Codex cls, and why clearing context is a real skill

What people mean by codex cls

Two different things hide behind the query codex cls. One is muscle memory from Windows: people type cls expecting a blank terminal and get nothing useful. The other, and the one worth writing about, is wanting to reset the agent's conversation so it stops dragging a failed approach forward.

For the first, the shell answer works. clear on macOS and Linux, or Ctrl+L in most terminals, tidies the display. It does not touch what the model remembers. That distinction trips up more people than it should. A clean screen with a polluted conversation is the worst of both, because you can no longer see the wrong turn that is still influencing every answer.

For the second, use the session's own clear or new-conversation command rather than a shell command. Check the in-app help for the exact word, since these commands change between releases and we would rather you read the current list than trust a page written months ago.

When to clear and start again

Clearing is not a tidiness habit. It is a recovery tool with specific triggers:

• Two consecutive fix attempts that did not move a failing test closer to passing. The model is now anchored on its own wrong theory.
• You changed the goal mid-session. The old task's assumptions are still in the window and they leak.
• The agent keeps referencing a file it edited and then reverted. Check whether the session is reasoning from stale state.
• Responses have become slower or less relevant; inspect context and service behavior before assuming a cause.
• You are moving to a different part of the codebase. Frontend conventions bleeding into backend work produces odd hybrids.

Clear the branch too

A cleared conversation with a half-edited working tree is a trap. The model no longer knows what it changed, and the files still carry the changes. Inspect git status and the diff before starting over. Save a patch or coherent checkpoint, preserve unrelated edits, and deliberately keep or revert the relevant changes. Do not use git checkout . as a generic conversation reset: it discards tracked working-tree changes.

Before clearing, take the thirty seconds to write down what you learned. The missing config file, the test that needs an env var, the module that actually owns the behaviour. Paste that into the fresh session as the opening brief. A restart that carries forward the findings but drops the failed attempts is usually the fastest path to a working diff, and it beats another twenty minutes of nudging by a wide margin.

The habit worth building

Long sessions feel productive because there is history to point at. For a first recovery trial, start a new session from the saved checkpoint and a brief containing the known failure and test command. Compare its result with the previous attempt. Treat a session like a branch. Cheap to create, cheap to throw away, never precious.