Portable Or Native: The Endpoint Choice Is A Runbook Decision, Not An SDK Preference

A dev.to post on building a first in-app chatbot that reviews logistics code changes argues for starting with an OpenAI-compatible endpoint and keeping the provider boundary narrow enough to replace later, and frames the pick as an operations decision disguised as an SDK decision ^[1][3]. The reason that framing matters: the dangerous outcome is not a slightly awkward client call but a deploy that silently drops the system prompt, replays a review, or changes the JSON shape when the team switches models ^[5].

The workload is mundane. The bot takes a diff, applies warehouse and dispatch rules, and returns structured findings ^[4]. The author's invariant is that a provider change must not alter the application's review request, finding schema, or retry policy, and that those three things live in your code while the remote chat protocol is treated as an adapter ^[6]. Findings carry stable fields such as severity, file, line, and explanation, and a model answer is evidence for a reviewer, not permission to merge ^[9]. The design test is the change you expect in month six rather than the demo you can finish this afternoon ^[7]: when dispatch/routing.go alters its fallback depot selection, the application should send the same repository path, diff, and policy version regardless of which model sits behind the endpoint ^[8].

Compatibility earns its keep narrowly. Existing chatbot samples and middleware can usually be reused, which reduces the adapter's surface area rather than eliminating it ^[10]. The author's case for the compatible contract for beginners is ecosystem breadth - examples, SDK support, middleware, migration paths - with Anthropic's native API chosen only when its distinct contract is a requirement you are prepared to own and test directly ^[2]. Pretending the semantics match is what the post calls portability theater: the types compile, but prompt placement, structured-output behavior, or error handling can still differ ^[11]. The stated rule is short: compatible contract when fast integration and future routing matter most, native API when provider-specific behavior matters more than replacement cost ^[13].

The operationally interesting part is the part no endpoint style fixes. A worker times out after the model accepts the request but before the application stores the findings; the queue delivers the job again; two successful model calls now exist and two workers may race to write the review ^[14]. The preventative boundary goes around the whole review record: derive a stable review ID from repository, commit SHA, policy version, and diff digest, store it before calling the model, and make the final database write conditional on the same ID ^[15]. Note that policy version shows up in both the request payload and the four-input key, which is a useful consistency check on whether your boundary is actually drawn ^[22]. A 429 is retryable after Retry-After, a malformed finding is not, and a client timeout is ambiguous, so the worker may call again but must never publish a second review record ^[16]. That is the at-least-once consumer reflex applied one layer above the model provider ^[17]. Without a retry budget and a stable local key, the shortest retry loop turns one delayed review into an unbounded queue backlog; the post puts an explicit limit in the client, for example three attempts, while the durable worker policy stays under application control ^[18].

What to watch is the evidence trail. Record the adapter name, requested model ID, policy version, prompt template version, response status, and raw response in access-controlled storage, and keep secrets and unredacted customer data out of general application logs ^[20]. For a logistics company handling personal data, retention and deletion need defining before the bot reaches production ^[21]. Measure your own queue visibility timeouts and diff sizes instead of copying a timeout from a sample ^[19]. And before a second provider goes in, run the replay set the author recommends: real, redacted diffs through both adapters, findings compared field by field ^[12].