Test results

OpenNext on Cloudflare Deploy Review: test results

Tested 2026-11-22, skill version 1.0.1 at the time of loading this page. We run every skill on a strong and a weak model before it is listed, and publish both verdicts, including where the weak one fails.

Verdicts

Date
2026-11-22
Strong · claude-sonnet-5-5 (Claude Code alias "sonnet")
Right on all 15 inputs on the 1.0.1 bytes, read by hand. It moved the caret ranges on next 16.4.0 with OpenNext 1.20.10 to the pair tried in November, pinned exactly; caught an OpenNext bump to 1.20.16 that left wrangler under 4.148 and next under 16.4.5; split a CI deploy on two custom domains into an upload and wrangler versions deploy; refused next upgrade --agent=security; read the 52 of 79 tail events as no proof of lost requests; called the DOQueueHandler build lines harmless; and answered OK on all five sound setups, including the one without a buildCommand. On the commit that deletes a Durable Object class it said the upload stops with an error naming the pending migration, so the job goes red and no version is made, and sent the commit to one full deploy from a terminal with a TTY. The version facts it used were measured on dates we give (11 October and 17 November 2026) and go stale, and it told the reader to check what is current.
Weak · claude-haiku-5-5 (Claude Code alias "haiku")
Right on all 15 inputs on the 1.0.1 bytes, read by hand. It moved the caret ranges to the tried pair with exact pins, named both version requirements of OpenNext for the bump to 1.20.16, split the CI job on custom domains into an upload and a promote step, kept the dashboard-only flag out of the full deploy by putting it in vars, and refused the security-upgrade prompt. On the commit that deletes a Durable Object class it said the upload refuses the pending migration, the job goes red and no version is made, and asked for one full deploy from a terminal. It answered OK on the default OpenNext build (no buildCommand, so npm run build runs) and on the correct upload-and-promote CI step, without asking for a smoke test. Its version numbers carry the measurement dates of the skill and go stale like them.

With and without the skill

Tested 2026-11-22.

Results with and without the skill, for Sonnet and Haiku
SonnetHaiku
withwithoutwithwithout
Inputs judged right (15 inputs)15/1510/1515/156/15

Same request on both sides, scored on the point in any words. Read by hand, Sonnet without the skill already knew the build-script bypass of a buildCommand, the loss of dashboard-only values (and keep_vars), lossy wrangler tail, the harmless DOQueueHandler lines, and that a version upload does not apply a Durable Object migration. It missed five measured facts: it approved caret ranges on next 16.4.0 with OpenNext 1.20.10, approved a CI deploy without a TTY on two custom domains and later moved a class deletion into that job, said yes to next upgrade --agent=security, and passed an OpenNext 1.20.16 bump with wrangler 4.140 and next 16.4.3. Haiku without the skill was right on six: the same misses, plus false alarms on four of the five sound setups. With the skill, run again after the wording fix, Haiku was right on all 15 and gave no false alarm on a sound setup.

Same cases and the same checks with and without the skill. The cases are ours, written around what the skill is for; with a handful of cases, a difference of one or two is within noise.

Note

Fifteen inputs written by us: eleven deploy setups pasted before a release (package.json with npm ls, a GitHub Actions job with the wrangler routes, open-next.config.ts, wrangler vars, a commit that deletes a Durable Object class), five of them sound, and four questions about a symptom (500 with loadManifest after npm update, SecurityFatalError, fewer tail events than requests, DOQueueHandler lines during the build). Code scores every answer, with identical checks for both sides: the verdict line for a review, and for each fault the point in any words (an exact pin, a tried pair, DNS together with versions deploy, the versions OpenNext needs, a full deploy from a terminal). Re-checked on 2026-11-22 in the installed sources: wrangler 4.156.0 sets override_existing_dns_record for custom domains when there is no TTY, and its version upload stops with an error while a Durable Object migration is pending; OpenNext runs the package manager's build script when no buildCommand is set. Owner-measured and not re-checked: the 500 of next 16.4.0 with OpenNext 1.20.10 (11 October 2026), next 16.4.6 paired with OpenNext 1.20.17 and serving production since 17 November 2026, the requirements of OpenNext 1.20.13 and 1.20.16 on wrangler and next, SecurityFatalError on 16.4.3, tail losing events, and the DOQueueHandler lines coming from the local proxy. Version facts go stale. Two checks were widened after the run for both sides: the migration check takes 'Neither applies Durable Object migrations' in the sentence after the upload commands, and the build-warning check takes 'comes from the Next.js build step' as the source. After the first run, the wrangler source showed that a pending Durable Object migration makes the upload stop with an error (CI goes red, no version is made), where the skill text had said the push goes green; the wording was fixed on 22 November 2026, with two more 'not a finding' lines (no buildCommand means OpenNext runs npm run build; a CI job without a smoke test). The skill side was then re-run in full on all 15 inputs (the earlier run with the old wording: Sonnet 15 of 15, Haiku 13 of 15). The bare side is unchanged. One run per model and input. Version 1.0.1, 22 November 2026: the version requirements were corrected from the adapter's own peer requirements on npm: wrangler 4.148 is needed from OpenNext 1.20.10 (the file had said 1.20.13), and 1.20.12 needs next 16.4.1, which the file did not say. Three sentences were reworded so that no published post on our blog shares a run of five words with the file. One check was widened for both sides: going back to any OpenNext from 1.20.10 to 1.20.15 meets the next requirement of that input. The side with the skill was run again in full on those bytes: Sonnet 15 of 15, Haiku 15 of 15. The side without the skill, re-scored with the same checks: unchanged, 10 and 6 of 15.

What was not measured

  • Models other than the two named above were not run.
  • Each verdict comes from the test run on the date shown; the skill may have changed since (check the version).
  • Full test inputs are not published here, only short excerpts of our own text.
  • Results on your own texts, languages and domains can differ.

Back to OpenNext on Cloudflare Deploy Review · Card (JSON)