_salary(compensation.get("summaryComponents") or [])
rows.append({
"source_id": str(j.get("id") or ""),
"title": title,
"location": (j.get("location") or "").strip(),
"url": url,
"posted": _parse_dt(j.get("publishedAt")),
"department": (j.get("department") or j.get("team") or "").strip(),
"description": strip_html(description),
"remote": j.get("isRemote"),
"salary": salary,
"salary_min": salary_min,
"salary_max": salary_max,
"currency": currency,
})
return rows
def _parse_workable(body: bytes, slug: str) -> list[Row]:
return _workable_rows(json.loads(body), slug)
def _workable_rows(data: dict[str, Any], slug: str) -> list[Row]:
"""One page of POST https://apply.workable.com/api/v3/accounts/{slug}/jobs
-> `{"total": int, "results": [...], "nextPage": token}`, as Hugging
Face's board answered on 2026-09-28. Ten results a page; the next page is
the same POST with `{"token": nextPage}`, and the last page has no
`nextPage` (see `_paged`). A result carries no URL and no text: the job page is
`https://apply.workable.com/{slug}/j/{shortcode}/`, and the description
is only on the per-job endpoint, one request per role, so `description`
stays empty here as it does for SmartRecruiters. `location` is an object
and `department` a list; `remote` is a flag beside `workplace`
(remote, hybrid, on_site)."""
rows = []
for j in data.get("results") or []:
title = (j.get("title") or "").strip()
shortcode = (j.get("shortcode") or "").strip()
if not title or not shortcode:
continue
loc = j.get("location") or {}
location = ", ".join(p for p in (loc.get("city"), loc.get("region"), loc.get("country")) if p)
dept = j.get("department") or ""
if isinstance(dept, list):
dept = ", ".join(d for d in dept if d)
rows.append({
"source_id": str(j.get("id") or shortcode),
"title": title,
"location": location,
"url": f"https://apply.workable.com/{slug}/j/{shortcode}/",
"posted": _parse_dt(j.get("published")),
"department": dept.strip(),
"description": "",
"remote": bool(j.get("remote")) or j.get("workplace") == "remote",
"salary": "",
})
return rows
def _parse_smartrecruiters(body: bytes, slug: str) -> list[Row]:
return _smartrecruiters_rows(json.loads(body), slug)
def _smartrecruiters_rows(data: dict[str, Any], slug: str) -> list[Row]:
"""One page of https://api.smartrecruiters.com/v1/companies/{slug}/postings
-> `{"offset", "limit", "totalFound", "content": [...]}`. At most 100 a
page, whatever `limit` asks for; the next is `?offset=` further on (see
`_paged`). The list omits the job description (only the
per-posting detail endpoint has it, and fetching one per row would be
hundreds of extra requests per board), so `description` stays empty here."""
rows = []
for j in data.get("content") or []:
title = (j.get("name") or "").strip()
job_id = str(j.get("id") or "")
if not title or not job_id:
continue
location = j.get("location") or {}
loc_text = (location.get("fullLocation")
or ", ".join(p for p in (location.get("city"), location.get("country")) if p))
rows.append({
"source_id": job_id,
"title": title,
"location": (loc_text or "").strip(),
"url": f"https://jobs.smartrecruiters.com/{slug}/{job_id}",
"posted": _parse_dt(j.get("releasedDate")),
"department": ((j.get("department") or {}).get("label") or "").strip(),
"description": "",
"remote": location.get("remote"),
"salary": "",
})
return rows
def _parse_recruitee(body: bytes) -> list[Row]:
data = json.loads(body)
rows = []
for j in data.get("offers") or []:
title = (j.get("title") or "").strip()
url = (j.get("careers_url") or "").strip()
if not title or not _is_web_url(url):
continue
salary = j.get("salary") or {}
salary_text = ""
if isinstance(salary, dict) and (salary.get("min") or salary.get("max")):
bounds = "-".join(str(v) for v in (salary.get("min"), salary.get("max")) if v)
salary_text = f"{bounds} {salary.get('currency') or ''}".strip()
salary_min, salary_max, currency = _yearly_amount(
salary.get("min"), salary.get("max"), salary.get("currency"),
salary.get("period"), {"year": 1, "month": 12})
rows.append({
"source_id": str(j.get("id") or ""),
"title": title,
"location": (j.get("location") or j.get("city") or "").strip(),
"url": url,
"posted": _parse_dt(j.get("published_at") or j.get("created_at")),
"department": (j.get("department") or "").strip(),
"description": strip_html(j.get("description") or ""),
"remote": j.get("remote"),
"salary": salary_text,
"salary_min": salary_min,
"salary_max": salary_max,
"currency": currency,
})
return rows
def _parse_personio(body: bytes, slug: str) -> list[Row]:
"""`{slug}.jobs.personio.de/xml` -> `<workzag-jobs><position>...`. No job
URL in the feed itself; Personio's own page is `/job/{id}`, confirmed
live against a real posting (2026-09-23)."""
