Opus 4.8 (1M context) tool-call XML serialization fails with stray "court" token when large system context + multiple screenshots accumulate

[kaduta]

Preflight Checklist

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

What's Wrong?

When using Claude Opus 4.8 (1M context) in a project with a large constantly-loaded system context (~92KB) AND accumulating screenshots via mcp__claude-in-chrome browser automation, tool-call XML serialization degrades and produces a stray token "court" with the antml: namespace dropped from tags, causing every subsequent tool call to fail.

Error message: "Your tool call was malformed and could not be parsed. Please retry."

Reproduction conditions (all 3 required):

1. Model: claude-opus-4-8 (1M context)
2. 2. Project with large always-loaded context (~92KB: CLAUDE.md + rules + memory files)
3. 3. Browser automation session with 5+ accumulated screenshots (mcp__claude-in-chrome__computer screenshot calls)
      Workaround: Switching to Sonnet 4.6 resolves the issue immediately. /clear does NOT help.

What Should Happen?

When using Claude Opus 4.8 (1M context) in a project with a large constantly-loaded system context (~92KB: CLAUDE.md + rules + memory files) AND accumulating screenshots via mcp__claude-in-chrome browser automation, tool-call XML serialization degrades and produces a stray token "court" with the antml: namespace dropped from tags.

Error message:

Your tool call was malformed and could not be parsed. Please retry.

Malformed output example (simplified):
The model outputs court as a stray token, followed by raw <invoke> / <parameter> tags without the antml: namespace prefix, making them unparseable.

Reproduction conditions (all 3 required):

1. Model: claude-opus-4-8 (1M context variant)
2. 2. Project with large always-loaded context (~92KB system prompt: CLAUDE.md + rules + memory)
3. 3. Browser automation session with 5+ accumulated screenshots (mcp__claude-in-chrome__computer screenshot calls)
      Timeline observed:

• Screenshots 1-4: tool calls work correctly
• • Screenshot 5+: tool-call XML starts malforming, stray "court" token appears
    What does NOT cause this:
• No XML/invoke-style fragments found in project files (grep confirmed clean)
• • Not a prompt injection issue
    Workaround confirmed:
• Switching to Sonnet 4.6 resolves the issue immediately under identical conditions
• • /clear does NOT help because large project context reloads immediately

Error Messages/Logs

Your tool call was malformed and could not be parsed. Please retry.

Model output contained stray token "court" followed by raw XML tags without antml: namespace prefix (e.g., <invoke name="..."> instead of the properly namespaced form). This repeats on every retry attempt.

Steps to Reproduce

1. Open Claude Code with a project that has a large always-loaded context (~92KB: CLAUDE.md + rules/.md + memory/.md)
2. 2. Set model to claude-opus-4-8 (1M context) via /model
3. 3. Start browser automation using mcp__claude-in-chrome MCP
4. 4. Take 5+ screenshots using mcp__claude-in-chrome__computer with action: screenshot
5. 5. Attempt any tool call (left_click, navigate, etc.)
6. 6. Observe: model outputs stray token "court" + raw XML without antml: namespace
7. 7. Error: "Your tool call was malformed and could not be parsed. Please retry."
8. 8. Retrying produces the same malformed output repeatedly
      Verification: Switch to Sonnet 4.6 → resolves immediately. /clear → does NOT fix it.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Unknown — issue was first observed with claude-opus-4-8 (1M). Sonnet 4.6 does not exhibit this behavior.

Claude Code Version

2.1.181 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Context size breakdown (always-loaded on session start):

• CLAUDE.md: 6KB
• • .claude/memory/misstake.md: 48KB
• • • .claude/memory/ai_memory.md: 3.7KB
• • • • .claude/skills/SKILLS.md: 11KB
• • • • • .claude/agents/AGENTS.md: 2.6KB
• • • • • • .claude/rules/*.md: ~19KB
• • • • • • • Total: ~92KB
              The "court" stray token appears consistently at the start of every malformed tool call. Switching to Sonnet 4.6 (/model command) within the same session immediately resolves the issue without any other changes. Claude Code version: 2.1.181

[github-actions]

Found 3 possible duplicate issues:

1. Malformed tool calls: stray 'court' token + dropped antml namespace prefix on tags #68719
2. Tool calls become malformed (stray token + missing antml: namespace) late in long sessions — Opus 4.8 #64977
3. Tool call intermittently emitted as literal "count" (antml: prefix dropped, "could not be parsed") in long Opus 4.8 1M-context sessions #64684

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code