Paths: File paths (shared/, references/, ../ln-*) are relative to skills repo root. If not found at CWD, locate this SKILL.md directory and go up one level for repo root.

Dependency Graph Auditor

L3 Worker that builds and analyzes the module dependency graph to enforce architectural boundaries.

Purpose & Scope

• Worker in ln-640 coordinator pipeline – invoked by ln-640-pattern-evolution-auditor
• Build module dependency graph from import statements (Python, TS/JS, C#, Java)
• Detect circular dependencies: pairwise (HIGH) + transitive via DFS (CRITICAL)
• Validate boundary rules: forbidden, allowed, required (per dependency-cruiser pattern)
• Calculate Robert C. Martin metrics (Ca, Ce, Instability) + Lakos aggregate (CCD, NCCD)
• Validate Stable Dependencies Principle (SDP)
• Support baseline/freeze for incremental legacy adoption (per ArchUnit FreezingArchRule)
• Adaptive: 3-tier architecture detection — custom rules > docs > auto-detect

Out of Scope (owned by other workers):

• I/O isolation violations (grep-based) -> ln-642-layer-boundary-auditor
• API contract violations -> ln-643-api-contract-auditor
• Code duplication -> ln-623-code-principles-auditor

Input (from ln-640)

- architecture_path: string    # Path to docs/architecture.md
- codebase_root: string        # Root directory to scan

# Domain-aware (optional, from coordinator)
- domain_mode: "global" | "domain-aware"   # Default: "global"
- current_domain: string                   # e.g., "users", "billing" (only if domain-aware)
- scan_path: string                        # e.g., "src/users/" (only if domain-aware)

# Baseline (optional)
- update_baseline: boolean                 # If true, save current state as baseline

When domain_mode=”domain-aware”: Use scan_path instead of codebase_root for all Grep/Glob operations. Tag all findings with domain field.

Workflow

Phase 1: Discover Architecture (Adaptive)

MANDATORY READ: Load references/dependency_rules.md — use 3-Tier Priority Chain, Architecture Presets, Auto-Detection Heuristics.

Architecture detection uses 3-tier priority — explicit config wins over docs, docs win over auto-detection:

# Priority 1: Explicit project config
IF docs/project/dependency_rules.yaml exists:
  Load custom rules (modules, forbidden, allowed, required)
  SKIP preset detection

# Priority 2: Architecture documentation
ELIF docs/architecture.md exists:
  Read Section 4.2 (modules, layers, architecture_type)
  Read Section 6.4 (boundary rules, if defined)
  Map documented layers to presets from dependency_rules.md
  Apply preset rules, override with explicit rules from Section 6.4

# Priority 3: Auto-detection from directory structure
ELSE:
  scan_root = scan_path IF domain_mode == "domain-aware" ELSE codebase_root
  Run structure heuristics:

  signals = {}
  IF Glob("**/domain/**") AND Glob("**/infrastructure/**"):
    signals["clean"] = HIGH
  IF Glob("**/controllers/**") AND Glob("**/services/**") AND Glob("**/repositories/**"):
    signals["layered"] = HIGH
  IF Glob("**/features/*/") with internal structure:
    signals["vertical"] = HIGH
  IF Glob("**/adapters/**") AND Glob("**/ports/**"):
    signals["hexagonal"] = HIGH
  IF Glob("**/views/**") AND Glob("**/models/**"):
    signals["mvc"] = HIGH

  IF len(signals) == 0:
    architecture_mode = "custom"
    confidence = "LOW"
    # Only check cycles + metrics, no boundary presets
  ELIF len(signals) == 1:
    architecture_mode = signals.keys()[0]
    confidence = signals.values()[0]
    Apply matching preset from dependency_rules.md
  ELSE:
    architecture_mode = "hybrid"
    confidence = "MEDIUM"
    # Identify zones, apply different presets per zone (see dependency_rules.md Hybrid section)
    FOR EACH detected_style IN signals:
      zone_path = identify_zone(detected_style)
      zone_preset = load_preset(detected_style)
      zones.append({path: zone_path, preset: zone_preset})
    Add cross-zone rules: inner zones accessible, outer zones forbidden to depend on inner

