AmbersightWarnings for Cloudflare Workers

ReferenceFrom docs/checks.md, copied

All 83 checks Ambersight runs on your Workers

The threshold, severity and fix of every check, from the product's own documentation. Link to any check by its ID, such as budget:free-daily-requests.

Ambersight is a free, open-source Worker you deploy in your own Cloudflare account. Every hour it reads your Workers' settings and metrics with a read-only API token, runs these 83 checks and gives a ranked list of what to fix.

Ambersight runs 83 checks. Each one below lists the ID you see in the dashboard and in config.json, how severe it is, exactly when it fires, and what to do about it. The numbers are the defaults. Most of them can be changed globally, per group or per Worker (see configuration.md).

Every finding in the dashboard carries the evidence behind it, a suggested fix and a link to the matching Cloudflare docs page. When the problem goes away, the finding resolves on its own at the next scan.

Severity decides where a finding goes. By default, critical findings are e-mailed right away, and everything else waits for the digest. See alerts.md.

Notes against Cloudflare's docs

The list below is the product's own text, copied from docs/checks.md. We checked its Cloudflare facts against Cloudflare's documentation on . Where they differ, Cloudflare's docs are right:

  • cron:cron-count counts against 250 Cron Triggers on Workers Paid. Cloudflare's limits page lists 1,000 per account on Workers Paid, and 5 on Workers Free.
  • drift:vars-near-limit counts against 64 variables and secrets per Worker. That is the Workers Free limit. Cloudflare lists 128 on Workers Paid.
  • cron:cron-cpu-near-limit uses 30 seconds on Workers Paid. That is Cloudflare's default. Cron Triggers on an interval of 1 hour or more get up to 15 minutes of CPU, and the limit can be raised to 5 minutes for shorter intervals.
  • The text says Cloudflare's budget alerts arrive a day late. Cloudflare's docs give no delay for budget alerts or for the email alerts of a spending limit.

Free plan budget

On the Workers Free plan, an account gets 100,000 requests per UTC day across all its Workers. Past that, every Worker answers with Error 1027 until midnight UTC. This check is hidden on Workers Paid.

Check Severity Fires when What to do
budget:free-daily-requestsWarning at 70%, critical at 90%The account has used 70% of today’s requests. The finding projects the time you hit the limit at today’s pace, or says you should stay under it.Find the Worker driving the traffic in the overview. If the traffic is real, move to Workers Paid.

Threshold settings: budgetWarn (0.7), budgetCritical (0.9).

Monthly budget (Workers Paid)

Cloudflare’s spending limit for Workers Paid covers only requests and CPU time and is not an exact maximum. Durable Objects have a spending limit of their own, and KV, R2 and other products are billed outside both. Set a monthly budget per account in Settings › Accounts and Ambersight compares the estimated bill with it after every scan. Cloudflare’s own budget alerts arrive a day late.

Check Severity Fires when What to do
budget:monthly-budgetWarning at 80% of the budget, or when the month-end forecast passes it (after 3 days of usage); critical past itWorkers Paid, with a monthly budget set for the account in Settings › Accounts. The estimate uses list prices and comes within the hour; Cloudflare’s own budget alerts arrive a day late.Open Analytics › Cost to see which Workers and products drive the spend, and look for runaway findings.

Free plan daily limits

The Free plan has more daily limits than requests, and every one is per account: spreading the work over several Workers does not help. Past a limit, calls of that kind fail until 00:00 UTC. Each check projects when the account runs out at today’s pace. The numbers come from the same GraphQL queries Ambersight already makes (D1 rows, KV operations by type), so most cost no extra API calls. Hidden on Workers Paid, except Workers AI.

