Prompt config is cheap until a UI edit reroutes production traffic

The typo is the least interesting part of the objection. The thing worth arguing about is that a Langfuse prompt version cannot tell you what kind of change it contains. Rewording an extraction instruction and swapping a small deployment for a large one both land as a version bump on the same record, and both go live the same way, by moving the label the code points at ^[1][4]. The post frames these as two lifecycles with different risk profiles sharing one editable object ^[17]. That undersells it a little: in most shops one of those changes has a reviewer and the other does not.

The registry example gives the gap a size. In the sample production config, `extractor-strong` resolves to an Azure deployment named `gpt-4o-prod` with an output ceiling of 4096 tokens and a default temperature of 0, while `summarizer-cheap` resolves to `gpt-4o-mini` with a ceiling of 1024 ^[9][10]. That is four times the output budget ^[14] on two differently priced deployments, and in the coupled design the value that decides between them is an unschema'd string in a text field ^[15].

The reason the alternative is a small change rather than a rewrite is a detail of how the client is built. The provider follows from which client type you construct and its endpoint, the deployment from the `GetChatClient` call, and by the time you hold an `IChatClient` it already knows what it is, so `ChatOptions.ModelId` barely gets touched ^[6]. Picking a model collapses into picking a pre-built object by name, which is what keyed dependency injection already does. Langfuse keeps the template, the variables and an alias; `appsettings.Production.json` decides what that alias resolves to, and the same alias can point somewhere cheaper in a lower environment ^[8].

Two things to be honest about. The redeploy-free promise does not survive whole, it survives for wording and variables: adding a model or retargeting an alias becomes an appsettings edit and a deploy ^[8]. And the audience the post concedes exists, the people who want a prompt version to fully describe and reproduce its own execution ^[11], gives that up completely. The version records intent, the deployed binary records what ran.

The scope is narrow and the post says so: a .NET codebase on Microsoft.Extensions.AI, against OpenAI or Azure OpenAI, with Langfuse supplying prompts ^[12]. The construction-time binding claim is made for those two providers, and the keyed-client pattern rests entirely on it. Worth noting too that the registry field is named `MaxOutputTokensCeiling` rather than `MaxOutputTokens`, and `ModelDefinition` carries a `[Required]` validation attribute ^[13], which reads like the prompt is permitted to ask for less and not more. The code that clamps it is not in the published excerpt.

So the boundary being proposed is not really about model strings. It is that the prompt store may say what work to do, and the deploy decides what executes it.