Codex Image Studio

Codex Image Studio

Overview

Use this skill to make Codex run a predictable image-generation workflow through an OpenAI-compatible API. The user should be able to ask for images in natural language while Codex handles prompt shaping, mode choice, API invocation, output naming, and result inspection.

Configuration

Use this configuration precedence:

1. Existing environment variables in the current process.
2. User-level private env file:
   • Windows: %USERPROFILE%\.codex\secrets\image_api.env
   • Unix-like: $HOME/.codex/secrets/image_api.env

The supported keys are:

• OPENAI_API_KEY: API key for the OpenAI-compatible provider.
• OPENAI_BASE_URL: provider base URL, for example https://example.com/v1.
• IMAGE_MODEL: optional model override. Default to gpt-image-2.

On first install or first use, create the secrets folder and template file if missing. Do not overwrite an existing file.

Template:

OPENAI_API_KEY=
OPENAI_BASE_URL=
IMAGE_MODEL=gpt-image-2

After creating the template, tell the user to fill in OPENAI_API_KEY and OPENAI_BASE_URL before generation.

Never print the API key. If the key is stored in a local file, read it without echoing it and set it only for the current process.

Engine

For text-to-image generation, use the repository's dependency-free HTTP adapter first:

<skill-root>/scripts/generate_image.py

Example:

python <skill-root>/scripts/generate_image.py `
  --prompt-file prompt.txt `
  --model "$env:IMAGE_MODEL" `
  --size 1024x1024 `
  --quality high `
  --out output/imagegen/output.png

The adapter posts directly to /images/generations, intentionally omits n, retries one time for transient 429/502/503/504 responses, and accepts common base64, data-URL, nested, and returned-URL response shapes. It reads existing process environment variables before filling missing values from the user-level env file and never prints the API key.

Use the bundled system image generation CLI only as a fallback for image editing or provider-specific batch features that the adapter does not cover:

$HOME/.codex/skills/.system/imagegen/scripts/image_gen.py

Mode Selection

• Use generate for text-to-image.
• Use edit for image-to-image, multi-reference generation, or mask-guided edits.
• Use generate-batch for multiple different prompts or a production run.

For multiple variants of one prompt, run the HTTP adapter once per output instead of sending n; this provider's compatibility layer may reject n=1. For multiple distinct prompts, use separate adapter calls or the system CLI's generate-batch only when that provider path is known to work.

Text-To-Image

Generate from text only. Always specify:

• model
• size
• quality
• output path
• prompt

The default text-to-image path is scripts/generate_image.py. Do not route this path through OpenAI().images.generate() unless the provider has been verified to return a populated data[].b64_json response.

Default model order:

1. IMAGE_MODEL if set.
2. gpt-image-2.
3. Provider alias such as image-2 only after the first model name fails.

Image-To-Image

Use edit and pass each reference image with repeated --image arguments. Use --mask only for localized edits. For gpt-image-2, do not pass --input-fidelity; the bundled CLI treats image inputs as high fidelity for that model.

For reference-guided generation, explicitly describe each input image's role in the prompt, such as "Image 1 is the shape reference" or "Image 2 is the material reference."

Batch Generation

Use JSONL where each line is one job. Keep jobs provider-neutral and include at least:

{"prompt":"...","out":"output-name.png","model":"gpt-image-2","size":"1536x1024","quality":"high"}

If the bundled CLI requires its own batch schema, follow the script's help output rather than inventing a new schema.

Prompt Rules

Write prompts as concise production specs:

• Asset type
• Primary request
• Subject
• Style or medium
• Composition and framing
• Lighting
• Materials and palette
• Constraints
• Avoid list

For product renders, add "no text, no logo, no watermark" unless the user asks for text.

Output Rules

• Save final images under the current workspace unless the user specifies another folder.
• Use descriptive filenames and avoid overwriting existing files unless requested.
• Inspect generated images when possible.
• Report saved paths and briefly mention the final prompt intent.

Failure Recovery

• If authentication fails, ask the user to verify OPENAI_API_KEY and OPENAI_BASE_URL.
• If the HTTP adapter receives a transient gateway response, let its single retry finish before reporting failure.
• If no image payload is found, report the HTTP status and sanitized response detail; do not silently fall back to an empty SDK response.
• If model lookup fails, retry once with image-2 when the provider is known to use aliases.
• If generation succeeds but the output misses core constraints, iterate with one targeted correction.
• If the provider does not support edit or batch endpoints, explain the limitation and fall back to text-to-image only.