Check Limit per day Severity What to do
budget:free-daily-d1-rows-read5,000,000 rows readWarning at 70%, critical at 90%Find the query reading the most rows (see usage:d1-inefficient-query). Usually a missing index.
budget:free-daily-d1-rows-written100,000 rows writtenSameLook for a write in a hot path or a loop.
budget:free-daily-kv-reads100,000 readsSameCache in memory or with the Cache API.
budget:free-daily-kv-writes1,000 writesSameKV is built for reads. Batch writes, or move write-heavy data to D1 or Durable Objects.
budget:free-daily-kv-deletes1,000 deletesSameUse expiring keys (TTL) instead of deleting.
budget:free-daily-kv-lists1,000 list requestsSameKeep an index key instead of listing.
budget:free-daily-do-requests100,000 Durable Object requestsSameLook for chatty clients or alarms that fire too often.
budget:free-daily-queue-ops10,000 Queue operationsSameBatch messages; check for a consumer that retries every message.
budget:free-daily-workflow-steps3,000 Workflow stepsSameCombine small steps.
budget:free-daily-ai-neurons10,000 Workers AI neuronsSame on Free. On Paid, info once the free allocation is used, since the rest is billed.Use a smaller model or cache answers.

Threshold settings: budgetWarn (0.7), budgetCritical (0.9), shared with the request budget. The Workers AI dataset is not in Cloudflare’s documentation, so that check may say the data is unavailable.

Storage caps

Read once a day with the inventory: the size of each D1 database, KV namespace and R2 bucket, and the Durable Object total.

Check Limit Severity What to do
budget:storage-d1-db500 MB per database on Free, 10 GB on Paid (a hard cap)Warning at 80%, critical at 95%A full database rejects writes. Delete old rows, split the data, or move large values to R2.
budget:storage-d1-account5 GB on Free, 1 TB on PaidSameSame.
budget:storage-kv-account1 GB on FreeSameDelete what you no longer need, or use expiring keys.
budget:storage-do-account5 GB on FreeSameDelete old objects’ storage with deleteAll().
budget:storage-r2-account10 GB freeWarningPast 10 GB, R2 is billed rather than blocked. Set lifecycle rules.

Runaway usage

These compare a Worker’s last full hour with its median hour over the past 7 days, or with the median of the same hour of day over those days when that is higher, so a nightly batch or a daily peak is judged against its own hour. A spike that no deploy explains names the binding removal or audit log change in the 6 hours before it. A Worker needs 72 hours of history before they apply, so a new install stays quiet for its first three days. Each one also has a floor, so a Worker that goes from 3 requests to 30 does not raise an alarm.

Cloudflare’s spending limit for Workers Paid covers only requests and CPU time and is not an exact maximum. Until it lands, a loop, a retry storm or a crawler is billed for every hour it runs, which makes these the checks that save money.

Check Severity Fires when What to do
runaway:traffic-spikeCriticalRequests in the last hour are 5x the usual hour, and at least 500.Look for a loop, a retry storm or a crawler.
runaway:cpu-spikeCriticalCPU time in the last hour is 5x the usual hour, and at least 60 seconds.Same. Also check for a new slow path in the latest deploy.
runaway:subrequest-spikeCriticalSubrequests in the last hour are 5x the usual hour, and at least 2,000.Look for a fetch inside a loop, or retries without a backoff.

Threshold setting: runawayMultiplier (5).

Idle loops and new bursts

The 5x baseline needs three days of history and cannot see a loop that starts on day one, or work that runs while nobody uses the app. These cover both.

Check Severity Fires when What to do
usage:idle-loop-do, -queues, -d1, -kv, -r2Critical on Paid, warning on FreeFor each of the last 3 full hours, a Durable Object namespace, queue, database, namespace or bucket did at least 2,000 operations while the Workers bound to it served fewer than 1 request per 100 operations. On Paid the finding gives the cost per hour.This is the shape of an alarm that always sets the next alarm, a queue consumer that fails and retries every message, or a cron that never finishes. Find it before it runs all night.
runaway:new-worker-burstCritical on Paid, warning on FreeA Worker with under 72 hours of history served at least 50,000 requests in the last hour.If the traffic is not expected, look for a loop or a crawler.
usage:new-resource-burst-<kind>SameA resource with under 72 hours of history had at least 100,000 operations in the last hour.Look for a loop reading or writing in a hot path.

Threshold settings: newWorkerHourlyRequests (50,000), newResourceHourlyOps (100,000).

Storage and log costs

The same 5x rule as runaway usage, applied to storage operations and Durable Objects. The finding names the resource and the first Worker bound to it. Cloudflare’s own usage notifications do not cover these operations.

