icsf-cli 0.1.12

ICSF – Intelligent Code Security & Fixing Platform (CLI)

ICSF – Intelligent Code Security & Fixing Platform

ICSF is a full-stack, AI-powered platform that automates the discovery, analysis, and remediation of security vulnerabilities in Java/Maven codebases. It combines a multi-agent cognitive fixing pipeline with an autonomous self-healing testing framework (Atlas) to deliver a closed-loop system: from vulnerability report → verified, PR-ready fix.

---

Table of Contents

• Architecture Overview
• End-to-End Application Flow
• Backend Deep Dive
  • FastAPI Main (main.py)
  • Configuration & Credentials
  • Pydantic Data Models
  • Services Layer
  • Agents Layer (Cognitive Fixing Loop)
  • Atlas Subsystem (Self-Healing Testing)
• Frontend Deep Dive
• RAG (Retrieval-Augmented Generation)
• AI / LLM Integration
• Input Requirements
• Technical Stack
• Getting Started
• Project Structure
• API Reference

---

🏗️ Architecture Overview

ICSF follows a layered, modular architecture:

┌─────────────────────────────────────────────────────────────────────┐
│                    Frontend (Streamlit – 5676 lines)                │
│   Premium dark-mode dashboard · Real-time progress · Lineage graph │
├─────────────────────────────────────────────────────────────────────┤
│                   Backend (FastAPI – 1636 lines)                    │
│         REST API · WebSocket/SSE · Request ID middleware            │
├────────────┬──────────────────────┬─────────────────────────────────┤
│  Services  │       Agents         │      Atlas Subsystem            │
│ (14 files) │  (Cognitive Loop)    │  (Self-Healing Testing)         │
│            │  5 agents + helpers  │  14 sub-packages                │
├────────────┼──────────────────────┼─────────────────────────────────┤
│            │   AWS Bedrock (LLM)  │  SQLite RAG Store               │
│            │  Claude 3.5 Sonnet   │  Titan Embeddings               │
│            │  Titan Embeddings    │  Cosine Similarity Search       │
└────────────┴──────────────────────┴─────────────────────────────────┘

Key Design Principles

Principle	Implementation
Single AI Provider	All LLM/embedding calls route through AWS Bedrock only
Multi-Agent Pipeline	5 specialized agents, each with a single responsibility
Self-Healing	Atlas auto-repairs build failures and test regressions
Cross-Repo Awareness	Dependency analysis spans across multiple repositories
Cost Control	CostGuardService enforces per-run budget limits
Resilience	Retry with exponential backoff + circuit breakers

---

🔄 End-to-End Application Flow

flowchart TD
    %% Entry & Configuration
    U((User)) -->|1. Setup| UI[Streamlit Dashboard]
    UI -->|2. Upload CSV| BE[FastAPI Backend]

    subgraph "Phase I: Discovery & Mapping"
        direction TB
        BE -->|3. Fetch Projects| GH[GitHub API]
        GH -->|4. List Repos| REPO[(Repository Store)]
        VS[Vulnerability Mapper] -->|5. Map Assets| FT[(Local Workspace)]
    end

    BE --> VS

    subgraph "Phase II: Quality Baseline – Atlas"
        direction LR
        BASE[Lightweight Baseline] -->|6. Verify Build| COV[Capture Coverage]
    end

    VS --> BASE

    subgraph "Phase III: Impact Analysis"
        direction TB
        DS[Dependency Service] -->|7. Map Blast Radius| MAP[Call Tree & Usage]
    end

    COV --> DS

    subgraph "Phase IV: Cognitive Fixing Loop"
        direction TB
        CC[Code Context Agent] --> FS[Fix Strategy Agent]
        FS --> CF[Code Fixer Agent]
        CF --> SV[Safety Validator Agent]
    end