root = ET.fromstring(body)
rows = []
for position in root.findall("position"):
job_id = (position.findtext("id") or "").strip()
title = (position.findtext("name") or "").strip()
if not job_id or not title:
continue
description = " ".join(
(value.text or "") for value in position.findall("jobDescriptions/jobDescription/value"))
salary_info = position.find("salaryInformation")
salary_min = salary_max = None
currency = ""
if salary_info is not None:
salary_min, salary_max, currency = _yearly_amount(
salary_info.findtext("min"), salary_info.findtext("max"),
salary_info.findtext("currencyCode"), salary_info.findtext("type"),
{"yearly": 1, "monthly": 12})
rows.append({
"source_id": job_id,
"title": title,
"location": (position.findtext("office") or "").strip(),
"url": f"https://{slug}.jobs.personio.de/job/{job_id}",
"posted": _parse_dt(position.findtext("createdAt")),
"department": (position.findtext("department") or "").strip(),
"description": strip_html(description),
"remote": "remote" in (position.findtext("office") or "").lower() or None,
"salary": "",
"salary_min": salary_min,
"salary_max": salary_max,
"currency": currency,
})
return rows
#: host -> parser. Every one takes the raw response body, plus the slug for
#: the two hosts whose feed does not carry the URL or the company itself.
_PARSERS: dict[str, Callable[..., list[Row]]] = {
"greenhouse": lambda body, slug: _parse_greenhouse(body),
"lever": lambda body, slug: _parse_lever(body),
"ashby": lambda body, slug: _parse_ashby(body),
"workable": _parse_workable,
"smartrecruiters": _parse_smartrecruiters,
"recruitee": lambda body, slug: _parse_recruitee(body),
"personio": _parse_personio,
}
# --------------------------------------------------------- the classifier
#: The fixed set of functions, and the keyword rules that map a title
#: (preferred) or a department to one of them. Ported from `cmd/seedgen/
#: main.go`'s `rules` — first match wins, and "product" is genuinely listed
#: twice: the first, narrower entry is checked before
#: recruiting/design/marketing/support/sales/data, and the second, which also
#: matches the bare word "product", is checked only after all of those, so a
#: title like "Data Product Manager" keeps classifying as data/product by the
#: same rule order the Go tool used.
#:
#: Not a straight port any more: the Go tool matched substrings, which let
#: " it" (meant to catch "IT" as a department) match inside "Italy" and
#: "Italia" — a European department like "Commercial Italy" filed as
#: engineering. `_classify_text` now matches on word boundaries. A handful of
#: short tokens (`_WHOLE_WORD`) need a boundary on *both* sides, because they
#: are themselves the start of an unrelated word ("it" of "Italy", "seo" of
#: "Seoul", "qa" of "Qatar"); every other keyword keeps a leading boundary
#: only; most of them are deliberate stems ("manag", "engineer", "recruit")
#: meant to also catch "management", "engineering", "recruiting", and a
#: trailing boundary on those would stop them matching the plural and
#: gerund forms real postings use ("Software Developers", "Account
#: Managers").
FUNCTIONS = ("engineering", "data", "product", "design", "sales", "marketing",
"customer success/support", "recruiting/people", "finance/legal/ops")
