17 KiB
Plan: a module per host, new hosts, and the list of companies checked by hand
Planned 2026-09-29. Three waves of work packages with disjoint files, built in parallel worktrees made from main, merged by the planner. This file and docs/plan-hosts/ are folded into decisions.md and deleted when the work lands.
Why
164 of the 362 companies have no board attached: 9 on Teamtailor whose career host is not found yet (the other 23 were attached today), and 155 on a host the app cannot read or on no host it could name. They include LEGO, Novo Nordisk, Vestas, Danske Bank, SimCorp, Ørsted, Klarna and Zalando. A company the app cannot read is shown as "not readable yet" with a link that nobody can edit, so the owner has no list to work through by hand.
What the survey found
Nine Sonnet agents read every one of the 155 career sites on 2026-09-29. Their notes are in docs/plan-hosts/: workday.md, successfactors.md, otherhosts.md, own-1.md to own-4.md, and survey-own.json with one record per company (verified jobs_url, class, host, identifier, listing URL, roles seen, evidence). Raw responses for fixtures are in the session scratchpad, /tmp/claude-1000/-home-claude-cboxcode-jobboard/1734d897-7509-48fc-844c-efe768acbf7a/scratchpad/, and a builder copies the ones it needs, trimmed, into tests/fixtures/ats/.
| Route | Companies | Examples |
|---|---|---|
| Board on a host the app reads already | 49 + 9 Teamtailor | Bosch, Wise, Ubisoft, Checkout.com, Lunar, DFDS, Voi, Remote |
| SuccessFactors | 20 | Novo Nordisk, Vestas, Carlsberg, Pandora, Coloplast, Scania, SAP, Tryg |
| Workday | 19 | LEGO, Zeiss, SimCorp, Zendesk, Unity, Saxo Bank, Red Hat |
| Oracle | 4 | Danske Bank, Nokia, Sinch, Milestone |
| Phenom | 3 | Philips, Just Eat Takeaway, King |
| Eightfold | 3 | Microsoft, Ericsson, Infineon |
| Eleven small hosts | 15 | Klarna (Deel), Booking.com (Jibe), sennder (Gem), Einride (Jobylon) |
| Sitemap and structured data | 8 | Arm, Mambu, Odoo, Aiven, AUTO1, ASML |
| Own JSON API | 8 | Ørsted, Nordea, Amazon, Vinted, Dynatrace |
| Own HTML list | 10 | Zalando, Netcompany, Google, Bolt, CHECK24 |
| By hand | about 25 | Revolut, Meta, Maersk, Bang & Olufsen, Visma, HashiCorp |
The 49 were missed because the slug is not the company's name (BoschGroup, jimdo.com, workatbackbase), the board is on a custom domain, the board is on Lever's EU host, or the careers page refuses a plain request.
Decisions
A host is a module. plugins/ats.py holds eight hosts in six parallel tables, and app/jobs/discover.py holds three more, so two builders adding a host edit the same lines. Each host moves into plugins/atshosts/<host>.py, and ats.py and discover.py build their tables from the package. The package is a sibling of ats.py and not plugins/ats/, since desk.sources.registry globs plugins/*.py and the tests patch names on the ats module.
A host is handed the fetch. A host module imports nothing from plugins.ats. Its functions receive get, which ats.py binds to its own _get at call time, so the allow-list, the redirect rule, the byte cap and the deadline stay in one place and a test that patches ats._get reaches every host.
A description read per role is read once. Workday, SuccessFactors and BambooHR state the description only on the role's own page. A host with a detail function has it called for roles not yet read in detail, at most DETAIL_CAP (200) per board per pull; the ids read are kept in ats:detailed:<target>. A known role is sent without a description on later pulls, which upsert_role ignores since an empty value never overwrites a filled column. Completeness is decided by the listing alone.
robots.txt is honoured. A host module names the page whose path must be allowed. The shared code reads the site's robots.txt at most once a day per board and fails a board that disallows it for *, with the reason "robots.txt disallows it". This takes Novonesis, Darktrace, Criteo and Bang & Olufsen off the automatic route. The SuccessFactors reader uses /sitemap.xml, /search/ and /job/ only, never /services/ or the career*.successfactors.eu back end.
