[BUG] Turns frequently fail to render output: leaked "call" text, "tool call could not be parsed", or empty response despite token usage

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?

Claude Code turns frequently fail to produce a valid response. Three
symptoms appear that seem to share one root cause — a turn that consumes
tokens but never renders a parseable output:

1. Leaked tool-call markup. Mid-response, the internal tool-call

(function call) markup gets corrupted, so the call is NOT executed and
instead leaks into the chat as raw text. On screen I see the word "call",
then content that is normally hidden (file paths, YAML, code fragments,
"true", etc. — different every time, since it's the parameters of
whatever tool was being called), then the output freezes.

1. Explicit parse failure. Other times the turn stops with:

"The model's tool call could not be parsed (retry also failed)."

1. Empty turn. Sometimes after I send a message the turn churns and

consumes tokens (UI shows e.g. "Churned for 51s") but no output is ever
displayed — the turn ends silently/blank.

These recur intermittently, often right after /clear and on long,
tool-heavy turns.

What Should Happen?

A turn should always end cleanly: either (a) normal visible text output,
or (b) tool calls executed silently with only their results shown. It
should never leak raw tool-call markup into the chat, never freeze
mid-stream, and never end empty after consuming tokens.

Error Messages/Logs

Parse-failure message:
The model's tool call could not be parsed (retry also failed).

Example of leaked tool-call markup shown as plain text
(the part below "call" differs every time):

call

/home/<user>/projects/koe/.github/workflows/ci.yml
      - uses: actions/checkout@v4
      ...
true

Steps to Reproduce

Note: this is intermittent; I have not found a deterministic repro. It
happens during normal use under these conditions:

1. Use Claude Code in a long session, or run /clear and immediately send

a new message.

1. Ask for something that triggers tool calls (reading/editing files).
2. Intermittently the turn either (a) prints "call" plus raw tool-call

content and freezes, (b) shows "The model's tool call could not be
parsed (retry also failed).", or (c) churns for many seconds and
returns nothing.

More likely on long conversations / large context (1M-context Opus model).

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.168 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

WSL (Windows Subsystem for Linux)

Additional Information

• Model: Opus 4.8 with the 1M context window. I have also seen these same

symptoms since Opus 4.7, so the issue may track the model version rather
than a specific Claude Code version.

• Running Claude Code inside WSL2 (Linux) on Windows.
• Frequency: all three symptoms recur, often right after /clear and on

long, tool-heavy turns. Likely correlated with large context size.