Claude Code auto mode blocks routine in-repo edit with no reason, then allows the identical command on retry
junhyeokchoi-prothentia · ghdev · 2026-09-29
A Claude Code user (VS Code extension, auto mode, Windows 11) had the agent run a routine, explicitly requested edit to a tracked file in their own repo — a single Bash command with no network calls, deletions, or git operations. The server-side auto mode classifier denied it with only "judged this action dangerous" and no explanation. When the user asked for a retry, the byte-identical command executed without any prompt. The issue highlights two problems: classifier decisions are non-deterministic, and denials carry no reason, leaving neither the user nor the agent able to tell what triggered the block. The author argues routine working-tree edits should never be flagged and every denial should state its rationale.
More from coding & agent
- AsideAI cuts compaction/dreaming token use 7x, doubles cache hit rate — garrytan · 2026-09-29
- PageIndex hits 36.7k stars: vectorless, reasoning-based RAG document indexing — VectifyAI · 2026-09-29
- dbx: a 25MB database client for 100+ databases with built-in AI and MCP hits 21.6k stars — t8y2 · 2026-09-29
- One env var change broke ticket tagging: GLM 5.3 no longer accepts thinking off — sweetcake_1530 · 2026-09-29
- Nobody knows what their production AI agents are doing: monitoring remains a blind spot — Standard-Procedure73 · 2026-09-29
- Stripe launches MCP server in public preview, letting AI agents manage payments end-to-end — jeff_weinstein · 2026-09-29