fix: strip markdown code fences from LLM JSON responses (fixes #959)

[willtwilson]

Problem

Some LLM providers (Claude, models via LiteLLM/OpenRouter) wrap JSON output in markdown code fences:

\
\\json
{ "query": "...", "sources": [...] }
\
\\

The \streamObject()\ paths in both OpenAI and Ollama providers pass accumulated text directly to \partial-json's \parse(), which fails on the fence characters. This affects all users running Claude or any model behind LiteLLM/OpenRouter that wraps JSON in markdown.

Solution

Added \stripMarkdownFences()\ and \safeParseJson()\ utilities in \src/lib/utils/parseJson.ts\ that strip markdown code fences (\\json ... \\ and \\ ... \) before parsing. Applied to:

• *\streamObject()* in both OpenAI and Ollama providers — strips fences from accumulated partial JSON text before passing to \parse()\
• *\generateText()* tool call argument parsing in the OpenAI provider

The existing \generateObject()\ paths already use
epairJson()\ with \�xtractJson: true, which handles this case, so those are left unchanged.

Testing

Tested with Claude 3.5 Sonnet via LiteLLM and OpenRouter — JSON responses with and without fences are now parsed correctly. The fix is a no-op for models that already return clean JSON.

Fixes #959

---

Summary by cubic

Strip markdown code fences from LLM JSON responses to prevent parse errors in OpenAI and Ollama streaming and OpenAI tool-call paths. Fixes #959.

• Bug Fixes
  • Added stripMarkdownFences() and safeParseJson() in src/lib/utils/parseJson.ts.
  • Applied to streamObject() for OpenAI and Ollama before calling parse from partial-json.
  • Used for OpenAI generateText() tool-call arguments; generateObject() unchanged (already handled by @toolsycc/json-repair with extractJson: true).

^{Written for commit 19f4057. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 3 files

---

Since this is your first cubic review, here's how it works:

• cubic automatically reviews your code and comments on bugs and improvements
• Teach cubic by replying to its comments. cubic learns from your replies and gets better over time
• Add one-off context when rerunning by tagging @cubic-dev-ai with guidance or docs links (including llms.txt)
• Ask questions if you need clarification on any suggestion

[Copilot]

Pull request overview

Fixes a regression where some LLM providers return JSON wrapped in markdown code fences, causing partial-json / JSON.parse failures during streaming object parsing and tool-call argument parsing.

Changes:

• Added stripMarkdownFences() and safeParseJson() utilities to sanitize fenced JSON before parsing.
• Applied fence stripping to streamObject() parsing in the OpenAI and Ollama providers.
• Applied safe JSON parsing to OpenAI generateText() tool-call argument parsing.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
src/lib/utils/parseJson.ts	Adds shared helpers to strip markdown code fences and safely parse JSON.
src/lib/models/providers/openai/openaiLLM.ts	Uses fence stripping/safe parsing for streamed object parsing and tool-call argument parsing.
src/lib/models/providers/ollama/ollamaLLM.ts	Uses fence stripping for streamed object parsing.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.