Reviews pasted code that calls a language model (JavaScript, TypeScript or Python; REST, @google/genai, google-genai, or an OpenAI-compatible client for Groq), usually through a chain of models or providers with retries. It finds where the call path goes silent or stops while another model would answer (an empty 200 taken as the answer, a 429 or 5xx that ends the call, a retired model's 404, a Groq schema 400 that drops the provider, one provider's stop that ends all) and what costs more or hides the cause (retries on top of SDK retries, a missing thinking setting, a thinking level the model refuses, a small output limit, a moving -latest alias, errors without the model name, unguarded JSON, a model picked from the listing, several keys of one project, a weak model that publishes, a quality gate that cannot fail, a fixed price for open-ended reasoning). Each finding has a fixed code, the place, the reason and a fix, then one verdict. Use to review LLM fallback code, check a Gemini 503 retry, a Gemini empty response or a Groq json_validate_failed handler before the code runs unattended.
LLM API Fallback Review: Gemini, Groq, Chains is a tested SKILL.md that reviews pasted code that calls a language model (JavaScript, TypeScript or Python; REST, @google/genai, google-genai, or an OpenAI-compatible client for Groq), usually through a chain of models or providers with retries; an agent buys it once for $0.05 over x402.
Running the code or seeing the live API: it reads pasted code only, so it cannot confirm which model an alias points to today. It does not choose a provider for you, judge your prompts, or cover image and audio models. For OpenAI and Anthropic it applies only the general rule for 4xx and 5xx.
Tested, honestly
Tested 2026-10-16 with a strong and a weak model.
With and without the skill
Results with and without the skill, for Sonnet and Haiku
Sonnet
Haiku
with
without
with
without
Code samples reviewed right (31 samples)
Code samples reviewed right (31 samples)
31/31
27/31
30/31
23/31
Same request on both sides, scored on the concept in any words. Read by hand, Sonnet without the skill already treated an empty answer as a failure, added a second model to a job that only postpones, sent a 404 or a 500 on to the next model, kept Groq running after a Gemini 403, knew that thinking costs output tokens and that a moving alias should be pinned. It went wrong in four places: it advised sending an ordinary 400 on to the next model (the vendor says do not retry it), did not know the Python SDK already repeats 429 and 5xx calls, trusted the models listing to prove a name works, and called two closed lists that share the value other sound. Haiku without the skill missed the same four, the Groq schema 400, the missing thinking setting and the minimal level that 3.8 Flash refuses.
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.
SonnetStrong model, claude-sonnet-5-5
Right on all 31 pasted code samples, read by hand: it named the empty answer saved as a result (two cases), the single-model job that only postpones on a 503 and the chain whose 500 throws, the 404 that never reaches the alias placed last, the Groq schema 400 that drops the provider (TypeScript and Python), the Gemini 403 that ends the call before Groq, the plain 400 walked through every model, the own retry loop around SDK retries, the missing thinking setting (two cases), the 60-token limit on a thinking model, the minimal level on 3.8 Flash and on 3.1 Pro preview, the alias first or alone (two cases), errors without the model name, unguarded JSON in a queue consumer, the model picked from the listing, four keys of one project, a lite model that publishes Bulgarian text unchecked, two closed lists that share a value, a length-only quality gate and a fixed price for high reasoning, and in the combined case both faults despite a planted comment asking for no findings. Where the input left a thrown network error uncaught it added that as well, which is true of those inputs. It answered No findings. for all five sound samples.
HaikuWeak model, claude-haiku-5-5
Named all twenty-six planted faults, read by hand, with the verdict line last and no fence, but it missed one sound sample: on the Groq client in strict mode it reported a schema 400 that strict mode does not return and a parse problem that the code already handles. In eleven of the fault answers it also added small extra findings, most often a missing response message reported as an error that does not name the model.
Our own test text, before and after the skill ran. Excerpts only.
English · claude-sonnet-5-5
Before
=== worker/summarize.ts ===
// scheduled job: one summary per product, saved to D1
const MODELS = ["gemini-3.6-flash", "gemini-3-flash-preview"];
async function ask(env: Env, prompt: string): Promise<string> {
let last = "none";
for (const model of MODELS) {
const res = await fetch("https://generativelanguage.googleapis.com/v1beta/models/" + model + ":generateContent", {…
After
[EMPTY-ANSWER-FINAL] worker/summarize.ts, `ask` (the `res.ok` branch) and `scheduled`: a 200 with no candidates, no text part or a non-text first part returns `""` through `?? ""`. The loop then writes `""` to `products.summary`. Because the job selects `WHERE summary IS NULL`, that product is never retried, so it keeps a blank summary for good and the second model is never tried.…
What is in the file
The answer
What the paste decides
How a response is sorted
The codes
Rules
Work in this order
How to check a fix
Short example
Languages
Any language. Tried in: English.
License
Perpetual, non-exclusive; use and modify for yourself incl. paid work; no resale or republishing. Holder: Georgi Kalchev, aiskills402.com. Full terms.
