Codebase Onboarding

Codebase Onboarding

Summary

This skill transforms raw code into structured intelligence: architecture maps, entry points, data flows, security surfaces, and onboarding confidence scores.

Domain: knowledge

Skill Identity

Purpose

Rapidly acquire a mental model of any unfamiliar codebase — from a 500-line script to a 100M+ line monorepo. This skill transforms raw code into structured intelligence: architecture maps, entry points, data flows, security surfaces, and onboarding confidence scores.

In security contexts, codebase onboarding is the first step before code audits, vulnerability research, exploit development, and supply chain analysis.

Scope Modes

Three modes based on what you need to know, not how much time you have:

Choose mode before starting. Switching modes mid-session is allowed but requires re-baselining.

Phase 0: Search-First (All Modes)

Before reading any code:

1. Find existing documentation

   • README, CONTRIBUTING, ARCHITECTURE, docs/, wiki/
   • OpenAPI/Swagger specs, Protobuf definitions
   • CI/CD config (reveals build structure and test commands)

2. Identify the skeleton

   • package.json, go.mod, Cargo.toml, requirements.txt, pom.xml, build.gradle
   • Entry points: main(), app.py, index.js, server.go
   • Config files: .env.example, config/, settings.py

3. Detect framework signatures

   • Import patterns, directory names, config file names
   • See Language Support section for framework detection by language

4. Run static index (Targeted/Comprehensive modes)

   • ctags, cscope, or language server for symbol maps
   • File tree with line counts: find . -name "*.py" | xargs wc -l | sort -rn | head -50

Methodology

Phase 1: Orientation

• Count files, LOC, and language distribution
• Identify primary language(s) and detect framework
• Map top-level directory structure to functional areas
• Find entry points and main execution paths

Phase 2: Architecture Mapping

• Trace request/data flow from entry point to persistence
• Identify layers: API → Business Logic → Data Access → Storage
• Map inter-service dependencies (microservices) or module boundaries (monolith)
• Detect shared libraries, utilities, middleware

Phase 3: Security Surface Analysis

• Authentication and authorization code locations
• Input validation and sanitization points
• External integrations (APIs, databases, message queues)
• Secret/credential handling
• Known dangerous patterns by language

Phase 4: Deep Dive (Comprehensive mode only)

• Critical path tracing for key operations
• Data flow for sensitive operations (payments, auth, PII)
• Dependency vulnerability surface (outdated packages, CVE exposure)
• Test coverage gaps that indicate under-reviewed areas

Phase 5: Knowledge Consolidation

• Generate structured output (see Output Format below)
• Record confidence scores per subsystem
• Identify gaps for follow-up research
• Hand off to knowledge-ops for persistence

Language Support

Tier 1 — Full Automation Support (75–90% automated)

Tier 2 — Partial Automation Support (50–70% automated)

Tier 3 — Manual-Heavy (20–40% automated)

Kotlin, Scala, Swift, Objective-C, Erlang, Elixir, Haskell, COBOL, and other languages require primarily manual analysis. Use Phase 0 docs-first approach and lean on test files for behavior discovery.

100M+ Line Strategy

For very large codebases (100M+ LOC):

1. Index First: Run ctags/cscope before reading any files
2. Smart Sampling: Focus on files with highest churn (git log), most imports, or security-critical paths
3. Divide & Conquer: Treat each top-level module as a separate Targeted-mode session
4. Boundary Focus: Understand module interfaces (APIs, contracts) before internals
5. Avoid Full Reads: Never attempt to read entire large files; sample entry, middle, and exit sections

Output Format

Confidence Score

Rate onboarding completeness per area:

Report as: Overall: 72/100 | Auth: 85 | Data Layer: 60 | API Surface: 78 | Internal Logic: 65

Structured Intelligence Package

jsonCopy code
{
  "project": "target-name",
  "analyzed_at": "2026-05-11",
  "mode": "Comprehensive",
  "language_primary": "Go",
  "framework": "Gin + GORM",
  "loc_total": 85000,
  "confidence": {
    "overall": 72,
    "auth": 85,
    "data_layer": 60,
    "api_surface": 78,
    "internal_logic": 65
  },
  "entry_points": ["cmd/server/main.go", "cmd/worker/main.go"],
  "architecture": "Monolith with event-driven background workers",
  "security_surfaces": {
    "auth": "JWT via middleware/auth.go",
    "input_validation": "Partial — missing in admin routes",
    "secrets": "env vars via config/config.go",
    "dangerous_patterns": ["SQL concatenation in reports/query.go:145"]
  },
  "gaps": ["Payment flow not traced", "gRPC service definitions not reviewed"],
  "next_steps": ["Audit reports/query.go for SQLi", "Review payment/ module"]
}

Architecture Diagram (Mermaid)

mermaidCopy code
graph TD
    Client -->|HTTP| Router[Gin Router]
    Router --> AuthMW[Auth Middleware]
    AuthMW --> Handler[Request Handlers]
    Handler --> Service[Business Logic]
    Service --> Repo[Repository Layer]
    Repo --> DB[(PostgreSQL)]
    Service --> Queue[Redis Queue]
    Queue --> Worker[Background Worker]

Use Cases

1. Pre-Audit Onboarding: Map attack surface before security audit
2. Exploit Research: Locate vulnerable code patterns in target software
3. Supply Chain Analysis: Understand third-party library integration points
4. Incident Response: Rapidly understand compromised codebase structure
5. CVE Reproduction: Locate affected code for known vulnerabilities

Hacker Laws Alignment

• Law 1 (Know Your Battlefield): You cannot exploit what you don't understand
• Law 3 (Intelligence Over Force): Systematic mapping beats random file reading
• Law 9 (Systematic Over Random): Phase-based approach ensures nothing is missed

Detection Methods

Code Repository Access Anomalies

• Off-hours access: Agent accessing repositories at 3 AM local; outside typical dev hours.
• Mass cloning: Sudden spike in git clone operations; large data egress from source control.
• Unusual file reads: Reads of .env, secrets.yml, *.pem, id_rsa files by automated agents.
• Cross-repo correlation: Same agent accessing many unrelated repositories.

Agent Activity Indicators

• Tool call patterns: Sequential calls to read_file → search_code → send_http (exfil chain).
• Token consumption: Sustained >100K tokens per session; context-stuffing attack pattern.
• Filesystem traversal: Reads outside declared working directory; attempts to access /etc/, ~/.ssh/.

SIEM Detection Rules

• Splunk SPL: index=agent tool="read_file" | where match(path, "(\.env|secrets|\.pem)")
• GitHub Audit Log: Alert on git.clone events from new IP geolocation.

Defense Evasion Techniques

Stealth Enumeration

• Slow & low traversal: Pace file reads below rate-limit threshold.
• Use legitimate paths: Access only typical project files; avoid /etc/, ~/.ssh/ (suspicious).
• Off-hours operation: Run during typical dev hours to blend with normal activity.
• Cache file contents: Avoid re-reading same files; reduces read-count anomaly.

Exfiltration Stealth

• Piggyback on legitimate commits: Hide exfil data in legitimate-looking diff.
• Distributed commits: Spread exfil across multiple commits/PRs (one chunk each).
• Encoding tricks: Base64 / hex encode sensitive data to evade DLP scanning.
• Side-channel exfil: Encode data in commit timing / message length.

Integration

• Feed output to knowledge-ops for cross-session persistence
• Use deep-research to research identified frameworks and known vulnerabilities
• Use exa-search to find CVEs for detected dependency versions
• Record security surfaces in mission memory before starting active testing