Qwen3 failure modes: text tool calls + infinite loop

[cboettig]

Summary

Two distinct qwen3 failure modes observed on 2026-04-05, traced to the geo-agent framework layer. Cross-repo issues filed:

• Handle text-format tool calls from qwen3 (XML <tool_call> as content) geo-agent#121 — qwen3 emits <tool_call> XML as plain text; framework doesn't parse it; agent terminates with garbage output
• Agent loop doesn't terminate: qwen3 makes 12+ tool calls with no summary geo-agent#122 — qwen3 loops indefinitely (12+ tool calls, 18K prompt tokens) with no termination/summary; also observed path hallucination (datadata_0.parquet instead of data_0.parquet)

App-level workaround (optional)

Until the framework fixes land, the system-prompt.md could add guidance like: "Once you have sufficient data to answer the user's question, stop querying and write a clear summary." This is a Layer 4 workaround for what is fundamentally a Layer 3 (framework) issue.

Default model

Changed default model to minimax-m2 in k8s/configmap.yaml as a mitigation — minimax-m2 handles multi-step spatial joins correctly with proper termination.