Check Severity Fires when What to do
usage:ops-spike-d1CriticalD1 queries on one database in the last hour are 5x the usual hour, and at least 1,000.Look for a query running in a hot path or a loop.
usage:ops-spike-kvCriticalKV operations on one namespace: same rule.Same.
usage:ops-spike-r2CriticalR2 operations on one bucket: same rule.Same.
usage:ops-spike-queuesCriticalQueue operations on one queue: same rule.Check for a consumer that fails and retries every message.
usage:ops-spike-doCriticalDurable Object requests: same rule.Look for an alarm or a client that reconnects in a loop.
usage:logs-full-samplingInfoA Worker has logs on at 100% sampling and serves at least 10,000 requests a day. The finding estimates GB a month, and on Paid the cost past the included 50 GB.From 1 December 2026, logs and traces are billed per GB. Lower head_sampling_rate if you do not need every request.
usage:logs-free-allowanceInfo at 70% of 0.5 GB a day, warning above itFree plan only. The estimated log volume of all Workers is close to or over the 0.5 GB a day that Free accounts ingest from 1 December 2026. Past it, logs stop being stored until 00:00 UTC; the finding says around what time.Lower head_sampling_rate on the busiest Workers, so the logs you need are kept. The estimate assumes about 1.5 KB per event.
usage:d1-inefficient-queryInfo, warning at 100 million rows read a dayOnce a day: a D1 query read at least 1 million rows in 24 hours, and at least 100 rows for every row it returned. The finding shows the query’s shape: parameters and every literal value appear as ?, and comments are removed.Add an index on the columns in its WHERE or ORDER BY. D1 bills rows read.
usage:logpush-onInfoLogpush is on for a Worker.Make sure the destination, and what it costs, is still wanted.

Errors and limits

Check Severity Fires when What to do
errors:error-rateWarning at 2%, critical at 10%At least 2% of a Worker’s requests failed in the last 24 hours. Workers with fewer than 50 requests are skipped.Open the Worker in Workers Observability and look at the most frequent exception.
errors:error-baselineWarningThe 24-hour error rate is 3x what the Worker had in the 6 days before, with at least 10 errors and a rate of at least 0.5%. Not raised when error-rate already fired.Something changed. Check recent deploys, then the most frequent exception. This catches a Worker that normally runs at 0.1% and is now at 0.8%, which the fixed 2% line would miss.
errors:resource-limitWarningAny invocation hit the CPU or memory limit in the last 24 hours.Profile the slow path, or move heavy work to a Queue or a Workflow.
errors:cpu-near-limitInfoFree plan only. The p99 CPU time is at least 80% of the 10 ms limit, and nothing has hit the limit yet.Requests over the limit fail with Error 1102. Trim CPU work before traffic grows.
errors:subrequests-near-limitWarningFree plan only. Over 24 hours, invocations made 40 or more subrequests on average (Free allows 50 per invocation).Only the average is known, so some invocations are probably over the limit already, and the 51st fetch throws. Batch calls, cache results, or move fan-out work to a Queue.

Threshold settings: errorWarn (0.02), errorCritical (0.1), errorBaselineMultiplier (3), cpuShareWarn (0.8).

Cron health

Ambersight reads each Worker’s cron expressions and counts how many runs should have happened in the last 24 hours. After a deploy that changes a schedule, it expects runs only from the next full hour.

Check Severity Fires when What to do
cron:missed-runsWarning. Critical when 2 or more runs were missed and the last 2 slots both went wrong.Fewer runs happened than the schedule expects.Check that the Cron Trigger is still attached and the scheduled() handler is deployed.
cron:failed-runsWarning. Critical when 2 or more runs failed and the last 2 slots both went wrong.One or more scheduled runs ended in an error.Look at the scheduled invocations in Workers Observability for the exception.
cron:cron-cpu-near-limitWarningA scheduled run used at least 80% of its CPU limit: 10 ms on Free, 30 seconds on Paid.Runs over the limit are cancelled. Split the job into smaller steps or move it to a Queue.
cron:cron-countInfo at 80% of the plan limit, warning at the limitThe account uses 4 of 5 Cron Triggers (Free) or 200 of 250 (Paid).At the limit, new schedules are rejected. Combine schedules into one Worker or remove unused ones.

If Cloudflare returns no scheduled runs at all, not even for Ambersight’s own cron, the cron checks pause and keep their current findings instead of raising a missed run for every Worker. See troubleshooting.md.

