dsh-palate
不断成长的审美：为 DSH agents 积累的设计品味。一个品味语料库和编码化原则，会随着你提供的每个示例而不断磨砺——依据习得的判断力，而非固定标尺审视设计。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:guo6x/dsh-palate#e0272f596382af07e893a0197b3bc2e8dba784adPalate 以 dsh-palate 发布，当前版本为 0.6.2。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
中文说明 · DeepSeek Harness plugin
Design-audit tools measure with a fixed ruler. dsh-palate trains an eye that grows.
Most design-review plugins ship a static ruleset and apply it forever — use them once or a thousand times, the judgment is identical. dsh-palate is the opposite: it keeps a taste corpus that accumulates. Every example you feed it and every principle you distill sharpens the judgment your agent draws on. The more you use it, the better its eye gets.
Taste is not a gift — it's pattern recognition built from exposure. See enough good and bad design, and the rules emerge. dsh-palate turns that into a mechanism an agent can actually use:
| Tool | What it does |
|---|---|
palate_intake | Stage a structured visual analysis as pending example/principle candidates; it never changes taste by itself |
palate_candidates | Inspect pending, accepted, or rejected visual-training candidates and their source sessions |
palate_decide | Apply the user’s explicit accept/reject decision; this is the only candidate-to-palate mutation path |
palate_review | Assemble the accumulated taste (principles + relevant past examples) as context, so the agent critiques grounded in learned judgment |
palate_feedback | Use a review_id to record whether a critique helped and which principles were accepted or rejected; only accepted principles gain evidence |
palate_add | Feed an example (good/bad/note + reason + tags) into the corpus — grows the palate |
palate_learn | Distill a new principle from experience and add it to the codified taste |
palate_packs | Inspect opt-in visual-reference packs and whether they are already applied |
palate_seed | Apply one or more visual-reference packs exactly once, without overwriting existing taste |
palate_list | Browse the accumulated corpus |
palate_principles | List the codified principles, ordered by evidence |
palate_effectiveness | See which principles were accepted or rejected in real review feedback |
palate_stats | How much taste has accumulated: examples studied, principles distilled |
Ships with a starter palate of 12 foundational principles plus four transparent teaching examples (good and bad dashboards, a readable table, and generic landing-page boilerplate), so the first review has concrete evidence — then it grows from there.
The four starter examples are inserted only when the local taste database is empty. Installing or upgrading the plugin never overwrites an existing palate.
dsh-palate also ships two opt-in visual-reference packs:
apple-product-storytelling — one product subject per viewport, proof-led imagery, restrained CTA choices, and a calm sequence of mini-campaigns.x-direct-utility — high-contrast identity, a decisive primary route, ranked secondary actions, and almost invisible supporting detail.They are transparent abstractions of public pages observed on 2026-08-27, not scraped assets, brand copy, or templates to imitate. Start by calling palate_packs, then explicitly apply one or both with palate_seed. Use tag: "apple" or tag: "x" in palate_review: the tag filters both matching examples and style-scoped principles, while universal principles remain available.
For example, ask an agent:
Inspect
palate_packs, then applyapple-product-storytellingandx-direct-utilitywithpalate_seed. Review our product-launch page with tagapple, and separately review our sign-in entry flow with tagx. Cite the evidence; do not copy either brand’s assets, copy, or identity.
palate_intake is deliberately a staging tool, not a hidden auto-learning button. First have the agent inspect a screenshot or page with a browser/vision capability. Then it records a compact, structured analysis: hierarchy, typography, color, spacing, interaction, and any other relevant dimension. The intake creates one example candidate plus any proposed principles; none are added to the corpus yet.
It can also record an explicit comparison with Apple, X, or future reference packs:
aligned — the observed evidence supports named abstract reference principlesconflicts — the evidence conflicts with named reference principlesinsufficient_evidence — the screenshot/page does not show enough to judgeComparing against a pack that has not been seeded is allowed for research, but the record is marked reference-only and does not activate or inject the pack. This keeps analysis separate from style adoption.
