One deduplicated, newest-first job feed from 110+ job boards in 60+ countries, 40+ ATS career systems and any company career page. Deepest in Switzerland, Germany and Austria. Daily new-jobs-only runs, radius filter, agency flag, and MCP for AI agents.
The status message now says "billed by the board" for rows over maxResultsTotal ("2 over maxResultsTotal, billed by the board") and for rows skipped by onlyNewJobs ("2 already delivered, skipped (billed by the board)"), as it already did for rows outside the radius.
New run_summary.rows_over_max_results_total: the number of jobs left out because of maxResultsTotal, including rows left over from the fair share between companies. It is left out when there are none.
A job taken from a company's own career page or website (companyUrls, careerSites) no longer gets an agency signal when the company only matches the staffing licence list. Before, a plain "Swisscom" job showed is_agency_signal: true and agency_type: "staffing". Staffing agencies with a high or medium match are unchanged on these rows.
On job board rows, a licence_only match now has agency_type null. agency_confidence stays licence_only and is_agency_signal stays true. excludeAgencies is unchanged and never drops these rows.
2026-11-18 (0.9.205)
With several companies or career sites, maxResultsTotal is now shared fairly. Each gets up to its equal share while results arrive; any share left unused is filled from the others at the end. Before, the company that finished first could take the whole total. The status message and run_summary.max_results_total_split show the split. One company, or no total, behaves as before.
When maxResultsTotal is set, each company or career site is now asked for no more jobs than the total, so you are not charged for rows that could never be delivered.
The status message no longer shows "0 already delivered, skipped".
The same career-site job found through a company website and through its careers page (for example a Workday posting) is now recognised as one job: it gets matching job_id_aliases from the posting URL and the job requisition number, is merged within a run, and is skipped by onlyNewJobs when the other path delivered it before. The job_id itself is unchanged.
2026-11-18 (0.9.204)
run_summary.rows_limited_by_max_charge and the status note "max charge limits some boards to fewer rows" now also list a board whose sub-actor stopped at its max charge and delivered fewer rows than wanted, for example jobware-de with 1 row of 10 on a $0.10 max charge. It uses the same stop as the "sources stopped at max charge" count, so the two always agree. A board that simply ran out of matches is still not listed.
jobscout24-ch, jobbasel-ch and ostjob-ch now count a start price of $0.001, as the boards charge. Their job price stays $0.003. The minimum max charge for these boards is slightly higher.
stellenanzeigen-de now runs with 256 MB and meinestadt-de with 128 MB. Results and prices are unchanged.
Internal price checks now accept the price of any plan tier, so they no longer report a false difference for accounts on a paid plan.
2026-11-18 (0.9.203)
run_summary.rows_limited_by_max_charge and the status note "max charge limits some boards to fewer rows" now count the rows a board really delivered, not the rows planned at start. A board that delivered its full page is no longer listed, and the note is left out when no board came up short.
arbeitsagentur-de prices are now $0.005 per start, $0.003 per job and $0.006 per job detail. The minimum max charge for this board is lower, and more of a small budget is left for its rows.
Reasons and status messages no longer show ".;" or ". ." where two parts are joined. A reason now reads "0 jobs delivered. jobup-ch covers ...".
With a radius, a job that lists several cities (for example "Düsseldorf, Berlin, Bremen") is now kept when any of these cities is inside the radius. Names that are not places, such as "mobiles Arbeiten", are ignored. The job's own city field is not changed.
2026-11-18 (0.9.202)
With a max charge, the budget for boards is now split by what a full page of each board costs, not equally. A board with a dear start price, such as arbeitsagentur-de, is no longer skipped while cheaper boards leave most of their share unused.
A board whose share is too small now starts later: after the first boards have finished it gets one more try with what is left. If it still cannot start, its reason says "not enough budget left after the other boards finished (needs at least $X)". "max charge reached" is only used when nothing is left.
When a board starts with less budget than a full page, the status message ends with, for example, ; max charge limits some boards to fewer rows (arbeitsagentur-de 8, stepstone-de 6). The boards are also listed in run_summary.rows_limited_by_max_charge.
Coverage warnings (outside_service_country, outside_service_region, location_ignored, outside_board_country) are now also listed in run_summary.warnings, and the status message says how many boards are outside the search place. A board with such a warning that delivers 0 jobs now carries the warning in its per_platform reason.
