/correlate

Correlate Static + Dynamic Findings

Connect static code findings with dynamic runtime observations to prioritize which issues are real vs theoretical.

Prerequisites

You MUST run /agent-inspector:setup BEFORE proceeding.

This is NOT optional. The setup command will:

1. Check if agent-inspector is already running
2. Auto-detect your LLM provider (OpenAI/Anthropic)
3. Start the server in background if needed
4. Verify MCP connection is working

DO NOT skip this step. If you proceed without running the Preflight Check, MCP tools will fail.

Correlation States

State	Meaning	Priority
VALIDATED	Static finding + runtime evidence match	Highest - actively exploitable
UNEXERCISED	Static finding, never triggered	Test gap, needs coverage
RUNTIME_ONLY	Dynamic issue, no static counterpart	Different fix approach
THEORETICAL	Static finding, safe at runtime	Lower priority

Correlation Workflow

1. Get Workflow State

get_agent_workflow_state(agent_workflow_id)

Verify both static AND dynamic data exist:

• If STATIC_ONLY: Inform user to run dynamic tests first
• If DYNAMIC_ONLY: Run /agent-inspector:scan first
• If COMPLETE: Proceed with correlation

2. Get Static Findings

get_findings(agent_workflow_id, status="OPEN")

3. Get Tool Usage from Runtime

get_tool_usage_summary(agent_workflow_id)

4. Get Correlation Data

get_agent_workflow_correlation(agent_workflow_id)

5. Determine Correlation for Each Finding

Tool-related findings: Check if tool was called at runtime

• Tool called -> VALIDATED
• Tool never called -> UNEXERCISED

Prompt findings: Check if code path was exercised

• Function/route called at runtime -> VALIDATED
• Never executed -> UNEXERCISED

Secret/Data findings: Check if file was loaded at runtime

• File accessed but safe in practice -> THEORETICAL
• Actively used -> VALIDATED

6. Update Each Finding

update_finding_correlation(finding_id, correlation_state="VALIDATED", correlation_evidence={
  "tool_calls": 47,
  "session_count": 15,
  "runtime_observations": "Tool called 47 times across 15 sessions"
})

7. Report to User

Correlation Complete!

Cross-referenced 5 static findings with 25 runtime sessions.

VALIDATED (2) - Active risks confirmed at runtime:
- Tool without constraints: Called 47 times across 15 sessions
- Hardcoded secret: Used in all sessions

UNEXERCISED (3) - Static risks, never triggered:
- Prompt injection in handle_request(): Code path never executed
- Missing validation in process_input(): Function never called
- Shell command in admin_action(): Admin route never accessed

Prioritize fixing VALIDATED issues first - they're actively exploitable.

To fix the most critical: /agent-inspector:fix REC-001

View correlation in UI: http://localhost:7100/agent-workflow/{id}/static-analysis

Example Scenarios

Scenario 1: VALIDATED

Static Finding: TOOL_DANGEROUS_UNRESTRICTED in tools.py
- Function: execute_shell()

Runtime Data: execute_shell called 47 times across 15 sessions

Result: VALIDATED
Evidence: "Tool called 47 times in 15 sessions - active risk!"

Scenario 2: UNEXERCISED

Static Finding: PROMPT_INJECT_DIRECT in agent.py
- Function: handle_request()

Runtime Data: No calls to handle_request observed

Result: UNEXERCISED
Evidence: "Code path never executed in 25 sessions - add test coverage"

Scenario 3: THEORETICAL

Static Finding: SECRET_API_KEY in config.py

Runtime Data:
- config.py loaded at runtime
- But environment variable overrides hardcoded value

Result: THEORETICAL
Evidence: "File loaded but value safely overridden by env var"

After Correlation

Suggest next actions based on results:

• VALIDATED issues -> /agent-inspector:fix REC-XXX
• UNEXERCISED code -> Add test coverage
• Check gate status -> /agent-inspector:gate