Cowork/Claude Code: repeated "course" filler tokens flood inline output

Status Open
Maintainer reply None cached
Activity 3 comments · opened Jul 23, 2026

Summary: In Cowork mode desktop app, during turns with many sequential tool calls, the assistant repeatedly emits the literal word "course" many times in a row as filler between tool calls. It appears inline in the chat and floods the transcript, making it unreadable. It recurs within a session even after being reported.

Environment: Claude Cowork mode / Claude Code, desktop app. Remote cloud sandbox. Client OS Windows 11. Date 2026-07-23 JST.

Steps to reproduce: 1. Run a multi-step Cowork task with many back-to-back tool calls. 2. Watch the narration between tool calls. 3. See the word "course" repeated many times inline.

Expected: No filler between tool calls.

Actual: course course course repeated many times floods the output.

Impact: Severely degrades readability and reads as a malfunction.

Note: All personal and identifying information has been masked

View original on GitHub ↗

3 Comments

jmylot · 1 month ago

Possibly the same root cause as #60584 (now closed): that report described the literal token "court" / "<invoke>" leaking into the text stream around tool_use blocks instead of a valid tool call. This issue is the "course" variant — the bare word repeated dozens-to-hundreds of times as inline narration between back-to-back tool calls in Cowork mode. Switching the session to a different model stopped it.

euyis1019 · 23 days ago

I also encountered

jmylot · 18 days ago

Thanks for confirming you're also seeing this — could you share a couple more details when you get a chance? Specifically: which model was active in the session, whether you were in Cowork mode or plain Claude Code, and roughly how many back-to-back tool calls had run before the "course" flood started?

For context, this looks like the same failure mode reported in #60584 (closed) — a filler token ("court"/"<invoke>" there, "course" here) leaking into the narration between tool calls instead of a valid tool call. In both that thread and the earlier comment on this one, switching the session to a different model stopped the repetition, so it may be worth trying as a workaround in the meantime.