A row with a city and a wrong country, such as Zürich labelled DE, is now placed by its city when the city exists in only one of CH, DE, AT and LI, so the radius filter can drop it. Names that exist in several of these countries stay as unknown.
nicejob-de now runs with 256 MB memory. Results and prices are unchanged.
2026-11-18 (0.9.201)
nicejob-de is now a Swiss board that also serves Vorarlberg (Austria). It no longer belongs to the German list, and a search with countries: ["DE"] never starts it. With countries: ["CH"] and includeRegionalBoards: "all" it is added with the other Swiss regional boards. The board id is unchanged.
A search place in Bregenz or Dornbirn no longer gets an outside_service_country warning for nicejob-de.
When the country was taken from your location because you set no countries or platforms, the run status message now ends with, for example, ; country CH inferred from Lausanne.
Lower default memory for two boards started by job-feed: arbeit-swiss runs with 256 MB and arbeitsagentur-de with 128 MB. Results and prices are unchanged.
2026-11-18 (0.9.200)
When your search place is in a different country than a national board, the board still runs but now gets a warning in its run_summary.per_platform entry and a coverage_warnings entry with reason outside_board_country, for example "stepstone-de is a DE board; the search place Lausanne is in CH, most rows will be outside the radius". This applies to boards job-feed picked from countries and to boards you named in platforms. No board is skipped for this reason.
An input with a location but no countries or platforms now takes the country from the location (Lausanne gives CH, Berlin gives DE) instead of defaulting to DE. The input warnings say, for example, "no countries or platforms set; country CH inferred from location Lausanne". If the location is not recognised, the DE default still applies.
2026-11-18 (0.9.199)
When your search has a place (a location that is a city or region, with enforceRadius on), job-feed no longer starts boards it picked itself that cannot serve that place, so they bill nothing. This covers glassdoor, which ignores the city and returns country-wide rows, regional boards whose regions your radius does not reach (for example jobbasel-ch for a 10 km search in Zurich, or westjob-at, which covers Tirol and Vorarlberg, for Wien), and nicejob-de and myjob-ch for places outside their country. A board with regions is skipped only when no part of your radius reaches them. Searches without a location, or with only a country name, are not affected.
Skipped boards are listed in run_summary.platform_selection.skipped_outside_coverage (platform_id, reason and the regions or countries the board serves) and appear in run_summary.per_platform with status SKIPPED_COVERAGE. They do not count as failed, and the status message names them.
A board you name yourself in platforms always runs. If it cannot serve your place, it gets a warning in its run_summary.per_platform entry, and run_summary.coverage_warnings lists every such board with platform_id, reason and message.
When a website could not be read because it needs a browser, run_summary.per_platform reasons and the run status message now name the input that fixes it: hiringBrowserFallback in companies mode, or adding website-job-extractor-browser to platforms otherwise. In companies mode the companies_not_read detail no longer also suggests adding that board to platforms.
2026-11-18 (0.9.198)
jobscout24-ch now receives your radiusKm. With a location and radiusKm above 0, the board keeps jobs within that distance of the place. A radius above 200 is sent to the board as 200, its maximum, and job-feed's own radius filter still applies your full radius. Without a location, with radiusKm 0, or with a country name such as "Switzerland" as the location, no radius is sent. The region mapping is unchanged.
In run_summary.companies_not_read, the advice to turn on the browser now names a job-feed input. In outputMode: "companies" it says hiringBrowserFallback. In jobs mode it says to name website-job-extractor-browser in platforms. Before, it named the extractor's own enableBrowserFallback input, which you cannot set in companies mode.
run_summary.per_platform[].dropped_too_old is now an explicit 0 for every board that ran, like dropped_off_location. It stays absent for a board that never ran, such as one skipped for budget.
2026-11-17 (0.9.197)
In outputMode: "companies", a company whose homepage was fetched fine but has no careers, jobs or job-system link now gets its own hiring_statusno_careers_page (with hiring_evidenceno_careers_page). hiring_open_roles_count is null, hiring_roles and hiring_roles_detail are empty and not_read_reason is null. It no longer counts as no_open_roles, is no longer listed in run_summary.companies_not_read, and run_summary.companies_mode.no_careers_page counts it on its own. The status message says "N without a careers page". no_open_roles now means only that a job system or a real careers page was read and held no open roles. Jobs mode is unchanged: such a company is still listed in companies_not_read with reason not_a_career_page. Companies that could not be read because a sub-run failed or extraction calls failed are still listed in both modes.