No reader is built on a key lifted from a site's script. Maersk's API, New Work's and Welcome to the Jungle's search index answer only with a key that the page's JavaScript carries. They go on the by-hand list.
The by-hand list is the companies table. A company is on it when it has no shown board and is not known from a search only. careers_url and notes exist already; migration 15 adds hand_reason and checked_at. No second table, since a company has one careers link and the list needs no history of checks.
A link typed by hand is kept. add_discovered and the seed fill careers_url only when it is empty.
A company has one board. DFDS has eleven Teamtailor boards by country, and Onfido, Tink and Glovo sit inside their owners' boards. DFDS gets its Danish board; the other three go on the by-hand list with reason group.
Contracts between packages
Host module, plugins/atshosts/<host>.py, as built by WP0. The docstring of plugins/atshosts/__init__.py holds the same contract and is the one to keep current. A module imports desk and ._shared only, never plugins.ats.
NAME = "workday" # the host of a target `workday:<slug>`; equals the module's file name
LABEL = "Workday" # the Sources page's select
PATTERNS = (r"https?://...",) # discovery: regexes over a URL, compiled with IGNORECASE
RANK = 100 # optional int, default 100; lower comes first
KEEP_CASE = True # optional, default False; keep a discovered slug's case
def slug_ok(slug: str) -> bool: ... # strict, re.fullmatch
def allowed(slug: str) -> Callable[[str], bool]: ... # host names a fetch may reach
def careers_url(slug: str) -> str: ... # the page a person opens
# one of the two
def feed_url(slug: str) -> str: ... # the whole board in one answer
def parse(body: bytes, slug: str) -> list[Row]: ...
def page(get: Get, slug: str, cursor: Any) -> Page: ... # a paged board, one page a call
def rows(decoded: Any, slug: str) -> list[Row]: ...
# optional
def fallback_url(slug: str) -> str: ... # feed hosts: a smaller feed, read once when feed_url is over the byte cap
def robots_page(slug: str) -> str: ... # the URL robots.txt must allow; on a host allowed(slug) admits
def detail(get: Get, slug: str, row: Row) -> dict[str, Any]: ... # fields from the role's own page
def probe_url(slug: str) -> str: ... # discovery: answers 2xx only for a board
def probe_ok(body: bytes) -> bool: ... # discovery: when 2xx is not enough
def probe_slugs(guess: str) -> tuple[str, ...]: ... # the slugs a guessed name is probed as; default (guess,)
Get, Row and Page are in plugins/atshosts/_shared.py, with the helpers several hosts share: only(host) for allowed, path_slug_ok and label_slug_ok for slug_ok, parse_dt, is_web_url, num, yearly_amount and pay_line. plugins/atshosts/__init__.py lists the modules with pkgutil.iter_modules, checks each against the contract at import (an ImportError names the module and what it lacks), and orders HOSTS by (RANK, NAME). The eight first hosts hold ranks 10 to 80, which keeps the order they had before the modules: the select opens on Greenhouse, and a guessed slug on two hosts goes to the one probed first.
Order and names. RANK orders the Sources page's select, the plugin's description and discovery's probes. NAME equals the file name. A host that discover.py names in its own _UNREADABLE table is taken from its module as soon as the module exists, so a wave 1 package does not edit discover.py.
Patterns. The slug is group 1, or, when a pattern has more groups, the groups that matched joined by "/" (https?://([a-z0-9-]+\.wd\d+)\.myworkdayjobs\.com/(?:[a-z]{2}-[A-Z]{2}/)?([A-Za-z0-9_-]+) gives lego.wd103/LEGO_External). Only group 1 is checked against discovery's words that name no board (api, jobs, embed, ...). The slug is lowercased unless the module sets KEEP_CASE = True, which a host with case-sensitive paths needs, since robots.txt paths are case-sensitive too. A pattern without a group names the host with no slug.
