# Changelog of Multi-ATS Job Scraper & API (`dalleyne/ats-job-scraper`) Actor - **URL**: https://apify.com/dalleyne/ats-job-scraper/changelog.md - **Full Actor documentation**: https://apify.com/dalleyne/ats-job-scraper.md ## Changelog ### Unreleased #### Note for consumers of the `salary` field The fixes below change what `salary` contains for Greenhouse records. This is not a schema change — `salary` was `object | null` and still is — but two value domains widen, so downstream code may need a look: - **`min`/`max` can now be fractional.** An hourly range like `$28.10 — $34.50` is reported as `28.1`/`34.5`. Every amount used to be an integer, so anything storing these as integers should be checked. - **`currency` gained `SGD`, `MXN`, `NZD` and `HKD`.** The previously reachable set was `USD`, `CAD`, `AUD`, `GBP` and `EUR`. Treat it as an open set — the parser now reports whatever ISO code a posting states, from a fixed allowlist that may grow. Greenhouse `salary` goes from populated on 13% of jobs to 63%. No record in that sample loses a salary it previously had — a measurement across 1893 jobs on 19 boards, not an invariant of the code. The failure was concentrated rather than uniform: boards using Greenhouse's structured `pay-range` widget extracted nothing at all, while boards writing the range inline in a sentence were largely unaffected — which is why the actor appeared to be extracting salary. - Fix Greenhouse salary extraction on boards that use the structured `pay-range` block. `src/ats/greenhouse.js` passed the API's `content` field straight to `parseSalaryFromText`, but Greenhouse serves that field as an HTML document that has itself been escaped, so a pay range arrives as `<span>$115,000</span><span>&mdash;</span><span>$194,000 USD</span>` — markup between the two amounts, and the separator an `&mdash;` entity. The regex requires the amounts be separated by whitespace and a hyphen or en dash, so it matched nothing. This is why the bug was invisible: the escaping only defeats the regex when markup sits *between* the two amounts, which is exactly what the `pay-range` block does, so those boards yielded 0% while boards writing the range inline in prose kept working. Fixed with two changes that are only effective together: a new `htmlToText` helper that flattens the content (decode → strip tags → decode, since the content is doubly encoded), and an em dash added to the separator class. Measured across 1193 live jobs on 10 Greenhouse boards, salary extraction went from 17 to 727 records; applying either change alone still yielded 17. Verified end-to-end against the gitlab board (0 → 75 of 188), with no change to any record that already parsed. - Accept `to` as a range separator (`"$144,000 to $162,000"`). Some boards write every range this way — Discord yielded salary on 0 of 44 jobs before this. Unlike a dash, `to` requires whitespace on both sides, so it cannot be read out of the middle of a word. A live posting advertising `"managing $1M to $5M+ in annual paid media budgets"` returns `null`, because the magnitude suffix sits between the amount and the separator — but note that requiring two symbol-prefixed amounts is not on its own enough to exclude prose; see the magnitude guard below. Only lowercase `to` is accepted; no capitalised form occurs anywhere in the 1893-job sample. - Accept amounts written with cents (`"$221,000.00 - $300,000.00"`). Vercel and SoFi write annual figures this way and the trailing `.00` previously blocked the separator, losing the range entirely. The cents are captured rather than discarded, so an hourly `"$28.10 — $34.50"` reports 34.5 and not 34, and a fractional `"$1.5k"` scales to 1500 and not 1000. - Combined effect of the two items above, measured across 1893 live jobs on 19 Greenhouse boards: 1100 -> 1187 records with salary, with **zero** records lost or changed except one strict improvement — a Gusto posting listing three regional ranges previously matched only the trailing abbreviated `"$170-$190 CAD"` and reported a salary of 170; it now matches the leading `"$185,000 to $200,000"`. - Derive currency from a code stated immediately after the range, falling back to the previous +/-200 character context guess only when none is stated. `£` and `€` name their own currency and are untouched; only `$` needed the work. 74% of live postings state a code, so this replaces a guess with the poster's own answer in most cases. Two classes of error are fixed: a posting listing several regional ranges used to attach a neighbouring region's currency to the wrong range (a Gusto posting quoting USD for Denver and CAD for Toronto tagged the Denver range CAD), and currencies the context heuristic never knew about were silently reported as USD — Airbnb's `"Mexico Monthly Pay Range $100,000 — $125,000 MXN"` was emitted as 100000 USD, and Coinbase's Singapore ranges as USD rather than SGD. The accepted codes are an explicit allowlist rather than any three-letter token, because non-currency tokens really do follow ranges in live postings: `USA` (Affirm's next heading), `OTE` (on-target earnings) and `CAN` all occur, and matching `[A-Z]{3}` would read each as a currency. Measured across the same 1893 jobs: 9 records change currency (5 to SGD, 3 to MXN, 1 to USD), all of them corrections verified against the code the posting states, and no record changes amount. - Guards added after an independent review of the above (GLM-5.2 and Claude Fable 5, run against the same diff). All were verified against live board data before being acted on: - **Reject amounts a magnitude word shows are not a salary.** `"$1 to $5 million"` parsed as a salary of 1 to 5. Requiring two symbol-prefixed amounts excludes the *suffixed* form (`"$1M to $5M"`, where the `M` breaks the separator) but not the spelled-out one, and the code comment claiming otherwise was wrong. - **Rank candidates instead of taking the first match.** Greenhouse appends its `pay-range` block at the END of the content, so an earlier money range won outright: a live Instacart posting emitted `"an average deal size of $80k–$120k"` as the salary in place of its real `$153,000 — $161,500 USD`. A range followed by a currency code, or near pay wording, now wins; ties keep the earliest. - **Reject inverted ranges** rather than emitting `min > max`, falling through to a later valid range in the same document. - **Do not silently drop precision** — `"$34.505"` reported `34.5`. - **Decode numeric character references.** A board serialising its divider as `—` instead of `—` would have returned silently to zero yield, reintroducing the very bug this release fixes one encoding variant away. - **Keep literal angle brackets in prose.** Stripping any `<...>` deleted the text between them once entities were decoded, so `"< $100,000 - $200,000 >"` flattened to an empty string. HTML comments are now dropped outright, since superseded pay figures are exactly what hides in them. - Fix `parseSalaryFromText` reporting k-suffixed salary ranges 1000x too small (`"$120k - $150k"` yielded `min: 120, max: 150`). The `[kK]?` quantifiers sat outside the regex capture groups, so the scaling branch in `parseAmount` never fired. Regression introduced when the previous `num < 1000` heuristic was removed — that heuristic had been masking the regex flaw on the k path while wrongly inflating plain sub-1000 amounts. Both paths are now correct. - Add a unit test suite for `parseSalaryFromText` (`npm test`, via the built-in `node --test` runner — no new dependencies). - Known limitation, unchanged: the range regex backreferences the currency symbol, so `"$120,000 - 140,000"` (symbol on one side only) does not match and returns `null`. Left as-is pending a decision; see the characterisation test in `test/normalize.test.js`. ### 1.0.0 - Initial release. - Scrape Greenhouse, Ashby, and Lever job boards from a single `urls` input with automatic ATS detection per URL. - One normalized output record across all sources (tagged with an `ats` field). - Filter by `departments` / `teams` (ATS-specific values) and `daysBack` before storing; `maxJobs` cap per board. - Composes the standalone `greenhouse-job-scraper`, `ashby-job-scraper`, and `lever-job-scraper` logic.