When a website-job-extractor chunk stops at its charge budget while other chunks left budget unused, job-feed now re-reads the stopped companies in one follow-up run, capped by what is left of your max charge. If the leftover is below the start minimum, no follow-up run starts and those companies keep the reason charge_limit. With a small max charge, job-feed also uses fewer, larger chunks so each chunk gets at least twice the start minimum.
Every entry in subruns now has started_at, finished_at and duration_s, and build_number is filled from the build tag when you pin an exact version such as 0.3.11.
Companies the website-job-extractor browser fallback has read (browser_fallback_result, superseded_by_browser) count as covered: they are no longer listed as not read, in jobs mode and in companies mode.
For UK jobs, a city written with a postcode district, such as "Stockport (SK1)" or "Manchester M1", now shows the town only ("Stockport", "Manchester"). Job ids do not change.
The browser retry is no longer counted as a board in the status message or in the skipped-for-budget list. When no board could start because the max charge is too low, the status message now starts with "No boards ran: max charge too low (...)" instead of "FAILED:", because the run ends SUCCEEDED.
Each run_summary.per_platform[] entry now shows dropped_off_location, which includes rows outside your radius, and dropped_too_old, the rows that board delivered older than postedWithinDays. The status message says the radius drops are "billed by the board, not by job-feed". The README now states that job-feed's own event is charged only for delivered jobs, while a board may already have billed rows it delivered outside the radius or older than postedWithinDays. Boards that honour location, a tighter location and the board's own posted-within filter reduce that.
With onlyNewJobs, run_summary.only_new has a new seen_skipped_rows count. seen_skipped still counts distinct jobs, and a job returned by several boards is one job but several rows. The status message now says "N already delivered, skipped" with the row count, so rows from boards = unique jobs + merged + skipped and dropped rows.
The README now explains that a board or extractor skipped for budget appears in run_summary.per_platform with status SKIPPED_BUDGET, a reason and floor_usd, which is the board's own start price plus one result at its standard (free-plan) price.
2026-11-17 (0.9.195)
The minimum budget needed to start a job board now includes that board's own start price (memory-based where the board uses Apify's start event), plus one result at the price the board charges for your input (the detail price when you ask for job details). job-feed no longer starts a board with less budget than its start charge. For example, arbeitsagentur-de needs 0.058 (0.05 start plus one 0.008 result) and jobs-ch needs 0.00387. A board with no known start price keeps a 0.01 floor.
Corrected the per-result prices job-feed uses for seven boards to match their live prices: stepstone-at (0.003), welcome-to-the-jungle (0.001), naukri (0.012), e-fellows (0.0015), tes-uk (0.004), charityjob-uk (0.0014) and glassdoor (0.0005). Cost reports and budget checks for these boards now use them.
The country list now shows 90 national boards (it was 91) after the Guardian Jobs board was removed.
2026-11-17 (0.9.194)
job-feed no longer downloads each sub-actor's full actor record at run time to read prices. It uses a built-in price table instead, so runs make fewer API calls on your account. In cost_report, price_basis is now "registry" (was "actor_fallback") when job-feed priced a sub-run from that table.
The minimum budget needed to start a career-site extractor now covers its first company and its first job, so a nearly used-up max charge no longer starts an extractor that cannot return a job.
The status message now reads "(incl. N nationwide rows kept, no place given)" instead of "+ N nationwide rows kept (no place given)". Those rows are already inside the unique count, so the old "+" looked like an extra amount.
The Guardian Jobs board (guardianjobs-uk) is no longer supported because its actor is no longer published. It stays in the platforms list marked as no longer supported, so existing inputs still work: if you list it, it is skipped and shown in run_summary.per_platform with status SKIPPED_UNSUPPORTED and the reason "guardianjobs-uk is no longer supported, skipped", and the status message says so. It is not part of any country preset.
A board or browser retry that is not started because your max charge is used up now shows the reason ("max charge reached, ...") and the start floor it needed (floor_usd) in its own run_summary.per_platform entry. The browser retry shows "max charge reached, browser retry not started".
2026-11-15 (0.9.192)
LinkedIn jobs in Switzerland written as "Town, Canton, Country" now show the town as the city and the canton code in region. For example, "Therwil, Basel-Country, Switzerland" gives city Therwil and region BL, and "Winterthur, Zurich, Switzerland" gives city Winterthur and region ZH. Before, a search for Basel or Zürich could replace the town with that city. Job ids do not change.