_RULES: tuple[tuple[str, tuple[str, ...]], ...] = (
("product", ("product manag", "product management", "product operations",
"program manag", "technical program")),
("recruiting/people", ("recruit", "talent", "people ops", "people &",
"people team", "people operations", "human resources",
"hr", "hrbp")),
("design", ("design", "ux", "user experience", "creative")),
("marketing", ("marketing", "growth", "content", "communications", "brand",
"seo", "social media", "demand gen", "events")),
("customer success/support", ("customer success", "customer support",
"customer experience", "customer service",
"support", "success", "implementation",
"onboarding", "professional services",
"solutions consult", "technical account",
"customer engineer", "deployment")),
("sales", ("sales", "account executive", "account manager", "account director",
"business development", "bdr", "sdr", "revenue", "partnership",
"go-to-market", "gtm", "solutions engineer", "solution engineer",
"pre-sales", "presales", "channel",
# "Commercial" is what French, Italian and Spanish org charts
# call the sales function ("Commercial Italy" is a sales
# department; see the docstring above), and the rest are the
# sales words of the seed's other languages that "sales"
# itself does not already cover (German, Dutch, Danish,
# Norwegian, Swedish, French) — a modest set, not a full port;
# `extract` is the rest.
"commercial", "vertrieb", "verkauf", "verkoop", "salg",
"försäljning", "vente")),
("data", ("data", "analytics", "machine learning", "ml",
"applied science", "research scientist", "statistic")),
("product", ("product manag", "product management", "program manag",
"product operations", "technical program", "product")),
("finance/legal/ops", ("finance", "financ", "accounting", "legal", "counsel",
"compliance", "operations", "ops", "administrat",
"office", "procurement", "strategy", "corporate",
"facilities", "risk", "audit", "tax", "treasury",
"payroll", "executive assistant", "chief of staff",
"workplace", "supply chain", "logistics", "fraud",
"trust & safety", "trust and safety", "policy")),
("engineering", ("engineer", "software", "developer", "infrastructure",
"platform", "devops", "sre", "site reliability", "security",
"information technology", "it", "technology",
"technical staff", "hardware", "qa", "quality assurance",
"architect", "research", "science", "scientist", "systems",
"network", "cloud", "mobile", "frontend", "backend",
"full stack", "fullstack", "web",
# The developer/engineer words of the seed's other
# languages, same reasoning as sales' above.
"entwickler", "ingenieur", "ontwikkelaar", "udvikler",
"utvikler", "utvecklare", "ingenjör", "ingeniør",
"développeur", "developpeur", "ingénieur")),
)
#: Keywords that must match a whole word, not just its start: each one is
#: itself the first letters of an unrelated word that would otherwise false-
#: match ("it" of "Italy", "seo" of "Seoul", "qa" of "Qatar"). Every keyword
#: not listed here keeps the old prefix behaviour on purpose — see the
#: `_RULES` docstring above.
_WHOLE_WORD = frozenset({
"it", "hr", "hrbp", "ux", "ml", "qa", "seo", "web", "tax", "ops",
"sre", "bdr", "sdr", "gtm",
})
def _keyword_pattern(keyword: str) -> re.Pattern[str]:
trailing = r"\b" if keyword in _WHOLE_WORD else ""
return re.compile(r"\b" + re.escape(keyword) + trailing, re.IGNORECASE)
#: `_RULES`, compiled once: `(function, patterns)` in the same order.
_COMPILED_RULES: tuple[tuple[str, tuple[re.Pattern[str], ...]], ...] = tuple(
(function, tuple(_keyword_pattern(keyword) for keyword in keywords))
for function, keywords in _RULES)
def _classify_text(text: str) -> str:
stripped = (text or "").strip()
if not stripped:
return ""
for function, patterns in _COMPILED_RULES:
if any(pattern.search(stripped) for pattern in patterns):
return function
return ""
def classify(department: str, title: str) -> str:
"""A title (preferred) or a department, to one of `FUNCTIONS`, or `''`
when neither names one — `''` and not `"other"`, which would claim a
knowledge the classifier does not have (`extract` may still fill it from
reading the posting; see `Store.set_extracted`).
Title before department: a department is often a generic umbrella
("Operations", "Engineering & Product") that says less about the role
than its title does. A hand-checked sample of 60 roles where the two
orders disagreed (docs/decisions.md) found title-first right about twice
as often as department-first, mostly because a specific title beats a
broad department name, so title now wins; department is the fallback
when the title matches nothing."""
return _classify_text(title or "") or _classify_text(department or "") or ""
def is_known_target(target: str) -> bool:
"""`host:slug`, with `host` one of `_BOARDS` and a non-empty slug. What
the target validator refuses: `app companies add` and the pull job
both call this before touching the network."""
host, sep, slug = target.partition(":")
return bool(sep) and host in _BOARDS and bool(slug.strip())
def careers_url(target: str) -> str:
"""The company's own careers page for `host:slug`, or "" for an unknown
host. What `app companies add` stores alongside the board."""
host, _, slug = target.partition(":")
if host not in _BOARDS:
return ""
return _BOARDS[host][1].format(slug=slug)
#: Pages read from one paged board at most, and the pause between two of
#: them. Workable answers ten roles a page and SmartRecruiters a hundred, so
#: the cap is 1,000 and 10,000 roles; the pause keeps Workable, which answers
#: a burst with 429, from seeing one.