    MAP -->|8. Start Fix| CC
    CC <-->|Rich Context| DS
    CF -->|9. Apply Fix| FT

    subgraph "Phase V: Self-Healing Pipeline – Atlas"
        direction TB
        BM[BuildMechanic] --> TH[TestHealer]
        TH --> TG[AI Test Generator]
        TG --> VR[Validation Report]
    end

    SV -->|10. Final Verify| BM
    BM -->|Self-Heal Build| FT
    FT --> TH

    subgraph "Phase VI: Delivery & Sync"
        direction TB
        VR --> PR[Batch PR Manager]
        PR -->|11. Create Sync PR| GH
        PR -->|12. Update UI| UI
    end

    %% RAG Knowledge Loop
    TG -.->|Save Success Patterns| RAG[(SQLite RAG Store)]
    RAG -.->|Context Enrichment| CC

Phase-by-Phase Walkthrough

Phase	What Happens	Key Service
I. Discovery	Upload CSV → fetch repos from GitHub API → match vulnerability file paths to repositories using intelligent path normalization	VulnerabilityService, GitHubService
II. Baseline	Run mvn compile test on the unmodified code to establish Ground Truth coverage & build health	AtlasService.run_baseline_only()
III. Impact Analysis	Parse Java files, build global dependency graph, find cross-repo callers of the vulnerable method	DependencyService
IV. Cognitive Fixing	4-agent pipeline: Analyze context → Plan strategy → Generate fix → Validate safety	FixOrchestrator + 4 Agents
V. Self-Healing	BuildMechanic auto-repairs compilation; TestHealer fixes broken tests; AI generates new security-targeted tests	Atlas pipeline
VI. Delivery	Aggregate all fixes into a single PR with rich markdown body, push to GitHub	BatchPRService, PRManagerService

---

🔧 Backend Deep Dive

1. FastAPI Main (main.py) — 1636 lines

The central orchestration hub. Defines the REST API, middleware, and all endpoint routes.

Startup & Middleware

Component	Purpose
_startup_validation()	Smoke-checks AWS + GitHub credentials on boot
RequestIDMiddleware	Injects a UUID X-Request-ID header into every request for log correlation
CORS middleware	Configurable via ALLOWED_ORIGINS env var

Pydantic Request/Response Models (inline)

Model	Fields	Used By
GitHubRepoRequest	username, email, token	POST /api/github/repos
Repository	id, name, full_name, clone_url, language, etc.	All repo endpoints
RepositoriesResponse	username, total_repos, repositories[]	Repo listing
TestingRequest	repo_url, repo_path, fixed_files, create_pr, vulnerability, etc.	Testing pipeline
Vulnerability	file_name, line_no	Vulnerability mapping
MappedVulnerability	repo: Repository, vulnerabilities[]	Mapping results

API Endpoints

Method	Route	Description
GET	/	Root welcome
GET	/api/health	Health check for Docker/LB probes
GET	/api/credentials/github	Retrieve loaded GitHub credentials
GET	/api/credentials/verify	Debug credential loading
POST	/api/github/repos	Fetch repos (POST with body)
GET	/api/github/repos	Fetch repos (GET with query params)
POST	/api/vulnerabilities/map	Upload CSV + map vulnerabilities to repos
POST	/api/dependencies/analyze	Analyze dependencies for a single vulnerability
POST	/api/dependencies/batch-analyze	Batch dependency analysis for multiple vulnerabilities
POST	/api/fix/orchestrate	Run the full multi-agent fixing pipeline
POST	/api/pr/create	Create a single PR with fixed code
POST	/api/testing/start	Start async testing pipeline job
GET	/api/testing/job/{job_id}	Poll job status
GET	/api/testing/stream/{job_id}	SSE event stream for real-time progress
GET	/api/testing/runs	List recent pipeline runs
POST	/api/testing/run	Legacy sync testing endpoint
POST	/api/fix/batch	Batch fix multiple vulnerabilities
POST	/api/pr/merge	Merge PR with conflict resolution
POST	/api/pr/check-mergeability	Check PR mergeability
POST	/api/pr/create-batch	Create single aggregated PR for all fixes