Because the city is now the town, these jobs are matched with the same job from other boards and career sites by town: LinkedIn Therwil merges with the Therwil copy, and LinkedIn Winterthur no longer merges with a Zürich ad. A job already delivered with onlyNewJobs is not delivered again.
The radius filter measures these jobs from the town, not from the search city.
When the radius filter keeps jobs that name only a country (for example "Deutschland") and are not remote, the status message now says "(incl. N nationwide rows kept, no place given)", and run_summary.radius_filter has the count in kept_nationwide.
The website field in the dataset schema has a more realistic example.
Fixed the status message count: when two boards listed the same job and both copies were outside the radius, only one was counted in "outside the radius", so the numbers were off by one. Each dropped row is now counted. The status message now uses the singular for a count of 1 ("1 row from boards", "1 unique job", "1 nationwide row kept").
2026-11-15 (0.9.190)
The dataset now has one view, Jobs, that shows every field of the output, with the most useful columns (title, company, location, posted date, apply URL) first.
Every field in the output schema now has a description and an example value.
The README has new Pricing and Support sections.
A location such as "bundesweit", "deutschlandweit" or "schweizweit" no longer shows up as a city. The city stays empty, and the country is filled in when the wording names it.
A location such as "Lörrach, bei Basel" now gives the city Lörrach. A place after "bei", "nahe", "near" or "in der Nähe von" is only a reference point, and is used as the city only when it is the sole place named.
The radius filter now finds the search location in any country. For example, searching "Basel" with only Germany selected now applies the radius around Basel (CH), so nearby German towns are kept and distant ones are dropped. Your country selection still decides which boards are used.
2026-11-14 (0.9.189)
Village names that are part of a combined municipality name (for example Räterschen in Elsau-Räterschen) are now recognised for the radius filter.
Rows that name only a broad Swiss region (for example Ostschweiz or Innerschweiz) are now checked against the radius: kept when the region is near the search location, dropped otherwise.
When a source says a job has no city (for example only a department name is given), the city field now stays empty instead of repeating that text.
A job that names only a Swiss region or canton (for example Mittelland or Kanton Aargau) now has an empty city and the region in the region field (the canton code for a canton, such as AG). Job ids and the radius result do not change.
2026-11-14 (0.9.187)
postedWithinDays is now passed to meinestadt-de, ostjob-ch and wellfound, so older jobs are skipped by the board itself instead of being fetched and dropped later.
The memory planning log mentions the account-limits fallback only when that fallback is actually used.
memory_peak_used_mb is now documented: memory reserved for concurrent sub-actor runs, not measured usage.
2026-11-14 (0.9.186)
The "memory budget resolved" log line now gives one clear reason when the budget is the 4096 MB fallback, including that the run's own memory is not the account's free memory. To set the budget yourself, use the existing maxTotalMemoryMbytes input.
Tracking parameters are now removed the same way from apply links and hiring page links: utm_*, gclid, gbraid, wbraid, dclid, fbclid, msclkid, igshid, yclid, mc_cid, mc_eid, _hsenc, _hsmi, ref_src, trk, trackingId and rltr. All other parameters stay in place, and job ids do not change.
The run status message now separates its parts with a colon instead of a long dash. The wording of the row counts and warnings is unchanged.
When you set a radius but it could not be applied (for example the search location is not in our place list, or it is a region or country), the status message now says so and gives the reason.
2026-11-14 (0.9.185)
Company names in the run summary now match the job rows. A companyUrls entry without a name shows the same name in companies_not_read and companies_partially_read as on its jobs (your name first, then the company name found on its postings, then the built-in directory name, then the name derived from the host, with CamelCase hosts split into words). The status line now says "1 company not read" and "1 career site only partly read" in the singular. job_id and dedup_key are unchanged.
The radius filter now also reads a row's own coordinates when a board sends them nested (a location, coordinates or geo object, or arbeitsagentur-de work locations), not only as top-level latitude and longitude. Values outside the valid range are ignored, and a row with its own coordinates is measured from them instead of from the city name.
The enforceRadius filter now drops a row whose country is known, is not the search location's country and is not one of its land neighbours (for example a US or UK row in a Zurich 20 km run), even when its city cannot be placed on the map. Rows without a country, and rows from a neighbouring country (France, Italy, Germany, Austria and so on) with an unknown city, are still kept, because a town just over the border can lie inside the radius.
