メニュー
Deterministic Accounts-Payable processing pipeline. Walks Extract → Validate → Post in sequence via ADK's SequentialAgent. No LLM at the workflow level — the order is code, not prose. Use as a sub-skill of ap-orchestrator; not intended for direct user invocation.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Conversational front door for the Accounts-Payable workflow. Chats with the user about invoices and AP policy, then transfers to the deterministic `ap-pipeline` SequentialAgent when an invoice needs processing. Never reproduces extracted invoice fields itself — the pipeline owns that.
Invoice field-extraction specialist for Accounts Payable. Reads the already-parsed content of an invoice document and returns clean, typed business fields (vendor, line items, totals, GL codes). Operates on a second layer above the AILANG Parse / DocParse step that runs at ingest time — does NOT re-parse the document.
Action specialist for Accounts Payable. Takes a validated invoice and either posts it to the ERP ledger or routes it to a human approval queue, then emits the final Invoice Review Card to the workspace pane. Final stage of the `ap-pipeline` SequentialAgent.
Grounded validation specialist for Accounts Payable. Checks extracted invoice fields against the vendor master, open purchase orders, and the tax/approval policy indexed in Vertex AI Search. Detects duplicates, PO mismatches, out-of-policy line items, and incorrect tax. Use after extraction, before any posting decision.
| name | ap-pipeline |
| description | Deterministic Accounts-Payable processing pipeline. Walks Extract → Validate → Post in sequence via ADK's SequentialAgent. No LLM at the workflow level — the order is code, not prose. Use as a sub-skill of ap-orchestrator; not intended for direct user invocation. |
| display_name | AP Pipeline (Workflow) |
| metadata | {"author":"aitana","version":"0.1","agentType":"sequential","subSkills":["invoice-extractor","ap-validator","ap-poster"]} |
This skill has no instruction body. A SequentialAgent has no model — it
walks its sub_agents deterministically. The "workflow" lives in code
(see google.adk.agents.sequential_agent.SequentialAgent), not in a
prompt that could be ignored.
The intake gate (backend/adk/agent.py::_make_intake_gate) short-circuits
the pipeline with a friendly message when session state has no document
loaded — graceful degradation, no confused output when a user invokes
the pipeline without context.
See docs/design/forks/gde-ap-agent/multi-agent-workflow-pipeline.md for the full design rationale, protocol-stack mapping, and submission narrative.
