OpenAI warns prefix edits silently drop agent reasoning, ships detection tools
OpenAIDevs · x · 2026-11-20
OpenAI's developer account advised that when using GPT-6.2 Astra or Sol in the Responses API, developers should keep conversation history unchanged and append new turns rather than editing earlier messages, so agents preserve reasoning across turns and build on previous work.
Adjusting reasoning effort or adding tools can unintentionally cause prefix edits. To help detect this, OpenAI shipped the inputtransformations object in the Responses API and a new Reasoning removals section in the Usage dashboard.
Related event: OpenAI Responses API Lets Agents Reuse Reasoning History to Cut Tokens(3 posts)→
More from coding & agent
- Nous Research partners with Microsoft to bring Hermes Enterprise Agent into Teams and Azure — NousResearch · 2026-11-20
- Vercel CEO: outcome-oriented engineers are now 1000x, and coding is AI's #1 killer app — tlakomy · 2026-11-20
- AssemblyAI Adds Circuit Breaker to LLM Gateway: Skip Degraded Models Before Your Request Fails — AssemblyAI · 2026-11-20
- LangChain Celebrates Fundraise With Mini Launch Week: Trace Governance, Tool Gateway, LangSmith for Coding Agents — hwchase17 · 2026-11-20
- Techtree Packages Skills, Evals and Training Envs to Provably Upgrade Business Agents — seanwbren · 2026-11-20
- Open-source F1 MCP Server: telemetry, tire wear and Monte Carlo strategy — modelcontextprotocol · 2026-11-20