Agency detection: Evoke (Evoke Staffing, evokestaffing.com) is now recognised as a staffing agency (is_agency true). A company domain whose name ends in staffing, recruitment, recruiting, personnel, interim or uitzend (for example acme-staffing.com) now also marks the row as an agency; software vendors such as recruitmentsoftware.com do not.
The minimum budget needed to start a source now uses the highest price tier on that source's rate card, so a source whose price lists only some tiers is no longer treated as free, and a start that cannot buy one row is skipped for budget instead of being refused by Apify. When a source lists several start events, the dearest one is used.
The "memory budget resolved" log line now says why the account memory limit could not be read (the failing request and its HTTP status, for example when the run's access token may not read account limits). The memory budget itself is unchanged: 4096 MB when the limit cannot be read, or the value you set.
2026-11-14 (0.9.184)
The status message no longer counts rows the dataset could not save. When a dataset push fails, the "unique jobs" figure and unique_jobs in the run summary now leave those rows out, and the line names them: "3 not saved: dataset push failed". Normally a row is charged only after it lands; close to your max charge the charge is made before the push, so a row that then fails to save can still be charged once (it is named in this bucket).
New input includeRemoteWithoutRegion (default off), in the career-site section. When you filter career sites by country, remote jobs that say only "Remote" are left out by default. Turn this on to keep them (they may not be open to applicants in your country). It is passed to career-site-jobs only when on.
A companyUrls entry without a name now gets the company's name from the built-in company directory when its domain is listed there ("Jet Aviation" instead of "Jetaviation", "Beyond Gravity" instead of "Beyondgravity"). A name you give always wins, and domains not in the directory keep the name derived from the host. Only the displayed name changes; job_id and dedup_key stay the same.
2026-11-14 (0.9.183)
The status message now also counts rows that were not delivered after deduplication, so the numbers add up: "2 over maxResultsTotal", "1 not accepted: max charge reached" (charge-stop modes only) and "3 not delivered: max charge reached" (rows held back when the next charge would pass your max charge). A bucket shows only when above 0.
career-site-jobs rows with workplace_type "on_site" (and other underscore spellings such as "fully_remote") now fill location.remote; additional_locations from career-site-jobs are added to locations[] after the main locations, without duplicates. job_id and dedup_key are unchanged.
Lifecycle: when a duplicate listing on another board is merged before the job's own row is recorded in the same run, that board now still counts as a source of the job, so a later run does not close the job early.
2026-11-13 (0.9.182)
Fixed a duplicate: a posting that jobs.ch (or jobup.ch, jobscout24.ch) lists under a uuid, and that arbeit-swiss delivers twice under two job-room ids, came out as two rows when the first row had the employer career-site link (Swiss Life "AI Software Engineer"). A row whose JobCloud uuid is already in the group now always joins it; it gets no "-xxxxxxxx" id any more. A row with a new uuid from the same board and another source id still stays separate.
2026-11-13 (0.9.181)
arbeitsagentur-de now gets your radius as radius (km, at most 200) when a location is set and radiusKm is above 0. The enforceRadius filter still runs afterwards.
The status message and run_summary.dedup_summary now count rows dropped after the board returned them, so the numbers add up: "75 rows from boards → 47 unique jobs (13 merged: indeed ×13, 15 outside the radius)". Shown only when above 0: outside the radius, below the relevance score, older than postedWithinDays, not matching the title filter, outside the location, in another country, from excluded agencies.
description_md no longer shows HTML entities ("&" shows as "&"); description_html stays raw.
Swiss rows written with ae/oe/ue now show the umlaut when the place is in the registry: "Baden-Daettwil" as "Baden-Dättwil", "Muenchenstein" as "Münchenstein". Display only; job_id is unchanged.
Board rows with a per-posting link (jobs.ch, jobup.ch and jobscout24.ch by posting uuid, other boards by apply URL) now answer to an apply-URL id in job_id_aliases, so a retitled posting keeps its row in later runs. Search and list pages get no alias. Rows from earlier runs do not have it.
2026-11-13 (0.9.180)
The same posting listed on two JobCloud boards under different posting ids (for example jobscout24-ch and an arbeit-swiss row with a jobup.ch link) is now one job when title, company and city are exactly equal and the boards differ. The later row's job_id goes into job_id_aliases. Two ids from the same board, or different cities, stay separate.
