Files
jobboard/docs/plan-readers.md
T

19 KiB

Plan: readers as plugins with a declared contract

Planned 2026-09-30. Turns every source of roles (the 22 ATS hosts, the 3 search providers, the company-specific readers of plan-hosts-wave-2.md) into a reader module with one documented API, and one runner that reads what each reader declares and handles paging, filters, per-role text, limits and failures for all of them. Pattern taken from ../news (desk/sources/spec.py, docs/plugin-rework.md): the plugin declares, the application does. Folded into decisions.md and deleted when the packages land.

This plan supersedes WP9 (provisional title, as far as the contract goes) and WP11 (board filters) of plan-hosts-wave-2.md, and covers items 2, 7, 17 and the "Plugin contract" and "Filtering at the source" sections of review-ingest.md.

How far we are

About two thirds of the way for boards, none of the way for searches.

Part Today Gap
One module per host yes, 22 in plugins/atshosts/, checked at import by problems() none
Documented contract docstring of plugins/atshosts/__init__.py covers hosts only
Capabilities inferred from which functions exist (page means paged, detail means per-role text) not declared; nothing about filters, caps, fields, date trust, places
Runner plugins/ats.py: guarded get, robots.txt, paging, per-role reads, failure state fused with the ats Source; search does not use it
Search providers three functions in plugins/search.py with their own getter no robots.txt, no size cap, no host allow-list, not modules
Row a bare dict with fixed keys; ats.py builds raw from a fixed list a host cannot hand over country, seniority, employment type, several places
Date trust TRUSTED_POSTED_HOSTS in app/models.py a host fact kept outside the host
Filters at the source none for boards; searches have per-provider params built in app/jobs/sources.py:search_params no way for a reader to say what it filters, nor for the runner to pass it
Oversized boards read incomplete on every pull (MAX_PAGES 100) no reading across pulls
Per-role text 200 per board per pull, never-tried first, inside the pull not ordered by score

What a reader must announce

The list below is what the 25 readers do today, plus what the wave-2 readers need. The user's list (whole board, paged, list then text, location filter) is complete for the listing. It misses the traits under "also".

Trait Values Who, today
Scope board (one company, its whole list) or search (many companies, never a whole list) 22 hosts, 3 providers
Listing whole (one request) or paged whole: Greenhouse, Lever, Ashby, Recruitee, Personio, Deel, Pinpoint, BambooHR, Jobylon, Rippling. Paged: the other 12 and the 3 providers
Text listing (full text in the list), excerpt (a teaser in the list), detail (one request per role), none detail: Workday, SuccessFactors, Oracle, Phenom, Eightfold, Gem, Jobylon, BambooHR, Rippling. Excerpt: Jobindex, Adzuna. None: SmartRecruiters, Workable
Completeness stated (the host gives a total), end (an empty page ends it), whole (one answer is all), never (a search result is never a whole list)
Page size and host ceiling integers; the ceiling is where the host stops answering whatever we ask Workday 2,000; Platsbanken offset 2,000; Jobindex 200; Eightfold 10 a page
Filters per dimension (country, place, remote, query): none, one (one value per request), many (a list in one request), ids (values must be looked up on the host first) country one: SmartRecruiters, Eightfold. Many: Phenom. Ids: Workday, Oracle, SuccessFactors. Query: the 3 providers. Greenhouse, Ashby, Recruitee: none
Places one, many, copies (see below)
Also: fields stated pay (figures, line, none), posted (trusted, listed, none), remote, country, seniority, employment seniority: Personio, SmartRecruiters, Recruitee, Jobylon, none mapped
Access key name (Adzuna), a smaller fallback feed (Greenhouse), the robots.txt decision taken when the reader was built
Title final or provisional (a sitemap reader knows a URL before a title) none yet; WP10 of wave 2 needs it
Empty answer trusted or suspect: whether an unknown slug answers an empty board suspect: Ashby, SmartRecruiters, Gem

Rate limits (429) and the per-request limits stay runner rules, the same for every reader. No reader has a need that differs.

