Running controlled LLM experiments with Codex CLI: disable every tool via config
lateinteraction · x · 2026-11-04
- Researcher neuralavb shares hard-won lessons for running isolated model-behavior experiments with Codex: the API is expensive, and the Codex app exposes tools like bash that make results unshareable.
- Recommended approach: use Codex CLI with a config file that explicitly switches off every built-in tool and feature (shell, browser, image generation, multi-agent, skills, etc.), defaulting to "no tools"; add your own MCP server only when needed.
- The open-source lmunderstand-logs repo ships a complete 39-line codexconfig.toml (model = "gpt-6-luna", sandboxmode = "read-only", approvalpolicy = "never") you can copy; an attached screenshot shows a full codex exec command line with --strict-config and -c overrides.
More from coding & agent
- ClickUp Brain gets browser use: reads the news and builds a 10-slide deck in 3 minutes — testingcatalog · 2026-11-04
- ClickUp Brain browser use hands back control for logins, runs on Browserbase — testingcatalog · 2026-11-04
- ClickUp rolls out Brain² browser use for sites without official integrations — testingcatalog · 2026-11-04
- Cerebral Valley teams with Meta and Google DeepMind for SF hackathon with $11,000 in prizes — jocarrasqueira · 2026-11-04
- Meituan launches LongCat-2.5: 1.6T-param native multimodal agentic model — thisdudelikesAI · 2026-11-04
- Postman installer silently adds Codex/Claude Code plugins and tries to self-approve its hook — Future_AGI · 2026-11-04