ergouzi-image-gen

Ergouzi Image Gen

Use this Skill as a thin adapter between Codex and the existing Ergouzi image API. Codex prepares the model input; the scripts only handle credentials, local file conversion, request submission, task polling, and result download.

Prerequisites

• Require Python 3.10 or newer and network access to https://ergouzi.life.
• Use the initialized user config by default. ERGOUZI_MEDIA_API_KEY and ERGOUZI_MEDIA_BASE_URL are explicit runtime overrides; generic ERGOUZI_API_KEY and ERGOUZI_BASE_URL are accepted when no config exists.
• If credentials are missing, ask the user to initialize them locally with python scripts/configure.py. Never ask the user to paste a key into chat.
• If the optional ergouzi-media-mcp Codex plugin is installed, prefer its check_configuration, list_models, get_model_schema, create_prediction, get_prediction, cancel_prediction, and download_prediction tools for the API lifecycle. Keep the Python runner as the fallback when the MCP tools are unavailable.
• Accept JPEG, PNG, and WebP inputs as directly downloadable HTTPS URLs, data URIs, or local files that fit the existing 4 MiB JSON request limit. Use --image, --person-image, --garment-image, or --reference-pose for common local-file workflows. In input JSON, put { "$local_file": "C:/path/image.png" } in place of a media URL. The runner converts local files to data URIs; use an HTTPS URL when the expanded request would exceed the API limit.

Workflow

1. Submit a paid prediction only when the user explicitly asks to generate, edit, try on, upscale, or remove the background of an image.
2. Honor an explicit model choice. Otherwise let Codex select one of the six models from the task's required input/output capability; do not hide models or apply server-routing policy.
3. Read references/model-reference.md for the selected model's objective API contract. Use --prompt only when that model exposes a prompt field; use the e-rmbg options from its input JSON example rather than --prompt. Use --input-file, --input-json, or stdin for the complete model input object, without the outer { "input": ... } envelope. Convenience arguments override the corresponding JSON fields. The runner resolves $local_file objects before submission and otherwise leaves input values unchanged. Prefer a UTF-8 JSON file for structured input across operating systems; files and stdin may include a UTF-8 BOM. The runner submits e-rmbg through its fixed deployment version endpoint; callers still select it by the public model name and must not add a version field to the model input.
4. Prefer the MCP tools when available. Otherwise run scripts/run.py predict. Both paths create one logical task, reuse the same idempotency key for bounded transport retries, record the task_* ID, poll to a terminal state, and download successful outputs. If a submission is interrupted before the task ID is returned, retry the exact request with the printed --idempotency-key; do not start a new paid request with a new key.
5. Report the model, task ID, terminal status, and absolute saved paths.
6. If execution was interrupted or timed out, resume with status --wait --download. Do not create a replacement task unless the user explicitly asks.