Versions
Current version 1.0.0, updated 2026-10-16. Whoever bought an earlier version gets new ones free through the same re-download token.
v1.0.0 · 2026-10-16
First release: reviews pasted code that calls a language model (JavaScript, TypeScript or Python; REST, @google/genai, google-genai, or an OpenAI-compatible client for Groq), usually through a chain of models or providers with retries, and lists each problem with a fixed code, the place, the reason and a fix, then a verdict (broken, fix before use, no known problems). Broken: an empty 200 taken as the answer, a 429, 408, 5xx or a thrown network error that ends the call, a retired model's 404 that ends the chain, a Groq schema 400 that drops the provider, one provider's stop that ends all providers. Fix before use: an own error walked through every model, retries on top of SDK retries, a missing thinking setting, a small output limit on a thinking model, a thinking level the model refuses, a moving alias first, errors that do not name the model, unguarded JSON, a model chosen from the listing, several keys of one project, a weak model that publishes, two closed lists swapped, a quality gate that cannot fail, a fixed price for open-ended reasoning. Facts: the Gemini and Groq vendor pages read on 16 October 2026 and our own measurements of September and October 2026; the skill tells the reader that the current vendor page wins.
FAQ
An empty 200 looks like success. What is wrong with it?
We measured in September 2026 that a Gemini call can answer 200 with finishReason STOP and zero text parts, because the thinking used the whole output limit. Code that returns an empty string and saves it has stopped for good, while the next model would have answered. The skill treats an empty answer as transient and names the model, finishReason and thoughtsTokenCount to log.
Is a Groq 400 always my own error?
No. Groq documents that best-effort structured output can return a 400 when the generated JSON does not match the schema, and lists retrying as a fix. On 22 September 2026 the same prompt and model that had worked hours before returned json_validate_failed. The skill sends exactly that error to the next model of the same provider and keeps every other 400 as a stop.
Which facts can go out of date?
Model names, aliases, thinking levels and prices. The vendor pages were read on 16 October 2026, for example that 3.8 Flash has no minimal level and that a moving -latest alias is swapped with every release. When the current vendor page disagrees with the skill, the page wins, and the skill tells you which line to re-check.
Does it help Claude Sonnet?
Yes, in four places. Without the skill Sonnet reviewed 27 of our 31 samples right; on its own it caught empty answers saved as results, single-model jobs, retired names and unguarded parsing. Its four misses: it advised passing an ordinary 400 on to the next model, did not know the Python SDK already repeats 429 and 5xx calls, trusted the models listing as proof that a name works, and called two closed lists sharing the value other sound. Using the skill, Sonnet scored 31 of 31; Haiku rose from 23 to 30, adding small extra findings in eleven answers.
Our model chain stopped while another model would have answered. Which responses go to the next model, which end a provider, and where the obvious rule fails.
Reviews pasted code and configuration that spends money (paid AI or API calls, purchases, paid lookups, anything billed per use) and checks whether the limit on that spending really stands in the way of the spending. It flags a cap that sits beside the path instead of on it (a gateway limit while a script calls the provider directly, a counter only in the browser), a missing setting that means no limit, an alert or a budget email used as the protection, a cap that the caller sets through a header or a request field, a paid job that runs by default, a per-call cap with no running total, a total kept where it resets, a cap checked only after the money is gone, and a spend counter whose failure or missing value lets the call through. Each finding has a fixed code, the place, the reason and a fix, then one verdict. Use before deploying a Worker, script or job that pays per call, or when a cap exists and you want to know if it can fail silently.
Finds the loops behind a bill spike or an unexpected bill. Reviews a web app's code and config for loops and repeats that run up the bill on their own on Vercel, Turso, Cloudflare Workers and D1 and paid AI APIs. Finds two schedulers on one job, retries without a cap, work that triggers itself, client polling and React render loops, caches cleared on every write, middleware running on every static file, queues that resend failures forever, image settings that multiply transformations, health checks that count whole tables, alerts sent on every run and bots crawling endless filter URLs. Counts the runs per month, names the billed unit, ranks by cost and gives the change that stops each one. Use when asked why a Vercel, Turso or Cloudflare bill jumped, to find an infinite loop or a runaway cost, or to review a cron job, webhook, queue consumer, middleware or polling code before it ships.
Reviews pasted code or a design for a liveness watchdog, a staleness alarm or a daily heartbeat mail (sources that must keep answering, a pipeline that must keep publishing) and lists every way it stays silent on a dead source or floods on a healthy one. It checks whether silence is measured in the time the source was last asked or by the wall clock, whether a stateless alarm window repeats or never fires, whether the threshold is longer than the system's own cadence plus a tick, whether a new row starts with a last success of zero, whether the sweeper or the probe sits behind a daily limit, whether a failed send is remembered as sent, whether alert memory is purged by age, whether the subject names the dead source and the environment, whether a lamp still has a writer after a scheduler swap, and whether every check lives inside the thing it watches. Each finding has a fixed code, the place, the reason and a fix, then one verdict; sound code gets exactly No findings. Use to review a watchdog, a liveness check or a staleness alarm, to audit the design of a monitoring job, or to check an alert mail before you rely on it.
Reviews pasted Cloudflare Workers code and configuration (a Worker, a Next.js app on OpenNext, D1 queries, wrangler config) for platform pitfalls that pass every local test and then fail in production or quietly cost money. It flags the edge runtime under OpenNext, a fetch to your own Worker's workers.dev address (error 1042), D1 queries that bind more than 100 parameters as data grows, LIKE patterns over D1's 50-byte limit, reading rowsAffected where D1 returns meta.changes, interactive transactions D1 does not have, secrets kept in plain vars, outbound fetch code that treats only a thrown error as failure, a browser User-Agent that bot protection challenges, client hop-by-hop headers forwarded to fetch, cron triggers whose day of the week is written as numbers, and unbounded queries on the request path. Each finding has a fixed code, the place, the reason and a fix, then one verdict. Use to review a Cloudflare Worker before deploy, check D1 or wrangler code, or audit a Next.js app running on Cloudflare.