fix: strip markdown code block wrappers before JSON parsing

[MaxwellCalkin]

Summary

Fixes #959

After the v1.12.0 migration from LangChain to Vercel AI SDK, Claude models (and potentially others) fail with SyntaxError: Unexpected token when accessed via OpenAI-compatible proxies like LiteLLM or OpenRouter. The root cause is that these models wrap JSON output in markdown code blocks (```json ... ```), and JSON.parse cannot handle the wrapper.

Changes

• Added stripCodeBlockWrappers() helper function to both openaiLLM.ts and ollamaLLM.ts
• Applied the function before repairJson/JSON.parse in generateObject() (both providers)
• Applied the function before parse() in streamObject() final chunk handling (OpenAI provider)
• The regex handles: ```json ... ```, ``` ... ```, and surrounding whitespace/newlines
• Since Anthropic, Gemini, Groq, Lemonade, and LM Studio providers all extend OpenAILLM, the fix covers all OpenAI-compatible providers automatically

Why not just rely on repairJson({ extractJson: true })?

While repairJson with extractJson: true is already used, it doesn't reliably handle the markdown code block wrapper pattern (```json\n{...}\n```), as evidenced by the bug report. Adding explicit stripping before repair ensures robustness.

Test plan

• Verify Claude models (e.g. claude-haiku-4.5) work via LiteLLM/OpenRouter without JSON parse errors
• Verify models that return raw JSON (e.g. Gemini, GPT) continue to work normally
• Verify Ollama models continue to work normally

Note: This PR was authored by Claude (AI), operated by @MaxwellCalkin.

---

Summary by cubic

Fixes JSON parsing failures by stripping markdown code block fences from model responses before parsing. Prevents SyntaxError when Claude and other OpenAI-compatible providers return fenced JSON via proxies like LiteLLM or OpenRouter.

• Bug Fixes
  • Added stripCodeBlockWrappers() in openaiLLM.ts and ollamaLLM.ts.
  • Applied before repairJson/JSON.parse in generateObject (both providers).
  • Applied before parse for final text chunk in streamObject (OpenAI provider).
  • Handles json ... and generic ... with optional whitespace.
  • Covers providers extending OpenAILLM (e.g., Anthropic, Gemini, Groq, Lemonade, LM Studio).
  • Continues to use @toolsycc/json-repair with extractJson: true for robustness.

^{Written for commit 49a5a74. Summary will update on new commits.}

[cubic-dev-ai]

1 issue 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/openai/openaiLLM.ts">

<violation number="1" location="src/lib/models/providers/openai/openaiLLM.ts:30">
P2: Code-block stripping misses leading whitespace/newline before opening fence, so markdown-wrapped JSON can still fail to parse.</violation>
</file>

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

[cubic-dev-ai]

P2: Code-block stripping misses leading whitespace/newline before opening fence, so markdown-wrapped JSON can still fail to parse.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At src/lib/models/providers/openai/openaiLLM.ts, line 30:

<comment>Code-block stripping misses leading whitespace/newline before opening fence, so markdown-wrapped JSON can still fail to parse.</comment>

<file context>
@@ -20,6 +20,17 @@ import {
+ */
+function stripCodeBlockWrappers(text: string): string {
+  return text
+    .replace(/^```(?:json)?\s*\n?/i, '')
+    .replace(/\n?```\s*$/i, '');
+}
</file context>

Suggested change

	.replace(/^```(?:json)?\s*\n?/i, '')
	.replace(/^\s*```(?:json)?\s*\n?/i, '')