Migration 14 moves each non-empty agents.intent into agent_prompts as one
prompt named after the agent, with its cached vector, and drops the three
intent columns. A role's fit is the maximum over the agent's prompts of the
calibrated cosine, and scores.prompt names the prompt that won. The agents
API takes prompts and still takes intent; the role endpoint serves the
agent's score and prompt.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A stored `stage:extract` / `stage:summarise` in the state table overrides
extract.enabled / summarise.enabled; with none, the file applies. One
function, stages.stage_on, decides, and both tasks' finders ask it.
GET/PUT /api/stages and `app stages` show and set it with who would answer;
Runs has a "What the model reads" block, and a search agent's card says when
a stage is off and links to the switch.
Co-Authored-By: Claude Sonnet <noreply@anthropic.com>
_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>
scale() maps a cosine to clip((cos - lo) / (hi - lo), 0, 1) with constants
from CALIBRATION in app/config.py, keyed by model and replaceable under
embed.calibration. multilingual-e5-small: fit 0.75..0.88, taste -0.14..0.19
(taste from synthetic votes). An uncalibrated model falls back to min-max
and logs a warning once per run.
The extract and summarise bars default to 0.55 and 0.65, about 300 and 35
roles for the starter agent on the measured data; stored limits are kept.
An agent without an intent is left out of the model-stage scope, and each
score run names such agents in one score.no_intent notice.
Co-Authored-By: Claude Opus <noreply@anthropic.com>
The pull records a board that fails on every fetch under ats:failing:<target>, and set_board can now move a company to another board. The old target's record would have outlived the board and counted in `app companies check` for good.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
recency and the newest feed both counted from first_seen, when the app
itself first stored a role, so a newly added board's whole backlog scored
as freshly posted for two weeks: 10,223 of 15,415 board roles were posted
more than 30 days before they were first seen (average gap in days: Ashby
78, Greenhouse 143, Lever 226, Personio 333).
Add roles.posted_at (migration 9): the earlier of first_seen and a trusted
host's published_at, where TRUSTED_POSTED_HOSTS (app/models.py) records
which of the ats and search hosts report a creation date rather than one
that moves on an edit, checked against live boards and the live database.
A future or pre-2015 date is not trusted either. upsert_role sets it on
insert and keeps the earliest on update, so an edit on the board can only
move it earlier, never later. recency, the feed's newest order, the "new
since" days filter and _top_scope's tie-break now read posted_at;
first_seen is unchanged and still what the UI shows as when the app found
a role. The row's age in roles.js now shows the posted date instead.
A partial index (posted_at DESC, uid) WHERE closed_at IS NULL serves the
newest feed; sqlite3's bundled build has no STAT4, so its planner
otherwise sorts every open role instead of using it, measured 30x slower
on the live database, so Store.roles() forces it for that query shape.
Co-Authored-By: Claude Sonnet <noreply@anthropic.com>
A second pull (schedule, Run button, `app job pull` in another process) could
start while one was fetching and overwrite the stored run marker, so the first
run's boards failed their stamp check or a settle closed on the other run's id
lists. `claim_run` sets `pull:run` in one conditional statement, only while it
equals `pull:settled` or is older than an hour; a refused `pull` job queues no
stage and says so in its result and a notice, and the CLI exits 1. Batches
carry the marker in their payload, the plugins stamp with it, and the settle
reads it off its own batches' payloads before releasing the run.
Co-Authored-By: Claude Opus <noreply@anthropic.com>
`app.jobs.sources.set_board` attaches a board to an existing company,
replaces the one it has (closing the old board's open roles) or creates
the company, with row and subscription changed in one transaction by
`Store.set_company_board`. A board that lists zero roles is kept. `add`,
`discover --add`, the seed importer (moved out of app/cli.py) and the new
`POST|DELETE /api/companies/{key}/board` all go through it; a seed re-run
attaches a board to a row that never had one. A company's subscription is
rewritten in place, and detaching hides it.
`companies.ats_slug` (migration 9) keeps the slug of a host the app
cannot read yet, and discovery returns such a host before probing slugs.
The Companies rows get Set board / Change, the host select is rendered
from `plugins.ats.HOSTS`, and a row with nothing found reads "board not
found" and is in the "Not readable yet" filter.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Workable and SmartRecruiters page their lists, and both parsers read page one
only, so the settle closed every role past it (Delivery Hero: 980 stated, 100
open). `_paged` follows nextPage and offset, and the plugin stamps
`ats:complete` only when the distinct ids read equal the stated total; a 429
after page one keeps what was read and closes nothing. Adzuna completes on its
`count`, Jobindex counts raw items and raises on an answer with no channel.
A board whose fetch raises is recorded under `ats:failing:<target>` with the
first failure's time and the latest reason; after STALE_DAYS the settle closes
its roles, and its next answer reopens them. `/api/companies`, the Companies
list and `app companies check` show it.
