Choose Codex planning and reasoning around the task

Give Codex a complete outcome and room to investigate it. A small bug with a clear reproduction can proceed directly to a focused fix. A change with unsettled interfaces needs a concrete plan before implementation.

Planning effort and model reasoning settings are separate choices. One decides when the human settles the design; the other changes a model's supported reasoning configuration.

Plan when a decision needs to be made

For a search-field focus bug, the completion condition might already be clear: typing and deleting a phrase should keep focus in the same field. Ask Codex to locate the cause, reproduce it, fix it, and verify keyboard behavior.

For replacing the search state model across several components, first request:

Map how queries and results move through the components. Propose the state boundary and compatibility approach. Include affected files, alternatives, migration steps, and checks. Identify decisions that need my input before implementing the change.

Use the host's Plan mode when available; Codex's documented interactive workflow includes /plan. Approve a plan after it names actual code and decisions. A list of generic phases is not ready to guide implementation.

Inspect the actual reasoning control

Choose among the levels exposed for the selected model and host. Record the selected model and effort with a comparison run. If the host does not expose a control, say so; a request to "think harder" does not establish a configuration change.

Start with a setting suited to the task's uncertainty. Try a higher available level when reasoning about several interacting constraints is the obstacle. If the problem is missing source files or a broken test environment, fix that input problem first.

Compare the result on the same task before making a team default. Include elapsed time, corrections, and verification quality. A longer response is not a measure of a better change.

Keep a coherent implementation task

After the design is settled, ask Codex to carry the change through its required checks. Keep progress in a short task record when the work is long. Durable conventions belong in AGENTS.md or a skill.

Codex supports context compaction during ongoing work. Continue when the task and evidence remain coherent. Start a new conversation when the outcome or context has materially changed, carrying a checked handoff. Avoid a universal rule to restart after a fixed number of turns.

What goes wrong

Planning becomes repeated permission requests for routine work already in scope. An effort label is inferred from a prompt. The agent writes tests that exercise its implementation while missing the original failure. A successful build is mistaken for proof of correct browser behavior.

Name the decisions, inspect the actual controls, and keep the original acceptance condition visible throughout the work.

How to check

The final result should show the changed files, reproduction, actual commands, their outcomes, and unresolved limits. For the focus bug, try typing, deleting, and moving through results with a keyboard.

For a reasoning-setting comparison, run the same task with the same relevant inputs. Score the result before reading which setting produced it when practical. Keep failed runs and human review time in the comparison.