agentleFS
sugarforeversugarforever/01coder-agent-skills/skills/codex-cli/SKILL.md

codex-cli

Skill139 starsChanged 9 months agoMITView on GitHub

Use when an agent needs to delegate a task to the OpenAI Codex CLI from another agent environment such as Claude Code, OpenClaw, or similar. Covers checking whether Codex CLI is installed, running one-off Codex prompts with `codex exec`, resuming sessions, collecting outputs, attaching images or files as input with `-i`/stdin, and handling Codex image generation including finding and reporting generated image file paths.

23 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use opencode-cli, antigravity-cli and 6 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • julweberSkill

    opencode-cli

    Expert control of opencode via the command line. Use when running opencode TUI, headless server, web UI, or the `run` non-interactive mode; managing sessions, providers, agents, MCP servers, models, stats, import/export, GitHub integration, debugging, database tools, upgrades, or shell completions.

    Also in 5 projects that use this

  • yu-iskwSkill

    antigravity-cli

    Executes coding, review, research, and automation tasks using Google Antigravity CLI (`agy`). Uses Antigravity's native fine-grained allow/ask/deny permissions instead of Gemini CLI approval modes.

    Also in 3 projects that use this

  • julweberSkill

    claude-cli

    Expert control of the `claude` CLI (Claude Code). Use when running Claude Code interactively or non-interactively, managing sessions, configuring MCP servers, managing plugins, handling authentication, or piping prompts through claude for scripting.

    Also in 3 projects that use this

  • ihiteshguptaSkill

    terraform

    Expert in Infrastructure as Code with Terraform. Use for creating, managing, and versioning infrastructure across cloud providers.

    Also in 3 projects that use this

  • mrgoonieSkill

    claude-code

    Also in 4 projects that use this

  • GoogleSkill

    ponytail

    Forces the laziest solution that actually works, simplest, shortest, most minimal. Channels a senior dev who has seen everything: question whether the task needs to exist at all (YAGNI), reach for the standard library before custom code, native platform features before dependencies, one line before fifty. Supports intensity levels: lite, full (default), ultra. Use on ANY coding task: writing, adding, refactoring, fixing, reviewing, or designing code, and choosing libraries or dependencies. Also use whenever the user says "ponytail", "be lazy", "lazy mode", "simplest solution", "minimal solution", "yagni", "do less", or "shortest path", or complains about over-engineering, bloat, boilerplate, or unnecessary dependencies. Do NOT use for non-coding requests (general knowledge, prose, translation, summaries, recipes).

    Also in 4 projects that use this

The file

Show its rules and text

We split this skill into the rules it gives an agent

23 rules. 0 are common, 11 are distinctive, and 12 are very specific to this project.

  • 0

    Common

    Most of the 20 edited versions we read keep these.

  • 11

    Distinctive

    This author's own choices. Worth a look.

  • 12

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Use this skill when the user asks this agent to talk to Codex, consult Codex, delegate a coding/review/research task to Codex CLI, or ask …

    line 8
  2. Distinctive:

    Before using Codex CLI, verify it is available:

    line 20
  3. Very specific:

    If codex is missing, tell the user Codex CLI is not available in this environment and stop.

    line 28
  4. Very specific:

    If codex exec --help is available, use it as the source of truth for the installed CLI flags.

    line 30
  5. Distinctive:

    Use this for normal delegation:

    line 34
  6. Distinctive:

    If the prompt is long, pass it via stdin:

    line 58
  7. Very specific:

    When using shell commands from another agent, avoid dangerous flags unless the user explicitly asked for them.

    line 64
  8. Very specific:

    Keep the prompt scoped to one task.

    line 77
  9. Very specific:

    If the installed CLI supports a different resume syntax, follow codex exec resume --help.

    line 89
  10. Very specific:

    Put -i flags before the prompt text.

    line 110
  11. Distinctive:

