Codex Imagegen

codex-imagegen

Overview

This skill drives the Codex CLI agent (codex exec) and asks it to produce a PNG at a path the caller chooses. On this machine the agent does not invoke a built-in image_generation tool — instead it shells out to https://api.openai.com/v1/images/generations with model gpt-image-1, base64-decodes the response, and writes the PNG. The wrapper handles the friction: companion-env interception, sandbox/network confinement, path resolution, output verification.

When to use

Trigger when the user's request has both:

1. An image-generation intent — "生圖 / 畫一張 / generate / draw / create an image of …", and
2. A target file path — "存成 X.png", "save to X.png", or an output directory.

If only (1) is present, ask for the path before proceeding.

How to use

Run the bundled script. It takes the output path as the first argument and the description as the rest:

bash {SKILL_DIR}/scripts/imagegen.sh <output-path> <description...>

{SKILL_DIR} is this skill's folder. From inside Claude Code use the absolute path /Users/htlin/.claude/skills/codex-imagegen/scripts/imagegen.sh (or the dotfiles source /Users/htlin/.dotfiles/claude.symlink/skills/codex-imagegen/scripts/imagegen.sh).

The script:

• Errors out if codex is missing or codex login status does not say "Logged in".
• Resolves a relative output path against the current working directory and mkdir -p the parent.
• Calls codex exec -C "$(pwd)" -s workspace-write --skip-git-repo-check with a Chinese prompt that instructs codex to use the image generation tool and save to the absolute output path.
• Fails (exit 1) if no file was written, otherwise prints wrote <path> (<bytes> bytes).

Smoke test

The canonical smoke test from the article:

bash /Users/htlin/.claude/skills/codex-imagegen/scripts/imagegen.sh tmp/shiba.png 戴墨鏡的柴犬

Expected: a PNG (~1–3 MB, ~1024px square or larger) at tmp/shiba.png and a final line wrote /…/tmp/shiba.png (<bytes> bytes).

Notes

Things the wrapper does that aren't obvious from the article:

• Strips CODEX_COMPANION_SESSION_ID and CLAUDE_PLUGIN_DATA. When this skill runs from inside Claude Code, those env vars are inherited from the Codex Companion plugin. Left alone, the companion intercepts codex exec and pipes the existing companion-session prompt into it (you'll see user / reply with ok in the codex log). Unsetting both forces a fresh exec.
• Uses --dangerously-bypass-approvals-and-sandbox. The article uses -s workspace-write, but on this machine the codex agent does not have a built-in image_generation tool and instead curls https://api.openai.com/v1/images/generations itself. workspace-write blocks that egress, so the agent gives up after exploration. Full bypass lets it through.
• Requires $OPENAI_API_KEY. Despite the article's "ChatGPT login is enough" claim, the agent on this machine reads $OPENAI_API_KEY directly. The wrapper warns (does not abort) when missing, in case the user is running in a session where ChatGPT auth does work.
• Verifies the file landed. Codex sometimes claims success without writing — the wrapper exits non-zero unless the target file exists and is non-empty.
• First image-gen run can take 30–60 s; later runs are faster.