Extract Clean Structured Output After DeepSeek think

Because DeepSeek-R1 outputs its deliberation inside <think> tags before the final answer, instruct the model in a single zero-shot prompt to emit valid JSON immediately after the closing </think> tag so your parser can slice on </think> cleanly.

User Prompt: Analyze the dependency version conflicts in the following lockfile diff: {{DIFF}}. After your closing </think> tag, output strictly a JSON object with keys "conflicting_packages" and "resolution_commands".