Tool call corrupted ("could not be parsed") when a tool call immediately follows long-form text in the same turn

Summary

When the model produces long-form text and then makes a tool call in the same turn, the tool call is frequently corrupted, causing "The model's tool call could not be parsed" and a hung/dead response. Short text before a tool call does not trigger this.

Environment

• Model: Claude Opus 4.8 (1M context)
• Platform: Windows 11, Claude Code Desktop
• Many MCP servers connected (very large number of tool definitions)

Reproduction

• Reliably reproduces when long explanatory text is emitted immediately before a tool call in the same turn
• Does NOT reproduce with short text before a tool call
• Confirmed via repeated manual testing

Impact

• Tool call fails to execute; response hangs or dies
• Sometimes raw broken tool-call syntax leaks into visible output
• 2026-06-01: occurred 11 times in one session; recurred multiple times since

Workarounds tried

• Splitting long text and tool calls into separate turns: effective but hard to enforce consistently across sessions
• Lowering effortLevel (xhigh -> lower): no effect
• Documenting the rule in config: not reliably followed

Request

Please investigate why tool-call token generation becomes unstable specifically after long-form text output.