Automatically fixes and improves user prompts before they reach the model.
Supports two CLIs, with the same settings:
See your CLI's README for install steps and the details that matter.
Every prompt runs through up to three stages, each of which you can switch on or off:
original → [correction] → [translation] → [enhancement] → final
- Correction — fixes grammar and spelling whilst preserving intent, style, and punctuation
- Translation — translates non-English prompts to English; English passes through unchanged
- Enhancement — rewrites for clarity, specificity, and structure
Stages are optional — the plugin works with just correction, just enhancement, or any combination.
Settings live in a YAML file. The options are the same on both CLIs — only the path differs, so see your CLI's README.
| Setting | Default | What it does |
|---|---|---|
enabled |
true |
Master switch |
correction |
true |
Fix grammar and spelling |
correction_model |
haiku |
Model for correction |
translation |
false |
Translate non-English to English |
translation_model |
haiku |
Model for translation |
enhancement |
false |
Rewrite for clarity |
enhancement_model |
sonnet |
Model for enhancement |
audit |
true |
Keep a log of prompts and changes |
verbose |
false |
Show extra detail |
Models — haiku, sonnet, and opus work everywhere. OpenCode can also use the providers you've connected (fast, capable, powerful) or an explicit provider/model ID. See the OpenCode README.
With audit on, each prompt is logged as one line of JSON — the original text, each stage's output, the mistakes caught, and the models used. OpenCode adds token usage. Paths and exact fields are in the CLI READMEs.
MIT