Token loop/degeneration: model outputs infinite repeated tokens (spaces, 'and', closing parens)

Bug Description

During a long coding session using Claude Code (VS Code extension), the model repeatedly enters a degeneration loop where it outputs thousands of identical tokens instead of meaningful content.

Symptoms

• Output degenerates into infinite repeated characters: spaces/newlines, and, closing parentheses ), or other single tokens
• Happens mid-response, sometimes after a valid tool call result is returned
• Requires manual interruption (Escape) to stop
• Occurred 5+ times in a single session
• The model appears unaware it is looping — no self-correction

Environment

• Claude Code VS Code extension
• Model: claude-opus-4-6
• Platform: Windows 11 Pro 10.0.26100
• Session: long conversation with multiple tool calls (file reads, edits, bash commands)

Reproduction

1. Start a long coding session with many tool calls (file reads, edits, grep, bash)
2. Continue working for 20+ exchanges
3. At some point, the model response degenerates into repeated tokens

Observed Patterns

• More likely to occur as the conversation gets longer (context pressure?)
• Sometimes triggered right after reading a tool result
• The repeated token varies: sometimes whitespace, sometimes words like and, sometimes )
• Exit code and tool results before the degeneration are normal — the bug is purely in text generation

Impact

• Wastes significant tokens/cost (thousands of repeated tokens before user can interrupt)
• Breaks flow — user must interrupt and retry
• In severe cases, happens multiple times in quick succession making the session unusable

Workaround

• Interrupt immediately when loop starts
• Start a new session if it happens repeatedly (suggests context-length correlation)