Career-site rows (career-site-jobs, website-job-extractor) now also answer to two extra ids in job_id_aliases: one built from the apply URL (tracking parameters, "www .", scheme, trailing slash and fragment ignored) and one built from the company name without punctuation and one trailing legal form ("Huber+Suhner" = "Hubersuhner", "Baumer AG" = "Baumer GmbH"). job_id itself does not change.
Swiss rows without a declared language: a detectable description wins; otherwise a French title ((H/F), technicien, responsable, ...), a jobup.ch link, the canton VD/GE/NE/JU or a French-speaking city (Lausanne, Sion, Yverdon, Nyon, ...) gives "fr". Other Swiss rows stay "de".
Swiss rows (and rows without a country) show the local city name for bare English, French and Italian spellings: Zurich as Zürich, Geneva as Genève, Berne as Bern, Basle as Basel, Lucerne as Luzern. Display only; job_id and duplicate keys are unchanged.
Company names no longer show HTML entities ("&" shows as "&"); job_id still hashes the raw name.
apply_url no longer carries utm_*, gclid, fbclid, msclkid, dclid, igshid, mc_cid and mc_eid; other parameters and the fragment stay.
2026-11-04 (0.9.179)
When boards are skipped because the max charge is too low, the run status message now says "4 of 7 boards ran; 3 skipped because the max charge is too low (jobs-ch, glassdoor, indeed; raise the max charge to include them)" instead of "4/7 platforms succeeded", so skipped boards are not read as failures. At most 5 board ids are listed, then "and N more", and the skipped boards are no longer counted a second time at the end of the message. run_summary fields are unchanged, and the message is the same as before when no board was skipped for budget.
2026-11-04 (0.9.178)
A source that bills per row through Apify's dataset item event (arbeit-swiss) now has that price counted in its start floor and its row price, so an allocation that cannot buy one row is skipped for budget instead of being refused by Apify with max-items-must-be-greater-than-zero. A source with a named per-row event is unchanged.
2026-11-04 (0.9.176)
A source row can now carry company_legacy, the company value the source used before it changed (jobup-ch used the placeholder "Jobup"). job-feed adds the job_id built from company_legacy to job_id_aliases of that row, and the onlyNewJobs store treats that id as seen, so a change of company name does not make old jobs look new. The job_id itself and rows without company_legacy are unchanged.
2026-11-02 (0.9.174)
career-site-jobs now reports the employer name instead of the Teamtailor tenant slug as company (for example "SNF FNS SNSF" instead of "snsf"), which would change the job_id of every Teamtailor job. job-feed now adds the job_id built from the slug (company_slug when the row has it, otherwise the subdomain of the teamtailor.com address) to job_id_aliases of those rows, and the onlyNewJobs store treats that id as seen. Other rows and the job_id itself are unchanged.
2026-11-02 (0.9.172)
Companies mode without companyUrls now stops at the start with a clear message: outputMode "companies" needs companyUrls (one company website per entry); it does not search job boards, so countries, platforms and keywords are ignored. Use outputMode "jobs" for a board search, or add companyUrls. The message names the search fields that were set and ignored. Before, the run failed with "input requires at least one of: countries, platforms, ...", which was wrong because countries had been set (seen on two runs with countries CH, platforms jobs-ch and jobscout24-ch).
The status text of that error now ends with a single period and the hint "Set companyUrls or switch outputMode to jobs." instead of the generic input-tab hint. includeRegionalBoards is no longer listed among the ignored fields.
Jobs mode, a bare {} input and companyUrls lists where every entry is invalid are unchanged.
2026-11-02 (0.9.171)
Companies mode: hiring_jobs_url no longer carries tracking parameters. utm_, gclid, fbclid, msclkid, mc_cid, mc_eid, _hsenc, hsmi, yclid, igshid and ref_src are removed; every other parameter, the fragment and the letter case of the URL stay as they are. Seen on a Google Maps Hiring Signals row ("South Austin Dentist") that showed the practice homepage with utm parameters.
Companies mode: when the roles were read from a website (hiring_evidence careers_page), hiring_jobs_url is now the page the roles were read from, not the company homepage. Rows that come from a job system (ATS) are unchanged.
Gender markers are now removed from the title anywhere in it when job-feed builds the job_id and the duplicate keys, not only at the end. This covers m/w/d, H/F, "w, m, d", "m / w / d", (w/m/div), (gn), (m|w|d), (alle), (e), the forms with a middle dot (trice, euse, e), and "/in" only after "-er" ("Plug/in" and a bare "F/X" stay). Example: "Laborant (m/w/d) Chemie" and "Laborant Chemie" are now the same job. No double spaces or dangling dashes are left where a marker was.
