Dev & Engineering codex-cliclaude-code-plugincode-reviewrefactoringautomated-editingsubprocess-delegation

Codex Integration Skill (skill-codex)

Lets Claude Code invoke the OpenAI Codex CLI for code analysis, refactoring, and automated editing, with seamless session resume across turns.

FollowSkills review · FSRS-2.0
Not recommended
44/ 100 5-point scale 2.2 / 5
1 2 3 4 5 6
1Trust13 / 25 · 2.6/5

The skill defaults to a read-only sandbox and requires explicit user confirmation via AskUserQuestion before high-impact flags (--full-auto, danger-full-access, --skip-git-repo-check), showing good least-privilege and confirmation practice; README states credential prerequisites. Deductions: the example combines --full-auto --skip-git-repo-check, and step 4 says 'Always use --skip-git-repo-check', weakening sandbox boundaries; blanket 2>/dev/null stderr suppression reduces data-flow and error transparency.

2Reliability8 / 20 · 2.0/5

Abnormal paths are documented in detail: the stdin-not-closed hang symptom with the </dev/null fix, per-effort timeout tables, and stop-on-nonzero-exit with user consultation — internally consistent. Deductions: no test suite or CI evidence; the model/effort matrix (gpt-6-astra, ultra tiers) is statically unverifiable; resume flag constraints rely on the model following prose; no executable verification of failure feedback.

3Adaptability7 / 15 · 2.3/5

The description gives precise semantic triggers (running codex exec/resume or referencing OpenAI Codex) and clear scenarios (analysis, refactoring, automated editing), with defaults and fallbacks stated. Deductions: core function depends entirely on a locally installed Codex CLI plus OpenAI credentials/network reachability, which is a reachability risk for mainland-China users and is not disclosed with alternatives; no explicit non-fit boundaries.

4Convention7 / 15 · 2.3/5

Well-layered docs (README overview + detailed SKILL.md), quick-reference table, install instructions, MIT license, and error-handling section. Deductions: no version number, changelog, or known-limitations list; publisher identity unknown with no stated maintenance ownership or update path; the 2026 copyright year is questionable.

5Effectiveness5 / 15 · 1.7/5

If prerequisites hold, wrapping the Codex CLI with session-resume flow saves manual command assembly and has marginal value. Deductions: static review cannot verify output usability; 2>/dev/null silently discards errors, raising debugging cost on failure; users could run Codex directly, so value rests mostly on process encapsulation.

6Verifiability4 / 10 · 2.0/5

The repo offers auditable primary material: full command syntax, a codex --version prerequisite check, and detailed failure-symptom descriptions, with README and SKILL.md corroborating each other. Deductions: no third-party execution evidence, tests, or CI; key claims (model names, capability matrix) cannot be cross-checked, and nothing was reproduced in this static review.

Evidence confidence:Low Reviewed Sep 18, 2026 Reviewed revision 0cf3e4b6e801
Before you use it
  • Core function depends entirely on a locally installed Codex CLI and OpenAI service reachability; may be unusable from mainland-China networks with no alternative disclosed.
  • SKILL.md instructs to 'Always use --skip-git-repo-check' and stacks --full-auto by default in examples, which can bypass repository boundary protections; verify working directory contents before use.
  • Blanket 2>/dev/null suppresses stderr errors; explicitly request error output when debugging failures.
  • Model names and the capability matrix (gpt-6-astra, ultra tiers) are statically unverifiable and may not match the actual CLI version.
  • Publisher identity is unverified, with no versioning or update path; review the code and pin a version before adoption.
Review evidence [1][2][3]
See the full review method →

What does this skill do, and when should you use it?

This is an Agent Skill distributed as a Claude Code plugin that teaches Claude how to correctly run the Codex CLI (codex exec and session resumes). It codifies the fiddly details: model and reasoning-effort selection, sandbox modes, stdin redirection, stderr suppression, and background timeouts. It also embeds guidelines for critically evaluating Codex output — treat Codex as a colleague, not an authority. It fits developers who use Claude Code and Codex side by side.

When a user asks to run Codex CLI or references OpenAI Codex, Claude activates this skill: it asks the user which model (default gpt-6-astra) and reasoning effort (default high) to use via AskUserQuestion; selects a sandbox mode (default read-only); assembles and runs a codex exec command (appending 2>/dev/null to suppress thinking tokens by default, and </dev/null to prevent stdin blocking in non-interactive contexts); then summarizes the output and tells the user they can continue with "codex resume". It requires explicit user permission before using high-impact flags (--full-auto, danger-full-access, --skip-git-repo-check).

