A 6.8M-token Codex run survived a five-hour pause (tectontide.com)

Codex CLI version 0.128.0, released on April 30, 2026, introduced a groundbreaking feature called persisted goals, which allows Codex to maintain a goal state through terminal restarts and long pauses. This enhancement enables a seamless continuation of tasks without the need for user re-prompting, significantly improving the efficiency of long-running AI sessions. A demonstration of this feature involved a real-world TypeScript project where Codex processed approximately 6.8 million tokens over a six-hour session, demonstrating an impressive 94% cache hit rate and achieving a final status of TASK_COMPLETE without user intervention. The significance of this update lies in its redefined interaction model between developers and AI agents. With the new runtime continuation and goal management capabilities, Codex shifts from requiring constant oversight to functioning autonomously based on well-defined contracts. Developers can now create structured prompts and success criteria beforehand, allowing Codex to navigate complex projects with minimal monitoring. This development represents a major step forward for the AI/ML community, enhancing automation in software development and raising the bar for autonomous AI capabilities while emphasizing the importance of upfront prompt engineering and context management.