Codex CLI 0.124.0: TUI Reasoning Controls, Amazon Bedrock & Fast Tier Default

Codex CLI 0.124.0: Reasoning Controls, Amazon Bedrock & Production Hooks

Codex CLI version 0.124.0 released on April 23, 2026, delivering three additions that matter to everyday users and enterprise teams alike, alongside a batch of reliability fixes.

TUI Reasoning Level Controls

The terminal UI now exposes keyboard shortcuts for adjusting reasoning effort on the fly. Pressing Alt+, decreases the reasoning level and Alt+. increases it — letting developers dial the trade-off between speed and thoroughness without leaving the TUI or editing config files. When switching models, the reasoning level resets to the model's default rather than carrying over a prior setting, preventing unintended behavior when context changes.

Amazon Bedrock Integration

Codex CLI gains a built-in amazon-bedrock model provider. Teams running on AWS can configure it directly in config.toml with AWS profile support and AWS SigV4 signing for credential-based authentication. This makes it straightforward to point Codex at Bedrock-hosted models — including Amazon Nova and OpenAI-compatible models hosted on AWS — without needing a proxy or custom provider shim. The release also adds model_reasoning_effort support for Bedrock-hosted models.

Fast Service Tier Default

Eligible ChatGPT plan subscribers now default to the Fast service tier unless they explicitly opt out. This reduces latency for the majority of users without requiring any configuration change, and is transparent to existing workflows.

Stable Hooks

Hooks graduate from experimental to production-ready in 0.124.0. They are now configurable directly in config.toml and can observe a broader set of operations: MCP tool calls, apply_patch edits, and long-running command sessions. This gives teams a reliable mechanism to integrate custom automation — logging, notifications, guardrails — into agent workflows without relying on command-line flags or experimental config keys.

Bug Fixes

Several reliability issues are resolved in this release:

• Cloudflare cookies now persist correctly across approved ChatGPT hosts, eliminating repeated authentication failures.
• Permission mode drift is fixed so /permissions changes survive side conversations and accurately reflect MCP approval state.
• Remote app-server reliability improves with websocket events continuing to drain under load and graceful handling of worker exits on shutdown.
• wait_agent now returns promptly when mailbox work is queued, rather than timing out unnecessarily.
• Startup failures on edge cases are reduced; unknown feature requirements now generate warnings rather than hard failures.