[Bug] Token repetition loop in constrained narration with subagent-driven-development skill

Bug Description
[Summary]
During a response, a single short token (e.g. "course") repeats hundreds of
times — a degeneration / repetition loop. It happens regardless of context
length (occurs even in short sessions) and only stops when I manually
interrupt with Esc.

[Symptoms]

• The response text collapses into "course course course ..." (same word repeated)
• Occurs specifically when generating the short narration placed right before

or between tool calls

• Does not self-terminate; the turn produces no real work until interrupted

[Key triage info]

• NOT correlated with context length (reproduces in short sessions too)
• fast mode ON/OFF: (please fill in — if ON, fast/speculative decoding is a

prime suspect)

[Common conditions observed]

• Happens while the superpowers plugin's "subagent-driven-development" skill

is active

• That skill imposes a hard constraint: "narrate at most one short line

between tool calls"

• output style = learning mode (adds extra formatting directives such as

Insight blocks)

• Hypothesis: the trigger is the combination of

"strong short-output constraint × pre-tool-call narration × learning style",
not context size

[Environment]

• Claude Code (CLI/IDE) on macOS
• Model: Claude Opus 4.8 (1M context) / claude-opus-4-8[1m]
• fast mode: (ON/OFF)
• Plugins: superpowers (subagent-driven-development skill)

Environment Info

• Platform: darwin
• Terminal: iTerm.app
• Version: 2.1.179
• Feedback ID: 8cd03112-25f2-4a1e-af4b-dcb9fb5564ed

Errors

[]