Several places for one role

A role at "Berlin, Munich, Remote (DE)" arrives in three shapes today:

  • One place of several kept. Ashby, Lever, Personio, Greenhouse, Workable, Recruitee, Platsbanken read the first and drop the rest. SuccessFactors keeps the last.
  • All places joined in one line. Teamtailor, Deel, Workday, Phenom, Oracle, Eightfold, Jibe join them with "; ". country becomes EU over several countries, which the country gate does not expand.
  • One posting per place. The host itself lists copies with their own ids (1,741 groups, 5,021 roles). dupes.py leaves them apart.

The rule (owner, 2026-09-30): a role has a list of places. A reader never multiplies a row per place and never drops a place. Row.places is a list of Place(text, remote). Copies the host publishes per place are a host fact that a reader cannot undo. The reader declares places="copies", and dupes.py folds them by company and title into one role, keeping each copy's place with its own outbound link.

The contract test (WP A) feeds every reader that declares places="many" a fixture with at least two places and fails the reader if fewer come out.

Places and companies as symbols

Places and companies become rows that other rows link to, not text matched at query time. Today the place knowledge sits in code (_CITIES, _REGIONS, _COUNTRY_NAMES in app/models.py, regions and aliases in app/places.py) and is matched word by word against roles.place_fold for every query.

  • places(id, name, kind, parent_id) with kind one of city, region, country, continent: Copenhagen belongs to Hovedstaden, which belongs to DK, which belongs to Europe. Seeded from the dictionaries in code.
  • place_aliases(fold, place_id): kobenhavn, kopenhamn, copenhagen point to Copenhagen, and skane and skåne to Skåne. Typing either spelling finds the place.
  • role_places(uid, place_id, text, url): one row per place of a role. place_id is empty for a text the gazetteer does not know. That text is added to an unresolved list, which a person or a later job resolves, and it is linked once resolved. url is the copy's own link when copies were folded.
  • A country filter or gate is a join over the tree: DK expands to every place under DK, and Europe to every country under Europe. The EU fallback and the per-request word matching go. This also enables counts per place.
  • roles.place stays the displayed line: the first place, then "+2" when there are more.
  • company_aliases(fold, company_key): "Orsted" and "Ørsted" become one company without changing the primary key. Companies gain empty columns for enrichment (sector, stage, size band, founded, last funding, source and date of each fact), which the startupmap.one and The Hub readers fill later.

The API

The package plugins/atshosts/ becomes plugins/readers/, and the search providers move into it. The contract stays the package docstring, extended, with plugins/readers/_contract.py holding the types and problems(). plugins/README.md stays the list of sources and their measured limits.

# plugins/readers/workday.py
NAME = "workday"
LABEL = "Workday"
PATTERNS = (r"...",)

READS = Reads(
    scope="board",
    listing="paged", page_size=20, ceiling=2000,
    text="detail",
    complete="stated",
    filters=Filters(country="ids", remote="ids"),
    places="many",
    fields=Fields(posted="listed", pay="none", remote="stated"),
    robots="read: robots.txt allows /wday/cxs/ (checked 2026-09-29)",
    empty="trusted",
)

def slug_ok(slug) -> bool
def allowed(slug) -> Callable[[str], bool]
def careers_url(slug) -> str

# listing="whole"
def feed_url(slug, ask: Ask) -> str
def parse(body, slug) -> list[Row]
# listing="paged"
def page(get, slug, cursor, ask: Ask) -> Page
def rows(decoded, slug) -> list[Row]
# text="detail"
def detail(get, slug, row) -> dict
# filters with "ids"
def facets(get, slug) -> dict[str, dict[str, str]]     # {"country": {"DE": "<host id>"}}
# optional, as today
def fallback_url, robots_page, probe_url, probe_ok, probe_slugs

Ask is what the runner wants from one series of requests: countries, places, remote, query, params (a search's saved params). A reader reads only the dimensions it declared, and the runner never hands it a value it did not declare. With one, the runner makes one series per value and unions them by id. With ids, it calls facets once a day per board, keeps the result in state and puts host ids into Ask. Row becomes a TypedDict with places, and optional country, seniority, employment. The runner puts every stated key into raw, so a new field needs no runner change.

problems() checks the declaration against the functions: paged needs page and rows, text="detail" needs detail, an ids filter needs facets, scope="search" needs complete="never" or stated, key needs a config name. A module that fails refuses to import, as today.