Phase 2: Build Dependency Graph

MANDATORY READ: Load references/import_patterns.md — use Language Detection, Import Grep Patterns, Module Resolution Algorithm, Exclusion Lists.

scan_root = scan_path IF domain_mode == "domain-aware" ELSE codebase_root

# Step 1: Detect primary language
tech_stack = Read(docs/project/tech_stack.md) IF exists
  ELSE detect from file extensions: Glob("**/*.py", "**/*.ts", "**/*.cs", "**/*.java", root=scan_root)

# Step 2: Extract imports per language
FOR EACH source_file IN Glob(language_glob_pattern, root=scan_root):
  imports = []

  # Python
  IF language == "python":
    from_imports = Grep("^from\s+([\w.]+)\s+import", source_file)
    plain_imports = Grep("^import\s+([\w.]+)", source_file)
    imports = from_imports + plain_imports

  # TypeScript / JavaScript
  ELIF language == "typescript" OR language == "javascript":
    es6_imports = Grep("import\s+.*\s+from\s+['\"]([^'\"]+)['\"]", source_file)
    require_imports = Grep("require\(['\"]([^'\"]+)['\"]\)", source_file)
    imports = es6_imports + require_imports

  # C#
  ELIF language == "csharp":
    using_imports = Grep("^using\s+([\w.]+);", source_file)
    imports = using_imports

  # Java
  ELIF language == "java":
    java_imports = Grep("^import\s+([\w.]+);", source_file)
    imports = java_imports

  # Step 3: Filter internal only (per import_patterns.md Exclusion Lists)
  internal_imports = filter_internal(imports, scan_root)

  # Step 4: Resolve to modules
  FOR EACH imp IN internal_imports:
    source_module = resolve_module(source_file, scan_root)
    target_module = resolve_module(imp, scan_root)
    IF source_module != target_module:
      graph[source_module].add(target_module)

Phase 3: Detect Cycles (ADP)

Per Robert C. Martin (Clean Architecture Ch14): “Allow no cycles in the component dependency graph.”

# Pairwise cycles (A <-> B)
FOR EACH (A, B) WHERE B IN graph[A] AND A IN graph[B]:
  cycles.append({
    type: "pairwise",
    path: [A, B, A],
    severity: "HIGH",
    fix: suggest_cycle_fix(A, B)
  })

# Transitive cycles via DFS (A -> B -> C -> A)
visited = {}
rec_stack = {}

FUNCTION dfs(node, path):
  visited[node] = true
  rec_stack[node] = true

  FOR EACH neighbor IN graph[node]:
    IF NOT visited[neighbor]:
      dfs(neighbor, path + [node])
    ELIF rec_stack[neighbor]:
      cycle_path = extract_cycle(path + [node], neighbor)
      IF len(cycle_path) > 2:  # Skip pairwise (already detected)
        cycles.append({
          type: "transitive",
          path: cycle_path,
          severity: "CRITICAL",
          fix: suggest_cycle_fix_transitive(cycle_path)
        })

  rec_stack[node] = false

FOR EACH module IN graph:
  IF NOT visited[module]:
    dfs(module, [])

# Folder-level cycles (per dependency-cruiser pattern)
folder_graph = collapse_to_folders(graph)
Repeat DFS on folder_graph for folder-level cycles

Cycle-breaking recommendations (from Clean Architecture Ch14):

1. DIP — extract interface in depended-upon module, implement in depending module
2. Extract Shared Component — move shared code to new module both depend on
3. Domain Events / Message Bus — for cross-domain cycles, decouple via async communication