---

2. Configuration & Credentials

config.py — The Config Class

Attribute	Source	Default
AWS_ACCESS_KEY_ID	.env	—
AWS_SECRET_ACCESS_KEY	.env	—
AWS_REGION	.env	us-east-1
AWS_SESSION_TOKEN	.env	None
BEDROCK_MODEL_ID	.env	anthropic.claude-3-5-sonnet-20240620-v1:0
BEDROCK_EMBED_MODEL_ID	.env	amazon.titan-embed-text-v1

Key Methods:

• get_github_credentials(force_reload=False) — Reads credentials.yaml for GitHub PAT, username, email
• validate_bedrock_credentials() — Returns (is_valid, error_msg) tuple
• get_bedrock_config() — Returns dict with access_key, secret_key, region

credentials.yaml

github:
  token: ghp_xxxxx
  username: your-username
  email: [e-mail]

---

3. Pydantic Data Models (models/agent_models.py)

These 10 models define the complete data flow through the multi-agent pipeline:

flowchart LR
    VFR[VulnerabilityFixRequest] --> VA[VulnerabilityAnalysis]
    VA --> CC[CodeContext]
    CC --> FS[FixStrategy]
    FS --> CF[CodeFix]
    CF --> SV[SafetyValidation]
    SV --> FE[FixExplanation]
    VFR --> FOR[FixOrchestrationResult]
    VA --> FOR
    CC --> FOR
    FS --> FOR
    CF --> FOR
    SV --> FOR
    FE --> FOR

Model	Role	Key Fields
VulnerabilityFixRequest	Input to pipeline	vulnerability_type, file_path, line_number, repo_path
VulnerabilityAnalysis	Agent 1 output	severity, security_impact, root_causes, fix_category
CodeContext	Agent 2 output	code_snippet, class_name, dependent_files_intra/inter, data_flow
FixStrategy	Agent 3 output	fix_approach, code_changes_plan, files_to_modify_primary/secondary
CodeFix	Agent 4 output	fixed_code (Dict[path→code]), diff, change_summary, reasoning
SafetyValidation	Agent 5 output	validation_status, correctness_score, breaking_changes, issues_found
FixExplanation	Agent 6 output	vulnerability_summary, fix_explanation, markdown_report
FixOrchestrationResult	Complete result	Aggregates all agent outputs + overall_status, errors
VulnerabilitySeverity	Enum	CRITICAL, HIGH, MEDIUM, LOW, INFO
ValidationResult	Enum	APPROVED, REJECTED, NEEDS_REVIEW

---

4. Services Layer (services/)

The services layer contains 14 files providing the core business logic.

4.1 bedrock_service.py — AWS Bedrock LLM Wrapper (439 lines)

The primary AI gateway used by the Agents layer.

Method	Description
invoke_claude(prompt, model_id, max_tokens, temperature, system_prompt)	Synchronous Claude invocation via Bedrock invoke_model API
ainvoke_claude(...)	Async wrapper using asyncio.to_thread
invoke_llama(prompt, ...)	Llama 3 70B invocation (different payload format)
ainvoke_llama(...)	Async Llama wrapper
embed_text(text, embed_model_id)	Generate embeddings via Amazon Titan Embed
invoke_model(model_id, prompt, ...)	Generic dispatcher — auto-selects Claude/Llama based on model ID
test_connection()	Smoke test with simple prompt

Supported Model Constants:

Constant	Model ID
CLAUDE_3_5_SONNET	anthropic.claude-3-5-sonnet-20240620-v1:0
CLAUDE_3_SONNET	anthropic.claude-3-sonnet-20240229-v1:0
LLAMA_3_70B	meta.llama3-70b-instruct-v1:0

4.2 github_service.py — GitHub API Client (455 lines)

