test: Add CPT E2E tests for document extraction in tool call results

[Copilot]

Adds end-to-end CPT tests validating that the AI Agent connector correctly delivers PDF documents returned by mock tool workers to the LLM, and that the LLM can reason about their content.

New files

• ai-agent-e2e-document.bpmn — Process with 5 HTTP document tools (Download_Report, Search_Documents, Fetch_Report, Get_External_Document, Download_Corrupted_Report) + GetDateAndTime script task
• document-tool-call-results/*.pdf — Three PDF fixtures with distinct, queryable content (Project Zypherion / 847 employees / Dr. Kael Thrennix)
• AiAgentE2EDocumentTestIT.java — 7 test scenarios covering the document delivery surface

Test scenarios

#	Scenario	Assertion
1	Single CamundaDocumentReferenceModel from tool result	LLM quotes project name + date from PDF
2	List of documents from tool result	LLM summarises both PDFs
3	Documents in nested map (attachments list + metadata.cover)	LLM references all three document facts
4	InlineDocumentReferenceModel (no Zeebe fetch needed)	LLM reports inline text content
5	Mixed: text tool + document tool in same turn	LLM uses both tool results
6	No document tools used	Process completes normally, no document errors
7	Broken document reference (non-existent ID)	hasActiveIncidents() — connector exhausts retries

Document delivery mechanism

Mock workers store PDFs in embedded Zeebe via newCreateDocumentCommand(), then return the reference wrapped in CamundaDocumentReferenceModel — which serialises with the "camunda.document.type": "camunda" discriminator required by ObjectDeserializer. The Docker connector fetches content from embedded Zeebe via gRPC.

var response = camundaClient.newCreateDocumentCommand()
    .content(new ByteArrayInputStream(pdfBytes))
    .fileName(fileName).contentType("application/pdf")
    .send().join();

return new CamundaDocumentReferenceModel(
    response.getStoreId(), response.getDocumentId(), response.getContentHash(),
    new CamundaDocumentMetadataModel("application/pdf", null, null, fileName, null, null, null));

Tests are gated on OPENAI_API_KEY and run under the it-real-llm Spring profile, matching the existing AiAgentE2ETestIT pattern.

[Copilot]

Pull request overview

Adds a new CPT end-to-end test suite that exercises the AI Agent connector’s ability to ingest and reason over documents returned via tool call results (including single docs, lists, nested structures, inline docs, and failure cases).

Changes:

• Added a dedicated BPMN process (ai-agent-e2e-document) exposing multiple “document tools” (HTTP JSON tasks) plus a time tool.
• Added a new real-LLM gated integration test (AiAgentE2EDocumentTestIT) covering 7 document-delivery scenarios, including a broken-reference incident case.
• Added PDF fixtures under document-tool-call-results/ used by the mock tool workers as document content.

Reviewed changes

Copilot reviewed 2 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
connectors-e2e-test/connectors-e2e-test-agentic-ai/src/test/resources/ai-agent-e2e-document.bpmn	New E2E process definition that models document-returning tools for the AI Agent connector to call.
connectors-e2e-test/connectors-e2e-test-agentic-ai/src/test/java/io/camunda/connector/e2e/agenticai/e2e/AiAgentE2EDocumentTestIT.java	New CPT E2E ITs validating document extraction/propagation from tool call results and LLM reasoning over PDFs.
connectors-e2e-test/connectors-e2e-test-agentic-ai/src/test/resources/document-tool-call-results/project-launch.pdf	PDF fixture used to validate single-document and nested-document reasoning.
connectors-e2e-test/connectors-e2e-test-agentic-ai/src/test/resources/document-tool-call-results/headcount-report.pdf	PDF fixture used to validate multi-document reasoning (e.g., headcount fact extraction).
connectors-e2e-test/connectors-e2e-test-agentic-ai/src/test/resources/document-tool-call-results/author-info.pdf	PDF fixture used to validate nested metadata/cover document reasoning.