Invalid input no longer FAILS the run (2026-11-04, local, not built or pushed; actor.json version stays 0.2). Input that parse_input rejects (e.g. an unknown types value, mode: "weekly", a bad ISSN, maxItems: "abc") used to call Actor.fail (run FAILED, exit 1, status message = the reason). The run now ends SUCCEEDED (exit 0) with one free status row {"recordType": "status", "type": "status", "status": "error", "statusRow": true, "message": "Invalid input, nothing was fetched or charged: <reason>", "error": "<reason>", "charged": false, "scrapedAt": ...} and the same message as the terminal status message. No request is made (not even robots.txt), no work is charged and no RUN_SUMMARY is saved. The prefix is "nothing was fetched or charged" (not "...and no results were charged") because this Actor's pricing has no start event; the reason texts are unchanged.
- Wrong-type crashes now give the same row with the reason
a field has the wrong type (<python error>). (e.g. lookbackDays or maxItems = Infinity used to crash with an uncaught OverflowError, exit 91).
- A whole input that is not a JSON object (a string, a number, a list of objects) used to crash with
AttributeError (exit 91); it now gives the same row with the reason the input must be a JSON object, got <type>. An empty list or null input still means {}.
- Real upstream and runtime failures (API down, 403, robots.txt) are unchanged: they keep the existing
{"type": "status", "status": "warning"} fail-soft row.
requirements.txt: apify >= 4.0.0, < 5.0.0 pinned to apify >= 4.1, < 4.2.- Memory unchanged (default 256, min 256, max 2048, defaultRunOptions 256). Pricing, default input, prefill and
exampleRunInput unchanged.
- Files:
my_actor/main.py (new end_invalid_input, wrong-type wrap around parse_input, non-object check), README.md ("Invalid input" bullet under How it works), requirements.txt, CHANGELOG.md, tests/test_invalid_input.py (new).
Default maxItems lowered from 10 to 9 (2026-11-02) in the input schema default, the prefill, the code fallback (DEFAULT_MAX_ITEMS, used for omitted, null or blank maxItems) and exampleRunInput. {}, the prefill and the example now cost at most $0.009, under the $0.01 Store test limit (10 works were exactly $0.01, no margin). The first page asks for 29 rows instead of 30. QUICK_MAX_ITEMS still equals the default, so the small-run time budget now applies to maxItems 9 or less: a run with maxItems 10 uses the normal request timeouts and retries. actor.json version stays 0.2.
- Files:
my_actor/main.py (constant and comment), .actor/input_schema.json (maxItems default, prefill, description "Default 9 (about $0.009)"), .actor/actor.json (exampleRunInput.maxItems 9), pricing.json (_note and emptyInputAndPrefill, which still said 100 / $0.10), README.md lines 70, 108, 125 and 126 only. README.proposed.md is not touched (it has a pending edit).
- Tests:
tests/test_input.py (test_empty_defaults; test_blank_max_items_is_default_10 renamed ..._9), tests/test_main.py::test_empty_input_one_request (9 works, 29 rows), tests/test_polite_pool.py (3 tests, 9 pushed), tests/test_pricing.py (test_readme_price, test_small_default_run_about_one_cent), tests/test_gap12_fail_soft.py (test_success_has_no_status_item, test_store_that_cannot_be_opened_is_a_warning, test_default_is_ten_and_quick renamed test_default_is_nine_and_quick).
Optional Crossref polite pool. If the environment variable CROSSREF_MAILTO holds a valid email address, requests to api.crossref.org add it as the mailto query parameter, and the User-Agent ends with ; mailto:<address>), the form in Crossref's etiquette example. If it is unset or invalid, requests are unchanged (public pool).
The address is never logged, pushed to the dataset or saved to RUN_SUMMARY or the key-value store. The log says only "Crossref polite pool: on/off", once per run. RUN_SUMMARY has a politePool true/false flag. Error text and httpx request logs are redacted.
The rate stays at 1 request per second with 1 connection, inside Crossref's limits for list requests (since 21 July 2026: 1/s in the public pool, 3/s in the polite pool).
No address is hardcoded anywhere. Tests use a fake example.org address in mocked requests only.
Docs: the query description no longer says matches are "ranked by Crossref". Results are always sorted newest first by date, and Crossref keyword matching is loose, so the description suggests filters or narrower phrases for precision. No relevance-sort option was added.
Docs: the input schema and README now state the dedupe limit: at most 200,000 DOIs per state key, with the earliest-saved DOIs dropped first (SEEN_CAP in my_actor/main.py). No code change; actor.json version stays 0.2.
Docs: README rewritten for the Store (use cases, "Getting a complete daily feed", Limits). The daily volume now uses verified counts (about 27,000 to 35,000 new journal articles and 42,000 to 87,000 works of all types per weekday, 26 to 30 October 2026) instead of the wrong "46,000 per day"; the maxItems description says the same. Rate-limit text now matches Crossref's limits since 21 July 2026 (list requests 1/s public, 3/s polite). Store title, SEO title, SEO description and description updated. No code change.
