Observability

/api/health

The first thing to check, in every environment:

{
  "ok": true,
  "database": "d1",
  "ai": {
    "configured": true,
    "model": "gemini-3.5-flash-lite",
    "liveModel": "gemini-3.8-live",
    "thinkingLevel": "default",
    "gateway": "claimpilot-chat"
  }
}

Field	Tells us
database	Which store the Worker is using: d1, neon or none
ai.configured	Whether a model key is present. false means chat will answer with an error event
ai.model, ai.liveModel	The models actually in use, after wrangler.jsonc and secrets
ai.thinkingLevel	The deployment default, or default for the model's own
ai.gateway	The AI Gateway id, or null when calls go straight to Google

The Claims Data page (/data) and the unlisted /status page show the same, with what is loaded.

/api/health is not gated by EMBED_KEY, but a cross-origin caller not in ALLOWED_ORIGINS gets a 403.

AI Gateway

With the gateway on, every model call (chat, generation and voice) is logged in Cloudflare AI Gateway: the request, the response, tokens, cost, latency and errors. That is the place to answer "what did the model actually see" and "what did that cost".

The gateway log holds claim content

Prompts carry the claim brief and tool results carry claim detail, so the gateway log is a second copy of claim content. Fine for synthetic claims; a data-protection question for real ones. See Security.

The Worker log

observability.enabled is on in wrangler.jsonc, so Workers Logs collects the Worker's console output. What it writes:

When	What
A refused origin	Refused origin <origin> on <path> (warning)
An unhandled route error	The error, then a JSON 500
A chat turn fails mid-stream	chat turn failed with the error; the browser gets an error event
DATABASE_PROVIDER=none and a vote	One line: {"feedback": {…}}, or {"feedback": {"turnId", "rating": null}} for a withdrawal

Locally, the same lines appear in the wrangler dev terminal.

Usage per turn

Every done event carries usage (input, output and cache-read tokens, summed over the turn's rounds) and the model. The chat footer shows the last turn's tokens; the Prompt Lab trace shows them per run. Prompt caching shows up as cacheReadTokens: the stable prefix (system prompts, the claim brief and the tool list, shared tools first) is what makes it work.

Answer quality

The feedback report in the Prompt Lab is the quality signal: the helpful rate by capability and by prompt set, and the reasons behind each thumbs down. GET /api/feedback returns the same as JSON.

Front-end versions

Each build writes version.json. Open tabs poll it every minute and offer a refresh when it changes, so a long-lived tab does not keep running an old build against a new API.