Bad deploys

This is not a separate check. When errors:error-rate, errors:error-baseline or any runaway check fires, Ambersight looks for a deploy of that Worker in the 60 minutes before the problem started. If it finds one:

  • the finding says “It started right after a deploy.”
  • a warning becomes critical
  • the suggested fix names the deploy: “Compare the deploy of 2026-11-09 22:14 UTC with the one before it, or roll back.”
  • the deploy is flagged on the timeline

Threshold setting: deployWindowMinutes (60). Turn the timeline module off to stop it.

The timeline also shows Cloudflare’s own notifications (alert history, needs Notifications Read), audit log entries for Workers, routes, DNS, certificates, Access, members and tokens, each Pages project’s latest deploy, and, if you turn it on, Cloudflare status incidents. A deploy with no author takes it from the audit log. With recording of who made changes off, the timeline shows no authors.

Check Severity Fires when What to do
deploys:stuck-rolloutWarningA Worker’s latest deployment still splits traffic between versions (a gradual rollout) after 24 hours.Finish the rollout with wrangler versions deploy at 100%, or roll back. While traffic is split, errors are hard to compare.

Threshold setting: rolloutMaxHours (24).

Deprecations

Check Severity Fires when What to do
deprecation:service-worker-syntaxWarningThe Worker uses the Service Worker syntax (addEventListener('fetch', …)).Migrate to ES modules. Durable Objects, Workflows and most bindings need them.
deprecation:old-compat-dateWarningThe compatibility_date is 12 months old or more.Read the compatibility flags changelog since that date, update the date and test.
deprecation:legacy-usage-modelWarningThe Worker reports the bundled or unbound usage model.Both were replaced by Standard pricing. Remove usage_model from your Wrangler config and redeploy.
deprecation:workers-sitesWarningThe Worker has a __STATIC_CONTENT binding, meaning it serves files with Workers Sites.Move to Workers Static Assets: replace the [site] block with an assets directory.
deprecation:redundant-flagsInfoA compatibility flag does nothing at the Worker’s compatibility date because its behaviour became the default. Today this covers nodejs_compat_v2 from 2024-09-23, and nodejs_compat from 2026-08-04, when Node.js compatibility became the default.Remove the flag from compatibility_flags.

Threshold setting: compatAgeMonths (12).

Exposure

Check Severity Fires when What to do
exposure:previews-without-workers-devCriticalPreview URLs are public, although workers.dev is off. People often assume turning off workers.dev turns off previews too. It does not.Set preview_urls = false, or put previews behind Cloudflare Access.
exposure:preview-urlsWarningEvery version of the Worker gets a public preview URL.Set preview_urls = false unless you share previews on purpose. To fix the whole group at once: Workers & Pages › Protect all Workers › Enable Access › Previews only.
exposure:workers-dev-duplicateWarningThe Worker answers on workers.dev and also on a route or custom domain.Requests on workers.dev skip your zone settings, such as WAF rules and Access. Set workers_dev = false if you do not use it.
exposure:workers-dev-publicInfoThe Worker answers only on workers.dev.Fine if you meant it. Set workers_dev = false for internal Workers.
exposure:pages-dev-publicInfoA Pages project has a custom domain, and its *.pages.dev address still answers.Requests to pages.dev skip your zone’s WAF, bot and rate-limiting rules. Redirect it to your domain with a Bulk Redirect, or put it behind Cloudflare Access.
exposure:not-behind-accessWarning. Off by default.A public URL of the Worker (route, custom domain or workers.dev) is not covered by a Cloudflare Access application.Ignore it for public sites. For internal tools, add a self-hosted Access application for those hostnames.

Only Workers with a fetch handler count for the workers.dev checks. For the Access check, an application scoped to a path covers only that path, so it does not count, and *.example.com covers subdomains but not example.com itself. Turn the check on in Settings › Checks. It needs the Access: Apps and Policies Read permission.

Zombies and orphans

All of these are info-level housekeeping. They stay folded away in the overview until you open them.