Because of this, the job_id of a row whose title has a marker in the middle changes. The job_id the 0.9.170 rule would have given is added to job_id_aliases of that row (no duplicates), so a seen-set built from older runs still matches. The onlyNewJobs store also treats the old id as seen.
2026-10-31 (0.9.170)
job-feed now reads the price of each source from the price list that is in effect today. Before, it took the last entry of the source's price history, which can be a price change scheduled for a later date (a batch is planned for 2026-11-05). The entry with the latest start date that is not in the future is used; an entry without a start date counts as always valid, and if all entries start in the future the earliest one is used.
The price list of a source is fetched once per run. If several parts of the run asked for the same source at the same moment, each one loaded it; now the others wait for the first load and reuse it. The full source record is large, so this also cuts data transfer.
The smallest share job-feed starts a source with is now its start fee plus one result at the price of the source's per-row result event. The row event is not marked in the price list, so the most expensive non-start result event is used (add-on events can be cheaper). The 0.0001 minimum and the skip for budget from 0.9.169 are unchanged.
2026-10-28 (0.9.169)
A source whose share of the max charge cannot buy even one result after its start fee is no longer started and no longer fails. Before, such a start was rejected by Apify with "Maximum charged results must be greater than zero" (seen on jobscout24-ch with 0.0026 and arbeit-swiss with 0.0001) and the run said "failed: jobscout24-ch, arbeit-swiss". The smallest share job-feed starts a source with is now its start fee plus one result at the source's cheapest result price, and never less than 0.0001 (smaller amounts were sent as 0.0000). A source below that is skipped exactly like any other source that does not fit in the max charge: status SKIPPED_BUDGET, counted in subruns_skipped_for_budget, reported as "skipped: max charge reached", and its share goes back to the pool. If Apify still rejects a start with that error, it is treated the same way. Shares are never rounded up, and runs without a max charge are unchanged.
2026-10-27 (0.9.168)
The cost report now takes the final cost of each source from one place. When Apify's usage for a source kept growing after job-feed had settled its charge, the source's settled amount in the cost report stayed at the earlier, lower value while its uncapped amount showed the final one, so the sources did not add up to the sub-actor total. The last check at the end of the run now settles every source again with the final values (the larger of usage and booked events times price, never more than the source's allowance). Each source's settled amount, the sub-actor total and the total cost now agree. A WARN line "sub-run usage grew after settle" is logged when the final amount is higher. Charging during the run is unchanged.
After an Apify restart, the sources job-feed had already started before the restart are no longer forgotten. Every source start is now saved in the run's key-value store (SUBRUN_LEDGER). After a restart, job-feed loads it before it starts anything: those sources count in the max charge again and appear in the cost report. A source whose earlier run is still running or has succeeded is picked up again (job-feed waits for it and reads its results) instead of being started a second time, and rows already delivered are not charged again. Only a source whose earlier run failed, was aborted or timed out, or that was never started, is started again. Log lines: "restart: sub-run ledger loaded" and "restart: re-attached sub-run".
After a restart, a job that is the same posting as a row delivered before the restart (other board, same title, company and city) is now merged and not delivered and charged a second time. The pre-restart row cannot be changed, so the merge is only counted.
After a restart, rows_delivered in the run summary now includes the rows delivered before the restart, so it matches the number of rows in the dataset.
The per-source entries in the OUTPUT record (subruns) now carry each run's own rows and charge. Before, every entry of a board repeated the board's totals. They also have uncapped_usd and settled_usd, the same values as in the cost report (null when the run is not known to the charge model).
The run summary has memory_run_mb and memory_run_peak_mb, read from the run's own record. memory_budget_mb still describes the pool for sources (account limits, or 4096 when Apify does not let the run read them).
A source whose charge ended at its allowance without a stop message from the source is no longer counted in subruns_stopped_at_budget. It is counted in the new subruns_settled_at_allocation. Sources that say they stopped at the max charge are still counted as stopped.
After a restart, when a board has an earlier source run in the ledger but none that matches the new input, job-feed logs "restart: no re-attach match" with the board, the ledger hashes and the new hash, once per start. The restart seeding line also shows how many rows went into the merge index (merge_index_rows).
