Claude Code Bug Reported: Model Fabricates User Messages and System Notifications
hidari · ghdev · 2026-08-10
A developer reported a generation bug in Claude Code on GitHub. Occasionally, after finishing its turn, the model continues to generate and fabricates the next user message, system notification, or tool feedback (e.g., a fake git push completion notice).
These fabricated blocks are then read back as genuine context in the following turn. The reporter clarifies this is not an external prompt injection or security breach, but a generation-termination failure on the model side. Developers can distinguish real from fabricated content by inspecting the local session JSONL file: fabricated text found inside a type: assistant entry is generated by the model, whereas genuine user messages and notifications are separate type: user entries.
More from coding & agent
- DIRECT: retrieval-based next-question suggestions for conversational agents — burkov · 2026-10-16
- Claude Code and Codex now autocomplete the user's next action — tokenbender · 2026-10-16
- Hermes Agent ships official Reverse Engineer Anything plugin with one-command setup — Teknium · 2026-10-16
- MLExpert: 35B MoE Agent Trained to Autonomously Handle Machine Learning Engineering — Shangheng Du · 2026-10-16
- Cowork-Gym: Alibaba Synthesizes 64.9K Agent Work Environments, Boosting Qwen by 8.7 Points — Anni Zou · 2026-10-16
- MMSearch-OPD: Joint RL and On-Policy Distillation for Long-Horizon Multimodal Search Agents — PekingUniversity · 2026-10-16