dsh-palate

LIMITATIONS

Known limitations

- **No embedding-based semantic matching in the plugin itself** — it retrieves locally by tags, words, and Chinese word fragments; the model does the deeper reasoning from the assembled context. - **Feedback is explicit** — the plugin does not guess whether a user adopted a recommendation; call `palate_feedback` after a review to form effectiveness data. - **Training decisions are explicit** — `palate_intake` stages evidence but does not inspect a raw URL/image or learn automatically; `palate_decide` needs a clear human accept/reject decision. - **Markdown mirrors are read-only exports** (human edit-and-merge-back is planned). - **Vision is delegated** — pair with a vision tool to inspect screenshots, or a browser to inspect URLs, before staging evidence. - **Reference packs are not cloning kits** — they preserve observable layout and hierarchy lessons, not protected assets, copy, or a promise that every page from a referenced brand is appropriate for every product.