A deploy to Giant Context finishes and the platform files a changelog entry for it. Nobody writes it, nobody approves it, nobody edits it. It is public at giantcontext.com/en/changelog.

Brain, the platform's own agent with access to the codebase, is never told what to write about. It asks the changelog where the last entry stopped and drafts everything between that commit and the one just deployed, so two entries cannot leave a commit between them.

A commit subject is written for whoever runs git log. Brain resolves each commit to the pull request that carried it and reads that request's body, which is where the change was already explained. The entries are drafted from the explanation rather than from the label, which is why they read as descriptions of features.

Thirty commits or fewer draft in a single pass. A longer range is cut into chunks of twenty that draft in parallel and are joined by a second pass, so a deploy carrying two hundred commits does not have to fit into one request.

The version is not semantic. It is the number of the build that deployed, so 1.108.0 names a specific CI run rather than a release somebody chose to call a minor. The entry is filed against that number, and a redraft of the same deploy updates it in place.