Codex CLI 0.123.0: workflows that hold up

The situation

Codex CLI 0.123.0 is a workflow release. The changelog points to a tighter loop around instruction discovery, MCP visibility, headless execution, and reviewable output. That matters most for teams using Codex on real repositories, where the question is usually whether it can follow repo rules, verify the change, and leave a diff a reviewer can trust.

The release adds a built-in Amazon Bedrock model provider, more explicit MCP diagnostics, broader .mcp.json loading, and fixes that reduce friction in the terminal. It also makes code review skills clearer and fixes transcript and follow-up behavior that could confuse a run.

If you already run Codex in production codebases, this version is a good time to tighten the repo contract: what Codex reads, what it can connect to, how it verifies, and what reviewers should inspect before merging. That is the practical center of Codex CLI workflows.

Walkthrough

Start with the repo contract. Codex’s instruction chain is built around AGENTS.md, with nested scope and override files for local exceptions. In practice, the first thing to tighten is not a prompt, but a short instruction file that tells Codex how this repository wants changes made.

---
description: Repo rules for Codex CLI work
---

# AGENTS.md

- Prefer small, reviewable diffs.
- Run the project’s test or verification command after each code change.
- Do not edit generated files unless the task explicitly requires it.
- If a nested directory has its own AGENTS.md, follow the nearest file first.

Next, make the verification loop explicit. Codex CLI is most useful when it can move from edit to check to review without guessing. A practical loop is: inspect the task, edit the smallest relevant surface, run the repo’s test or lint command, then summarize what changed and what was verified. The changelog’s fix for queued follow-up text while a manual shell command is running is a reminder to keep that loop predictable when humans interleave commands.

For teams using MCP, 0.123.0 improves the debugging path. The new /mcp verbose command is for full diagnostics, including servers, resources, and templates, while the plain command stays fast. That split is useful in production because it keeps the default path light and gives you a deeper view when you are checking connector scope, missing resources, or unexpected tool behavior.

{
  "mcpServers": {
    "docs": {
      "command": "node",
      "args": ["./tools/mcp-docs.js"]
    }
  }
}

The changelog also says plugin MCP loading now accepts both mcpServers and top-level server maps in .mcp.json. That reduces migration friction when teams inherit older config shapes. The governance habit is simple: treat connector config as reviewable code. Check what each server can reach, whether the scope is least-privilege, and whether the repo has a clear owner for connector changes.

The release’s headless path matters too. codex exec now inherits root-level shared flags such as sandbox and model options, which makes automation less surprising when you move from interactive use to CI or scripted runs. That is the right place to enforce the same rules every time: sandbox boundaries, approval mode, and a verification command that produces a diff a reviewer can inspect.

A practical starter checklist:

• Add or tighten the root AGENTS.md.
• Mirror any special-case rules in nested AGENTS.md or AGENTS.override.md files.
• Keep .mcp.json small and reviewable.
• Use /mcp verbose only when diagnosing connector behavior.
• Run codex exec with the same sandbox and model defaults your team expects in production.
• Require a verification step before asking for review.

One methodology note: this is a Build step, not a prompt-writing exercise. The useful artifact is the repo rule file and the verification loop around it, not a longer instruction paragraph. That keeps the change auditable and easier to review.

Tradeoffs and limits

The new Bedrock provider and broader MCP loading make Codex more flexible, but flexibility increases configuration risk. More providers and more connectors mean more places to mis-scope access, more environment variables to manage, and more chances for a task to succeed in a sandbox but fail in a real deployment context.

The improved /mcp verbose path is helpful, but verbose diagnostics can become a crutch if teams use them instead of fixing unclear connector ownership or brittle server definitions. Keep the default path fast and reserve verbose output for debugging.

The CLI fixes around transcript handling, shell snapshots, and Unicode input reduce friction, but they do not remove the need for review discipline. A clean transcript is not the same as a correct change. Reviewers should still check the diff, the verification command, and any repo-specific instructions that shaped the run.

The biggest limit is organizational, not technical: Codex works best when the team agrees on the contract. If AGENTS.md, MCP scope, and verification commands are inconsistent across repos, the CLI will still function, but the workflow will feel unpredictable.

Further reading

• https://developers.openai.com/codex/changelog
• https://developers.openai.com/codex/cli
• https://developers.openai.com/codex/guides/agents-md
• https://developers.openai.com/codex
• https://modelcontextprotocol.io/introduction
• https://developers.openai.com/codex/configuration

One methodology lens

One useful way to read this through our methodology is the Plan step: delegate first-pass decomposition and dependency mapping, review the sequencing and assumptions, and keep ownership of scope and priorities. If that split is still fuzzy, the workflow usually is too.