Slugs. slug_ok uses re.fullmatch (a $ lets a trailing newline through) and refuses anything that could add a path, query, fragment, port or user to the URLs built from it. A slug of several parts joins them with "/": Lever's EU boards are lever:eu/<site>.
get(url, *, method="GET", payload=None, headers=None) -> bytes is ats._get bound to the board. A POST sends payload as JSON; headers are added to the default User-Agent and Accept and replace either where they name it. It fetches https on port 443 only, never an IP address or a local name (localhost, .local, .internal, .lan, .home.arpa), follows at most three redirects each checked against allowed(slug), cuts the body at MAX_BYTES and ends within DEADLINE. It returns the body and raises httpx.HTTPStatusError on a non-2xx answer, httpx.HTTPError on a timeout or a connection failure, and another exception, ValueError among them, for a refused host, an oversized body or a broken redirect. A module lets these through.
Page is (decoded, results, total, cursor): the decoded page handed to rows, its raw results as dicts with an id (counted by distinct id for completeness), the total the host states or None, and the next cursor or None after the last page. The first call gets cursor None. A board with a total is complete when as many distinct ids were read, one without when the read ended on an empty page; a 429 after the first page ends the read, incomplete. A paged module may also have parse(body, slug) for one page's body, for its tests.
A Row has source_id, title, location, url, posted (aware datetime or None), department, description (plain text), remote (True, False, a word or None) and salary_text, and, where the host states pay, salary_min, salary_max (yearly integers) and currency.
detail returns a dict of what the role's own page states. The shared code calls it for roles not yet read (ats:detailed:<target>), at most DETAIL_CAP (200) per board per pull, PAGE_DELAY apart, the roles never tried before those whose detail failed (ats:detailfailed:<target>). Only DETAIL_FIELDS are taken: description, location, remote, currency, salary_min, salary_max and posted, the keys whose columns survive a later pull that sends the listing's empty value (upsert_role's keep-filled columns, the salary figures, and posted_at, which keeps the earliest date). Any other key is dropped with a debug log line: title, url, department and salary_text cannot be filled from detail, since the listing rewrites them every pull. A role not read in detail on a pull is sent with an empty description. A 404 or 410 counts as read, any other failure is retried on a later pull, and a 429 ends the detail reads for that board and pull. Completeness is decided by the listing alone.
robots.txt. For a module with robots_page, the shared code reads https://<its host>/robots.txt before the listing, at most once a day per board (ats:robots:<target>), following a redirect to the host's www. twin or back and no other. It fails the board with "robots.txt disallows it" when the rules for * or for the app's own token disallow the page, matched per RFC 9309: the longest matching rule wins, Allow wins a tie, * and a final $ in paths, paths case-sensitive, an empty Disallow allows. A 4xx other than 429 means there are no rules, which allows. A 5xx, a 429, a timeout, a connection error or a redirect elsewhere fails the board with "robots.txt could not be read", and the next pull asks again.
Slugs: Workday lego.wd103/LEGO_External; SuccessFactors, Phenom and the sitemap reader take the site's hostname, with a path where the site has one (jobs.teradyne.com/MiR); Oracle <host>/<siteNumber>.
Company JSON, in GET /api/companies: CompanyOut gains domain, notes, hand_reason, checked_at and by_hand (bool). hand_reason is one of '', unread (readable, no parser yet), robots, blocked (the site answers a program with a challenge), key, group, gone, added (by hand or by an import, not looked at). from_search is false for a company with a hand_reason.
| Endpoint | Body | Answer |
|---|---|---|
POST /api/companies |
name, careers_url, notes, hand_reason (default added) |
201 CompanyOut; 409 when the key is taken |
PATCH /api/companies/{key} |
any of name, careers_url, notes, hand_reason |
CompanyOut; the key stays; 409 on name for a company with a board |
POST /api/companies/{key}/checked |
none | CompanyOut with checked_at now |
DELETE /api/companies/{key} |
none | 204; 409 when it has a board or open roles; its agent_companies rows go with it |
careers_url is http or https, as safeUrl demands on the page.
