Claude Code stalls on open models

This is probably not user error so much as an expectation mismatch: Claude Code is an opinionated agent stack, and the open models are being asked to behave like Claude without having the same instruction-following or tool-call discipline.

• –Claude Code is built around Anthropic’s own models and workflows, so Sonnet/Opus naturally fit its prompts, tool schema, and “do the thing” cadence better
• –Many open-source coder models are good at code generation but weaker at staying inside strict agent loops, which can show up as verbose analysis instead of action
• –If the API proxy or router is even slightly off, tool calls and role instructions can degrade fast, and the model may default to narrating instead of executing
• –The practical fix is usually model selection plus prompt/router tuning: use a coder-tuned model with solid function calling and make sure the Anthropic-compatible wrapper is preserving the full tool contract
• –Bigger picture: Claude Code is not a neutral shell for every LLM; it’s a workflow built around a specific family of models, and local swaps often expose that assumption