Add integration tests for agent with guardrails

[apetraru-uipath]

Integration Tests for Agent Guardrails

Overview

Adds comprehensive integration tests for agent guardrails, verifying proper invocation and enforcement at Agent, LLM, and Tool scopes.

Files Added

• tests/cli/test_agent_with_guardrails.py (1325 lines) - 8 integration tests
• tests/cli/mocks/joke_agent_with_guardrails.py (333 lines) - Mock agent with 6 guardrails
• tests/cli/mocks/joke_agent_uipath.json (139 lines) - Guardrail configurations
• tests/cli/mocks/joke_agent_langgraph.json (8 lines) - LangGraph config
• tests/cli/conftest.py (updated) - Shared fixtures for mocking

Test Coverage

Agent-Level Guardrails

Test	Guardrail	Stage	Action	Trigger	Validates
test_pii_guardrail_not_triggered	PII Detection	POST_EXECUTION	Block	No PII	Execution succeeds
test_pii_guardrail_triggered	PII Detection	POST_EXECUTION	Block	Email in output	Execution blocked

LLM-Level Guardrails

Test	Guardrail	Stage	Action	Trigger	Validates
test_prompt_injection_guardrail_triggered	Prompt Injection	PRE_EXECUTION	Block	Malicious prompt	LLM never invoked, execution blocked
test_llm_pii_escalation_guardrail_hitl	PII Detection	POST_EXECUTION	Escalate	Email in output	HITL triggered, user approves, execution continues
test_llm_pii_escalation_guardrail_rejected	PII Detection	POST_EXECUTION	Escalate	Email in output	HITL triggered, user rejects, execution stops

Tool-Level Guardrails

Test	Guardrail	Stage	Action	Trigger	Validates
test_tool_guardrail_filter_output	Custom (word match)	POST_EXECUTION	Filter	"donkey" in output	Field removed, execution succeeds
test_tool_guardrail_block_execution	Custom (word match)	PRE_EXECUTION	Block	"forbidden" in input	Tool never invoked, execution blocked
test_tool_pii_guardrail_triggered	PII Detection	PRE_EXECUTION	Block	Email in input	Tool never invoked, execution blocked

Guardrails Tested

Built-in Validators

• PII Detection (Agent, LLM & Tool scopes) - Detects Email, Address, Person with 0.5 threshold
• Prompt Injection (LLM scope) - Detects malicious prompts with 0.5 threshold

Custom Deterministic Guardrails

• Filter Action - Removes input_phrase field when containing "donkey"
• Block Action - Blocks tool execution when input contains "forbidden"

Actions Tested

• Block Action - Stops execution at Agent, LLM, or Tool scope
• Filter Action - Removes fields from outputs
• Escalate Action (HITL) - Triggers human approval with both approval and rejection flows