MAX_PAGES = 100
PAGE_DELAY = 0.5
#: SmartRecruiters' page size, which is also the most it serves per page.
SMARTRECRUITERS_PAGE = 100
#: A paged host's one page: `(client, slug, cursor)` -> the decoded page, its
#: raw results, the total it states, and the cursor of the next page (None
#: after the last). The first call gets cursor None.
Page = tuple[dict[str, Any], list[dict[str, Any]], Any, Any]
def _workable_page(client: httpx.Client, slug: str, cursor: Any) -> Page:
url = _BOARDS["workable"][0].format(slug=slug)
data = json.loads(_get(client, url, "workable", slug, method="POST",
payload={"token": cursor} if cursor else {}))
return data, data.get("results") or [], data.get("total"), data.get("nextPage") or None
def _smartrecruiters_page(client: httpx.Client, slug: str, cursor: Any) -> Page:
offset = int(cursor or 0)
url = (_BOARDS["smartrecruiters"][0].format(slug=slug)
+ f"?offset={offset}&limit={SMARTRECRUITERS_PAGE}")
data = json.loads(_get(client, url, "smartrecruiters", slug))
content = data.get("content") or []
return data, content, data.get("totalFound"), offset + len(content) if content else None
#: host -> (one page, that page's rows) for the hosts that page.
_PAGED: dict[str, tuple[Callable[[httpx.Client, str, Any], Page],
Callable[[dict[str, Any], str], list[Row]]]] = {
"workable": (_workable_page, _workable_rows),
"smartrecruiters": (_smartrecruiters_page, _smartrecruiters_rows),
}
def _paged(client: httpx.Client, target: str) -> tuple[list[Row], bool]:
"""Every page of a Workable or SmartRecruiters board, `PAGE_DELAY` apart,
until the host has no next page, the stated total is reached, or
`MAX_PAGES`. Returns the rows and whether the board is complete: as many
distinct results read as the host's stated total. Counted on the raw
results, so a row the parser drops does not make a board incomplete, and
on distinct ids, so a page that shifted under an offset and repeated a
role does not make it complete.
A 429 after the first page stops the read with what it has, incomplete;
one on the first page raises like any other failure."""
host, _, slug = target.partition(":")
page_of, rows_of = _PAGED[host]
rows: dict[str, Row] = {}
seen: set[str] = set()
total: Any = None
cursor: Any = None
for page in range(MAX_PAGES):
if page:
time.sleep(PAGE_DELAY)
try:
data, results, total, cursor = page_of(client, slug, cursor)
except httpx.HTTPStatusError as exc:
if not page or exc.response.status_code != 429:
raise
log.warning("ats: %s: 429 on page %d; kept %d of %s, not complete",
target, page + 1, len(seen), total)
return list(rows.values()), False
seen |= {str(r.get("id") or r.get("shortcode") or "") for r in results}
for row in rows_of(data, slug):
rows.setdefault(row["source_id"], row)
if cursor is None or (isinstance(total, int) and len(seen) >= total):
break
complete = isinstance(total, int) and len(seen) == total
if not complete:
log.warning("ats: %s: read %d of %s stated; not complete", target, len(seen), total)
return list(rows.values()), complete
def _fetch_rows(client: httpx.Client, target: str) -> tuple[list[Row], bool]:
"""One board, fetched and parsed: its rows and whether they are the whole
board. Raises on any failure — refused host, timeout, a feed shape none
of the parsers understand. A host that does not page answers with the
whole board, so it is complete whenever it parses; a paged one is
complete when `_paged` says so.
