chore: TDR: unstructured review

[valdis]

TDR describing how we handle code review issue collection and reporting when the model does not support structured responses (JSON schemas)

[github-actions]

QualOps Code Quality Analysis

Status: ✅ PASSED - No issues found

Summary

• Total Issues: 0
• Critical: 0 🔴
• High: 0 🟠
• Medium: 0 🟡
• Low: 0 🟢
• Files Analyzed: 1

No issues found in the analyzed code.

📊 Full Report

View detailed report

---

Powered by QualOps

[sebastianwessel]

The document says reasoning models and many OpenAI-compatible/local models often cannot use json_schema. That was more true in 2024/early 2025, but it is too broad for June 2026.

A May 2026 cross-provider test claims 199 of 244 models passed structured output on all native endpoints tested, i.e. 82%, and 14 of 23 providers had a 100% pass rate. LiteLLM specifically passed 143 of 159 SDK compatibility tests, about 90%. This is not perfect, but it contradicts the idea that the landscape generally requires a prose fallback.

https://www.requesty.ai/blog/structured-outputs-across-llm-providers-the-compatibility-mess

Given the current movement into agentic & tool calls and so on, and the fact, that inference providers & hosts will drop old models to get capacity for more recent ones, my personal opinion is, that they become irrelevant very quickly.

Even if there is the need to support such a model, it is imo the wrong approach, to build it into the core solution. It should be on inference serve/proxy level to "fix" this - like LMStudio and Ollama do it.

https://lmstudio.ai/docs/developer/openai-compat/structured-output

Imo, it is not worth to invest time here.
Let's define clear requirements, instead to try to build a solution "which fits all".