Bug: Tool call intermittently emitted in a malformed envelope ("court" + bare <invoke>), so it never executes

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Bug: Tool call intermittently emitted in a malformed envelope ("court" + bare <invoke>), so it never executes

Summary

Intermittently, instead of a well-formed tool/function call, the assistant emits a
malformed envelope: a stray literal token court followed by a bare
<invoke name="...">...</invoke> block that is missing the function_calls
wrapper and the antml: namespace prefix. The harness parser does not recognize
it as a tool call, renders it as literal text, and nothing executes — the turn
appears to hang / "crash".

Expected

The tool call is emitted in the canonical structured format and executed.

Actual

Output contains literal court + an unnamespaced/unwrapped <invoke>; no tool runs.
Re-sending sometimes works (the same session also produced many correct calls).

Likely root cause (observed, internals not visible)

Corruption of the tool-call control-token serialization:

1. antml: namespace prefix dropped (antml:invoke → invoke)
2. function_calls wrapper omitted
3. a spurious literal court inserted at the opening-tag position

Unrelated to command content / Docker / DB / project — purely the call envelope.

Repro conditions (correlated)

• Long conversation / large context
• Many tool calls in quick succession
• Right after the assistant described tool-call syntax in prose

Impact

Work stalls; user must resend. No data/code corruption (output-format only).

Suggested mitigations

• Model: validate/regenerate the tool-call envelope; self-repair malformed ones
• Harness: detect near-miss envelopes (e.g. court + bare <invoke>) and either

auto-correct or return an explicit "malformed tool call" error instead of
silently passing it through as text

Environment

• Claude Code (VS Code extension), model: Opus 4.8 (1M context)

What Should Happen?

The tool/function call should be emitted in the correct envelope (proper function_calls / invoke with the antml: namespace) and executed by the harness. Instead it is sometimes rendered as literal text — a stray court token followed by a bare <invoke> that is missing the wrapper and namespace — so nothing runs.

Error Messages/Logs

Steps to Reproduce

Note: this is intermittent/probabilistic — no deterministic minimal repro found.

1. Use Claude Code (VS Code extension) in a long session with many tool calls.
2. Have the assistant make several tool calls in quick succession (shell commands, repeated file edits, etc.).
3. Occasionally, instead of executing a tool call, the reply prints it as literal text:

a stray court token followed by a bare <invoke name="...">...</invoke>
(missing the function_calls wrapper and the antml: namespace).

1. The tool does not run and the turn appears to stall. Re-sending sometimes yields a correct call.

Observed correlates: long context; many rapid tool calls; right after the assistant
described tool-call syntax in prose. Model: Opus 4.8 (1M context). Claude Code 2.1.177.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.177

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

VS Code integrated terminal

Additional Information

_No response_