Safety classifier repeatedly stops replies on ordinary project work (hundreds of times over 3-4 days)

Summary

For the past 3-4 days, the safety classifier has been stopping Claude's replies in Claude Code on ordinary internal project work, hundreds of times in total. Each stop shows: "Your response above was stopped by a safety classifier — this is not a tool or API error." The work is routine and benign, and much of the paid usage is being spent on stopped replies instead of finished work.

Environment

• Claude Code, VS Code extension, Windows 11
• Model: Opus 5.5
• Example session ID: 3008e9e8-0149-450e-bd89-4e8713126a5b (2026-10-14)

Example (one session today)

Task: building an animated YouTube explainer about a public company's quarterly earnings results (public financial figures, explicitly "not investment advice"). Replies were stopped during completely routine steps:

• reading the project's own markdown and source files
• writing a Python file containing the narration text
• running a small Python script that computes four percentages
• running /debug to ask why the classifier was firing

After each stop, the instruction is not to repeat the withheld content, but nothing says what triggered it, so there is no way to adjust the work and avoid it.

Impact

• Normal internal projects can't be completed reliably.
• Paid usage is consumed by replies that are then withheld.
• No local log or setting explains the stops, and /debug shows nothing.

Requests

1. Review the classifier's false-positive rate for ordinary Claude Code development work.
2. Give users some reason or category when a reply is stopped, so they can work around it.
3. Credit usage lost to false-positive stops.