Skip to content

baml-language-source-9e09e8c19a52ac6a915d2aa1d05e43d678fa6ff6

baml-language-source-9e09e8c19a52ac6a915d2aa1d05e43d678fa6ff6: fix: reject non-data LLM output schemas (#4470)
@antoniosarosi antoniosarosi tagged this 18 Aug 02:11 
## Summary

- reject non-data realized types at the shared LLM render boundary with
catchable E0164 diagnostics
- report nested failures with their full class-field path and ignore
fields marked `@skip`
- reject realized generic classes before the current unsubstituted
renderer can silently erase their schema and proceed to provider I/O
- replace the low-level `unreachable!` fallback with a fallible
formatter error

## Root cause

`type._validate_renderable` guarded open interfaces, conflicting runtime
definitions, and empty enums, but did not recursively reject non-data
leaves. Those leaves reached `sys_ops::output_format`; `never` and its
family aborted the process, while the first fix made unsubstituted
generic-class fields return an error that throws-never helper callers
swallowed into an empty schema.

## Compatibility / strictness

This intentionally makes previously degraded renders fail loudly with
E0164: examples include a `uint8array` field that rendered as
`RESULT=[]`, a `type` field, and a `string | uint8array` field that
rendered only the string branch. Skipped fields remain ignored, matching
the renderer. Generic functions whose realized outputs are non-generic
remain supported.

## Deliberate deferrals

- Real substitution of realized generic class arguments into
`ClassDefinition.field_type` in `sys_ops::output_format` is a separate
migration follow-up. Until then, generic classes fail safely instead of
aborting or rendering without a schema.
- Static rejection is intentionally not implemented: this is the
runtime-only shared seam, matching E0159/E0161/E0162 precedent.
- The throws-never low-level output-format helpers are not routed
through `_validate_renderable`; their comment is narrowed because they
may still degrade an unsupported formatter error to an empty string.

## Tests

- root and nested non-data diagnostics, including `unknown` and `type`
- runtime-minted nested class definitions
- skipped non-data and open-interface fields
- generic-class rejection before provider I/O
- generic-function concrete-output positive control
- existing E0159/E0161/E0162 precedence coverage
- full pinned gate: 3,761/3,761 passed, 24 skipped; doctests clean; zero
unreferenced or pending snapshots

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **Bug Fixes**
* LLM output schemas with unsupported non-data types now return
catchable `E0164` diagnostics instead of crashing or being silently
ignored.
* Improved validation across nested types, unions, aliases, generic
classes, and skipped fields.
* Rendering now reports clear errors for unsupported types, including
`never`, `uint8array`, and `type`.
  * Valid data-class fields continue to render successfully.

* **Documentation**
  * Added changelog coverage for the updated diagnostic behavior.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
Assets 2
Loading