2026-10-26 (0.9.167)
The max total charge now also holds for the run's total cost, meaning job-feed's own charges plus the charges of its sub-runs. A finished source counts at its full allowance until Apify has settled its charge. Before, an early, too-low read of a finished source let job-feed deliver past the cap (0.9.164 to 0.9.166).
A source's cost in the cost report is now the larger of what Apify billed as usage and the booked events times their price, never more than the source's allowance. The cost report also lists every source with its uncapped amount, its allowance and whether it went over, plus a count of sources over their allowance.
Near the max charge, a source that would be skipped for budget now waits briefly (at most 60 seconds, never past the run's own time limit) for finished sources to settle their charge, and starts if that frees enough room. Capped runs keep reading the same boards as before the settle change.
Until a finished source's charge is settled, the share kept back for job-feed's own per-job fee is based on that source's allowance, not on the early charge read, so fewer jobs are held back at the end. It switches to the settled charge once known.
The cost report is now built after the background settling has finished (or after at most 5 seconds), so no settle message appears after it.
After an Apify restart (migration), job-feed now reads the charges made before the restart before it starts any source. Before, a source could start while that count was still zero, so it was given too large an allowance and job-feed had no room left under the max charge for its own rows. That could leave the restarted run with fewer rows delivered than the cap allowed, in the worst case none.
2026-10-26 (0.9.166)
career-site-jobs: when created_at is null, posted_date is now read from the new posted_at_text field (ISO, "25 Sept 2026", "Posted 3 Days Ago", "Posted Today" and similar are converted to a date).
career-site-jobs: if the date text still cannot be read, posted_date is now empty. Before, the raw text was passed through into posted_date. Other platforms are unchanged.
New dedup_key on every job row: the normalized title|company|city key job-feed uses to merge duplicates in a run, for matching jobs across runs. It can change when the normalizers change.
A source that finished on its own is no longer reported as stopped at max charge when Apify's charge counts were not settled yet at the time of the read. A source that says it stopped at its max charge is still reported.
Near the max charge, job-feed now works out locally how many jobs still fit, charges exactly those and delivers exactly those. A charge is never followed by a withheld row (0.9.165 could bill jobs it then did not deliver when Apify's charge counts were slow to update).
Apify's charge counts are now only logged and counted in charge_readback_unconfirmed when they differ from the charge sent. They no longer decide which jobs are delivered.
unique_jobs and orchestrator_billed_events in the run summary now count delivered jobs. Jobs left out because the max charge was reached are counted in charge_refusals.
2026-10-25 (0.9.165)
Fixed a double charge: the same posting fetched by several country queries of one board (seen on Glassdoor) came out twice, the second copy with a suffixed job_id, and both were charged. Rows from the same platform with the same apply URL (query ignored), title and city are now one job. The same role posted in different cities stays separate.
Fixed the max charge cut-off: rows past the max charge were delivered for free. Near the limit the event is now charged first and only accepted rows are delivered, so orchestrator_billed_events matches what Apify accepted. This is a bug fix, like the fleet max-charge fix.
charge_refusals now also counts a partly accepted charge. New charge_readback_unconfirmed in run_summary counts charges near the limit whose accepted count could not be confirmed; those rows are delivered.
onlyNewJobs now remembers only the rows that were actually delivered.
OUTPUT subruns: every sub-run of a board that runs once per country now has its own run_id and final status (before, only one entry per board was updated).
The status line no longer counts boards that stopped at the max charge under "career sites only partly read"; those are already reported as stopped at max charge.
A sub-run is no longer reported as stopped at its charge budget just because its start event costs about as much as its budget; start events are ignored in that check.
career-site-jobs: textual posted dates are converted to ISO dates (commit caf2d0768).
2026-10-24 (2)
Internal API requests now authenticate with a request header, and error messages no longer include request URLs.
Costs in cost_report now use the amount Apify actually billed for each sub-run.
A sub-run that reached its share of the max charge is reported as partly read (reason charge_limit).
2026-10-24
Sub-runs now respect the run's maximum charge. Before this release, the sub-actors that job-feed starts for you (website and career-site readers) were not limited by the max charge you set on the job-feed run, so the total could exceed it. Each sub-run now gets the remaining budget; sources that no longer fit are skipped and listed in the run summary.
Companies mode (hiring signals): new fields hiring_evidence and hiring_roles_detail; a homepage with no careers page now counts as no_open_roles; directory and social profile URLs are not read (directory_profile_url); needs_browser is only reported for JavaScript-rendered sites.