image
Generates raster images (icons, illustrations, textures) via the logged-in Codex CLI image tool. Use when an agent needs a real image but has no native image generation.
shipshitdev · AI review on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add shipshitdev/skills --skill imageAbout this skill
name: image
description: Generates raster images (icons, illustrations, textures) via the logged-in Codex CLI image tool. Use when an agent needs a real image but has no native image generation.
license: MIT
compatibility: Requires the codex CLI (logged in) plus python3 and base64; sips is optional for post-processing on macOS.
metadata:
tags: "codex, image-generation, gpt-image, cli, assets, app-icon"
when_to_use: "app icon, codex image generation"
disable-model-invocation: true
Codex Image Gen
Generate a real raster image from a text prompt when the running agent has no
native image generator. The mechanism is non-obvious: the headless codex exec
path genuinely calls the image tool, but it never writes the PNG to disk — only
the Codex desktop app has the plumbing that polls the async job and saves it.
The finished image is still recoverable, because its full base64 PNG is recorded
in the session rollout JSONL. Run the prompt, read the session id Codex prints,
open the matching rollout, and decode the largest result string to a PNG.
Contract
Inputs:
- An image-generation request, prompt, and selected local output path
Outputs:
- Generated image file and extraction evidence
Creates/Modifies:
- Prompt, diagnostic log, and image files at the selected local destination
External Side Effects:
- Calls the authenticated Codex CLI image service and reads the resulting session rollout
Confirmation Required:
- Before using an external provider or writing outside the requested output scope; preserve existing images unless replacement was authorized
- Loading the skill grants no additional authority. Existing explicit approval applies only to the same target and actions; preserve report-only restrictions.
Delegates To:
- None
When this is the right tool
- The agent needs a generated raster image (icon, illustration, texture, app icon, placeholder art) and has no built-in image-generation tool.
- A
codexCLI is installed and logged in, and shelling out to it is allowed. - A vector result is not required — this produces a PNG, not an SVG.
If a native image tool exists, prefer it. If the deliverable is a logo or crisp vector, prefer a vector workflow.
Why the naive approach fails
Running codex exec "draw a ..." and then watching ~/.codex/generated_images/
produces nothing: that folder is written by the desktop app's async-job poller,
not by codex exec. People conclude CLI image generation is impossible. It is
not — the completed image lives in the session rollout as the result field of
the image-generation response item. This skill reads it from there.
Pipeline
1. Write the prompt to a file
Avoid shell-escaping pain by putting the prompt in a file. Be explicit — the model has no other context:
- Subject and style ("flat vector mark", "glossy 3D glass icon", "hand-drawn").
- Composition: full-bleed vs. padded, centered, single object vs. scene.
- Palette and background (solid color, transparent intent, gradient).
NO text, NO letters, NO wordsunless you specifically want type.- Target aspect ratio and rough size.
REPO_TMP="$(git rev-parse --show-toplevel)/.tmp"
mkdir -p "$REPO_TMP"
cat > "$REPO_TMP/img-prompt.txt" <<'PROMPT'
A single app icon: a glossy translucent envelope on a soft blue-to-violet
gradient, Liquid Glass style, centered with even padding, no text, no letters,
1:1 square, high detail.
PROMPT
2. Run codex exec and capture stdout
codex exec -s read-only "$(cat "$REPO_TMP/img-prompt.txt")" 2>&1 | tee "$REPO_TMP/codex-run.log"
Run it in the foreground. exec returns once the turn completes; in practice
the result is already written to the rollout by then.
3. Parse the session id
Codex prints a session id: <uuid> line. Pull it from the captured log:
SESSION_ID=$(grep -oE 'session id: [0-9a-f-]{36}' "$REPO_TMP/codex-run.log" | awk '{print $3}')
echo "session: $SESSION_ID"
4. Extract the PNG from the rollout
The rollout lives at ~/.codex/sessions/YYYY/MM/DD/rollout-*<session-id>*.jsonl.
Walk it, find the largest result string (the base64 image), and decode it. Use
the bundled helper:
python3 scripts/extract-codex-image.py "$SESSION_ID" "$REPO_TMP/out.png"
5. Assert success
Confirm the file exists and is a real PNG before using it:
test -s "$REPO_TMP/out.png" && file "$REPO_TMP/out.png" # expect: PNG image data, 1254 x 1254
If extraction finds no base64 result, the turn did not actually generate an
image (e.g. the model answered in text). Re-run step 2 with a more explicit
"generate an image" instruction.
6. Post-process (optional)
Default output is roughly 1254×1254 PNG, RGB, no alpha. Resize / strip alpha
with sips on macOS:
sips -z 1024 1024 "$REPO_TMP/out.png" --out "$REPO_TMP/icon-1024.png" # downscale
sips -s format png "$REPO_TMP/out.png" --out "$REPO_TMP/flat.png" # normalize
Reference extractor
scripts/extract-codex-image.py (also reproduced here so the procedure is
self-contained):
import json, sys, base64, glob, os
session_id, out = sys.argv[1], sys.argv[2]
sess = max(
glob.glob(os.path.expanduser(f"~/.codex/sessions/**/*{session_id}*.jsonl"), recursive=True),
key=os.path.getmtime,
)
best = None
def walk(o):
global best
if isinstance(o, dict):
for k, v in o.items():
if k == "result" and isinstance(v, str) and len(v) > 100000:
if best is None or len(v) > len(best):
best = v
else:
walk(v)
elif isinstance(o, list):
for v in o:
walk(v)
for line in open(sess):
try:
walk(json.loads(line))
except Exception:
pass
if best is None:
sys.exit("no base64 image result found in rollout — the turn may not have generated an image")
open(out, "wb").write(base64.b64decode(best))
print("WROTE", out)
Read the complete source ↗