Co-Authored-By: Claude Opus <noreply@anthropic.com>
Measured on the live roles against the "Copenhagen engineering" agent's
intent: the English-only all-MiniLM-L6-v2 separates engineering from
non-engineering roles with an AUC of 0.68 overall but only 0.74 restricted
to Danish-country roles (a Danish engineering role beats the rest's
non-engineering median only 68% of the time). multilingual-e5-small reaches
0.70 overall and 1.00 on Danish-country roles, for about 3x the CPU time.
paraphrase-multilingual-MiniLM-L12-v2 measured worse than the English-only
baseline (AUC 0.60) and was dropped.
e5 needs "query: "/"passage: " prefixes; app/jobs/embed.py adds them by
matching the model name so a model change can't leave a stale prefix
setting behind. Added tests for the prefixing and for the two invariants
the model change relies on: embed drops the old model's vectors before
re-embedding, and an agent's cached intent vector is keyed by model so a
model change is a cache miss, never a stale-dimension read.
Co-Authored-By: Claude Sonnet <noreply@anthropic.com>
A require gate on an empty column gated the role ("country unknown"), and extract, the stage that fills the column, is bounded to the roles that rank, so 2,281 of 15,556 open roles could never come back: every board that writes "Remote", "Hybrid" or "Distributed" as its location. An empty column now neither passes nor fails a gate; the role ranks, extract reads it, the rescore decides.
`any` is a third gate mode: the any gates are one group a role must match one of, so a ranking can say "in Denmark, or remote", which require alone could not. The group is gated only when some member had evidence against it.
`remote_of(place)` fills `remote` from the location line the way `country_of` fills `country`, and the startup backfill does both, since Greenhouse writes "Hybrid" as the location and gives no flag.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Run one stage goes: the cascade is the unit of work and app job keeps the
per-stage entry points. The pull button is disabled while a pull runs and
the lede no longer hardcodes the cadence. Messages folds into Runs as
Problems. pull_batch checks for a cancel before every board, one plugin per
target, and both plugins now close their HTTP client.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
agents.limits holds extract_top, extract_min, summarise_top and
summarise_min; the scope query reads them per ranked row, and a row saved
before this behaves like a fresh agent. The four global config keys go.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
`app companies prune`, `/api/prune` and the Sources disclosure switched off every board with no open role in a region. It existed because the first seed was 186 US boards; the European seed makes it moot, and it was wrong on its own terms: a board with no European role today (a company between openings, or a fetch that timed out) was switched off for good. A company that hires nowhere near should not be in the list, and a ranking's country gate handles the rest. `EUROPE` in models had no other reader and goes with it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The seed was jobhunt-hq's 186 US boards, chosen because they could be read. It is now a hand-picked list of European tech employers, with `board` filled by the new `app.jobs.discover` from each company's own careers site and every board fetched once before it landed: 175 have one, and the rest keep the ATS host their site runs on (Teamtailor 32, Workday 12, SuccessFactors 9, ...) and a careers URL, as rows the app cannot read yet. Companies shows them as "not readable yet" with a link, behind a filter, so the next parser to write is the host that filter shows most.
`app companies discover NAME DOMAIN [--add]` reads the home page, its careers links and the usual paths for a link into a known host, one level down on a careers page, and probes Greenhouse, Lever, Ashby, SmartRecruiters, Recruitee and Personio by slug when the pages link nothing. It is what the LinkedIn flow will call.
Migration 7 adds companies.ats, country and city. The Workable parser now reads the v3 shape its list endpoint actually sends (location object, department list, shortcode and no URL or text); the first populated board showed it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The free stages run first and their score bounds the two model stages.
Extract reads only the best extract.per_agent roles of each agent at
extract.min_score or more, then score runs again so the gates read what it
filled; summarise gets a min_score of its own. The cascade steps by
position, since score is in it twice.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The per-agent migration becomes v6 behind main's v5. A vote still stamps
the role's read_at, tag counts hide the agent's hidden companies, and
/api/filters takes the agent like /api/roles.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Votes on roles and company affinity move to role_votes and agent_companies,
one row per search agent, so two agents rank the same roles differently. The
score gains taste, the cosine to mean(liked) - mean(passed) of the agent's
votes, as newsdesk ranks articles. The API takes agent on roles, votes and
companies and falls back to the default agent, the enabled one with the
lowest id.
Settings become Search agents (a list again, with New and Delete), Sources
(job boards, company boards and runs with the message log, one section per
plugin) and Account. Old paths answer 301 to their section.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The Roles tab takes newsdesk's shell, ported from its index.html and
app.css rather than rewritten: a filter rail (Show read, Best first or
Newest first, first seen, tag chips with counts, the old filters behind a
disclosure), the feed in newsdesk's row layout (score gutter, title,
company, place, badges, age, tags, the summary as excerpt, judge buttons)
and the reader as a third column that replaces the list on a phone. The
router pushes /roles/{uid}, which the server also renders with the reader
open, so the separate read page and role.js go. j/k, u/n/d, S, Enter,
Escape, o, /, f and c work as in newsdesk. The phone rules come with their
comments.
