[Bug] claude-opus-4-7: Malformed tool calls with stray text prefix in extended MCP sessions

Summary

In extended sessions (150k+ context tokens) with heavy MCP tool usage, claude-opus-4-7 occasionally prefixes tool call XML with stray text tokens (observed: court\n immediately before <invoke>), causing the runtime to reject the call as malformed.

Reproduction context

• Model: claude-opus-4-7 (claude-sonnet-4-6 does NOT reproduce)
• Session length: ~150k+ tokens, multi-hour session
• Tool density: 10+ MCP tool calls in the session (Gmail, Linear, Bash, Read, etc.)
• Language: Japanese-language responses with inline tool calls
• Frequency: ~5–6 occurrences in a single session (not every call, but clustered in later turns)

Symptom

Tool calls are emitted with a stray text prefix, e.g.:

court
<invoke name="mcp__gmail__search_threads">
  <parameter name="query">...</parameter>
</invoke>

Runtime error: Your tool call was malformed and could not be parsed

Impact

• Tool call silently fails; model must retry
• In a long session this happened 5–6 times, each requiring a manual retry by the user
• Disruptive for workflows that depend on reliable MCP tool execution

Workaround

Switching to claude-sonnet-4-6 eliminates the issue entirely.

Notes

• The stray token (court) appears to be a fragment leaking from the model's internal reasoning/narration before the tool call XML
• Does NOT appear to be a prompt injection or hook issue — the stray text originates from model generation, not from injected context
• CLAUDE.md / hooks cannot structurally prevent this since the error occurs at generation time, before any hook can intercept

Environment

• Claude Code CLI (latest)
• macOS darwin 25.5.0
• MCP servers: Gmail, Linear, Bash, Read (standard tools)