The run now finishes before its own timeout: no source is started in the last 10% of the run time (20 s to 2 min), what was read is delivered and charged, and the run ends SUCCEEDED with time_hit in SUMMARY and "run time limit reached" on the run card instead of TIMED-OUT. Runs that finish in time are unchanged: no prices, input fields, defaults or output fields changed.
maxItems (sent by AI agents) now also lifts default caps: leave maxResults, searchLimit and maxSources at their defaults and the run delivers up to maxItems rows in total; explicit caps still win. Add-ons follow the delivered rows; no prices, defaults or output fields changed.
Runs started without maxTotalChargeUsd (no spending limit) no longer fail after the first source is charged; they now behave like a capped run with a large budget. No prices, output fields, defaults or input fields changed.
Unknown input fields are ignored instead of failing the run (agents often add maxItems or proxyConfiguration); maxItems caps the rows delivered and charged; input problems end with an explanation and no charge instead of a failed run. No prices, output fields, defaults or input fields changed.
Initial release: public profile/app metadata, bounded requests, input validation, filters, global limits, caller-managed change hashes, pre-fetch spending checks and explicit failure summaries.