본문으로 건너뛰기
cccskills
무료GitHub에 공개

commerce-ui-tools

The reference presentation-tool contract, covering server-side enrichment, suggestion chips, the event stream, progressive rendering, both roles' built-in components, and adding a vertical component. Load when building or reviewing how a commerce agent's output reaches a frontend.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

설치 방법 보기

포함된 파일(1)

  • SKILL.md6.7 KB

SKILL.md(원문)

설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.

Presentation tools

Paths are in the reference repo: commerce_common/ is commerce-common/commerce_common/, shopping_agent/ is shopping-agent/core/shopping_agent/, merchant_agent/ is merchant-agent/core/merchant_agent/.

The contract

  • A component is a tool. The model's arguments carry its judgment: which ids, in what order, the reason or the note for each, what to compare on. The server joins every fact (title, price, image, metric value, change record) from records tools returned this session, so the frontend never renders a model-authored value.
  • One spec per component: PresentationComponent in commerce_common/presentation.py names the tool, the component string the host renders, the payload_model that validates the arguments (each role's tools/presentation.py), and the enrich hook (each role's enrichment.py, collected in PRESENTATION_COMPONENTS).
  • run_presentation in commerce_common/presentation.py is the one runner: validate, enrich, emit the ui event. The model reads the executor's displayed_text plus the hook's notes; the enriched payload goes to the host only.
  • An enrich hook drops ids without provenance and appends a note naming them; with nothing left it raises PresentationRefused, which comes back as a held call when it names a gate and as an error otherwise.
  • Shopping joins from ShoppingSessionState.seen_products, the cart, or an order read; merchant joins from MerchantSessionState.latest_snapshot, seen_series, seen_campaigns, seen_analyses, seen_listings, and seen_changes. enrich_change_preview also removes model text whose currency or weekday disagrees with the change record.
  • Composition rules (which component ends a turn, text carries the reasoning, cards carry the data) are static prompt material; each component's when-to-use line is its tool description (commerce-architecture's layer table).

Suggestions

  • present_suggestions is the one tool that carries the turn's chips, up to four; the model calls it in the same round as the turn's last component, or after its text when the turn has none, and a tapped chip is sent as the next user message. No other payload, built-in or extension, carries chips.
  • sanitize_suggestion_chips in commerce_common/fencing.py runs on PresentSuggestionsPayload: invisible and control characters out, whitespace collapsed, empties dropped, 80 characters each, four at most. The payload fails when no chip survives sanitizing.
  • A round of clean presentation calls that includes present_suggestions ends the turn (round_closes_turn in commerce_common/turn.py, close_on_presentation in the config): the Messages API loop stops there, and the Agent SDK runtimes stop through a PostToolBatch hook (close_on_presentation_hook in commerce_common/agent_sdk.py). Under host approval the merchant prompt says no chip approves or applies.

Streaming

The event list is the docstring of commerce_common/streaming.py; to_sse frames each event and a host ignores types it does not know. ui carries one enriched component, ui_partial the same while its call still streams, cart_update the whole cart after a cart write, and change_update a change record after it moved. outcome_events in commerce_common/turn.py stamps each ui event with its call id as stream_id and follows it with the tool_result, whose status is ok, error, or blocked. Every tool that is not a presentation tool may take a status argument first (with_status in commerce_common/execution.py): a few words for the person waiting, which the executor drops before the call runs and the runtime emits as the tool_call event's label; the web layer shows it as the activity line. The MCP servers build their tools without it. The Agent SDK runtimes run the same executor but return the payloads together in result.ui after the turn.

Progressive rendering

A spec with an enrich_partial hook is rendered while its arguments arrive: the request marks its tool eager_input_streaming (with_eager_input in commerce_common/prompt_assembly.py), the orchestrator parses the buffer with parse_partial_json (a string still being written is left out with its key), calls enrich_partial (commerce_common/presentation.py), and yields ui_partial whenever partial_signature changes (StreamedRound.frame in commerce_common/turn.py; every visible change with eager_partial_frames on the config); the final ui event carries the same stream_id and replaces it. Input that streams as text that is not JSON comes back to the model as an error result and the round goes on (StreamedRound). Partial hooks are synchronous and join from session state only (partial_products, partial_comparison, partial_plan, partial_guide in shopping_agent/enrichment.py; partial_metrics, partial_digest, partial_change_preview in merchant_agent/enrichment.py).

