anthropics/commerce-agents

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.

ソースを見る
リポジトリの原文

見出し、例、コード、表、リンク、参照画像を含む原文を表示しています。

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.

同じリポジトリから

関連する Skills

すべての Skills
anthropics
公式

commerce-architecture

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.

導入数
1
GitHub Stars
2997
更新日
9月11日
anthropics
公式

commerce-merchant-operations

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.

導入数
1
GitHub Stars
2997
更新日
9月11日
anthropics
公式

commerce-trust-safety

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.

導入数
1
GitHub Stars
2997
更新日
9月11日
anthropics
公式

inventory-operations

Stock, capacity, and availability monitoring, acting on low-stock and slow-mover alerts, triaging order exceptions such as delays, return spikes, and complaints, and the operator's daily briefing; any start-of-day or what-needs-attention rundown is this flow, presented as a digest. Not needed for performance questions with no operational action attached.

導入数
1
GitHub Stars
2997
更新日
9月11日