Work packages
Wave 0, in parallel, from today's main:
| WP | Branch | Model | Owns |
|---|---|---|---|
| 0 Host modules | wp0-hosts |
Opus | plugins/ats.py, plugins/atshosts/, app/jobs/discover.py, tests/test_ats.py, tests/test_discover.py, tests/test_atshosts.py |
| 1 By-hand backend | wp1-byhand-api |
Sonnet | app/models.py, app/storage/schema.py, app/storage/store.py, app/web/api.py, app/jobs/sources.py, app/cli.py, tests/test_storage.py, tests/test_web.py, tests/test_sources.py |
| 2 By-hand page | wp2-byhand-ui |
Opus | app/templates/settings/sources.html, app/static/byhand.js, app/static/companies.js, app/static/app.css, app/web/pages.py |
Wave 1, in parallel, from main after WP0 is merged. Each owns its modules in plugins/atshosts/, tests/test_ats_<host>.py and tests/fixtures/ats/<host>*, and nothing else:
| WP | Branch | Model | Hosts |
|---|---|---|---|
| 3 | wp3-workday |
Sonnet | Workday |
| 4 | wp4-successfactors |
Sonnet | SuccessFactors |
| 5 | wp5-oracle-phenom |
Sonnet | Oracle, Phenom |
| 6 | wp6-json-hosts |
Sonnet | Eightfold, Jibe, Pinpoint, BambooHR, Gem |
| 7 | wp7-small-hosts |
Sonnet | Deel, Jobylon, HR-ON, HR Manager, Rippling, HiBob |
| 8 | wp8-sitemap |
Sonnet | sitemap: job URLs from a sitemap, the role from the page's JSON-LD or microdata |
Wave 2 is decided after wave 1 by the roles each company would add: a module per company for the own JSON APIs (Ørsted, Nordea, Amazon, Vinted, Dynatrace, RELEX, Siteimprove, Bending Spoons) and the own HTML lists (Zalando, Netcompany, Google with DeepMind, Bolt, CHECK24). Until then these companies are on the by-hand list with reason unread.
WP0. The eight hosts move into modules with no change in behaviour: the existing tests pass with edits to imports only. careers_url, is_known_target, _allowed_host and _paged lose their Teamtailor branches to the module. _get takes headers. detail and the robots check are added to the shared code with tests against a fake host. discover.py takes READABLE, HOSTS, _PROBES and the body checks of probe from the modules; the hosts it names but cannot read stay in a table of its own. Discovery gains Lever's EU host. The tests that use workday: or successfactors as the unreadable host use taleo instead.
WP1. Migration 15, the four endpoints, CompanyOut's new fields, remove_company taking the agent_companies rows, careers_url filled only when empty in add_discovered and seed_companies, the seed reading hand_reason, and app companies add, edit, checked and remove in the CLI. The workday: and successfactors fixtures in its three test files become taleo.
WP2. A section "Check by hand" in the Sources rail, after Company boards: a count, a New company button, and rows sorted by checked_at, never checked first. A row shows the name, the affinity control, the reason, the careers link, the note, "checked 3 days ago", and Checked, Edit and Remove. New and Edit use the form pattern of the job boards section (sources.js: a <template>, the row becomes the form in place). The filter "Not readable yet" in Company boards is removed, since the section replaces it. Until WP1 is merged the page is built against the contract above.
WP3 to WP8. One module per host as the contract says, with parser tests from a saved response, paging and completeness tests against a fake get, and the allow-list and slug tests. Each host's notes name the listing, the paging quirks and the fields.
Planner, after the merges
- After WP1: write the survey into the live database and
seed/companies.json: the verifiedjobs_urlascareers_urlfor the 155, and thehand_reasonof each company that stays by hand. - Attach the 49 boards the survey found with
set_board, which fetches each before it lands, and find the career host of the 9 Teamtailor companies left. Not attached: Wallbox and Metacore (empty boards, possibly stale), Glovo, Cognigy. - After each host of wave 1: attach its boards, add it to
TRUSTED_POSTED_HOSTSwhere its date is a posting date, and name it inREADME.md. - Run a full pull on a copy of the live database before the live one.
- Read one Workday and one SuccessFactors board from the deployed server, since every Workday host is behind Cloudflare and the survey ran from this machine.
- Correct the seed: Yousign is Youtrust, Philips's link is the global site, Novonesis's tenant is
novonesis, Booking.com and Arm are not on iCIMS. - Fold this file into
decisions.mdand delete it withdocs/plan-hosts/.