Gary branding cleanup: UI copy, AI prompts, DOCX author, and metadata
zgbrenner swept Mike branding out of every user- and model-facing surface in one commit: the app wordmark, global-error title, AI self-description in prompts, tracked-change author in exported Word files, and the page metadata. No new functionality; all cleanup.
The diff covers 481 additions and 68 deletions across 29 files. The backend chat system prompt and the Review Table assistant prompt now name Gary explicitly, so the model identifies itself consistently with what users see. The DOCX tracked-change author and change-ID prefix switch to Gary as well - relevant if any document-processing logic downstream parses those fields.
Metadata got a proper rewrite. Page title, description, and Open Graph/Twitter cards all land on "Gary - Open-source legal AI for U.S. lawyers." More practically, metadataBase moves off a hard-coded production hostname (mikeoss.com) and onto NEXT_PUBLIC_SITE_URL, an env variable. That makes preview and staging deployments work correctly without manual edits.
zgbrenner was careful about claims. Neither the AI-identity prompts nor the metadata assert accuracy guarantees or bar-compliance. The README was rewritten at roughly 154 lines net, with a plain-English product overview, a step-by-step flow, a comparison table against MikeOSS, and a limitations section. Attribution and the AGPL-3.0 license are preserved.
A new branding.test.ts (210 lines) guards the rename going forward: it asserts that metadata uses Gary copy, checks the README preserves the upstream attribution block, and runs a terminology scan that would fail if old upstream-brand strings crept back in outside intentional exceptions. Two database columns and all internal package names were deliberately left as-is to avoid migration risk.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?