Tool Calling Leaks Into Chat Output

This looks less like “the model forgot how” and more like a serialization bug: if the tool call isn’t preserved as structured metadata, it gets flattened into chat text on the next pass.

• –Community replies point to chat template and parser mismatches, especially when using Qwen-family models with local runtimes
• –A correct tool-call flow needs separate assistant text and tool-call objects, plus a matching tool-result turn
• –If tool-call/result pairs are not replayed exactly, later generations can degrade into plain prose or raw XML tags
• –Fixes usually live in the orchestration layer: template alignment, history replay, and strict separation between narration and action
• –Once a bad turn poisons the loop, resetting conversation state often “fixes” it temporarily, which is a strong hint the bug is in state handling