Method	Description
verify_token_and_get_user(username)	Validate PAT + retrieve user info
get_user_by_username(username)	Public API user lookup
get_username_from_email(email)	Reverse email → username lookup
get_user_organizations()	List authenticated user's orgs
get_organization_repositories(org_name)	List all repos in an org (paginated)
get_all_repositories(username, include_private, include_orgs)	Aggregated repo fetch (user + org repos)
get_repository_details(owner, repo_name)	Single repo metadata
get_repository_file_tree(owner, repo_name, branch)	Recursive file tree via Git Tree API

4.3 vulnerability_service.py — CSV Parser & Repo Mapper (833 lines)

Parses vulnerability reports from Fortify, Checkmarx, SonarQube, Snyk, etc.

Method	Description
parse_csv_file(file_content, filename)	Parse CSV into DataFrame; auto-detects column names
extract_repo_name_from_url(url)	Handles HTTPS, SSH, .git suffix URLs
normalize_repo_identifier(repo_name, repo_url)	Lowercase normalization for matching
normalize_file_path(file_path)	Cross-platform path normalization
get_path_variations(file_path)	Generates multiple path format variations for fuzzy matching
match_file_in_repo(file_name, repo_files)	Intelligent file matching with early-exit optimization
clone_repository_and_get_files(repo_url, clone_dir)	Git clone + file tree extraction
map_vulnerabilities_to_repos(df, repositories, repo_files_map, clone_repos)	Core mapping: CSV rows → repository + file matches

4.4 dependency_service.py — Java Dependency Graph Engine (2037 lines)

The largest service file. Performs static analysis of Java source code and Maven POM files.

Method	Description
parse_java_file(file_path)	Extracts package, imports, classes, methods, interfaces, method calls via regex/AST parsing
parse_pom_xml(pom_path)	Extracts groupId, artifactId, version, dependencies, parent POM
find_java_files(repo_path)	Recursive .java file discovery
find_pom_files(repo_path)	Recursive pom.xml file discovery
build_global_dependency_graph(all_repos, artifact_index)	Builds both intra-repo and inter-repo dependency edges. Node identity = (repo_name, file_path)
build_intra_repo_dependencies(repo_path)	File-to-file dependencies within a single repo (import-graph)
find_maven_artifact_for_file(file_path, repo_path)	Map a .java file to its Maven artifact coordinates
find_cross_repo_dependent_files(...)	Inter-repo blast radius: finds files in other repos that depend on the vulnerable file
_build_cross_repo_dependency_chains(...)	Transitive dependency chain traversal across repos (up to max_depth=5)
build_maven_artifact_index(all_repos)	Maps (groupId, artifactId) → repository metadata

4.5 fix_orchestrator.py — Multi-Agent Pipeline Controller (564 lines)

Coordinates the sequential agent execution:

Agent 2 (Code Context) → Agent 3 (Fix Strategy) → Agent 4 (Code Fix) → Agent 5 (Safety Validator)

Method	Description
orchestrate_fix(request, stop_at_agent, validate_fix, max_validation_retries, all_repositories_info)	Main entry point. Runs agents 2→5 sequentially, with optional validation loop
get_orchestration_status(result)	Human-readable status summary
_create_skeleton_analysis(request)	Generates a default VulnerabilityAnalysis from request data

Supports stop_at_agent for incremental testing (e.g., run only agents 2-3).

4.6 batch_fix_service.py — Batch Vulnerability Processing (753 lines)

Processes multiple vulnerabilities in sequence or with controlled concurrency.

Method	Description
_process_vulnerability_fix(current_idx, vuln_idx, vuln, ...)	Process a single vulnerability with logging
_run_testing_agent(repo_path, repo_name, phase, fixed_files)	Run Atlas testing (baseline or validation phase)
fix_single_vulnerability(vulnerability, repo_path, ...)	Single fix with full orchestration
fix_batch_vulnerabilities(vulnerabilities, repo_path, ..., max_concurrent, auto_create_pr, run_tests_after_fix)	Main batch entry point. Runs baseline → sequential fixes → validation → optional PR

Workflow: Baseline → Fix each vulnerability → Run Atlas validation → Create aggregated PR

4.7 pr_manager_service.py — Git & PR Operations (1359 lines)

Complete Git workflow management.