  1. A developer with both Claude Code and Codex access who wants Claude to delegate standalone code-analysis tasks to Codex
  2. A user who wants a read-only repository review with improvement suggestions but doesn't know the codex exec flag surface
  3. Automated refactoring or editing of local code, with Claude picking the sandbox and full-auto settings and summarizing results
  4. Anyone running codex exec from scripts or background tasks who has hit the stdin-never-closed hang
  5. Teams iterating across many turns with codex exec resume --last to inherit model and configuration from the original session

What are this skill's strengths and limitations?

Pros
  • Bundles hard-won codex exec details: stdin redirection to prevent hangs, stderr suppression, and per-effort execution timeouts (150s–1800s) that are hard to assemble elsewhere
  • Clean session-resume design that inherits model and configuration, well suited to iterative multi-turn workflows
  • Built-in critical-evaluation guidance — Claude is told not to defer blindly and to resume the session as an identified peer AI when disagreeing
  • Offered both as a plugin (auto-updates) and a standalone skill folder; MIT licensed with no usage friction
Limitations
  • Tightly coupled to a Claude Code-only mechanism (AskUserQuestion); other Agent Skills-compatible clients would need edits to use it
  • Read-only sandbox is the default, and the resume syntax is flagged as "no flags allowed" in the quick reference, limiting flexibility
  • The README's example workflow still shows gpt-5.3-series models and a three-tier effort scale, inconsistent with SKILL.md's gpt-6/gpt-5.6 models and six-tier scale — potentially confusing
  • No test suite or cross-platform validation is shown in the repository, and the timeout table is not independently verified

How do you install this skill?

Option 1 (recommended, plugin): in Claude Code run /plugin marketplace add skills-directory/skill-codex then /plugin install skill-codex@skill-codex. Option 2 (standalone skill): git clone --depth 1 [email protected]:skills-directory/skill-codex.git /tmp/skills-temp && mkdir -p ~/.claude/skills && cp -r /tmp/skills-temp/plugins/skill-codex/skills/codex ~/.claude/skills/codex && rm -rf /tmp/skills-temp. Prerequisite: codex CLI installed and on PATH with valid credentials, verified via codex --version.

How do you use this skill?

Trigger it in Claude Code with natural language, e.g.: "Use codex to analyze this repository and suggest improvements for my claude code skill." Claude asks which model and reasoning effort (unless already specified), picks a sandbox mode, and runs something like codex exec -m <model> --config model_reasoning_effort="high" --sandbox read-only --full-auto --skip-git-repo-check "your prompt" 2>/dev/null. To continue the Codex session later, say "codex resume"; resumes inherit the original session's model, effort, and sandbox and take no config flags by default. The source does not document how to trigger this skill outside Claude Code.

How does this skill compare with similar options?

The README explicitly points to klaudworks/ralph-meets-rex as an alternative for more autonomous agentic workflows — worth examining if you want stronger orchestration rather than manual single-prompt delegation. The source names no other competitors.

FAQ

Besides the codex CLI, is there any cost or extra setup?
You need the codex CLI installed, on PATH, and configured with valid credentials (verify with codex --version). Codex usage costs depend on your OpenAI account; the skill itself is MIT-licensed and free. Note that per SKILL.md, the ultra effort level is the slowest and most expensive — reserve it for the hardest jobs.
Will it modify files or access the network without asking?
No. The default sandbox is read-only, and before using high-impact flags (--full-auto, --sandbox danger-full-access, --skip-git-repo-check) the skill requires Claude to ask your permission via AskUserQuestion.
What happens if a command fails or appears stuck?
The skill mandates stopping and reporting whenever codex --version or codex exec exits non-zero, and asking the user before retrying. A process with zero output, zero CPU, and an indefinite hang usually means stdin wasn't closed — SKILL.md gives the fix of appending </dev/null. Also note Codex produces no intermediate output; if killed early, the output file is silently empty.
Can I trust Codex's conclusions?
The skill explicitly says to treat Codex as a colleague, not an authority: Claude should trust its own knowledge, verify disagreements with search or docs, optionally resume the session identifying itself by its real model name, and let the user decide if ambiguity remains.

Related skills