marathon debugging session — AI mode, streaming, architecture pivot to headless browser

this one was long. started with wiring up the AI mode for sitesnapshot.org — the tier 3 public app built on top of the site-snapshot Claude skill we made a few days ago. the skill works great inside claude.ai chat: give it source files or screenshots, it rebuilds the page as a frozen HTML file. the plan was to expose that same capability via the API.

hit every wall. firebase-admin corrupted the service account JSON in netlify's env var UI. firebase-admin stomped node's global fetch. node-fetch broke SSL cert verification. netlify killed the connection after 15s of silence (fixed with claude's streaming API). then the real problem: claude sonnet maxes out at 64K output tokens per call. stripe.com needs 60-80K tokens of HTML to fully represent. even with streaming keeping the connection alive, claude just runs out of ink before finishing the page. tried compressing the prompt, smart approximations, aggressive HTML pre-cleaning — none of it changes the fundamental limit.

consulted another model on the architecture. the answer was obvious in hindsight: stop asking the LLM to rewrite pages from scratch. use a real headless browser (browserless.io) to visit the URL, render everything including JavaScript, inline all stylesheets, and serialize the complete DOM as a single file. it's a copy, not a recreation. no token limits, deterministic, $0.001 per capture vs $0.20 with claude. built the browserless integration — the capture ran but CSS inlining failed (content without styling). that's a code bug, not an architectural dead end.

shipped the greyed-out version: free mode works in production, AI and screenshot modes show "coming soon," mobile responsive added to all components. next session: debug the CSS inlining in the browserless capture code, potentially move to an async job pattern (firebase cloud functions) so there's no timeout pressure at all. the LLM still has a role — screenshot-to-HTML rebuilds and optional cleanup passes — but it's not the engine for URL captures anymore.