Fix sensitive_data_leakage tool context not reaching agent callback in Foundry path

[slister1001]

In the Foundry execution path, agent-specific context items (with tool_name fields like document_client_smode) were stored in SeedObjective.metadata but never propagated to the callback's context parameter. The agent received an empty context dict and could not recognize the injected tools, causing all sensitive_data_leakage objectives to score 0.0 (false negative).

Add a fallback in _CallbackChatTarget._send_prompt_impl() that extracts context_items from request.prompt_metadata when labels['context'] is empty. This matches the ACA runtime behavior where FunctionTool definitions are dynamically created from context items.

• Add prompt_metadata fallback for context extraction
• Add seed objectives with tool_name/context_type for E2E coverage
• Update E2E test to assert tool context delivery to callback
• Add unit tests for metadata fallback and labels precedence

[Copilot]

Pull request overview

This PR addresses a Foundry-path regression where agent tool context (e.g., tool_name/context_type context items used by sensitive-data leakage objectives) was not making it into the callback context parameter, leading to false-negative scoring. The changes attempt to propagate tool context by creating context SeedPrompts and extracting tool context from conversation history in _CallbackChatTarget.

Changes:

• Create context SeedPrompts for standard attacks so tool context can be recovered from prepended conversation history.
• Extract tool context from conversation history prompt_metadata in _CallbackChatTarget as a fallback when labels don’t contain context.
• Expand unit/E2E coverage (new seed objectives with tool context; assertions that callback receives tool context; PyRIT memory reset between E2E tests).

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/red_team/_foundry/_dataset_builder.py	Adds standard-attack context SeedPrompt creation and an objective SeedPrompt to drive Foundry context/tool propagation.
sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/red_team/_callback_chat_target.py	Extracts tool context from conversation history prompt_metadata and passes it via context['contexts'] to the callback.
sdk/evaluation/azure-ai-evaluation/tests/unittests/test_redteam/test_dataset_builder_binary_path.py	Adds unit tests verifying context SeedPrompt creation and sequencing expectations.
sdk/evaluation/azure-ai-evaluation/tests/unittests/test_redteam/test_callback_chat_target.py	Adds unit test verifying tool context extraction from prepended conversation history.
sdk/evaluation/azure-ai-evaluation/tests/e2etests/test_red_team_foundry.py	Adds an autouse fixture to reset PyRIT singleton memory; updates sensitive-data-leakage E2E to assert tool context delivery.
sdk/evaluation/azure-ai-evaluation/tests/e2etests/data/redteam_seeds/sensitive_data_leakage_seeds.json	Adds seed objectives that include tool context payloads for E2E validation.

[nagkumar91]

LGTM — two-part fix is clean. Context SeedPrompts + callback extraction solves the tool context gap. Good unit + e2e coverage.

[nagkumar91]

Re-reviewed after updates. Good improvements — stricter metadata checks, context filtering in _build_messages_from_pieces (was a gap in v1), is_objective tagging, and seed data fix. LGTM.