Two things the row lacked: roles.read_at, stamped when the reader opens a
role or a vote is cast, behind include_read on /api/roles; and role_tags,
written by summarise alongside the summary and served with counts by
/api/filters, filtered in SQL before LIMIT. Migration 5.
PyCharm run configurations in .run/ for serve, pull, seed and pytest.
core.autocrlf=input is set locally so a Windows-side checkout commits LF.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
app/jobs/summarise.py is a ModelTask beside extract: about 60 words per role
from title, company, place and the fetched page else the description (capped
at 8000 characters). The pull cascade gains a summarise stage after score that
takes only roles without a summary in the top summarise.per_agent of an
enabled agent's ungated scores, and finds nothing when summarise.enabled is
false or no model is reachable. A reply that leaves a role unanswered is
counted in the state table and the role is given up after three.
POST /api/roles/{uid}/text scrapes original_url once and stores
article_text; POST /api/roles/{uid}/summarise queues a summarise_batch for
that uid alone, deduped per role.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The top bar is Roles and Applications plus health's cog. Behind the cog,
/settings/{ranking,searches,companies,pulls,messages,account} share one row
of sub-tabs, an underlined row on a desktop and pills on a phone. The old
paths answer 301 with their query string.
Ranking edits the first enabled agent, or creates one, with the intent,
the weights and the gates under Advanced. The order menu on Roles reads
Newest and Best match with one enabled agent, and names them with several.
config.yaml keeps the Copenhagen starter; the other two are commented
examples in config.example.yaml.
roles.js is split into roles.js, applications.js, ranking.js and
companies.js; count, cell, debounce, the company list and the affinity
control live once in app.js. app.css gets the phone frame after health and
newsdesk: one-row bar, edge-to-edge panels, 44px targets and 16px fields on
a coarse pointer, hover only where there is one, safe areas. The app is
installable from a manifest and roledesk-coloured icons. Every lede touched
is written for the person using the app.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
What a pull reads can now be changed on a page instead of only the command
line: try a search and see its first ten results before saving it, edit or
remove one, follow a company's ATS board, switch any board or search on or
off, and preview and apply a prune by region. Each row shows its open roles
and the countries they are in.
The page and `app companies`/`app searches` call the same functions in
app/jobs/sources.py. Switching a target off closes its open roles, and the
pull's settle closes them again if a batch in flight reopened them. Roles
pulled before `country` existed are backfilled at startup.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The ats boards only read companies someone listed, and the seed list is
US-centred: 31 of 29,000 open roles were in Denmark. A `search` plugin runs
saved queries on boards that span companies, pulled in the same task as the
boards: Jobindex (Denmark, RSS, no key) and Adzuna (UK and most of the EU,
free key, built to its documented API). Jobnet and Jooble were checked and
left out.
`country` is now filled from the board's location line by a lookup, with EU
for Europe-wide roles, so country gates work without extract. `app companies
prune --in` unsubscribes boards with no open role in a region. The starter
agents and searches are Copenhagen and Europe.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With fit as (cos + 1) / 2 the real run spanned about 0.55 to 0.79, so a
prefer gate at weight 0.2 moved a score more than the whole spread of fit.
fit is now the cosine min-max scaled across the roles an agent scores in
that run: the closest is 1, the farthest 0, a single role 1, and a role a
require gate rules out stays out of the range. The starter agents' prefer
weights drop to 0.05, enough to break a tie. A regression test pins the
case: two close cosines that the old mapping let a 0.05 gate reorder.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
app/main.py assembles it in the boilerplate's order, auth.connect first,
then the queue with the scheduler and the pages. TABS is Roles,
Applications, Agents, Companies, Jobs, Schedules, Messages and Account:
the alpha page's three views become templates on the boilerplate's
base.html, driven by static/roles.js (the old app.js) through api() and
toasts.js, over the unchanged JSON routes. Companies lists every board with
its open roles and the affinity control. A 422 is one sentence, so a toast
never reads [object Object].
Config takes the boilerplate's shape: APP_CONFIG or config.yaml, a missing
file is the defaults, an unknown key is an error. roles.yaml is gone;
config.yaml is the committed default and carries the starter agents.
pull is now a job that asks for the pull cascade, because a schedule
enqueues a job kind; its default schedule, every 6h, is declared beside it
in app/jobs/cascade.py. pull_batch's targets are keyword-only so the Jobs
tab offers no batch button. `app worker` is the boilerplate's; `app job`
still drains a queue of its own and fires no schedules.
test_coupling holds pipeline, auth and desk apart (desk may import
pipeline) and plugins to desk; test_architecture keeps app's layers, now
with main.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The boilerplate's rule: the deploy kit's Dockerfile, the APP_CONFIG,
APP_PASSWORD and APP_DATA variables and its commands all assume a package
called app, so a project renames its display title and nothing else. The
console script is `app`, and the launch entry runs `uv run app serve --port
8790`. The domain noun stays `role`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>