Phase 4: Validate Boundary Rules

# Load rules from Phase 1 discovery
# rules = {forbidden: [], allowed: [], required: []}

# Check FORBIDDEN rules
FOR EACH rule IN rules.forbidden:
  FOR EACH edge (source -> target) IN graph:
    IF matches(source, rule.from) AND matches(target, rule.to):
      IF rule.cross AND same_group(source, target):
        CONTINUE  # cross=true means only cross-group violations
      boundary_violations.append({
        rule_type: "forbidden",
        from: source,
        to: target,
        file: get_import_location(source, target),
        severity: rule.severity,
        reason: rule.reason
      })

# Check ALLOWED rules (whitelist mode)
IF rules.allowed.length > 0:
  FOR EACH edge (source -> target) IN graph:
    allowed = false
    FOR EACH rule IN rules.allowed:
      IF matches(source, rule.from) AND matches(target, rule.to):
        allowed = true
        BREAK
    IF NOT allowed:
      boundary_violations.append({
        rule_type: "not_in_allowed",
        from: source,
        to: target,
        file: get_import_location(source, target),
        severity: "MEDIUM",
        reason: "Dependency not in allowed list"
      })

# Check REQUIRED rules
FOR EACH rule IN rules.required:
  FOR EACH module IN graph WHERE matches(module, rule.module):
    has_required = false
    FOR EACH dep IN graph[module]:
      IF matches(dep, rule.must_depend_on):
        has_required = true
        BREAK
    IF NOT has_required:
      boundary_violations.append({
        rule_type: "required_missing",
        module: module,
        missing: rule.must_depend_on,
        severity: "MEDIUM",
        reason: rule.reason
      })

Phase 5: Calculate Graph Metrics

MANDATORY READ: Load references/graph_metrics.md — use Metric Definitions, Thresholds per Layer, SDP Algorithm, Lakos Formulas.

# Per-module metrics (Robert C. Martin)
FOR EACH module IN graph:
  Ce = len(graph[module])                          # Efferent: outgoing
  Ca = count(m for m in graph if module in graph[m])  # Afferent: incoming
  I = Ce / (Ca + Ce) IF (Ca + Ce) > 0 ELSE 0      # Instability

  metrics[module] = {Ca, Ce, I}

# SDP validation (Stable Dependencies Principle)
FOR EACH edge (A -> B) IN graph:
  IF metrics[A].I < metrics[B].I:
    # Stable module depends on less stable module — SDP violation
    sdp_violations.append({
      from: A, to: B,
      I_from: metrics[A].I, I_to: metrics[B].I,
      severity: "HIGH"
    })

# Threshold checks (per graph_metrics.md, considering detected layer)
FOR EACH module IN metrics:
  layer = get_layer(module)  # From Phase 1 discovery
  thresholds = get_thresholds(layer)  # From graph_metrics.md

  IF metrics[module].I > thresholds.max_instability:
    findings.append({severity: thresholds.severity, issue: f"{module} instability {I} exceeds {thresholds.max_instability}"})
  IF metrics[module].Ce > thresholds.max_ce:
    findings.append({severity: "MEDIUM", issue: f"{module} efferent coupling {Ce} exceeds {thresholds.max_ce}"})

# Lakos aggregate metrics
CCD = 0
FOR EACH module IN graph:
  DependsOn = count_transitive_deps(module, graph) + 1  # Including self
  CCD += DependsOn

N = len(graph)
CCD_balanced = N * log2(N)  # CCD of balanced binary tree with N nodes
NCCD = CCD / CCD_balanced IF CCD_balanced > 0 ELSE 0

IF NCCD > 1.5:
  findings.append({severity: "MEDIUM", issue: f"Graph complexity (NCCD={NCCD:.2f}) exceeds balanced tree threshold (1.5)"})

Phase 6: Baseline Support

Inspired by ArchUnit FreezingArchRule — enables incremental adoption in legacy projects.

