Merchant agent

Paths are in the reference repo: merchant_agent/ is merchant-agent/core/merchant_agent/, merchant_agent_runtime/ is merchant-agent/runtime-messages-api/merchant_agent_runtime/, merchant_agent_sdk/ is merchant-agent/runtime-agent-sdk/merchant_agent_sdk/, commerce_common/ is commerce-common/commerce_common/.

What it does

The merchant agent works with an operator inside their back office over one MerchantBackend (merchant_agent/backend.py). Its five flows are the skills in merchant-agent/skills/: performance-insights, catalog-listings, inventory-operations, pricing-promotions, marketing-campaigns. The mechanisms in commerce_common/ are shared; the prompt, tools, gates, and executor are its own, and every tool result comes back inside MERCHANT_FENCE (merchant_agent/fencing.py).

The staged-change contract

1. Every write is a stage_* tool (stage_listing_update, stage_price_update, stage_inventory_action, stage_promotion, stage_campaign). Each returns the staged record and a note, emits a change_update event, and, with stage_shows_preview on (the default, merchant_agent/config.py), emits the change_preview card through the runner present_change_preview uses (MerchantToolExecutor, merchant_agent/executor.py). The MCP server turns it off because its executor events do not reach the operator, so the hosted agent calls present_change_preview; that tool also shows an earlier change again. Live state is untouched until apply_change, which performs the platform write.
2. Staging accepts only listing and campaign ids that tools returned this session, and a content edit also needs a get_listing read; apply_change and discard_change accept only change ids that staging or get_pending_changes returned (merchant_agent/gates.py). Showing the preview card marks nothing approved.
3. check_guardrails in merchant_agent/changes.py runs when a change is staged (ChangeLedger.stage, or your backend's equivalent) and again in check_apply_change (merchant_agent/gates.py) under the config in force at apply time. The limits are max_items_per_change, max_price_delta_pct, max_promotion_discount_pct, max_restock_quantity, max_campaign_budget, and max_listing_field_chars in merchant_agent/config.py; the defaults are demonstration values.
4. protected_fields can never be staged, price_bearing_fields are the fields the price cap reads, and listing_update_blocked_fields may not ride a free-form listing update. A domain that prices under another name (a nightly rate, a monthly fee, a tier price) appends to these tuples and never replaces them.
5. require_host_approval defaults to True: apply_change succeeds only for an id in MerchantSessionState.approved_change_ids, which only host code writes. The three surfaces in the repo are the portal's /changes/{change_id}/apply route (examples/demo_common/merchant.py), MerchantToolset.host_approve (merchant_agent_sdk/merchant_tools.py, prompted per change by merchant-agent/runtime-agent-sdk/main.py), and always_ask on apply_change in the hosted manifest; there the MCP server's config sets require_host_approval=False (its default_config does; a config you pass must too, or every apply is held). An approval typed into the chat sets nothing.
6. require_host_approval and approval_surface render into the static prompt and into refusals, so they are set per deployment and a change to either is a redeploy (commerce-prompt-caching). Name the real surface in approval_surface.
7. A change kind a system does not support raises ChangeNotApplicable (merchant_agent/changes.py) and the executor relays it; the tool stays registered. Applied and discarded changes stay queryable as the audit trail, each stamped with the operator from MerchantSessionContext.operator, which a production host derives from its authentication.

Grounding and follow-through

• A performance question forces get_business_snapshot, and a change request carrying an apply phrase with nothing staged this session forces get_pending_changes (GROUNDING_RULES in merchant_agent/grounding.py; lexicons and flags in merchant_agent/config.py, appended to and never replaced).
• A turn that matched change_requested and ended on bare text, with no stage_* attempt and no present_suggestions close, gets STAGING_FOLLOWTHROUGH_REMINDER once, as a user message (merchant_agent/gates.py; applied in merchant_agent_runtime/orchestrator.py and merchant_agent_sdk/agent.py); the reminder text is excluded from memory extraction.
• present_metrics joins each pick from the snapshot, a queried series, a campaign, or a recorded analysis and drops the rest with a note (resolve_metrics in merchant_agent/enrichment.py), so a card never carries a model-authored figure.

The analysis delegate

• enable_analysis (default off) registers run_analysis, a DelegateExtension built by build_analysis_delegate in merchant_agent_runtime/analysis.py. The delegate's tools are a submit tool, a progress tool, the ANALYSIS_READ_TOOLS (merchant_agent/analysis.py), and a query tool when the backend implements MerchantBackend.execute_analysis_query (analysis_sql_only then leaves the per-series reads off). It holds no staging tool; its result is recorded in seen_analyses and rendered as its own metrics card.
• check_analysis_sql refuses anything other than one SELECT without comments before the backend runs it; the backend owns the read-only role and merchant scoping; analysis_query_timeout_s and cap_analysis_table (max_analysis_rows, max_analysis_table_chars) bound each query, and analysis_timeout_s, max_analysis_iterations, and max_delegate_calls_per_turn bound the run.
• analysis_use_code_execution mounts the hosted sandbox and works on the Anthropic API only (including a Foundry deployment hosted on Anthropic); the query method works everywhere (docs/deployment.md). On the Agent SDK the same contract is a subagent whose tools are exactly the read tools (build_analysis_agent in merchant_agent_sdk/agent.py).
• A scheduled digest is one headless turn of the same agent (merchant-agent/managed-agents/scheduled-digest/run_morning_digest.py).

Memory and components

• Memory is keyed by merchant_id (memory_subject in merchant_agent/executor.py); the extraction prompt in merchant_agent/memory.py admits what the operator stated about running the operation and excludes anything from listings, reviews, buyer messages, or metrics and anything about an identifiable customer. The write filter, retention, delete, purge, and enable_memory are the shared rules (commerce-trust-safety).
• The built-in components (merchant_agent/enrichment.py) are present_metrics, present_digest, present_change_preview, and present_suggestions; under host approval the prompt says no chip approves or applies. A vertical adds its own (present_occupancy_calendar, present_plan_mix, present_event_pacing in the examples) as extensions (commerce-ui-tools).

Marketplaces

• Many sellers' content goes through the one fence; there is no per-seller label and no seller whose text escapes it. In a shopping deployment the seller is a search dimension (SearchFilters.attributes) that the components show, and search_policies results state whose terms they are when platform and seller terms differ.
• Provenance and caps are per session, so a session can stage against the listings its own tools returned and no others.
• Buyer and seller messages reach the model as fenced material; a message that says a change is approved sets no mark.
• The memory subject is the operator's own business, keyed by its merchant_id; a shopper's facts are keyed by the shopper.