Method	Description
_run_git_command(repo_path, command, timeout)	Safe subprocess wrapper for git commands
create_branch(repo_path, branch_name, base_branch)	Create and checkout new branch
commit_changes(repo_path, files_to_commit, commit_message, author_name, author_email)	Stage + commit with configurable author
push_branch(repo_path, branch_name, remote)	Git push to remote
create_pull_request(owner, repo, title, body, head_branch, base_branch)	GitHub API PR creation
_validate_compilation(repo_path, files_modified)	Best-effort Maven/Gradle compilation check
_clean_code_before_validation(code)	Removes markdown artifacts, separator lines from LLM output
_validate_java_code(code, file_path)	Basic Java structure validation (package, class, brace matching)
apply_fixed_code(repo_path, files_modified, fixed_code_map)	Write fixed code to files with validation
create_pr_for_fix(repo_path, repo_owner, ..., include_all_repo_changes)	Complete workflow: apply → branch → commit → push → PR

4.8 batch_pr_service.py — Aggregated PR Creation (429 lines)

Method	Description
_extract_files_and_code(fix_result)	Parse fix result into files_modified + fixed_code_map
create_single_pr(fix_result, ...)	Create PR for one vulnerability
create_batch_prs(successful_fixes, ...)	One PR per vulnerability
create_single_batch_pr(successful_fixes, ..., test_results)	Single aggregated PR combining all fixes + test results

4.9 atlas_service.py — Testing Pipeline Façade (430 lines)

Bridges the backend API to the Atlas subsystem.

Method	Description
_check_required_tools()	Validates git, mvn, java are on PATH
run_testing_pipeline(repo_url, create_pr, job_id)	Full pipeline on remote repo (clone → test → coverage → PR)
run_testing_pipeline_local(repo_path, repo_url, fixed_files, ...)	Full pipeline on already-cloned local repo
run_baseline_only(repo_path, repo_url)	Lightweight: build + existing tests + coverage — NO AI

4.10 fix_validator_service.py — Post-Fix Validation (277 lines)

Method	Description
validate_fix(repo_path, files_modified)	Run Maven build + tests on the fixed repo. Uses BuildMechanic for auto-repair
get_validation_feedback(validation_result)	Generate feedback string for retry loop

4.11 job_manager.py — Async Job & SSE Streaming (86 lines)

Method	Description
create_job()	Create UUID-identified job with asyncio.Queue
update_job(job_id, status, message, progress)	Update status + push to SSE queue
end_job(job_id)	Signal [DONE] to SSE stream
stream_job_events(job_id)	Async generator for StreamingResponse

4.12 run_history.py — SQLite Run Persistence (115 lines)

Method	Description
create_run(repo_url, repo_path)	Insert new run record
update_run(run_id, status, result_data, error_msg, cost)	Update with test/coverage/regression/quality gate reports
get_recent_runs(limit)	Fetch recent runs with JSON report parsing

Schema: pipeline_runs(run_id, repo_url, repo_path, status, start_time, end_time, total_cost, test_report, coverage_report, regression_report, quality_gate_report, error_message)

4.13 cost_guard.py — LLM Cost Limiter (50 lines)

Method	Description
start_run(run_id)	Initialize per-run cost tracking
add_cost(run_id, prompt_tokens, completion_tokens, model_id)	Accumulate cost; returns False if budget exceeded
get_run_cost(run_id)	Query accumulated cost

Pricing: Claude 3.5 Sonnet — $0.003/1K prompt tokens, $0.015/1K completion tokens. Default budget: $5.00/run.

