Scenario matrix: catalogue seed data sharing

You are in the deep end

Seed-data sharing between stages. The public sequence is the request lifecycle.

Source of truth: agent-data-plane/.../V1__dataplane.sql, agent-capability-registry/.../V1__registry.sql, agent-fabric-scripts/route-runs/job/jobs.json.

route_id	pattern	workflow / manifest	share kind	stage handoff (json_to_http)	notes
`fee_explain`	1	`fee_explain` manifest	`goal_only`	—	Single tool; ingress utterance in `goal`
`llm_pipeline`	2	`llm_pipeline`	`notes_to_llm`	—	Three LLM stages; no domain HTTP
`policy_memo`	2	`policy_memo`	`prefetch_pack`	prefetch → generate (packed chunks)	D6–D7
`account_notify`	2	`account_notify`	`goal_only`	—	`account_id` in job payload
`card_freeze`	2	`card_freeze`	`goal_only`	—	`card_id` / `account_id` in job payload today; identity/limit mocks return prose only
`dispute_intake`	2	`dispute_intake`	`goal_only`	—	`dispute_id` in payload
`purchase_refund`	2	`purchase_refund`	`json_to_http`	`extract_fields` → `match_purchase`: merchant, amount, date	D5 proof route; ingress has `doc_id`, `account_id` only
`pack_then_notify`	2	`pack_then_notify`	`prefetch_pack`	prefetch → notify	D6–D7
`pack_then_freeze`	2	`pack_then_freeze`	`prefetch_pack` + `goal_only`	prefetch then card HTTP chain	D6–D7
`pack_then_review`	2	`pack_then_review`	`prefetch_pack`	prefetch → ocr/score/memo	D6–D7
`clause_lookup`	2	`clause_lookup`	`notes_to_llm`	query_formulation → HTTP search	LLM writes query string
`template_retrieve`	2	`template_retrieve`	`notes_to_llm`	two query_formulation stages
`msa_risk_review`	2	`msa_risk_review`	`notes_to_llm` + `json_to_http` (future)	ocr JSON → score (optional D5 alt)	query stages use notes; score could consume OCR slot
`kyc_onboarding`	2	`kyc_onboarding`	`json_to_http` + `branch` + `human_gate`	risk_score slot → branch keys	D8–D9
`claims_adjudicate`	2	`claims_adjudicate`	`notes_to_llm`	forced retrieve then clause retrieve
`ticket_triage`	3	`ticket_triage`	`goal_only`	—
`product_explain`	3	`product_explain`	`prefetch_pack`		D6–D7
`narrow_review`	3	`narrow_review`	`notes_to_llm`
`contract_review`	3	`contract_review`	`notes_to_llm`
`due_diligence`	3	`due_diligence`	`notes_to_llm`
`fraud_investigate`	1 + agent	manifest + `start_contract_review`	`agent`	parent → child projection	D10
`ops_start_kyc`	1 + agent	manifest + `start_kyc_onboarding`	`agent`	parent → child projection	D10
`agent-chat`, `email_summarize`, `chat_session`, `agent-policy-qa`, `policy_chat`	0	prompt only	`none` / `prefetch_pack` (policy)		Pattern 0
`search_only`, `research_assistant`, `contract_investigation`	1	manifest	`goal_only`		Pattern 1 loops
`conversation` / `long_term` memory flags	—	—	out of plan	—	Shared Memory (future-enhancement)

D5 proof route: purchase_refund — match_purchase requires merchant, amount, and date from the extract_fields classify slot; they are not in the job payload.