codex-1m-context-window-setup

Run the apply mode of the codex_context_window.py script, which sets model_context_window in $CODEX_HOME/config.toml to the smaller of 1,000,000 tokens or your selected model's declared maximum. Start a new Codex session afterward, since running threads are not expanded.

Codex uses the model's catalog default context window unless model_context_window is overridden in the base config. The doctor mode reports the catalog default, the model maximum, and the current config values so you can see exactly why compaction triggers early.

No. The script reads the live model catalog and caps the request at the selected model's max_context_window, then applies the model's effective_context_window_percent for the usable session window. It refuses to guess or force 1M when the model declares a smaller limit.

No. The script changes only the top-level model_context_window and model_auto_compact_token_limit keys, preserving all other TOML content. It creates a timestamped backup, writes atomically, runs strict-config diagnostics, and restores the prior bytes if validation fails.

No. Configuration is read when a session starts, so the running thread keeps its existing window. After a successful apply, start a new CLI thread or open a new Codex Desktop thread to use the expanded context.

Run the script in verify mode. It re-reads the live model catalog and exits non-zero if the current base config no longer matches the model-aware recommendation, which can happen after Codex upgrades or default-model changes.