Migrate Worker to Cloudflare-native AI; fix truncated replies; auto-close threads on send

[Copilot]

The AI pipeline was truncating long multi-part answers (1 024-token ceiling), using Gemini for both agents (external API dependency + cost), and leaving threads open after autopilot auto-sent a reply.

Truncation fix

• Research Agent: maxOutputTokens 1 024 → 2 048
• Writing Agent: maxOutputTokens 1 024 → MAX_WRITING_TOKENS (4 096)

Writing Agent → @cf/zai-org/glm-4.7-flash

Replaced geminiGenerateContent call with cfAiGenerateText() via the Workers AI AI binding. Added [ai] binding = "AI" to wrangler.toml. No external API key needed.

Research Agent → Workers AI + HTMLRewriter

Gemini's proprietary urlContext tool replaced with manual URL fetching fed into the same CF model. Up to 5 deduplicated datasource URLs are fetched in parallel (5 s timeout, 3 000 chars/URL cap).

Script/style blocks are stripped with HTMLRewriter (Cloudflare's built-in streaming HTML parser) before a catch-all tag regex runs — avoids the js/incomplete-multi-character-sanitization CodeQL alerts that regex-only stripping triggers:

const withoutCode = new HTMLRewriter()
  .on('script', { element(el) { el.remove(); } })
  .on('style',  { element(el) { el.remove(); } })
  .transform(response);
const html = await withoutCode.text();

Confidence is URL_FETCH_SUCCESS_CONFIDENCE (0.7) when at least one URL returned content — same value as the previous Gemini URL-context fallback. Citations become the list of successfully fetched URLs.

Removed entirely

geminiGenerateContent, GEMINI_API_BASE, computeConfidence, extractCitations, GEMINI_API_KEY from Env, MAX_GEMINI_URLS, deduplicateForGemini. No external AI service or API keys required by the Worker.

Thread auto-close on send

After autopilot sends a reply, the thread is PATCHed to status: 'closed'. This is fire-and-forget (.catch() logs the error) so a Supabase hiccup at this step cannot retroactively revert a reply that was already delivered to the customer. Threads are already re-opened when a customer replies again, so the lifecycle is complete.

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
answerify	Ready	Preview, Comment	Mar 1, 2026 0:34am