Improve Risk Agent with Risk Report and better criteria generation prompt

[NISH1001]

Summary 📝

This PR upgrades the RiskAgent guardrail into a comprehensive safety diagnostic tool. Instead of returning a binary pass/fail or a simple score, the agent now performs per-risk detection and leverages a secondary RiskReportAgent to generate detailed, technically grounded reports when violations occur. This ensures that safety failures are actionable and backed by specific evidence from the evaluated content.

Details ⚙️

1. Active Harm Detection Prompting:

• Refactored RISK_SYSTEM_PROMPT to prioritize "Presence of Risk" over "Absence of Safety."
• Criteria are now designed to be unforgiving and failure-oriented, ensuring that subtle scientific or technical risks are not overlooked.

2. Automated Risk Reporting:

• Integrated a new internal _RiskReportAgent that parses failed High Importance criteria.
• The agent maps specific evidence (verbatim quotes/indices) from the risky_content to the risk definition, providing a structured "Criterion-Evidence-Analysis" report.

3. Adaptive DAG Verdict Logic:

• Implemented dynamic bucket generation for the weighted_ratio. The number of verdict buckets now scales with the number of risks evaluated (max(5, num_risks + 4)), providing more granular scoring resolution.

4. Resilient Verdict Extraction:

• Introduced a dual-layered extraction strategy in risk_agent.py.
• Primary: Direct extraction from DeepEval TaskNode._output.
• Fallback: A robust Regex-based parser that scans verbose_logs to reconstruct verdicts if node states are lost during DAG execution.

5. Type-Safe Refactoring:

• Updated the internal configuration and DAG builder to use RiskCategory enums consistently, improving maintainability and reducing string-parsing errors.

Bugfixes 🐛

• Node State Persistence: Resolved an issue where DeepEval's DAGMetric would occasionally fail to populate outputs on original node references, leading to "False Pass" scenarios.
• Weight Resolution: Fixed a bug in weight mapping where unspecified risks would default to zero weight; they now correctly default to 1.0.

Checks

• Closed Risk eval in guardrails.py and new "agent" that generates risk report for failed risks #249
• Tested Changes (549 tests passed)
• Stakeholder Approval

---

[github-actions]

✅ Tests passed

📊 Test Results

• Passed: 549
• Failed: 0
• Skipped: 23
• Warnings: 133
• Coverage: 78%

Branch: enhance/risk-agent-guardrail
PR: #298
Commit: f3225b8

📋 Full coverage report and logs are available in the workflow run.

[muthukumaranR]

Thanks for working on the risk report. Lgtm

[github-actions]

✅ Tests passed

📊 Test Results

• Passed: 549
• Failed: 0
• Skipped: 23
• Warnings: 134
• Coverage: 77%

Branch: enhance/risk-agent-guardrail
PR: #298
Commit: e10fea2

📋 Full coverage report and logs are available in the workflow run.

[github-actions]

✅ Tests passed

📊 Test Results

• Passed: 549
• Failed: 0
• Skipped: 23
• Warnings: 134
• Coverage: 78%

Branch: enhance/risk-agent-guardrail
PR: #298
Commit: ee9fb73

📋 Full coverage report and logs are available in the workflow run.

[github-actions]

✅ Tests passed

📊 Test Results

• Passed: 549
• Failed: 0
• Skipped: 23
• Warnings: 134
• Coverage: 78%

Branch: enhance/risk-agent-guardrail
PR: #298
Commit: 1a052fb

📋 Full coverage report and logs are available in the workflow run.

[github-actions]

✅ Tests passed

📊 Test Results

• Passed: 549
• Failed: 0
• Skipped: 23
• Warnings: 132
• Coverage: 78%

Branch: enhance/risk-agent-guardrail
PR: #298
Commit: d11357f

📋 Full coverage report and logs are available in the workflow run.

[NISH1001]

Note:

• I haven't included _confirm_risk_score() method from Risk eval in guardrails.py and new "agent" that generates risk report for failed risks #249 because now in this PR the verdicts and criteria are all generated from DAG metric instead of parsing verbose logs. So there's no need to recompute/revalidate.