    Use common raster formats: PNG (best for screenshots/UI, lossless) or JPEG.

    line 111
  12. Very specific:

    If CODEX_HOME is unset, use ~/.codex/generated_images/<codex-session-id>/.

    line 137
  13. Distinctive:

    After any Codex image-generation request, find the image path before replying.

    line 144
  14. Distinctive:

    If the session id is not known, list recent generated images:

    line 158
  15. Distinctive:

    When possible, sort by modification time:

    line 164
  16. Very specific:

    If the user requested a project asset, copy or move the selected image into the project after generation if permissions allow.

    line 171
  17. Distinctive:

    If the user requested a specific aspect ratio, verify dimensions before answering.

    line 175
  18. Distinctive:

    If Codex exits with network, auth, or model refresh errors, report the error plainly and include the relevant stderr lines.

    line 219
  19. Very specific:

    If Codex claims it generated an image but no image file exists under $CODEX_HOME/generated_images, say so explicitly and do not invent a …

    line 221
  20. Very specific:

    If the host agent cannot read $CODEX_HOME/generated_images, ask the user to grant access or run:

    line 223
  21. Distinctive:

    Do not pass secrets to Codex unless the user explicitly asks and understands the risk.

    line 231
  22. Very specific:

    Use read-only for consultation.

    line 233
  23. Distinctive:

    Do not claim Codex performed work solely from its final message.

    line 235

“Common” is measured: kept by at least 6 in 10 of the 20 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Codex CLI

Use this skill when the user asks this agent to talk to Codex, consult Codex, delegate a coding/review/research task to Codex CLI, or ask Codex to generate an image.

The core interface is:

codex exec "prompt"

For automation, prefer non-interactive codex exec over interactive codex.

First Check

Before using Codex CLI, verify it is available:

command -v codex
codex --version
codex exec --help

If codex is missing, tell the user Codex CLI is not available in this environment and stop. Do not simulate a Codex response.

If codex exec --help is available, use it as the source of truth for the installed CLI flags. Codex CLI changes over time.

One-Off Task Workflow

Use this for normal delegation:

codex exec -C "$PWD" -s read-only "Review this repository and identify likely bugs"

For tasks that may edit files, use the sandbox the user requested or the host agent allows:

codex exec -C "$PWD" -s workspace-write "Implement the requested change"

For a response file:

codex exec -C "$PWD" -o /tmp/codex-last-message.md "Summarize this project"

For machine-readable events:

codex exec -C "$PWD" --json -o /tmp/codex-last-message.md "Analyze this project"

If the prompt is long, pass it via stdin:

codex exec -C "$PWD" - < /tmp/prompt.txt

When using shell commands from another agent, avoid dangerous flags unless the user explicitly asked for them. Prefer read-only for review, planning, critique, image generation, and analysis.

Prompt Shape

For non-trivial tasks, shape the delegated prompt with OpenAI's Codex best-practices structure:

Goal:
Context:
Constraints:
Done when:

Keep the prompt scoped to one task. For complex implementation work, ask Codex for a plan first in read-only mode, then run a separate write task after the plan is accepted.

Session Resume

Codex emits a session id in normal startup output and in JSONL events as thread.started.

To resume:

codex exec resume <session-id> "Continue from the previous task"

If the installed CLI supports a different resume syntax, follow codex exec resume --help.

Image and File Input

Codex cannot browse or read image files on its own. To let Codex see an image, attach it explicitly with -i/--image. This is the input side; image generation (below) is the output side.

Attach a single image:

codex exec -C "$PWD" -s read-only -i screenshot.png "Explain the error shown here"

Attach multiple images. Both forms work — comma-separated or a repeated flag:

codex exec -i before.png,after.png "Compare these two UI states"
codex exec -i mock1.png -i mock2.png "Which layout is closer to the spec?"