TRUSTED_POSTED_HOSTS goes: fields.posted="trusted" says the same thing in the module.

robots.txt is decided when a reader is built, not checked at run time (owner, 2026-09-30). Whoever builds a reader reads the host's robots.txt, reports what it disallows, and the owner decides per reader whether to read the host anyway. robots is a required string that records the decision and the date. The runtime check, its state keys and its tests go. Jobindex is read with paging and area. HR Manager, left out because its robots.txt disallows reading, goes back to the owner for a decision.

The runner

plugins/readers/_runner.py, taken out of plugins/ats.py. plugins/ats.py and plugins/search.py stay as thin Source classes, so subscriptions and state keys (ats:*, search:*) keep their names and no migration is needed.

For one target, the runner:

  1. Loads the reader and its subscription.

  2. Resolves Ask. The reach is set in the front end by the agents that are switched on:

    • Base reach: the union of their any gates on country, town and remote, or everything when one agent has none.
    • Favourite companies (favourite = 2, which already skips the any gates): only the require gates, which today means Europe. Mistral or Databricks are read across Europe, so a role outside Denmark still shows up and can prompt a question to them.
    • Board override: params on the board's subscription.

    The reach is then narrowed to the filters the reader declares. A search's own params go in Ask.params. Boards under 300 roles get no filter.

  3. Applies the key.

  4. Reads the listing: whole, paged, or one series per filter value, unioned by id. Through one guarded get for every reader: host allow-list, https only, size cap, deadline, fallback feed. A teaser in the list is kept (Row.teaser).

  5. Sweeps across pulls. A paged read that hits MAX_PAGES saves its cursor and the ids seen, and the next pull continues from there. A sweep that reaches the end is complete and closes what it did not list. A host ceiling (Workday 2,000) is not swept: the reader must filter below it or stays incomplete.

  6. Decides completeness from the declaration, and applies the guards on closing: a filtered read closes roles outside the filter, except roles with a vote, an application or a note. A filtered or empty="suspect" read that returns under 10% of the last count counts as failed.

  7. Records failures (failing since and reason, 21 days to close), 429 handling and waits, as today.

The runner makes two runs (owner, 2026-09-30):

  1. Listing run (the pull). The agents hand the reach to the runner. The runner calls each reader with its filters and handles paging and sweeps. Readers return standard rows, with every place of a role. Then dupes.py folds copies, and embed and score run.
  2. Enrich run (new stage enrich). After scoring, the runner is handed the ranked roles. For each role that still needs a call (text through detail, and later other per-role facts), it fetches the highest-scored first, within a budget per board and per run. It drops the vector of each role whose text improved, and embed and score run again for those roles only.

The enrich run follows every rescore that changes the ranking materially: new roles from a pull, an agent created or edited, a company starred or unstarred. A vote on a role does not trigger it. A role never scored above the agents' thresholds gets no extra call. That is the budget which keeps the enrich run from reading every board in full.

Seniority: stated, then the model

43% of open roles have a level, all taken from title words. A sample of 59 without a level was 85% individual contributors. Personio, SmartRecruiters, Recruitee and Jobylon state a level that nothing reads. extract has never run.

A role keeps three sources of a level, and the displayed roles.seniority is the first that is filled, in this order:

  1. seniority_model: the model's verdict, from extract run as an afterburner. It reads roles whose model value is empty. The prompt includes the stated and title values as hints, and the model may overrule them.
  2. seniority_stated: the host's own word, mapped to the app's vocabulary (app.seniority.SENIORITIES) by a table per reader in _shared.py. The raw word stays in raw. A word the table does not know maps to empty and is logged, so the table can grow.
  3. seniority_title: seniority.from_title, as today.

The filters read roles.seniority. A later comparison of stated and model values measures how far each host can be trusted.

