Site health & failed jobs

The platform admin header runs seven health checks on every page load. A green dot means the platform is fully configured; amber or red means something needs attention. The /admin/jobs/failed page surfaces the failed-job side of the same picture.

The seven checks

Check	Severity	What it verifies
Failed jobs	Critical	Count of rows in failed_jobs. Anything > 0 turns the pill amber; > 50 turns it red.
Stripe	Warning	cashier.secret is set. Without it, no plan syncs and no checkout works.
LLM provider	Warning	At least one of: Cloudflare account + token, OpenAI key, OpenRouter key.
Vector store	Warning	Cloudflare Vectorize index name OR Qdrant URL.
Mail	Warning	Mail driver + sender configured. Needed for invitations, password reset, lead notifications.
Reverb	Warning	Reverb app key + secret set. Without it, the inbox doesn't update live.
Cache	Warning	Redis cache reachable. The hot path caches retrieval results here; a slow cache means slower responses.

Score & label

The header pill aggregates: score = % of checks passing. label goes from Strong (≥ 90%) to Stable (≥ 70%) to Watchlist (≥ 50%) to Critical (anything below 50%). The pill color follows: green for Strong, amber for Stable / Watchlist, red for Critical.

Each notification in the dropdown has a deep link to the relevant config page so you can fix it in two clicks.

Failed jobs

Open /admin/jobs/failed for the full list. Each row shows:

• Job name — the queue + class.
• Connection.
• Failed at — when the last attempt blew up.
• Exception — first line of the trace, expandable.
• Retry — re-queue the job with the same payload.
• Forget — drop the row.

The page also has Retry all and Flush all buttons. Use Retry all after fixing a transient outage (the LLM provider went down; jobs failed; provider is back). Use Flush only when you've decided the failures are unrecoverable.

Common failure shapes

By queue:

• crawl — usually a 4xx/5xx from the upstream site or a Browserless rate-limit. Retry once; if it persists, the source URL is dead.
• index — usually an LLM embedding rate-limit or a Vectorize quota issue. Retry after the rate window resets.
• default — anything else (usage events, gap detection, webhook delivery). Look at the exception.

Webhook delivery failures

The lead-captured dispatcher (SignedDispatcher) is single-attempt — the lead is already persisted, so a failed delivery surfaces in the workflow run log rather than blocking the visitor. Workflow-step webhooks (DispatchWebhookJob) retry up to 3 times via Laravel's queue retry mechanism; after the third failure the job lands in failed_jobs with the destination URL + signature in the payload, so you can replay manually after the receiver is fixed.

Horizon

If you've enabled Laravel Horizon, /horizon shows real-time queue throughput, wait times, failed-job rates, and per-job-class histograms. Recommended to keep open in a tab during deploys.

Site Health pill — what each color means

Color	Score	Meaning	Action
Green	≥ 90% (Strong)	Everything healthy.	Nothing.
Amber	50% – 89% (Stable / Watchlist)	One or more warnings; product still functional.	Click the pill, fix what's broken when you have time.
Red	< 50% (Critical)	Multiple checks failing or a critical-severity check (e.g. failed-jobs surge).	Drop everything. The product may be visibly broken for some customers.