Impact analysis agent

Instructions

Purpose

Identify the likely impact of a proposed change before implementation. This skill maps the change request to graph nodes, identifies affected files, functions, APIs, and tests, and produces a reviewable impact map that gates the Build Agent's context.

---

Trigger conditions

Use this skill when:

• A change request targets an existing system.
• The change could affect multiple files or modules.
• Regression risk exists (the change touches shared code).
• The user asks "what will this impact?" or "what could break?"
• The Red Team needs a pre-change risk map.
• Gate 0 (system-understanding-agent) has passed.

---

Inputs

- change_request.md                            required
- .agile-v/phases/00-understanding/system_overview.md    required
- .agile-v/phases/00-understanding/normalized_graph.json optional but strongly preferred
- .understand-anything/diff-overlay.json       optional
- existing requirements (if any)               optional

---

Outputs

All outputs are written to canonical .agile-v/phases/01-impact/.

impact_map.md
affected_components.json
regression_test_candidates.md
change_risk_assessment.md
impact_confidence.md

---

Required behavior

Step 1: Map change request to graph nodes

1. Read the change request.
2. Identify key concepts: changed APIs, functions, modules, behaviors.
3. Search the normalized graph for nodes whose name, path, or summary match.
4. Classify each matched node as directly or indirectly affected.

Step 2: Identify direct impact

Direct impact = nodes that must be modified, added, or removed to satisfy the change request.

For each directly affected node, record:

• Component ID
• File path
• Symbol name
• Reason for impact
• Confidence: High / Medium / Low

Step 3: Identify indirect impact

Indirect impact = nodes that import, call, extend, test, or document a directly affected node.

Walk the graph one or two hops from each directly affected node. Flag indirect nodes with lower confidence unless the relationship is explicit.

Step 4: Identify regression test candidates

For each affected node:

• Find test nodes connected by tests or covers edges.
• Find test files that import or reference the affected file path.
• Mark each test with priority: High (directly tests affected behavior), Medium, Low.

Step 5: Generate risk assessment

For each risk identified:

• Assign Risk ID, description, severity (High/Medium/Low).
• Propose a mitigation.
• Specify a verification approach.

Step 6: Record confidence and assumptions

• State overall confidence: High/Medium/Low.
• List assumptions made during impact analysis.
• Flag low-confidence nodes explicitly.

---

Output template: impact_map.md

See integrations/understand-anything/examples/impact_analysis_example.md.

---

Output schema: affected_components.json

[
  {
    "component_id": "<node id>",
    "path": "<file path>",
    "symbol": "<function or class name, or null>",
    "impact_type": "modify|add|remove|review|test-only|doc-only",
    "reason": "<why this component is affected>",
    "confidence": "high|medium|low"
  }
]

---

Anti-patterns to avoid

• Do not include the entire graph in the impact map. Focus on the change.
• Do not generate implementation code.
• Do not mark files as affected just because they are large or important.
• Do not assign High confidence to indirect dependencies without a clear relationship.
• Do not skip the risk assessment step for L3/L4 changes.

Inherited obligations: apply agile-v-core halt, decision-log, and applicable typed-lineage rules; if AI materially influences these outputs at any risk level, create/update .agile-v/aibom/<task_id>/AI_RUN_MANIFEST.yaml per agile-v-aibom.

---

Gate 1 evidence contribution

The impact map informs Human Gate 1 but does not redefine it. Gate 1 reviews the persisted requirement revision plus independent findings; this supporting evidence must be reviewed where applicable:

• Impact map reviewed and accepted.
• Interface contract reviewed.
• Regression-test plan accepted.

---

Compatible tools

Claude Code, Cursor, Copilot, Codex, Gemini CLI, OpenCode, Cline.