[BUG] False positive safety block during autonomous agentic work on custom game engine

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Request ID: req_011CbKoHtVUN3qDtvfm6diBM, req_011CbKpvToZj7HroLuUPtwdZ, ...

Description:

Claude Code was working autonomously on a legitimate task - programming a custom Minecraft-clone game engine in Rust (voxel renderer, chunk system, mob AI, networking). After a long agentic session with no harmful content, the final user message was simply "pokracuj" (Czech for "continue"). This triggered a safety block.

To recover, the user asked Claude to write its progress into two continuity files (CLAUDE.md and TODO.md) and started a fresh thread. The new thread also immediately triggered the same safety block.

Exact trigger message in the new thread (written by Claude itself as continuation instruction):

"V novem vlakne: Read TODO.md + CLAUDE.md na zacatku. Pokracovat W47.10 (sky render - smallest safe) nebo W47.8 (place/break). Pending bug fixes: redstone halo per-element split, flame stroboskop (W46.8), mob AI. Save je v test3 slugu - muzes testovat save/load pipeline."

This message was generated by Claude in the previous thread as a handoff note. Claude then blocked itself on its own output in the new thread.

Attached files: CLAUDE.md, TODO.md (written by Claude itself to enable continuation)

CLAUDE.md
TODO.md

Error message:

Claude Code is unable to respond to this request, which appears to violate our Usage Policy.

Expected behavior:

No block. The task is standard game engine development in Rust with no harmful content.

Actual behavior:

Entire conversation thread blocked and killed. Reproduced in a fresh thread using only Claude-generated continuity files and a Claude-generated handoff message.

Root cause hypothesis:

The classifier is evaluating runtime-generated code during agentic sessions, not just the final output. The block likely triggered on in-progress generated code (e.g. inventory system + net/dispatch combination from W47.7 planning). Since the continuation message and context files were all written by Claude itself, the classifier is effectively blocking Claude's own output - making the block impossible to avoid or rephrase around.

Impact:

• Loss of entire in-progress agentic session
• Cannot continue work even in a new thread
• Cannot rephrase the trigger - it was Claude's own handoff note
• The continuity files and trigger message were all written by Claude itself

Steps to reproduce:

1. Start a long agentic Claude Code session on Rust game engine development (voxel renderer, chunk streaming, net dispatch, inventory system)
2. Ask Claude to summarize progress into CLAUDE.md and TODO.md
3. Open a new thread, load those files
4. Send Claude's own handoff note as the first message
5. Observe safety block

Error Messages/Logs

Request was blocked
This request triggered safety guardrails. Rephrase your prompt or rewind to continue.
View details
API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy (https://www.anthropic.com/legal/aup). Try rephrasing the request or attempting a different approach.

Request ID: req_011CbKpvToZj7HroLuUPtwdZ