[BUG] Malformed tool-call emission (stray "court" + bare <invoke>/<parameter>) in long tool-dense sessions; rendered as prose, no validation error, self-reinforcing

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?

In a long, tool-call-dense session (hundreds of Read/Edit/Bash calls, large
accumulated context), the assistant intermittently emitted literal text imitating
a tool call instead of a valid invocation: a stray token court followed by a
bare <invoke name="..."> / <parameter ...> block — missing the proper
function-calls wrapper/namespace.

The harness rendered this as plain assistant text: no tool executed, and no
validation error was returned to the model, so it got no corrective signal. The
failure was self-reinforcing — once the malformed pattern was in context (and
when it was quoted back), it repeated across several consecutive turns, while
valid tool calls still succeeded intermittently in between. Not a hard capability
loss; looks like formatting drift + context-echo amplification.

What Should Happen?

Either a valid, executed tool call — or, if the model emits an invalid
invocation, the harness should detect the malformed/attempted tool call (stray
sentinel token, <invoke outside a valid function-calls block) and return a
structured "invalid tool call" error to the model so it self-corrects, instead of
passing it through as user-visible prose with no signal.

Error Messages/Logs

No error surfaced — that is part of the bug. The malformed invocation rendered as
plain assistant text, e.g.:

court
<invoke name="Read">
<parameter name="file_path">...</parameter>
</invoke>

No debug log was produced for the session (/debug reported "No log file exists yet").

Steps to Reproduce

1. Run a long Claude Code session (VSCode extension) with many sequential tool

calls (Read/Edit/Bash) and large accumulated context.

1. Observe occasional tool calls emitted as plain text: court + `<invoke ...>

<parameter ...></invoke>` instead of a real invocation.

1. No tool runs and no error appears.
2. The malformed pattern recurs over subsequent turns once it is in the

conversation (especially if quoted back), while some valid calls still succeed.

Non-deterministic; correlated with session length / tool-call density.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

2.1.168

Claude Code Version

2.1.168

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

• Self-reinforcing via context echo: malformed output (and user re-quotes of it)

act as few-shot examples that increase recurrence.

• The stray literal "court" suggests a corrupted/mis-decoded opening sentinel for

the tool-call block.

• Suggested mitigations: (1) harness-side validation of tool-call delimiters,

returning an error on a malformed invoke; (2) investigate the "court" sentinel
corruption; (3) periodic context compaction in very long sessions to reduce
drift and self-imitation.

• A full transcript excerpt can be attached on request — the malformed calls are

visible verbatim.

<img width="400" height="492" alt="Image" src="https://github.com/user-attachments/assets/876d047c-7947-48c1-80d3-86ed4a049414" />