[BUG]

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?

Title: Tool calls leak as literal <invoke> text in long sessions; desktop hard-fails with "tool call could not be parsed"

Environment

• OS: Windows 11 Pro (build 10.0.26200)
• Claude Code CLI: 2.1.153
• VS Code extension: anthropic.claude-code 2.1.153 (win32-x64)
• Also reproduced in the Claude Code desktop app
• Account: official paid subscription (not a proxy / third-party endpoint)
• Model shown: Opus 4.7 (1M context), effort "Extra high", Fast mode
• Config: ~6 remote MCP servers + many skills/plugins loaded globally

Summary
In long-running agentic coding sessions, the model stops emitting structured
tool_use blocks and instead prints the tool call as literal text, e.g.:

<invoke name="Edit">
<parameter name="file_path">...</parameter>
<parameter name="old_string">...</parameter>
...
</invoke>

A short literal token ("court" or "call") precedes each leaked block — looks
like a render artifact of a stripped <function_calls> wrapper.

Behavior differs by surface (same 2.1.153):

• Desktop app: HARD FAILS. Shows "An API error occurred / The model's tool

call could not be parsed (retry also failed)." and the turn breaks.

• VS Code extension: the parser usually RECOVERS — the tool still executes

(e.g. TodoWrite todos render as completed) and the turn continues, but the
raw <invoke> text is still shown inline.

Trigger
Strongly correlated with LONG conversation context (multi-phase runs:
build → repackage EXE → verify, all in one thread; large tool outputs; big
file reads). Starting a fresh conversation (/clear or new window) reliably
makes tool calls work again. Once one text-shaped tool call lands in history,
subsequent turns imitate it and degrade further (pressing "Continue"/"Retry"
makes it worse — later calls become degenerate, e.g. <invoke name="Edit">
<parameter name="file_path">test</parameter></invoke>).

Ruled out

• NOT model strength: reproduced on Opus 4.7 at Extra high effort.
• NOT a fake/proxied model: official subscription, no ANTHROPIC_BASE_URL,

no router installed; the CLI on the same machine/version works fine.

What Should Happen?

Title: Tool calls leak as literal <invoke> text in long sessions; desktop hard-fails with "tool call could not be parsed"

Environment

• OS: Windows 11 Pro (build 10.0.26200)
• Claude Code CLI: 2.1.153
• VS Code extension: anthropic.claude-code 2.1.153 (win32-x64)
• Also reproduced in the Claude Code desktop app
• Account: official paid subscription (not a proxy / third-party endpoint)
• Model shown: Opus 4.7 (1M context), effort "Extra high", Fast mode
• Config: ~6 remote MCP servers + many skills/plugins loaded globally

Summary
In long-running agentic coding sessions, the model stops emitting structured
tool_use blocks and instead prints the tool call as literal text, e.g.:

<invoke name="Edit">
<parameter name="file_path">...</parameter>
<parameter name="old_string">...</parameter>
...
</invoke>

A short literal token ("court" or "call") precedes each leaked block — looks
like a render artifact of a stripped <function_calls> wrapper.

Behavior differs by surface (same 2.1.153):

• Desktop app: HARD FAILS. Shows "An API error occurred / The model's tool

call could not be parsed (retry also failed)." and the turn breaks.

• VS Code extension: the parser usually RECOVERS — the tool still executes

(e.g. TodoWrite todos render as completed) and the turn continues, but the
raw <invoke> text is still shown inline.

Trigger
Strongly correlated with LONG conversation context (multi-phase runs:
build → repackage EXE → verify, all in one thread; large tool outputs; big
file reads). Starting a fresh conversation (/clear or new window) reliably
makes tool calls work again. Once one text-shaped tool call lands in history,
subsequent turns imitate it and degrade further (pressing "Continue"/"Retry"
makes it worse — later calls become degenerate, e.g. <invoke name="Edit">
<parameter name="file_path">test</parameter></invoke>).

Ruled out

• NOT model strength: reproduced on Opus 4.7 at Extra high effort.
• NOT a fake/proxied model: official subscription, no ANTHROPIC_BASE_URL,

no router installed; the CLI on the same machine/version works fine.

Error Messages/Logs

Steps to Reproduce

Steps to reproduce (best effort)

1. Configure several MCP servers (large tool surface) + several skills.
2. Run a long multi-step agentic task with many tool calls and large outputs.
3. After context grows large, observe tool calls being emitted as literal

<invoke>...</invoke> text instead of executing.

Expected: tool calls are always parsed/executed as structured tool_use.
Actual: under long context they leak as text; desktop errors with
"tool call could not be parsed (retry also failed)".

Impact
Breaks long agentic sessions on desktop; forces frequent /clear. VS Code is
usable but noisy.

Possible mitigations on client side: more robust tool-call parsing /
recovery (match VS Code's tolerant behavior on desktop), and/or guard against
the text-tool-call feedback loop once one leak enters history.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.153

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

_No response_