GPT-5.5 tool calls fail due to self-referential namespace bug in built-in tool routing
GPT-5.5 has a bug where every tool call fails with "unsupported call: exec_commandexec_command" because the model sends a self-referential namespace on built-in tool calls. The model retries indefinitely but every attempt fails identically. GPT-5.4 works fine with the same binary, repo, and sandbox settings. The root cause was confirmed via rollout logs and source trace, showing the raw function_call items contain a duplicated namespace issue.
Key quotes
Every tool call GPT-5.5 makes is rejected by the local tool router
GPT-5.5 retries indefinitely; every attempt fails identically.
gpt-5.4 works with the identical binary, repo, and sandbox settings.
error=unsupported call: exec_commandexec_command, retried until exhausted.
Root cause (confirmed via rollout logs + source trace)
From the article
You might also wanna read
The Global Namespace Risk: Universal Bucket Hijacking Technique for Cloud Data Exfiltration
The Global Namespace Risk: Universal Bucket Hijacking Technique for Cloud Data Exfiltration
Comparing Approaches to LLM Tool Descriptions: Direct Tool Calls vs. Shell Command Integration
The article discusses evolving industry practices in 2026 for structuring tool descriptions within LLM context windows. It compares two main
Reflections on debugging with AI coding agents
The author reflects on their experience using AI coding agents since late last year, noting the ironic dynamic where AI agents make mistakes
Reflections on debugging with AI coding agents
The author reflects on their experience using AI coding agents since late last year, noting the ironic dynamic where AI agents make mistakes
Build Your Own LLM From Scratch: A Hands-On GPT Training Workshop
A hands-on workshop and GitHub repository that guides users through building their own GPT training pipeline from scratch, inspired by Andre
GitHub·2mo ago
GPT-5 Released for Developers: A State-of-the-Art Coding Model
GPT-5, the latest model from the API platform, is designed for coding and agentic tasks, excelling in benchmarks like SWE-bench Verified (74
openai.com·11mo ago
A Skeptical Look at GPT5's Coding Capabilities
The article is a critique of the capabilities of large language models (LLMs), specifically GPT5, in generating functional code. The author
overengineer.dev·11mo ago
Comments
Sign in to join the conversation.
No comments yet. Be the first.