Check Severity Fires when What to do
zombies:no-trafficInfoA Worker with only a fetch handler had no requests in the last 30 days (Ambersight keeps each Worker’s last hour with a request). It waits until it has tracked the account that long, and never uses a window longer than its own history retention.If nobody uses it, delete the Worker and any storage only it uses.
zombies:never-deployedInfoThe Worker was created but never deployed.Deploy it or delete it.
zombies:missing-serviceInfoA service binding points at a Worker that no longer exists.Calls through it fail. Remove the binding or deploy the missing Worker.
zombies:missing-tailInfoA tail consumer no longer exists.Remove it from tail_consumers.
zombies:orphan-routeInfoA route points at a Worker that no longer exists.Delete the route in the zone’s Workers Routes.
zombies:orphan-domainInfoA custom domain points at a Worker that no longer exists.Remove the custom domain.
zombies:unbound-kvInfoA KV namespace is not bound to any Worker. When storage operations are tracked, the finding adds when it was last used (from any source, not only Workers), its size and whether it is empty. If it had operations in the last 48 hours it says so and gives no delete command: something outside Workers may still use it.Check Last used. “None in 30 days” means nothing touched it, from Workers or anywhere else; export what you need and delete it.
zombies:unbound-d1InfoA D1 database is not bound to any Worker.Same.
zombies:unbound-r2InfoAn R2 bucket is not bound to any Worker.Same. R2 is often used through its S3 API, so check before deleting.
zombies:unbound-queuesInfoA queue has no producer, no consumer and is not a dead-letter queue.Same.
zombies:bound-idle-kvInfoA KV namespace is bound to at least one Worker but had no operations in the last 30 days (only once Ambersight has tracked that long). Storage bound only to one unused Worker is listed on that Worker’s finding instead.Remove the binding from each Worker’s Wrangler config and redeploy, then delete it.
zombies:bound-idle-d1InfoThe same for a D1 database.Same.
zombies:bound-idle-r2InfoThe same for an R2 bucket.Same.
zombies:bound-idle-queuesInfoThe same for a queue (not one with a consumer).Same.

Is it safe to remove? “Not bound” alone does not prove it: storage can be used through the API, an S3 client or a public URL. So Ambersight also keeps the last hour each KV namespace, D1 database, R2 bucket and queue had operations, from any source, and shows it as Last used with the size. This needs the usage checks on (the default) and costs no extra API calls: it comes from the operations the scan already reads. “No operations in 30 days” is only said for a kind (KV, D1, R2 or queues) whose operations Ambersight has read without a gap for those 30 days: a failed dataset, a kind skipped because the scan ran out of API budget, the usage checks being off or two days without a scan start the count again, and until then the finding makes no claim about use. The stored time is updated at most every 6 hours per resource, so it costs a handful of D1 rows a day.

Cleaning up a Worker. zombies:no-traffic and zombies:never-deployed list under Leaves behind the storage that only that Worker binds, and How to remove it adds the commands for it after wrangler delete. Storage that still had operations in the last 48 hours is listed but gets no command. Bindings are read from every Worker in the account, so storage another (out-of-scope) Worker uses is never listed.

Cleaning up after a deleted Worker. Deleting a Worker leaves its KV namespaces, D1 databases, R2 buckets and queues in place. After the next inventory refresh (once a day by default) they show up under Housekeeping as not bound to any Worker. Each of these findings has a How to remove it section with the exact command, such as npx wrangler d1 export old-db --remote --output=old-db-backup.sql and then npx wrangler d1 delete old-db. Ambersight is read-only and never runs them. Routes and custom domains left pointing at deleted Workers get the dashboard or API step instead.

What it frees. Under the Housekeeping heading, a line shows how much storage removing the unused and idle storage would free, per product and as a share of the account allowance, with the monthly saving when the cost estimate is on, and how many housekeeping findings were resolved in the last 7 days. The weekly digest has the same count.

Resources page. Resources lists every KV namespace, D1 database, R2 bucket and queue with the Workers that bind it, its size, when it was last used and a status: In use, Idle, Not bound or Used outside Workers. Check it before deleting anything.

Export. Export .md next to Needs attention (critical and warning) and next to Housekeeping (info) saves the findings as a Markdown checklist. The Housekeeping export ends with every cleanup command in one sh block, set to the right account with CLOUDFLARE_ACCOUNT_ID. The export covers the selected group, whatever the severity filter or search shows.