A Greenhouse board whose `?content=true` body still exceeds `MAX_BYTES`
gets one retry without it: the description is lost, but the titles land
rather than the whole board being skipped."""
if not is_known_target(target):
raise ValueError(f"{target!r} is not host:slug for a known ATS host")
host, _, slug = target.partition(":")
if host in _PAGED:
return _paged(client, target)
url = _BOARDS[host][0].format(slug=slug)
try:
body = _get(client, url, host, slug, method="GET")
except _TooLarge:
if host != "greenhouse" or "?" not in url:
raise
log.warning("ats: %s: body exceeded %d bytes with content=true; "
"retrying without it", target, MAX_BYTES)
url = url.split("?", 1)[0]
body = _get(client, url, host, slug, method="GET")
return _PARSERS[host](body, slug), True
def fetch_board(target: str) -> tuple[list[Row], bool]:
"""Fetch and parse one board directly: no `Source`, no `self.state`, just
"does this board answer, what does it list, and is that all of it". For
`app companies add|check`, which ask that question about one board at a
time outside any pull. Raises the same way `_fetch_rows` does."""
with httpx.Client() as client:
return _fetch_rows(client, target)
def failure_reason(exc: BaseException) -> str:
"""One short line for why a board did not answer: the status of an HTTP
error, else the exception's class and the first line of its message."""
if isinstance(exc, httpx.HTTPStatusError):
return f"{exc.response.status_code} {exc.response.reason_phrase}".strip()
lines = str(exc).strip().splitlines()
text = f"{type(exc).__name__}: {lines[0]}" if lines else type(exc).__name__
return text[:200]
# -------------------------------------------------------------- the plugin
class ATSSource(Source):
name = "ats"
description = "Company job boards: Greenhouse, Lever, Ashby, Workable, " \
"SmartRecruiters, Recruitee, Personio."
#: The raw entries under `sources.ats.boards` are `{"url": target, "label":
#: company}` — `url` carries the `host:slug` target and `label` the
#: company name, which is how `Source.targets()`'s generic dict reading
#: hands both back without a plugin-specific override.
subscription_field = "boards"
spec = PluginSpec(
api_version=API_VERSION, id=name, version="1", label="ATS boards",
targets=TargetSpec(mode="editable", kind="board", defaults=[]),
profiles={"default": Profile(
kind="role",
tasks=Tasks(
summary=SummaryPolicy(enabled=False),
tagging=TaggingPolicy(enabled=False),
),
presentation=ARTICLE_READER,
)},
)
def fetch(self) -> Iterator[Item]:
with httpx.Client() as client:
# Handed in by the batch that runs this plugin, as the marker of
# the pull run that queued it (see `app/jobs/pull.py`), and never
# read from state, where a later run may have replaced it. A
# plugin run on its own — a test, a dry run — has no run to stamp
# itself against, so it makes one up.
run_marker = str(self.options.get("run") or "") or datetime.now(UTC).isoformat(
timespec="seconds")
for target, label in self.targets():
host, _, slug = target.partition(":")
failing = f"ats:failing:{target}"
try:
rows, complete = _fetch_rows(client, target)
except Exception as exc:
log.warning("ats: %s failed (%s)", target, exc)
# The first failure in a row sets `since`, later ones only
# the reason; `app.jobs.pull` closes the board's roles
# once `since` is `STALE_DAYS` old.
try:
since = json.loads(self.state.get(failing, "") or "{}").get("since")
except ValueError:
since = None
self.state[failing] = json.dumps({
"since": since or datetime.now(UTC).isoformat(timespec="seconds"),
"reason": failure_reason(exc)})
continue
self.state.pop(failing, None)
seen_ids = []
for row in rows:
# The compound id, matching `roles.source_id` — what
# `Store.mark_closed` compares against, not the board's own
# bare job id.
compound_id = f"{host}:{slug}:{row['source_id']}"
seen_ids.append(compound_id)
yield Item(
source_id=compound_id,
location=row["url"],
title=row["title"],
excerpt=(row["description"] or "")[:300],
body=row["description"],
published_at=row["posted"],
subscription=label,
kind="role",
raw={
"company": label,
"place": row["location"],
"department": row["department"],
"function": classify(row["department"], row["title"]),
"remote": row["remote"],
"salary": row["salary"],
# Only Ashby, Greenhouse, Lever, Recruitee and
# Personio's parsers set these (see "structured
# pay" above); the rest leave them at row.get's
# default of no figure.
"salary_min": row.get("salary_min"),
"salary_max": row.get("salary_max"),
"currency": row.get("currency", ""),
"board": host,
"slug": slug,
},
)
# Stamped only once the board has parsed without raising, so a
# board that blew up half way through is neither marked ok nor
# has a partial id list treated as the whole of it. `complete`
# says whether the ids are the whole board, which the pull
# needs before it closes what they leave out.
self.state[f"ats:ok:{target}"] = run_marker
self.state[f"ats:complete:{target}"] = "1" if complete else ""
self.state[f"ats:ids:{target}"] = json.dumps(seen_ids)