Bug Fix·Pushed May 1, 2026·S

Structured output now works with reasoning models

Gitpulse's AI summarizer now handles reasoning models like DeepSeek-R1 and MiniMax, which prefix responses with thinking traces. A robust JSON extractor strips these blocks automatically, enabling structured output across more model types.

Some AI models think out loud before answering. Models like DeepSeek-R1 and MiniMax prefix their responses with blocks containing their reasoning process — but this broke gitpulse's structured output pipeline. The previous implementation relied on LangChain's withStructuredOutput, which uses tool-calling under the hood. When the parser received a response starting with raw thinking text instead of a clean tool-call, it failed. A custom JSON extractor now handles the heavy lifting. It strips blocks, removes markdown code fences, and narrows to the outermost {} even when prose wraps the JSON. The same Zod schema validation happens afterward, so output integrity is preserved. The summarizer now works across reasoning and non-reasoning models — no configuration changes required.

Technical description

## Changes to [[code ref=1]]action/src/llm.ts[[/code]] The core issue: reasoning models (DeepSeek-R1, MiniMax M2.x) prefix responses with thinking traces. The previous approach used LangChain's [[code ref=2]]StorySchema[[/code]] with [[code]]withStructuredOutput[[/code]], which relies on tool-calling. When these models include raw text before any structured response, the parser fails. ### The Fix Replaced tool-calling with a custom extraction pipeline: ````typescript file=action/src/llm.ts // Before: structured output via tool-calling const structured = llm.withStructuredOutput(StorySchema, { name: 'gitpulse_story' }); const result = await structured.invoke([]); // After: plain chat + manual extraction const response = await llm.invoke([]); const raw = typeof response.content === 'string' ? response.content : extractText(response.content); const json = extractJson(raw); const parsed = JSON.parse(json); return StorySchema.parse(parsed); ```` ### JSON Extraction The [[code ref=3]]extractJson[[/code]] function handles three cases in sequence: 1. **Strip blocks**: Uses regex [[code]]/[\s\S]*?<\/think>/gi[[/code]] to remove thinking traces 2. **Extract from code fences**: Matches ```json blocks and extracts the content 3. **Narrow to JSON**: Finds the outermost [[code]]{}[[/code]] using indexOf/lastIndexOf This approach is resilient — it works whether the model outputs clean JSON, JSON in fences, or JSON wrapped in explanatory prose. ### Schema Update The Zod schema was simplified to remove inline descriptions (now in the system prompt instead), but validation rules remain identical. ### Files at a Glance - [[code]]action/src/llm.ts[[/code]] — LLM integration with reasoning model support