fix: strip markdown code fences before JSON parsing in structured output

[VibhorGautam]

Summary

Fixes #959

Some LLM providers (notably Claude via OpenAI-compatible APIs and certain Ollama models) wrap their structured output in markdown code fences:

```json
{"key": "value"}
```

This causes JSON.parse and partial-json to throw parse errors, breaking generateObject() and streamObject() calls with these models.

Changes

• Added a stripMarkdownFences() helper that removes ```json ... ``` wrappers before the content reaches repairJson / parse
• Applied to both OpenAI and Ollama provider implementations
• Covers both generateObject() (full response) and streamObject() (streaming partial + final)
• Other providers (Groq, Gemini, Anthropic, LMStudio, Lemonade) extend OpenAILLM so they inherit the fix automatically

Why repairJson({ extractJson: true }) isn't enough

The existing repairJson with extractJson handles some cases but doesn't reliably strip markdown fences in all edge cases — particularly when the fence includes a language tag (json) or has varying whitespace. The explicit strip before repair makes the pipeline more robust.

How to test

1. npm install && npm run dev, open http://localhost:3000
2. In settings, pick a model that's known to wrap structured output in code fences (e.g. Claude via an OpenAI-compatible proxy, or an Ollama model like deepseek-r1)
3. Send a query that triggers the search pipeline (e.g. "latest news about TypeScript") — this exercises generateObject() in the classifier
4. Confirm the search completes without a JSON parse error in the terminal logs
5. Also test with a standard OpenAI model (e.g. GPT-4o) to make sure the no-fence path still works — stripMarkdownFences should be a no-op when there are no fences

[cubic-dev-ai]

2 issues found across 2 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/lib/models/providers/ollama/ollamaLLM.ts">

<violation number="1" location="src/lib/models/providers/ollama/ollamaLLM.ts:23">
P2: Fence stripping is too strict for streamed partials and misses common fence header variants, causing repeated parse failures / empty partial outputs.</violation>
</file>

<file name="src/lib/models/providers/openai/openaiLLM.ts">

<violation number="1" location="src/lib/models/providers/openai/openaiLLM.ts:272">
P2: Streaming partial parsing still breaks for fenced JSON because fence stripping only works after a closing fence is present, so most deltas parse as invalid and fall back to `{}`.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

1 issue found across 2 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/lib/models/providers/openai/openaiLLM.ts">

<violation number="1" location="src/lib/models/providers/openai/openaiLLM.ts:32">
P2: Regex change regressed fence stripping by requiring a newline after opening backticks, so same-line fenced JSON is no longer handled.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}