claude-code - 💡(How to fix) Fix Tool calls intermittently emitted as literal text (stray 'court' + raw <invoke>) instead of executing — mostly Edit/Read (Opus, CLI)

Summary

In a long Claude Code CLI session (Opus, large-context), the agent intermittently emitted tool calls as literal text in the transcript instead of executing them. The leaked text is prefixed with a stray court token followed by the raw function-call XML, e.g.:

court
<invoke name="Edit">
<parameter name="file_path">/path/to/file</parameter>
<parameter name="old_string">...</parameter>
<parameter name="new_string">...</parameter>
</invoke>

Instead of the Edit (or Read) being dispatched, the <invoke>…</invoke> block rendered verbatim in chat. The stray leading court token is the consistent signature.

Frequency / pattern

• Happened ~6+ times across one session.
• Almost exclusively on Edit and Read tool calls. Bash calls were unaffected.
• Sometimes the underlying tool still executed despite the leaked text; other times the turn was wasted and the call had to be re-issued.

Impact

• Heavy transcript noise and operator confusion.
• Wasted turns re-issuing calls.
• Looks like the agent is malfunctioning mid-task.

Environment

• Surface: Claude Code CLI
• Model: Opus (large-context)
• OS: macOS

Notes

The court-prefixed raw <invoke> strongly suggests a function-call serialization/tokenization corruption rather than a tool-side error (the tools themselves work when the call is well-formed). Filing for visibility; happy to provide more redacted transcript context if useful.