How to Use Goals in Codex

Goals are a persistent objective attached to a Codex thread, allowing tasks that are hard to finish in a single prompt to continue across multiple turns.

Suitable tasks include the following.

• profiling
• patching
• benchmarking
• flaky test reproduction
• evidence-based audit

A good Goal is not a simple request but a completion contract. It should specify the outcome, verification surface, constraints, boundaries, iteration policy, and blocked stop condition together, and the result must be judged by numbers and evidence.

This usage is available starting from Codex 0.128.0. Use /goal <result> to set a goal, and control its state with /goal pause, /goal resume, and /goal clear. Stop conditions include success, pause, removal, abort, reaching the budget limit, and encountering a blocker that requires user input.

Completion is judged not by the model's guesswork but by evidence such as tests, benchmarks, logs, and file changes. Goals are saved within the scope of the thread and only continue while the thread is idle and budget remains. It's emphasized that a Goal is not boundaryless autonomous execution, but persistence under user control.

As a research reproduction example, Deep Hedging is presented. If exact replication isn't possible, partial reproduction, approximate results, and blockers should be reported separately; this isn't suited for one-line edits or single-shot questions, but for tasks that require multi-turn investigation and verification.