Image generation

Image Generation

Generate or edit an image using the Codex CLI based on $ARGUMENTS.

Parse $ARGUMENTS as one of:

• "<prompt>" — generate a new image
• "edit <path> <prompt>" — edit an existing image
• Optional inline hints: size:<WxH>, format:<png|jpg|webp>, out:<path>, style:<hint>

Examples:

• "a watercolor moonlit library"
• "edit assets/hero.png replace background with white"
• "a minimalist OCaml logo size:512x512 out:docs/logo.png"

Step 1 — Parse and Resolve Arguments

Extract from $ARGUMENTS:

• MODE: generate or edit
• RAW_PROMPT: the user's original description
• INPUT_PATH: (edit mode only) path to source image — verify it exists; abort if not
• SIZE: default 1024x1024; override if size:WxH present
• FORMAT: default png; override if format:X present
• OUTPUT_PATH: if out:<path> present, use it; otherwise ask the user where to save the result

If OUTPUT_PATH is missing and $ARGUMENTS provides no hint, ask:

"Where should I save the image? (e.g., assets/hero.png)"

Do not proceed until OUTPUT_PATH is resolved.

Step 2 — Prompt Refinement

Before invoking Codex, rewrite RAW_PROMPT into a detailed, visually precise prompt:

• Add composition, lighting, mood, and style details the user left implicit
• Incorporate SIZE as an aspect ratio hint (e.g., "square 1:1 composition")
• Incorporate FORMAT only if it has visual implications (e.g., transparency for PNG)
• Keep the user's intent intact — do not substitute a different concept

Store the result as REFINED_PROMPT. Show the user both the original and refined version:

Original: <RAW_PROMPT> Refined: <REFINED_PROMPT>

If the user rejects the refinement, use RAW_PROMPT as-is.

Step 3 — Generate

Set ATTEMPT=1, MAX_ATTEMPTS=3.

Generate mode:

codex exec --full-auto "Generate a <SIZE> <FORMAT> image of: <REFINED_PROMPT>. Save the result to <OUTPUT_PATH>."

Edit mode:

codex exec --full-auto "Edit the image at <INPUT_PATH>: <REFINED_PROMPT>. Save the result to <OUTPUT_PATH>."

Capture both stdout and stderr. On completion, proceed to Step 4.

Rate Limit Handling

If stderr contains rate limit, 429, or quota:

• Notify the user: "Rate limit hit — waiting <WAIT>s before retry."
• Wait: WAIT = 15 * ATTEMPT seconds (15s, 30s, 45s)
• Retry the same command. Do not increment ATTEMPT for rate limit retries — at most 3 rate limit waits per attempt.
• If rate limit persists after 3 waits, abort and report: "Rate limit not cleared after 3 retries. Try again later."

Fatal Error Handling

Abort immediately (do not retry) if stderr contains:

• command not found / codex: not found → "Codex CLI not installed. Run: npm install -g @openai/codex"
• authentication / unauthorized / 401 → "Codex authentication failed. Check your credentials."
• not found on INPUT_PATH (edit mode) → "Source image not found at <INPUT_PATH>."

Step 4 — Validate

4a. File existence check

Verify OUTPUT_PATH exists and is non-empty (size > 0 bytes).

If missing or empty:

• This counts as a failed attempt. Increment ATTEMPT.
• If ATTEMPT <= MAX_ATTEMPTS, refine the prompt further (add more explicit detail) and go to Step 3.
• If ATTEMPT > MAX_ATTEMPTS, abort: "Codex did not produce an output file after 3 attempts."

4b. Partial file check

Read the file. If the image cannot be decoded (corrupted header, truncated):

• Treat as failed attempt. Same retry logic as 4a.

4c. Vision quality check

Inspect the generated image visually:

• Does the image depict what RAW_PROMPT requested?
• Are there obvious artifacts, blank regions, or clearly wrong content?
• Does the composition match the aspect ratio implied by SIZE?

Score the result: pass / soft-fail / hard-fail

Result	Condition	Action
pass	Intent clearly satisfied	Proceed to Step 5
soft-fail	Mostly correct but minor issues	Show user, offer retry or keep
hard-fail	Wrong content or unusable	Auto-retry with corrected prompt

On soft-fail, present the image path and ask:

"The image was generated but may not fully match your intent. Keep it, or retry with adjustments?"

On hard-fail, identify the specific mismatch, adjust REFINED_PROMPT to correct it, increment ATTEMPT, and go to Step 3 if ATTEMPT <= MAX_ATTEMPTS.

If ATTEMPT > MAX_ATTEMPTS on hard-fail:

• Save the last result anyway
• Report: "Could not fully satisfy the prompt after 3 attempts. Saved best result to <OUTPUT_PATH>. Issues: <mismatch description>."

Step 5 — Log

Append a JSON entry to .imagelog.json in the project root (create if absent):

{
  "timestamp": "<ISO 8601>",
  "mode": "<generate|edit>",
  "raw_prompt": "<RAW_PROMPT>",
  "refined_prompt": "<REFINED_PROMPT>",
  "output_path": "<OUTPUT_PATH>",
  "size": "<SIZE>",
  "format": "<FORMAT>",
  "attempts": <ATTEMPT>,
  "outcome": "<pass|soft-fail|hard-fail|aborted>"
}

If .imagelog.json exists, parse it as an array and append. If it does not exist, create it as a single-element array.

Step 6 — Report

On success:

"Image saved to <OUTPUT_PATH> (attempt <ATTEMPT>/<MAX_ATTEMPTS>). Logged to .imagelog.json."

On abort: summarize what failed and what the user should do next.

Rules

• Never invoke Codex without a resolved OUTPUT_PATH.
• Never skip vision validation — even if the file exists and is non-empty.
• Never hard-retry more than MAX_ATTEMPTS times total (rate limit retries are separate).
• Never overwrite an existing file without confirming with the user first.
• Never expose raw Codex stderr to the user — translate errors into plain language.
• If the user explicitly says "skip validation", proceed directly to Step 5 after Step 3.
• .imagelog.json must always be updated, even on failure (log the abort).