Use palate_candidates to show the evidence and candidate IDs to the user. Only after they clearly say accept or reject should the agent call palate_decide. Accepted records preserve their training-session provenance; rejected records remain in training.md so a team can revisit the judgment later.
inspect screenshot / URL with vision or browser
│
▼
palate_intake (structured observations + pack comparison)
│
▼
pending example / principle candidates ──▶ palate_candidates ──▶ user explicitly accepts or rejects
│ │
└──────────────────────────── palate_decide ◀─────────────────────┘
│
accepted only ─────────┼───────── rejected stays auditable in training.md
▼
taste corpus + codified principles
│
palate_review (a design) ──▶ review_id + learned evidence ──▶ agent writes grounded critique
▲ │
└── palate_feedback (accept/reject + why) ──▶ effectiveness + accepted-principle evidence ─┘
node:sqlite (built into Node ≥ 22) at $DSH_HOME/palate/, plus human-readable taste.md / principles.md / feedback.md / training.md mirrors. Zero runtime dependencies.palate_review snapshots its evidence; palate_feedback records the outcome, while feedback.md and the panel show actual acceptance/rejection data.modlens_read_image) or pages with a browser, then pass the evidence to palate_intake or palate_review. The plugin never fetches, screenshots, or claims to see a raw URL by itself.This is accumulated retrieval + codified principles + explicit decisions and feedback, not model fine-tuning. The plugin supplies learned taste as context; the model renders the critique. palate_intake does not count as learning: only a user-confirmed palate_decide adds a candidate, and only a user/agent-confirmed palate_feedback adds effectiveness evidence. That keeps judgment auditable through taste.md, principles.md, feedback.md, and training.md without retraining anything.
# GitHub is the supported release channel.
dsh plugin --profile web add github:guo6x/dsh-palate
Restart a running dsh web process, then refresh the page. Installation is complete when a 👁️ button appears at the bottom of the sidebar. Click it to see the starter palate, its principles, feedback history, and any staged training candidates.
Requirements: the DeepSeek Harness web profile and Node ≥ 22. The plugin itself uses only local SQLite storage — it does not need an account, API key, or embedding service. The natural-language chat demo still uses the model configured in Harness, so that provider must have a valid credential.
Developing from a checkout instead? Run dsh plugin --profile web add . from the repository directory. The repository commits the lib/ entrypoints, so GitHub and path installs can start immediately without running an install-time build script.
Use this short path to verify the install before teaching the palate:
Install with the command above, restart dsh web, and open a new chat. If the composer asks for a workspace, click Choose workspace, select or create one, and continue the one-time notice if it appears.
Make sure the selected Harness model has a valid provider credential, then paste this prompt:
Call
palate_stats, then usepalate_reviewto critique “a dashboard with twelve equal KPI cards, one primary revenue metric, and a small trend chart”. Tell me which stored principles and examples you used, and return thereview_id.
Confirm the response contains the starter principles, grounded evidence, and a review_id. Open the 👁️ panel: the review should also appear there.
This proves the plugin path — host discovery, local storage, retrieval, and Web rendering. The plugin does not make a remote model request itself, but the chat response does depend on the selected Harness provider. If the response reports an authentication error, verify the 👁️ panel and local counters first; the plugin can be installed correctly while the model credential is invalid. To see the corpus grow, follow the 90-second loop below and add palate_feedback only after you have actually judged the recommendation.
For a copyable keyless transcript, Web-panel success signals, and a failure matrix, see the first-run demo.
Start a new chat and paste this safe, local-first task:
Build our first taste record for a dense analytics dashboard. Use
palate_addto save one bad example: “all 12 KPI cards have equal visual weight, so the decision signal is buried”; tag itdashboard, hierarchy. Then usepalate_reviewto critique “an analytics dashboard with twelve equal KPI cards, one primary revenue metric, and a small trend chart.” Explain which learned principles you used.
The response should name the matched record and starter principles instead of applying a generic checklist. Open the 👁️ panel to see the example count grow and the new review appear. If you adopt a recommendation, ask the agent to record palate_feedback for that review; only confirmed helpful principles gain evidence.
After the agent has actually inspected a screenshot or page, paste this task:
Analyze the inspected product landing page with
palate_intake. Record at least hierarchy, typography, color, spacing, and interaction observations; stage one example candidate and up to two concrete principle candidates. Compare it withapple-product-storytellingasaligned,conflicts, orinsufficient_evidence, citing the exact reference principle(s). Show me the pending candidates and do not callpalate_decideuntil I explicitly choose accept or reject.
The 👁️ panel should show a new training session and pending count, while the example and principle totals stay unchanged. Once you make an explicit decision, the agent can call palate_decide; training.md preserves both the analysis and the result.
dsh plugin --profile web list dsh-palate.dsh web process after installing; a browser refresh alone cannot load new host code.pnpm install
node build.mjs # esbuild → lib/index.js (host ESM) + lib/client.js (ModuleLoader bundle)
node tests/smoke.mjs # pure-logic checks (no browser needed)
MIT licensed. Ideas and examples welcome — open an issue.
palate_feedback after a review to form effectiveness data.palate_intake stages evidence but does not inspect a raw URL/image or learn automatically; palate_decide needs a clear human accept/reject decision.