Design decisions¶

Four choices shape the code. Keep them unless you mean to change the decision itself.

No built-in default model¶

Without --model or a saved default, genimg stops and says how to pick a model. A hardcoded default could name a model a fresh key cannot reach, and the very first command would fail. genimg setup offers to save a default, so the guided path still ends in a command with no flags.

No organisation-specific defaults¶

genimg never falls back to a hardcoded GCP project or credential. The Vertex project comes from --project, then the profile's project setting, then GOOGLE_CLOUD_PROJECT, then the service-account file or gcloud. If none is set, genimg says what to set. A hardcoded fallback would be wrong for every user but its author.

Providers are plugins, auth lives in profiles¶

Each backend is a Provider that owns its auth modes, model capabilities, prices and model-id inference, so the CLI, the grid and Draw Studio read the same tables. A few provider-name branches remain outside the providers: OpenAI pricing in cost.py, the OpenAI --mode batch error in cli.py, the suggested model in setup.py and Draw Studio's offline fallback in draw.py.

Each provider and auth mode pair is an AuthProfile. Users pin one as a [profiles.NAME] table in config.toml, which never holds secrets. Without one, genimg uses the first mode whose detection succeeds: an API key in the environment, a gcloud ADC token, or a Codex login.

The full records are in internal/adr/ in the repository.

Skills install through npx skills¶

npx skills add lucharo/genimg is the only install path; genimg skills prints that command and genimg skills path shows where the bundled copies live. genimg once had its own install, update and uninstall verbs, but they managed the same directory as the user's own skills, such as genimg-preferences, and could delete them. The skills CLI installs per agent and per project, and leaves unrelated skills alone.