Best practices:

  • Put -i flags before the prompt text.
  • Use common raster formats: PNG (best for screenshots/UI, lossless) or JPEG. Keep images reasonably small; very large images slow the request.
  • Combine images with a clear text instruction — the image alone is rarely enough context.
  • For visual review/analysis, pair with -s read-only so Codex inspects but does not edit.

For non-image files there is no attach flag. Two options:

  • Let Codex read the file from the workspace: point it at the directory with -C "$PWD" and name the file in the prompt ("Review src/auth.ts"). Use this for files already in the repo.
  • Pipe file contents via stdin. With a prompt also present, stdin is appended to the prompt as a <stdin> block:
codex exec -C "$PWD" "Summarize the attached log" < /tmp/build.log

Image Generation Workflow

Codex can generate images through its built-in image generation tool when asked through codex exec.

Example:

codex exec "Generate a black banana image with aspect ratio 16:9"

Important behavior:

  • Built-in image generation commonly saves files under $CODEX_HOME/generated_images/<codex-session-id>/.
  • If CODEX_HOME is unset, use ~/.codex/generated_images/<codex-session-id>/.
  • Codex may not print the path in its final response.
  • The agent using this skill must proactively locate and report generated image paths after image generation.
  • The PNG metadata may not expose the exact image model. Do not claim a specific backend model unless Codex output, CLI logs, or metadata clearly show it.

Finding Generated Images

After any Codex image-generation request, find the image path before replying.

First extract the Codex session id from output if available. It may look like:

session id: 019e8cbf-5054-7131-9440-ee592f0d8a17

Then check:

ls -l "${CODEX_HOME:-$HOME/.codex}/generated_images/<session-id>"

If the session id is not known, list recent generated images:

find "${CODEX_HOME:-$HOME/.codex}/generated_images" -type f -maxdepth 3 -print

When possible, sort by modification time:

find "${CODEX_HOME:-$HOME/.codex}/generated_images" -type f -maxdepth 3 -print0 \
  | xargs -0 ls -lt

If the user requested a project asset, copy or move the selected image into the project after generation if permissions allow. Never leave a project-referenced asset only under $CODEX_HOME/generated_images.

Verifying Aspect Ratio

If the user requested a specific aspect ratio, verify dimensions before answering.

On macOS:

sips -g pixelWidth -g pixelHeight /path/to/image.png

Portable fallback:

file /path/to/image.png

Report actual dimensions and whether they match approximately or exactly. Many generated images use nearest integer dimensions, so a 16:9 request may produce dimensions like 1672 x 941, which is effectively 16:9 but not mathematically exact.

Reporting Results

For normal Codex tasks, report:

  • command used, summarized if long
  • Codex status or final answer
  • any files Codex changed, if known
  • any limitations, errors, or missing permissions

For image tasks, always report:

  • generated image path
  • actual dimensions if available
  • whether the path is in $CODEX_HOME or copied into the project
  • whether the exact image model is known or not

Example response:

Codex generated the image here:
/Users/me/.codex/generated_images/019e.../ig_....png

Actual size: 1672 x 941, effectively 16:9.
The output does not expose the exact image backend model, so I cannot verify whether it was gpt-image-2.

Failure Handling

If Codex exits with network, auth, or model refresh errors, report the error plainly and include the relevant stderr lines.

If Codex claims it generated an image but no image file exists under $CODEX_HOME/generated_images, say so explicitly and do not invent a path.

If the host agent cannot read $CODEX_HOME/generated_images, ask the user to grant access or run:

find "${CODEX_HOME:-$HOME/.codex}/generated_images" -type f -maxdepth 3 -print0 | xargs -0 ls -lt

Safety Notes

Do not pass secrets to Codex unless the user explicitly asks and understands the risk.

Use read-only for consultation. Use workspace-write only when Codex is expected to edit files. Avoid danger-full-access unless the user explicitly requested it in a controlled environment.

Do not claim Codex performed work solely from its final message. When file changes matter, inspect the filesystem or Git diff after the run.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.