fix: drop ua-based markdown switching

[pamelachia]

Summary

The middleware served markdown based on user-agent matching, but markdown responses only declare Vary: Accept, so the CDN cache key ignores the user agent and whichever variant is computed first gets cached for every client. I observed this live as deterministic variant poisoning: byte-identical ChatGPT-User requests received text/html on GET and text/markdown on HEAD for the same cached URL. This change removes the UA branch so content varies only by the Accept header, which makes the declared Vary truthful and the poisoning structurally impossible. Clients that want markdown still get it via Accept negotiation or explicit .md URLs; separately, the ChatGPT-User agent's reader currently hard-fails on markdown it receives via UA matching (details in GROWTH-973), so consistently serving it the fully server-rendered HTML also fixes readability in that assistant.

Changes

• Remove LIVE_FETCH_UA_RE/isLiveFetchAgent from src/start.ts; bare URLs serve markdown only when the Accept header prefers it, explicit .md URLs unchanged

Testing

Tested locally:

• Dev-server matrix: ChatGPT-User and Claude-User UA GETs return text/html; Accept: text/markdown returns text/markdown; GET and HEAD agree
• Claude's real fetcher keeps receiving markdown (it sends Accept: text/markdown, text/html, */*, which wins negotiation)

Verify in production post-deploy (dev harness cannot exercise these):

• GET/HEAD content-type consistency on a previously cached blog URL
• Explicit .md URLs still serve markdown

Linear

• fixes GROWTH-973

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
multigres-site	Ready	Preview, Comment	Jul 9, 2026 9:41am