Codex Sangpye

When to use

When the user wants a Korean e-commerce product detail page ("상세페이지" / "상폐") asset set — 13 emotional-journey sections (Hero → Pain → Problem → Story → Solution → How → Proof → Authority → Benefits → Risk → Compare → Filter → CTA) plus a single 1080×12720 combined PNG — generated from 1–14 product photos and a Korean brief.

Prefer this skill over invoking image generation by hand, because:

• It runs the full analysis (gpt-5.5) + 5-bundle parallel image generation (orchestrator + image_generation tool) + slicing + vertical composition pipeline.
• It uses your Codex OAuth session — no API key, no per-token billing.
• One command, one JSON result.

How it talks to Codex (0.3.0+)

Codex CLI 0.130 removed the codex responses subcommand, so this skill no longer shells out. The sangpye CLI now reads the OAuth tokens written by codex login (~/.codex/auth.json) and POSTs directly to https://chatgpt.com/backend-api/codex/responses — the same endpoint the old subcommand used. Wire format, model (gpt-5.5 + image_generation tool), and ChatGPT-subscription billing are unchanged. You no longer need any specific version of the codex binary on PATH; only the auth file matters.

Preconditions

1. ~/.codex/auth.json exists with a ChatGPT OAuth session. If missing or expired, run codex login once and pick the ChatGPT option (not API key).
2. sangpye --version succeeds (install via uv tool install git+https://github.com/NewTurn2017/codex-sangpye-skill).
3. 1–14 product image files exist locally.

If any precondition fails, tell the user how to fix and stop.

Command path

sangpye (installed globally via uv tool install).

Parameters

Layouts: flat vs cards

--layout flat (default) stacks the 13 sections with zero gap into the canonical 1080×12720 PNG — the standard full-bleed e-commerce 상폐 look.

--layout cards places each section as a rounded card on a unified dark background with even margins (left/right --card-side, between cards --card-gap, page top/bottom --card-pad). This removes the abrupt background-color jumps at section seams and gives a tidy "card rhythm" — well suited to lecture / content detail pages (e.g. the /video-use-summary pipeline passes --layout cards). The combined height is no longer exactly 12720 because each card is width-inset and gaps are added. The --card-* flags only take effect with --layout cards.

Basic usage

sangpye \
  --image ./photos/earbuds_01.jpg \
  --prompt "무선 이어폰, ANC 탑재, 30시간 배터리, IPX5 방수" \
  --output ./out

Example with explicit options

sangpye \
  --image ./photos/earbuds_01.jpg \
  --image ./photos/earbuds_02.jpg \
  --image ./photos/earbuds_lifestyle.jpg \
  --prompt "프리미엄 무선 이어폰. 30시간 재생, ANC, IPX5 방수, 인체공학 디자인. 20-40대 직장인 대상." \
  --category electronics \
  --quality high \
  --output ./out

Expected result

stdout (single JSON line):

{"job_id":"a1b2c3d4","output_dir":"/abs/...","combined":"/abs/.../combined.png","sections":["/abs/.../01_hero.png", "..."],"plan_path":"/abs/.../analysis.json","elapsed_sec":252.4,"failed_bundles":[],"reused_bundles":[]}

Show the user:

1. The absolute path to combined.png (the main deliverable).
2. The job_id so they can find the artifacts again.
3. Optionally, the list of 13 individual section PNGs.

Partial success (exit code 5)

If failed_bundles is non-empty, the CLI exited with code 5: combined.png was still produced, but some sections use dark placeholders. Tell the user:

• How many bundles failed + which ones (e.g. "B2_OPENING")
• They can retry only the failed bundles by re-running with the same --output and --job-id — the skill auto-resumes from the saved analysis.json and existing bundles/*.png.

Auto-resume

When output_dir/{job_id}/ already contains analysis.json:

• Step 1 (gpt-5.5 analysis) is skipped — the stored plan is reused.
• Individual bundles with a non-empty bundles/{id}.png on disk are reused.
• Only missing bundles are generated; then slice + compose run over the full set.

This makes a failed run trivial to recover from without re-burning quota. Suggest this to the user whenever they see exit=5 or a bundle failure.

Troubleshooting

• error: ~/.codex/auth.json not found → Run codex login and pick the ChatGPT/OAuth option.
• error: ~/.codex/auth.json has no ChatGPT OAuth tokens → The current auth.json is API-key mode. Run codex logout && codex login, choose ChatGPT.
• error: ChatGPT OAuth rejected the request (HTTP 401) → The access token in ~/.codex/auth.json expired (lifetime ~10 days). Run codex login to refresh.
• error (codex): HTTP 500: ... or HTTP 503 → Upstream/ChatGPT outage. Retry in a few minutes.
• responses network error → Local network problem reaching chatgpt.com. Check connectivity.
• error (codex): rate_limit / 429 → ChatGPT subscription is throttling. Wait, retry, or pass --quality standard.
• Frequent server overloaded retries visible in stderr → Lower parallelism by running with SANGPYE_MAX_CONCURRENCY=1 sangpye ... (default is 2).
• The model 'gpt-5.5' does not exist or you do not have access to it → Your ChatGPT subscription tier does not yet include gpt-5.5. Set SANGPYE_MODEL=gpt-5.4 as a temporary fallback, or surface the error verbatim.

Runtime expectations

• Typical: 5–10 minutes for a full 13-section run.
• Under load: up to ~15 minutes; the CLI transparently retries server overloaded and rate_limit events with exponential backoff (10s/30s/60s/90s/150s × 1.5 for overload).
• analysis.json is persisted to output_dir/{job_id}/analysis.json immediately after Step 1 (Codex analysis), so it survives later image-gen failures.
• Stderr shows per-bundle lifecycle events (▶ B2_OPENING generating..., ⟲ B2_OPENING overloaded, backing off 45s, ✓ B2_OPENING done in 67.3s) so progress is visible during long runs.

Agent rule

After a successful run, ALWAYS show the user the absolute path to combined.png and the job_id. Do not silently retry on errors — surface them so the user can decide. Do not invoke this skill more than once per user request unless explicitly asked.

Source: NewTurn2017/codex-sangpye-skill — distributed by TomeVault.