Writing
Short technical notes from real debugging sessions. No tutorial padding.
A 404 that was really a provisioning gap
Three paths returned 404 on a Firebase Hosting domain while DNS resolved correctly to the anycast address. The interesting part: the wildcard resolves whether or not a site exists, so a resolving hostname proves nothing. What actually settled it was listing hosting sites per project and finding the site name already claimed, with only a placeholder deployed. Diagnostic order that worked: DNS type A → HTTP status on /, /robots.txt, /sitemap.xml → hosting:sites:list → compare against the deploy target in .firebaserc.
Drafting a conformance ruleset for MCP servers
The MCP ecosystem has directories but no neutral verifier. A workable first ruleset splits into two halves. Conformance is mechanical: handshake and protocol version, capability claims that match actual methods, valid JSON Schema on every tool, declared-required fields the handler actually enforces, errors that come back as JSON-RPC errors rather than success payloads, and stdout that stays clean on stdio transport. Security is about the text an agent reads before it acts: instruction-override patterns inside tool descriptions, zero-width and bidi characters, unbounded arguments where a constrained type belongs, and list tools with no limit parameter.
The rule that matters most is the boring one: every rule needs a fixture from a real server, or the false-positive rate makes the whole report worthless.
Consent Mode is a measurement bug, not a legal checkbox
When conversions fall after a consent banner ships, the usual suspects are ordered wrong. Check the defaults sent before interaction, then whether the consent state is actually propagated to every tag, then whether ads data redaction is on, then whether the banner blocks the tag entirely instead of degrading it. In most audits the loss is not legal compliance — it is a tag firing before consent resolution, which looks identical to a reporting bug and gets "fixed" by changing attribution models.