Built-in components

  • Shopping (shopping_agent/enrichment.py): present_products, present_comparison, present_plan, present_guide, present_order_status, checkout (renders the cart; charges nothing), present_suggestions, and present_disclosure, registered only with enable_disclosures and filled from StorefrontBackend.get_disclosure.
  • Merchant (merchant_agent/enrichment.py): present_metrics, present_digest, present_change_preview, present_suggestions.

Adding a vertical component

  • Build a PresentationExtension (commerce_common/presentation.py): the tool description and input_schema the model sees, the payload_model that validates the same shape, the component name, and the enrich hook.
  • Pass it as extra_presentation_tools to ShoppingAgent or MerchantAgent (examples/travel/api/main.py); on the Agent SDK and MCP paths, subclass the toolset or pass executor_class= so the executor receives it as extensions. build_tools rejects a name that collides with a built-in, and the extension's bytes join the cached tool list (commerce-prompt-caching).
  • The seven in the repo: examples/travel/api/itinerary.py, examples/travel/api/occupancy.py, examples/telecom/api/plan_matrix.py, examples/telecom/api/plan_mix.py, examples/entertainment/api/venue_map.py, examples/entertainment/api/hold_view.py, examples/entertainment/api/event_pacing.py. Each keeps its types and its card in the vertical; nothing is added to a core package.

리뷰

아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.

댓글

로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.

아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.

보내면 바로 공개됩니다. 이메일 주소나 전화번호 같은 개인 정보는 쓰지 마세요. URL은 2개까지입니다.

0 / 1,000자

이 양식은 reCAPTCHA로 보호되며 Google 개인정보처리방침 및 서비스 약관이 적용됩니다.

같은 저장소의 스킬

Creating and improving listing content, covering titles, descriptions, attribute completeness, categorization fixes, image callouts written as text, edits written from material the operator supplies, and content-quality audits and bulk fixes across many listings. Not needed for questions about how a listing is performing.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

anthropics/commerce-agents3.3천2026년 10월 2일 업데이트

How the reference commerce agents of either role are structured, covering the loop, where each rule lives, skills, the backend interface, delegates, and model fields. Load when designing or reviewing the structure of a shopping or merchant agent.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

anthropics/commerce-agents3.3천2026년 10월 2일 업데이트

Authoring and running behavioral evals for a shopping or merchant agent, covering the case shape, authoring rules, code graders and judges, the run pattern, and poisoned fixtures. Load when writing eval cases or rubrics or deciding how a suite runs.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

anthropics/commerce-agents3.3천2026년 10월 2일 업데이트

The reference merchant agent, covering its flows, staged changes and host approval, metrics grounding, the analysis delegate, store memory, components, and marketplace rules. Load when building or reviewing an agent for the operators of a store, property, subscriber base, or venue.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

anthropics/commerce-agents3.3천2026년 10월 2일 업데이트

The reference agents' cache-stable request assembly, covering the static system and per-request context split, the fixed tool list, the rolling conversation breakpoint, which config fields are prompt bytes, and verification. Load when writing or reviewing a commerce agent's prompt assembly, when cache reads are zero, or when a turn's latency or cost is the question.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

anthropics/commerce-agents3.3천2026년 10월 2일 업데이트

The rules the reference agents enforce in code for third-party content, writes, grounding, identity, and memory, each with its module, plus two adversarial-eval rules. Load when handling untrusted tool results, guarding a write tool, or scoping what an agent remembers.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

anthropics/commerce-agents3.3천2026년 10월 2일 업데이트

anthropics의 스킬 모두 보기

이 스킬의 문제 신고