Counting operations stopped by Claude Code's auto mode from conversation logs

Agents & DevPublished: Source: Zenn (Claude Code)

#Anthropic #Claude Code

Key point

Claude Code's auto mode uses another model (classifier) to check each operation and stop dangerous ones. An analysis of about one month of logs (19,952 tool calls) showed 62 stopped cases (approx. 0.3%).

Instead of proceeding with work without asking for permission, Claude Code's auto mode uses another model (classifier) to look at each operation one by one and stop dangerous ones. While users may have the impression of "being stopped often," the actual number of times cannot be understood from the screen.

Stopped operations remain in the conversation logs along with their reasons. Using a small script to analyze about one month of logs on a Windows environment with Claude Code 2.1.296, it was found that out of 19,952 tool calls, 62 cases (approx. 0.3%) were stopped. Conversation logs are located in ~/.claude/proj...

This item is a summary. The full, checkable version is at the original source:

Read the original at Zenn (Claude Code) ↗