Repeated pre-tool-call filler text (recurring stray "court" token) accumulates in long sessions, degrading throughput

Environment:

• Claude Code — version unknown at time of filing (will add claude --version in a comment)
• OS: Windows 11
• Shell: Git Bash + PowerShell
• Session type: very long single session (many hours, hundreds of tool calls, multi-repo work)

Summary:
In a long-running session, the assistant began emitting large runs of repeated filler text immediately before tool calls — dozens of short lines such as "Issuing now", "Let me issue the call", "Now", and a recurring stray token that renders as the literal word "court" — before finally emitting the actual tool call. The tool calls themselves executed correctly every time; no tool result was wrong or corrupted. The problem is purely the volume of pre-call output, which wasted a large amount of the session and slowed every action.

Observed characteristics (reproduction signals):

1. Worsens with session length. Rare early in the session; by late in a multi-hour session it preceded nearly every tool call.
2. Correlates with specific tool types. Most severe before Playwright/browser MCP calls, git push, and some Bash commands. Less frequent before lighter calls (TodoWrite, Read, Grep).
3. A specific recurring artifact: a stray token that prints as "court" (sometimes on its own line) repeatedly appears as a prefix before the intended tool call.
4. Results were never affected — once the tool call was finally emitted it ran correctly, so this appears to be an output-generation/decoding pattern rather than a tool-execution failure.
5. Restarting the session was the effective workaround.

Expected behavior:
The assistant should emit the tool call directly (optionally with a brief one-line rationale), not dozens of repeated filler lines before it.

Impact:
Severe throughput loss and wasted tokens in long sessions; each tool call became expensive. Did not cause incorrect results, but made sustained multi-step work very slow.

Root cause:
Undetermined — not speculating. Could be context-length-related, a decoding/sampling artifact, or harness-side. Reporting the observable symptom plus the reproduction signals above so it can be diagnosed.

Repro difficulty:
Emerged organically in a very long, tool-heavy session; no minimal deterministic repro. The strongest signals are session length, the specific tool types above, and the recurring "court" token.