Qwen 3.6 draws tool-calling complaints

This looks less like a pure “bad model” verdict and more like an agent integration failure mode: once tool calls, templates, and shell quoting get brittle, the whole workflow collapses.

• –Windows, PowerShell, and local-serving wrappers add extra chances for malformed JSON or quoting bugs to break the tool loop
• –Commenters in the broader discussion point to chat templates, `qwen3_coder` parsers, and `preserve_thinking` settings as the real pressure points
• –If a model can edit C++ more reliably than HTML/CSS, that hints the issue is often prompt/tool formatting and harness behavior, not raw coding ability
• –For coding agents, reliability beats benchmark chatter; a model that loops for minutes on file writes is a productivity hazard
• –The takeaway for local deployers is practical: test the exact serving stack, parser, and template before blaming the weights