The unbound checks wait until Ambersight has read every Worker’s bindings at least once. On a large fleet that takes a few scans (see how-it-works.md), so they may appear a little after the others.

Threshold setting: zombieDays (30).

Pages

Read once a day with the inventory (needs Pages Read). Pages Functions count toward the same 100,000 daily requests as Workers; the overview’s Pages section shows each project’s function requests.

Check Severity Fires when What to do
pages:deploy-failedWarning for production, info for previewsThe latest deploy of a Pages project failed.Read the build log in Workers & Pages. The site keeps serving the last deploy that worked.

Security

Check Severity Fires when What to do
security:member-changeWarning (info for a removal)The audit log shows an account member added, changed or removed in the last 7 days.If you did not expect it, check Manage Account › Members. Acknowledge it once you have.
security:token-createdInfoThe audit log shows a new API token in the last 7 days.If you did not expect it, review its permissions under API Tokens.
security:cert-expiringWarning at 7 days, critical at 2An active edge certificate expires soon. Managed certificates renew about 30 days ahead, so this one is not renewing.Check its validation in SSL/TLS › Edge Certificates. Upload a renewed one if it is a custom certificate.
security:cert-failedCritical when expired, otherwise warningA certificate pack is expired or timed out, and no other active pack covers its hostnames.Visitors see TLS errors. Check DNS validation records, CAA records and whether the hostname is still proxied.
security:cert-pendingWarningA certificate is stuck in validation with an error.Fix the error shown, usually a missing TXT record or a CAA record that blocks the certificate authority.
security:dangling-pagesCriticalA DNS record points at <name>.pages.dev, and the account has no Pages project of that name.Anyone can create a project with that name and serve content on your hostname. Delete the record.
security:dangling-workers-devWarningA DNS record points at <worker>.<your-subdomain>.workers.dev, and that Worker no longer exists.Requests to the hostname fail. Delete the record or deploy the Worker again.

The audit log is read every 6 hours (it needs Account Settings Read, which the token already has). Certificates and DNS records are read per zone on a rotating cursor, a few zones per scan, each zone at least once a day (SSL and Certificates Read, DNS Read). The dangling Pages check waits until the Pages project list has loaded once, so a missing permission cannot make every record look dangling.

Config drift

Check Severity Fires when What to do
drift:binding-removedWarningA binding, variable or secret that a Worker had at the last read is gone, and still missing. Raised for 7 days.Usually a deploy dropped it: variables and secrets set only in the dashboard are replaced by the next wrangler deploy unless they are in your config (or kept with --keep-vars).
drift:env-mismatchInfoA Worker named like api-staging, api-dev, api-preview or staging-api has binding names that api does not, or the other way round.Compare the two configs. Secrets are compared by name only; their values are never read.
drift:vars-near-limitInfo at 58, warning at 64A Worker has close to the 64 environment variables and secrets allowed.Group related settings into one JSON variable, or move configuration to KV.

Packages

Off by default. For each Worker you map to a GitHub repository in Settings › Checks › Packages, Ambersight reads package-lock.json once a day, takes the direct dependencies and devDependencies with their locked versions, and compares them with the latest release on the public npm registry. The Packages page lists them as locked → latest.

Check Severity Fires when What to do
packages:outdatedWarning if a direct dependency is a major version behind, otherwise InfoOne or more direct dependencies whose lockfile entry shows they came from the public npm registry are behind its latest release.Run npm outdated in the repository and update. Read the changelog of each major version first.
packages:lockfile-unreadableWarningThe lockfile could not be read: no package-lock.json at that path, no access, the repository moved, a lockfile older than npm 7 (lockfileVersion 1), or a file over 8 MB, or a repository outside GITHUB_REPO_ALLOWLIST.Check the repository, folder and branch in Settings, and that GITHUB_TOKEN can read the repository.

What is compared. Only a package whose lockfile entry was resolved from registry.npmjs.org under its own name is looked up there. A package from a private registry, a Git URL, a local path or an alias, or one with no recorded source, is listed as not compared: its name is not sent to npm and it gets no update advice, because a package of the same name on npm is not an update to it. Rows stored by an older version have no source yet: until the lockfile is read again, which the first scans after the upgrade do, they keep their comparison with the version already stored, and nothing new is looked up for them.

