New government tenders and contract notices from the official EU TED and UK Find a Tender and Contracts Finder APIs, filtered by keywords, CPV codes, countries and value, with only-new monitoring and Slack-ready summaries.
Store title changed to lead with the plain job, not monitoring (2 Nov 2026; see docs/research/2026-11-02-store-crowding.md). The Actor name in its URL and all behaviour are unchanged.
Test fixture for the change-notice lookup replaced with a real capture (2 Nov 2026); the live drift check confirmed the lookup works as coded: no code change.
FAQ explains the catch-up after a big first window: held-back notices arrive up to 1,000 a run, still as new (independent review, 31 Oct 2026, finding 8).
0.1.3 (2026-10-30)
Cost hardening from the cost-abuse audit (docs/research/2026-10-30-cost-abuse-feeds.md) and the template's new guards. The audit's five failing tenders scenarios are regression tests in tests/test_cost_abuse.py; same events and prices.
Cost guard (template CostGuard, MeteredBiller, RunContext): a run stops starting new work once its compute and row writes are $0.02 ahead of 80% of what it has charged, checked before every request and retry wait and every 10 notices in the parse and filter loops. Sources not checked share one free skipped row. A run that earns nothing now costs at most about $0.024 at 1 GB, not $0.20 an hour until the timeout. RUN_SUMMARY.cost shows the estimate.
Spending limit: a source starts only when the limit covers one check and one notice for it and for the sources already running. A $0.002 limit used to read every source in full and charge nothing; now it reads nothing.
Inputs: at most 10 source lines per run (one free row for the rest, bad lines in one push); at most 100 keywords, exclude keywords, CPV codes and countries (a longer list is a free input row, never cut silently); each keyword list is one compiled pattern (1,001 exclusions cost 0.14 CPU-s per notice); 3 failed sources in a row end the run; error rows quote at most 5 problems.
Bounded scans: TED is always asked for 250 notices per page (a cap of 20 used to mean pages of 20, up to 750 requests); a normal run stops a source after 10 pages in a row without a matching notice; monitors read Find a Tender and Contracts Finder from their last complete, fully delivered read minus one day instead of the whole window; every page is capped at 20 MB.
Open notices only in monitors: monitoring runs neither send nor remember notices whose deadline has passed, and a remembered tender is forgotten when its deadline passes (notices without a deadline, such as awards, 60 days after they were last seen, as before), so the memory no longer grows with closed tenders. Normal runs still list closed notices.
CPU: contact details are scrubbed out of descriptions only for the notices returned, not for every notice read (94% of a no-match scan's CPU).
Network: TED's POST now goes through the template's fetch like every GET: 10 s to connect, 60 s per request, the private-address checks and the size cap; all requests share the run's per-host limiter (2 in flight per host, a host left alone after 10 failures in a row).
Rate limits (field test): when Find a Tender or Contracts Finder answer 429 (or 503) with a Retry-After longer than 30 s, as they did with Retry-After: 120 when a run was started again a minute later, the source gets a free skipped row ("the service asked to wait 120 s; it will be read next run") instead of an error row, is not counted as a problem in the status message, and its memory is untouched.
Failed, not-found and throttled sources are never charged (tested). PRICING.md has a new Cost protection section with the worst case per run; the README FAQ explains early stops, rate limits and the input limits.
24-hour cap per Apify account: the cost guard's $0.02 per-run allowance now comes out of $0.20 of unpaid compute per account per 24 hours, across runs (template AllowanceLedger, one record in the fixed store public-tenders-monitor-limits), so runs started back to back cannot add up to a large unpaid bill. A run stopped by it says so in plain words in its free skipped row and status message ("this Apify account has used today's allowance...; it resets within 24 hours"). If the record cannot be read, the run goes ahead.
0.1.2 (2026-10-30)
Terms audit: Open Government Licence v3.0 attribution and TED source acknowledgement added to the README and dataset views; a Retry-After longer than 30 seconds now ends that source for the run with a free row instead of an early retry.
0.1.1 (2026-10-30)
Fixes from the pre-publication code review (each has a regression test in tests/test_review_regressions.py):
Monitoring memory records when each notice was last seen, and forgets a notice 60 days after it was last seen instead of 60 days after it was first seen. A long-running tender updated after two months is no longer sent and charged again as new. Memories saved by 0.1 fall back to the first-seen date.
TED change notices for contract awards and planned tenders that name the changed notice by publication number (seen live) are matched to that notice: the number is resolved to its notice identifier from the pages already read, or with one batched TED search. One award and its correction are one notice, not two new rows. If the search fails, the source still succeeds and those change notices are tracked as before.
Privacy: sentences in descriptions are dropped when they hold 9 or more digits once spaces, dots, dashes and brackets are removed (CPV codes, notice numbers, dates and amounts with a currency are set aside), a phone label followed by a number in more languages (Tel.-Nr., Telefonnummer, Telefoon...), contact labels (Kontakt:, Persona de contacto, Referente:, RUP...), a capitalised "Contact Jane Smith" or a title before a name (Herr, dott....). Sentences are no longer split after abbreviations such as Tel. or dott.. Normal descriptions with quantities, codes, lots and prices are kept whole (tested both ways).
One out-of-range date (0001-01-01T00:00:00+01:00) no longer fails a whole source, and any record the parser cannot read is skipped and counted (recordsSkipped in the run summary, and in the status message) instead of failing the read.
New rows go out oldest first, so when maxNoticesPerSource holds notices back, the ones closest to leaving the window go first; the status message says how many were held back. README and input schema now say what happens when more notices arrive per run than the cap.
When a tender has several versions in one window, the newest (by publication date, then notice number, not the API's order) is chosen before filtering, so an older version with a superseded deadline or value is never shown when the newest fails a filter.
Default memory raised to 512 MB for broad monitors with large memories. README and input schema recommend a window of 3 days or more for scheduled monitors.
0.1 (2026-10-29)
First release: tenders, planned tenders and contract awards from EU TED (Search API v3), UK Find a Tender and UK Contracts Finder (OCDS), in one row shape with title, buyer, country, deadline, value, CPV codes, procedure and the official link.
Filters: keywords (pushed into TED's full-text search), exclusions, CPV prefixes, buyer countries, minimum value in euros across currencies, notice type, publication window.
Monitoring of new notices and of tenders whose deadline or value moved (TED change notices are matched to their procedure, UK updates to their OCDS process); nothing is ever reported removed, because each run reads a rolling window. Optional quiet first run.
Find a Tender is read without its stages filter, which returns only notices under the old regulations, so Procurement Act 2023 notices are included.
No personal data: contact points are never read, and sentences with e-mail addresses, phone numbers or named contacts are removed from descriptions.
Safe billing from the Studio template: a source check is charged only after its rows are delivered; problem rows are free; an invalid filter stops the run before anything is fetched. Every row has a one-line summary, and the status message lists the first new tenders for Slack alerts.
