Codex Fork: Auto Next-Turn Closed Loop

This repository is a custom Codex fork focused on one enhancement:

• after a turn completes, run an intelligent next-step decision;
• if the result is safely continuable, inject the next turn automatically;
• extend execution from single-turn completion to a longer closed loop.

What Changed

• Added notify_next_turn hook directive parsing and next-turn queueing in core.
• Added per-session sidecar startup via notify_next_turn_service.
• Added in-session warning visibility for decision reasoning in TUI/CLI.
• Added local scripts:
  • scripts/autonext_hook.py
  • scripts/autonext_service.py
• Added optional local turn logging (out.1) controlled by env switch, off by default.

Runtime Flow

1. Codex finishes one turn.
2. notify_next_turn hook is triggered.
3. Hook sends payload to local sidecar decision service.
4. Sidecar calls configured model/provider and returns JSON decision.
5. If need_next_turn=true, Codex injects a new user turn in the same session.

Build And Run

cd codex-rs
cargo build -p codex-cli --release
./target/release/codex

Config

Set in ~/.codex/config.toml:

notify_next_turn = ["python3", "/ABSOLUTE/PATH/TO/codex/scripts/autonext_hook.py"]
notify_next_turn_service = ["python3", "/ABSOLUTE/PATH/TO/codex/scripts/autonext_service.py"]

Local Logging (out.1)

out.1 writing is disabled by default.

Enable it only when needed:

export next_turn_local_log=1
# or
export NEXT_TURN_LOCAL_LOG=1

Optional output filename/path:

export AUTO_NEXT_OUTPUT_FILE=out.1

When local logging is enabled, hook writes:

• last assistant output
• decision trace line ([auto-next] ...)

Decision Model Source

By default, decision service reads the same provider/model settings from your Codex config and auth. You can still override provider/model by env vars (see docs/auto_next_turn.md).

Docs

• docs/auto_next_turn.md
• docs/install.md
• docs/contributing.md

This repository is licensed under the Apache-2.0 License.