Which repositories. Anyone signed in can map a Worker to any repository GITHUB_TOKEN can read, and the Packages page then shows that repository’s dependency names. Give the token only the repositories you mean to map. To enforce it from the deployment, set GITHUB_REPO_ALLOWLIST to owners or owner/name pairs: anything else is not read, what was stored for it is dropped, and the Worker gets a packages:lockfile-unreadable finding. Out-of-scope Workers are never read.

Why this needs a separate token: Cloudflare stores your Worker as bundled code, without package.json or the lockfile, so the Cloudflare API cannot tell which versions you use. Workers Builds knows the repository, but its API needs an Edit permission, which Ambersight never takes. Only package-lock.json v2 and v3 are read (npm 7 and later); pnpm and yarn lockfiles are not supported yet. Lockfile reads and npm lookups count against the scan’s subrequest budget, so on the Free plan a large fleet is covered over several scans.

Ambersight’s own health

They have no switch in Settings › Checks: a monitor that has gone blind should say so.

Check Severity Fires when What to do
health:token-expiryWarning at 14 days, critical at 3Ambersight’s API token has an expiry date that is close.Extend the token or create a new read-only one, then update CF_API_TOKEN.
health:token-scopeCriticalAt an inventory refresh, the token has a permission outside the list of read permissions Ambersight has a use for (below). The comparison is against that list, not against what a name looks like, so an extra read permission counts too.Remove the permission, or create a new read-only token. If you added a read permission on purpose, switch this check off from the finding.
health:token-scope-unverifiedInfoThe token’s permissions could not be read at the last inventory refresh, usually because it lacks API Tokens Read. A grant found earlier stays reported as health:token-scope.Add API Tokens Read, or check the token yourself under My Profile › API Tokens.

After 3 failed scans in a row, Ambersight also sends one alert a day on the channel set for critical alerts, and the overview shows the last error. A scan that fails cannot store a finding, so this alert has none behind it.

Switching checks off

There are three levels:

  • A whole group. Settings › Checks has a switch for usage, errors, cron, deprecations, exposure, zombies, the deploy timeline, cost, Cloudflare notifications, Pages, security, config drift, Cloudflare incidents and packages.
  • One check. From any finding group in the overview, switch that check off. It shows up under Settings › Checks › Switched off from the overview, where you can turn it back on. In config.json this is the mutedChecks list, using the IDs above.
  • One finding. Acknowledge it, or snooze it for 7, 30 or 90 days. Acknowledge all and Snooze all work on a whole group.

When you switch a check off, its open findings resolve, the same as if you had fixed them.

Which permission each group needs

Group Token permission (all Read)
EverythingWorkers Scripts, Account Analytics
Free plan budgetBilling, to detect the plan. Without it, set the plan in Settings › General.
Storage and log costs, unbound storageWorkers KV Storage, D1, Workers R2 Storage, Queues
Exposure, orphaned routes and domainsWorkers Routes, Zone
Access coverageAccess: Apps and Policies. Off by default and not in the pre-filled token link; add it by hand when you turn the check on.
Checking the token’s own permissions (setup wizard and daily)API Tokens Read
PagesPages Read (key page)
CertificatesSSL and Certificates Read
Dangling DNSDNS Read
Audit logAccount Settings Read
Cloudflare notifications on the timelineNotifications Read. It has no key for the pre-filled token link, so add it by hand.
PackagesNo Cloudflare permission. A separate GITHUB_TOKEN secret: GitHub fine-grained token, Contents: Read on the mapped repositories.

If a permission is missing, the scan skips what it cannot read, carries on with the rest, and lists what it skipped. See troubleshooting.md.

Read next

  • Error 1027

    What Cloudflare error 1027 means on Workers Free, who it stops and how to see it coming.

  • Workers Free limits

    The Workers Free daily, per-request and storage limits, and when Ambersight warns about each.

  • The Workers Paid bill

    Runaway usage, idle Durable Object and Queue work, the estimated bill and Cloudflare's spending limit.

  • The warnings chart

    The home page: a demo fleet on the chart, how Ambersight works and how to install it.

Put your own Workers on the chart.

One Worker in your account, a read-only token and about 5 minutes. The first scan starts on its own.