docs(example): add extract -> query -> act dynamic-workflow walkthrough

examples/invoice-extraction-workflow.md demonstrates data_extractor (spatial
to file) -> query_extraction (low-confidence + region slices) -> act via
ai_redactor/document_signer, keeping the full payload out of context.

🔮 View transcript: https://nutrient-agentlogs.dev/s/duk4x9tr3rmlnxta7c4bwk6o

Author: jdrhyne

README.md

@@ -223,6 +223,8 @@ Use `query_extraction` to pull just the elements you need from a saved spatial f
 
 > **Note:** Extracted content returned inline (Markdown output, or `query_extraction` results) enters the conversation and may be logged by the host. For sensitive documents, prefer spatial output to a file plus scoped `query_extraction` calls.
 
+For a worked extract → query → act walkthrough, see [examples/invoice-extraction-workflow.md](examples/invoice-extraction-workflow.md).
+
 ## Usage Examples
 
 These examples assume your files live inside the configured sandbox and that you use explicit paths.

examples/invoice-extraction-workflow.md

@@ -0,0 +1,79 @@
+# Dynamic workflow: extract → query → act
+
+This example shows how an AI agent chains the Data Extraction tools with the
+existing document tools to process an invoice **without ever loading the full
+extraction into context**. It is the pattern dynamic workflows are built on:
+extract structured data, branch on it, then act.
+
+**Prerequisites**
+
+- `NUTRIENT_EXTRACTION_API_KEY` (Data Extraction API key, starts with `pdf_live_`) for `data_extractor`.
+- `NUTRIENT_DWS_API_KEY` (or OAuth) for the `ai_redactor` / `document_signer` "act" steps.
+- `SANDBOX_PATH` set to a directory containing `invoice.pdf`.
+
+## Step 1 — Extract structured elements to a file
+
+The agent calls `data_extractor` in `understand` mode with spatial output. The
+element list (with coordinates and confidence) is written to a file; only a
+compact summary comes back.
+
+```jsonc
+// tool: data_extractor
+{ "filePath": "invoice.pdf", "mode": "understand", "format": "spatial", "outputPath": "invoice.elements.json" }
+```
+
+```
+Extracted 142 elements across 2 page(s) and wrote the full spatial JSON to invoice.elements.json (38217 bytes).
+Element types: paragraph: 96, table: 2, keyValueRegion: 18, picture: 1.
+Low-confidence elements (confidence < 0.6): 7.
+Retrieve specific elements with query_extraction ...
+```
+
+The agent now knows the shape of the document — and that **7 fields are
+low-confidence** — without 142 elements entering the conversation.
+
+## Step 2 — Branch on the result with `query_extraction`
+
+The summary flagged low-confidence elements, so the agent pulls just those to
+decide whether the document needs human review:
+
+```jsonc
+// tool: query_extraction
+{ "filePath": "invoice.elements.json", "minConfidence": 0, "elementTypes": ["keyValueRegion"], "limit": 50 }
+```
+
+It can also grab a specific region — e.g. the totals box in the bottom-right of
+page 2 — to read the amount due:
+
+```jsonc
+// tool: query_extraction
+{ "filePath": "invoice.elements.json", "pages": [1], "region": { "x": 1200, "y": 2000, "width": 600, "height": 400 } }
+```
+
+Only the handful of elements the agent actually needs — with their text and
+coordinates — enter context.
+
+## Step 3 — Act with the existing tools
+
+Branching on what it found, the agent acts:
+
+- **Low-confidence or sensitive fields →** redact before sharing:
+
+  ```jsonc
+  // tool: ai_redactor
+  { "filePath": "invoice.pdf", "criteria": "Bank account and routing numbers", "outputPath": "invoice-redacted.pdf" }
+  ```
+
+- **Clean and approved →** sign it:
+
+  ```jsonc
+  // tool: document_signer
+  { "filePath": "invoice.pdf", "outputPath": "invoice-signed.pdf", "signatureOptions": { "signatureType": "cms" } }
+  ```
+
+## Why this is the workflow primitive
+
+The agent reasons over **structure and coordinates** (counts, confidence,
+regions) rather than a wall of text, retrieves only the slices it needs, and
+hands off to deterministic document operations. The large, sensitive payload
+stays on disk; the conversation stays small and auditable.