baseline_path = docs/project/dependency_baseline.json

IF file_exists(baseline_path):
  known = load_json(baseline_path)
  current = serialize_violations(cycles + boundary_violations + sdp_violations)

  new_violations = current - known
  resolved_violations = known - current

  # Report only NEW violations as findings
  active_findings = new_violations
  baseline_info = {new: len(new_violations), resolved: len(resolved_violations), frozen: len(known - resolved_violations)}

  IF input.update_baseline == true:
    save_json(baseline_path, current)

ELSE:
  # First run — report all
  active_findings = all_violations
  baseline_info = {new: len(all_violations), resolved: 0, frozen: 0}
  # Suggest: output note "Run with update_baseline=true to freeze current violations"

Phase 7: Score + Return

MANDATORY READ: Load shared/references/audit_scoring.md for unified scoring formula.

penalty = (critical * 2.0) + (high * 1.0) + (medium * 0.5) + (low * 0.2)
score = max(0, 10 - penalty)

Note: When baseline is active, penalty is calculated from active_findings only (new violations), not frozen ones.

{
  "category": "Dependency Graph",
  "score": 6.5,
  "total_issues": 8,
  "critical": 1, "high": 3, "medium": 3, "low": 1,
  "architecture": {
    "detected": "hybrid",
    "confidence": "MEDIUM",
    "zones": [
      {"path": "src/core/", "preset": "layered"},
      {"path": "src/features/", "preset": "vertical"}
    ]
  },
  "graph_stats": {
    "modules_analyzed": 12,
    "edges": 34,
    "cycles_detected": 2,
    "ccd": 42,
    "nccd": 1.3
  },
  "cycles": [
    {
      "type": "transitive",
      "path": ["auth", "billing", "notify", "auth"],
      "severity": "CRITICAL",
      "fix": "Apply DIP: extract interface in auth, implement in notify"
    }
  ],
  "boundary_violations": [
    {
      "rule_type": "forbidden",
      "from": "domain",
      "to": "infrastructure",
      "file": "domain/user.py:12",
      "severity": "CRITICAL",
      "reason": "Domain must not depend on infrastructure"
    }
  ],
  "sdp_violations": [
    {
      "from": "domain",
      "to": "utils",
      "I_from": 0.2,
      "I_to": 0.8,
      "severity": "HIGH"
    }
  ],
  "metrics": {
    "users": {"Ca": 3, "Ce": 5, "I": 0.625},
    "billing": {"Ca": 1, "Ce": 7, "I": 0.875}
  },
  "baseline": {"new": 3, "resolved": 1, "frozen": 4},
  "findings": [],
  "domain": "users",
  "scan_path": "src/users/"
}

Critical Rules

• Adaptive architecture — never assume one style; detect from project structure or docs
• 3-tier priority — custom rules > architecture.md > auto-detection
• Hybrid support — projects mix styles; apply different presets per zone
• Custom = safe mode — if no pattern detected, only check cycles + metrics (no false boundary violations)
• Internal only — exclude stdlib, third-party from graph (only project modules)
• Baseline mode — when baseline exists, report only NEW violations
• Cycle fixes — always provide actionable recommendation (DIP, Extract Shared, Domain Events)
• File + line — always provide exact import location for violations

Definition of Done

• Architecture discovered (adaptive 3-tier detection applied)
• Dependency graph built from import statements (internal modules only)
• Circular dependencies detected (pairwise + transitive DFS + folder-level)
• Boundary rules validated (forbidden + allowed + required)
• Metrics calculated (Ca, Ce, I per module + CCD, NCCD aggregate)
• SDP validated (stable modules not depending on unstable)
• Baseline applied if exists (only new violations reported)
• If domain-aware: all Grep/Glob scoped to scan_path, findings tagged with domain
• Score calculated per audit_scoring.md
• Result returned to coordinator

Reference Files

• Boundary rules & presets: references/dependency_rules.md
• Metrics & thresholds: references/graph_metrics.md
• Import patterns: references/import_patterns.md
• Scoring algorithm: shared/references/audit_scoring.md

Version: 1.0.0 Last Updated: 2026-02-11