Decisions taken by the owner, 2026-09-30

  1. Filters come from the agents that are switched on. Favourite companies are read across the require reach (Europe), and a board can override its filter.
  2. robots.txt is decided per reader at build time. Nothing checks it at run time. Jobindex is read in full.
  3. Places are a list per role. Copies per place are folded, each keeping its own link. Places and companies become symbols with aliases.
  4. A teaser is kept. A short description in the list (Phenom and Oracle: 1,066 roles, plus Jobindex and Adzuna excerpts) goes into Row.teaser and roles.teaser. Until the full text arrives, it is shown and embedded, and a reader that sees one must hand it over. When the full text arrives, it replaces the teaser for embedding and the vector is dropped, as today.
  5. Seniority has a stated value from the plugin and a model value from an afterburner that fills what is empty and may overrule it.

Work packages

Each package owns the files listed and their tests. The planner makes each worktree from local main (see memory: isolation worktrees branch from stale origin), merges, and removes the xfail marks a package fixed.

WP Model Builds Owns Needs
A Contract Opus rename atshosts to readers; _contract.py with Reads, Filters, Fields, Ask, Place, Row (with places, teaser), problems(); the extended docstring with the robots.txt rule; a conformance test over every module (declaration against functions, places="many" and teaser fixtures) marking today's gaps xfail per module plugins/readers/__init__.py, _contract.py, _shared.py, every import of atshosts, tests/test_readers_contract.py
B Runner Opus _runner.py out of ats.py; runtime robots.txt check removed; Ask from params; series union; sweep; closing guards; ats.py and search.py thin; raw built from every stated Row key plugins/ats.py, plugins/search.py, plugins/readers/_runner.py, app/jobs/pull.py, tests/test_ats.py, tests/test_search.py, tests/test_pull.py A
C Providers Sonnet jobindex.py, adzuna.py, platsbanken.py as reader modules with teaser and places; search_params built from the declarations the three modules, app/jobs/sources.py, their tests A
D1 Hosts a to l Sonnet READS, all places, teaser, host-stated seniority mapped to seniority_stated, date trust in fields, robots decision string for Ashby to Lever those modules and tests A
D2 Hosts o to w Sonnet the same for Oracle to Workday those modules and tests A
E1 Places and teaser storage Opus places, place_aliases, role_places, roles.teaser, the three seniority columns, migration seeded from the dictionaries in code; upsert_role writes them; gates and the Place filter as joins over the tree; TRUSTED_POSTED_HOSTS removed; embed falls back to the teaser app/storage/*, app/models.py, app/places.py, app/jobs/embed.py, their tests A
F Reach and filters Opus reach from the agents and favourites into Ask; facets and Ask for SmartRecruiters, Eightfold, Workday, SuccessFactors, Phenom, Oracle; board override on the Companies page; each reader's declaration shown on the Sources page app/jobs/reach.py, those six modules, app/web/api.py, app/static/companies.js, app/static/sources.js B, D1, D2, E1
G Enrich run Opus enrich as its own stage over the ranked roles, highest score first; triggered after a pull's scoring, after an agent is saved and after a star changes, not after a vote; re-embed and rescore only the roles it filled app/jobs/detail.py, app/jobs/cascade.py, app/jobs/stages.py, their tests B
E2 Copies and place list Sonnet dupes.py folds copies per place into one role with per-place links; the reader shows the place list with links; the teaser is shown when there is no text app/dupes.py, app/static/roles.js, app/templates/*, their tests E1
I Seniority Sonnet seniority_stated, seniority_model, seniority_title and the precedence; extract as an afterburner over roles without a model value, with the hints app/seniority.py, app/jobs/extract.py, their tests (the columns come from E1) E1
H Companies as symbols Sonnet company_aliases, alias folding on insert and lookup, empty enrichment columns app/storage/* after E1, app/jobs/sources.py after C, their tests E1, C

Order: A first. Then B, C, D1, D2 and E1 in parallel. Then F, G, E2, H and I in parallel; H adds a migration after E1's. After the merges, the planner attaches Infineon and Microsoft on Eightfold with a country filter, updates plugins/README.md, and folds this plan into decisions.md.