apiwatch

All changes

How apiwatch works

How a record gets into the feed, how it stays correct, and what the free GitHub Action does with it.

Where a change comes from

A daily job reads vendor changelogs. A new entry there is a lead: a prompt to research, never a record. Model deprecation pages are also parsed into drafts on demand; a draft is published only after the review below. A record is written from the vendor's own primary source: its docs, changelog, release notes, status page or official repository. Blogs, news, aggregators and model memory don't count.

How a record is reviewed

A record is published only after it passes two checks. The reviewer is an AI agent that did not write the record; no human step is required.

  1. An independent review. A reviewer that did not write the record opens the vendor's source and checks every field: the exact date the change takes effect (the shutdown date, not the announcement date), the severity, the summary and fix, and the identifiers the scan looks for. Anything the source doesn't say comes out. The verdict is pass, fix with exact edits, or reject.
  2. An automated source check. Each record carries two to five short strings copied word for word from the source, such as model ids and dates. A program fetches the source and confirms they are there.

Severities mean:

A dated retirement within 180 days counts as breaking.

The daily source check

Every day, before the site is built, the source check runs again on every record. If a vendor edits or removes the page and a record's quoted strings are gone, that record is left off the site while its source no longer supports it, and the run is flagged so the record gets reviewed again. A page that fails to load is reported but doesn't pull the record, so a flaky vendor site can't take it down for a day.

What the free Action does

Add the two steps shown at the end of this page to a GitHub Actions job. Each run scans your code for the identifiers in the records bundled with the Action version you use: model ids, endpoint paths, tool and field names. Matching is token aware, so gpt-4 does not match gpt-4o. Records that only apply to some code (a dependency version, some files) are only reported there. The run then:

Optional inputs keep one report comment up to date on a pull request once there is a hit (comment, needs pull-requests: write) and gate a pull request only on the files it changes (only-changed). The scan itself makes no network calls: your code never leaves your runner.

What it doesn't do

apiwatch only knows the changes in its feed. No hit is not the same as safe. Records summarize vendor changes in our own words and link the source: confirm details that matter before acting.