claude-code - 💡(How to fix) Fix Malformed tool-call syntax leaks raw parameter fragments into the visible transcript

Summary

When the assistant emits a tool call with incorrect tag syntax (e.g. bare invoke/call instead of the namespaced antml:invoke/antml:parameter), the UI shows Your tool call was malformed and could not be parsed, but it also renders the half-parsed fragments — parameter values such as 1, 1.3, true, and the text body — to the user as plain markdown content.

Expected

Unparseable tool-call output should be suppressed or cleanly errored, not displayed to the user as assistant content.

Actual

The malformed call's internal fragments leak into the rendered transcript, e.g.:

⏺ call

  <message text here>
  1
  1.3
  true

Environment

• Claude Code, model claude-opus-4-8 (1M context)
• Platform: macOS
• Reproduced repeatedly when invoking an MCP tool (mcp__voicevox__speak) mid-response.

Notes

The trigger is model-side (the model emitting the wrong tag namespace). The reported defect is the rendering/robustness behavior: a tool call that fails to parse should not surface its raw fragments to the user.