| name | gpt-image |
| description | `/omg:gpt-image` explicit-only ChatGPT Images generation. Activate only when that command loads this skill; never activate from ordinary image, drawing, or design requests. |
GPT Image (ChatGPT Images web)
This is an explicit command-only capability: /omg:gpt-image <prompt>. It uses the user's logged-in ChatGPT subscription through a local Chromium CDP session and only https://chatgpt.com/images/. It has no API, backend, generic-chat plus-menu, browser launch, auto-login, or dependency-install fallback.
The prompt is sent to the external ChatGPT service. Treat it as an upload/privacy boundary: do not submit secrets, private source, credentials, or material the user may not send to OpenAI. A user must manually maintain a dedicated logged-in browser profile. Do not run /omg:insane-review or any other CDP automation concurrently.
Resolve the engine
Resolve the exact non-symlink engine from the current suite-root binding. Do not select a plugin
cache glob or an old-identity binding for this new capability.
GI="$(python3 - <<'PY'
from pathlib import Path
import os
import stat
import sys
cwd = Path.cwd()
home = Path.home()
asset = Path("bin/gpt_image_web.py")
bindings = [
cwd / ".gjc/runtimes/oh-my-gajae-code/root",
home / ".gjc/agent/runtimes/oh-my-gajae-code/root",
]
for binding in bindings:
try:
info = binding.lstat()
if binding.is_symlink() or not stat.S_ISREG(info.st_mode) or stat.S_IMODE(info.st_mode) & 0o077:
continue
lines = binding.read_text(encoding="utf-8").splitlines()
if len(lines) != 1 or not os.path.isabs(lines[0]) or any(ord(ch) < 32 for ch in lines[0]):
continue
root = Path(lines[0])
canonical = root.resolve(strict=True)
if str(root) != str(canonical):
continue
candidate = canonical / asset
if candidate.is_symlink() or not candidate.is_file():
continue
print(candidate)
raise SystemExit(0)
except (OSError, UnicodeError):
continue
checkout = cwd / "plugins/oh-my-gajae-code" / asset
try:
if not checkout.is_symlink() and checkout.is_file():
print(checkout.resolve(strict=True))
raise SystemExit(0)
except OSError:
pass
print("gpt-image engine not found; rerun the hardened OMG installer", file=sys.stderr)
raise SystemExit(1)
PY
)" || exit 1
빈 값이거나 검증에 실패하면 중단한다.
Procedure
- Require a nonempty exact prompt. Clarify it before executing; do not silently embellish it.
- Resolve
$GI with the hardened binding resolver and run the read-only prerequisite check:
python3 "$GI" --check-env
It requires POSIX deadline enforcement, Python Playwright, and a real local Chrome/Chromium CDP endpoint. It also requires the endpoint to be proven bound to the dedicated profile — via the listener process's --user-data-dir argv (Chrome 136+, which no longer writes a receipt) or the legacy DevToolsActivePort receipt when present. The user must manually start/login the dedicated browser; do not install, launch, or log in for them.
- Generate only after a successful check:
python3 "$GI" --output-dir .gpt-image -- "$ARGUMENTS"
- Report the emitted PNG path and its adjacent provenance JSON. The engine uses the Images composer, exact prompt, a fresh conversation, one completed new assistant image, and the Images fullscreen Save action. It rejects ambiguous assets and never downloads a thumbnail, screenshot, signed URL, or backend response.
ChatGPT currently exposes the fullscreen Save/Download control behind the selected image's
share-labeled viewer action. Opening that viewer is allowed only to reach the original download;
never click Copy link, X, LinkedIn, Reddit, or any other publication control.
Output is atomically/exclusively written to the current project's non-symlink .gpt-image/ directory (0700); PNG and provenance files are 0600. Provenance includes prompt, conversation URL, SHA-256, size, dimensions, timestamp, and route, never cookies or signed asset URLs. Failures remove temporary downloads and partial output.
