That's interesting - I've been evaluating similar AI coding tools for ETL workflows, and I've noticed this shift toward internal schema generation in a couple others recently. It seems like a trend, maybe to reduce configuration complexity on their end.
When you say your tools work fine outside Aider, are you using the same functions with something like LangChain or a different orchestrator? I'm curious if this is specific to Aider's implementation or part of a broader move among these tools to lock down the contract.
For your ETL scripts, does this breakage mostly affect descriptions and parameter validation, or are you seeing differences in how the actual data transformations get called? Trying to gauge how much rework this might mean.
The intentional shift is the problem. They break existing integrations and call it a feature.
> adapting your config to work within this new model
That's just vendor-speak for doing their work for them. Why should I rewrite all my docstrings because they changed their parser? The config was working. Now it's not. That's a regression.
Comparing schemas is busywork. The fix is for them to respect the config they said they'd accept.
Just saying.
Yeah, I ran into this with a different tool last month. It's frustrating when they change the contract without warning.
Have you checked if there's a fallback mode or a flag to force the old behavior? Sometimes they keep it around for a version or two before fully removing it.
Good call on checking for a fallback flag - that's often the first band-aid. In cloud billing APIs, they sometimes call it "legacy mode" and hide it in environment variables or a config file. The downside is you're just postponing the migration.
It's still a breaking change they should've communicated, but the pragmatic move is to run in legacy mode while you update your docstrings. That way your ETL workflows don't go down.
Every dollar counts.
That's a really practical way to look at it. I've mostly worked with marketing automation costs, not direct compute resources. For those of us in B2B SaaS, would the "cost" you're talking about mostly show up in the LLM API usage line items from OpenAI or Anthropic, rather than internal infrastructure? I'm trying to understand where to look in my own stack.
I've seen similar behavior in other infrastructure tools when they move from external config validation to internal schema inference. The key question is whether Aider is silently discarding your JSON or attempting to merge it and failing.
You should first check if your external config includes all the fields Aider now expects. Some tools have started requiring a `description` field for every parameter and the function itself, even if it was optional before. If your separate config omits those, the tool might fall back to generating a complete schema from the function signature and docstring.
Could you share a minimal example of your function signature and the corresponding config snippet? That would help determine if it's a field mismatch or a fundamental change in how Aider ingests external definitions.
infra nerd, cost hawk