Claude Code hits local Qwen /compact failures

The interesting part here is that `/compact` is a harness feature, not just a normal chat turn, so local models have to cooperate with Claude Code’s prompt format, tool schema, and session state management.

• –Anthropic’s docs show `/compact` is an explicit context-management command with custom focus instructions, which means the failure surface includes the CLI wrapper, not just the model
• –The most common local failure mode is mismatched tool-call formatting or weak function-calling support, especially when the endpoint or template is off
• –Qwen variants differ a lot in tool-use reliability; older or poorly configured setups can look like “the model refuses,” when the real problem is the protocol bridge
• –Debugging usually requires inspecting raw request/response payloads from the server or router, because the visible Claude Code UI hides the exact failure point
• –Practically, this is a reminder that Claude Code works best with models trained and tuned for agentic tool use, not just general chat completion