---

5. Agents Layer (agents/) — Cognitive Fixing Loop

5.1 code_context_agent.py — Blast Radius Mapper (642 lines)

Purpose: Understand the full context around a vulnerability — local code, dependent files, data flow.

Method	Logic
_read_file_with_context(file_path, line_number, context_lines)	Extract code snippet + surrounding context. Includes class/method even if vulnerability is in imports
_extract_class_and_method(file_path, line_number, code_content)	Regex-based Java class/method extraction
_analyze_data_flow_and_usage(code_snippet, vulnerability_type, ...)	LLM call: Analyze how user input flows from source → vulnerable sink
_analyze_method_usage_in_dependents(vulnerable_class, ..., dependent_files)	Check if the fix will break dependent files by analyzing their imports and usage
_discover_other_repositories(current_repo_path)	Scan temp_cloned_repos/ directory for cross-repo analysis
analyze(request, vulnerability_analysis, all_repositories_info)	Main entry: reads file, finds dependents (intra + inter-repo), runs LLM data flow analysis

5.2 fix_strategy_agent.py — Surgical Planner (633 lines)

Purpose: Design a backward-compatible fix plan.

Method	Logic
_get_available_java_files(repo_path, max_files)	Inventory of Java files for validation
_analyze_file_imports_and_usage(repo_path, file_path)	Static import analysis to find related files
_build_strategy_prompt(request, analysis, context)	Constructs a detailed LLM prompt with vulnerability info, dependents, constraint rules
_parse_strategy_response(response_content)	JSON extraction from LLM response
analyze(request, vulnerability_analysis, code_context)	LLM call: Generate fix strategy; categorizes files as Primary (logic change) or Secondary (impacted usage)

Key Decision Logic:

• If a method is called by 50+ files → force backward-compatible fix (overloaded method, not breaking change)
• Uses FrameworkDetector for framework-specific recommendations (Spring Security, Jakarta, etc.)

5.3 code_fix_agent.py — Multi-File Code Generator (1071 lines)

Purpose: Generate actual fixed Java code across multiple files.

Method	Logic
_read_file(file_path)	Read source file
_find_nearest_pom_xml(repo_path, file_rel_path)	Walk up directories to find pom.xml
_project_allows_spring_security(repo_path, file_rel_path, original_code)	Check if Spring Security dependencies exist before generating SS code
_dependency_constraints_text(repo_path, ...)	Generate constraint text for LLM prompt
_postprocess_for_project_dependencies(code, ...)	Deterministic safety net: strip Spring Security constructs if project doesn't include it
_generate_diff(original, fixed, file_path)	Unified diff generation
_clean_generated_code(code, file_path)	Aggressive cleanup: removes markdown, <thinking> blocks, ensures valid Java
_generate_fixed_code(original_code, request, ...)	LLM call: Generate complete fixed file with prompt-chain reasoning
fix_code(request, ..., fix_strategy)	Main entry: fixes ALL files in files_to_modify_primary, runs post-processing

Uses ImportManager.add_missing_imports() and SyntaxValidator.validate() for post-processing.

5.4 safety_validator_agent.py — Logic Gate (371 lines)

Purpose: Verify the fix is correct, introduces no regressions.

Method	Logic
_format_fixed_code(fixed_code)	Format code dict for display
_format_dependent_files_for_validation(code_context)	Format dependent files context
_build_validation_prompt(request, ..., code_fix)	Comprehensive validation prompt
_parse_validation_response(response_content)	Extract structured validation data
_normalize_validation_data(parsed)	Ensure correct types for downstream consumption
validate(request, ..., code_fix)	LLM call: Returns APPROVED/REJECTED/NEEDS_REVIEW with correctness_score (0-1)

5.5 codebase_analysis_agent.py — Repository Intelligence (594 lines)

Purpose: Deep structural analysis of the codebase (similar to AI coding assistants).

Method	Logic
analyze_codebase_structure(repo_path, focus_file)	Full repo analysis with in-memory cache (TTL=300s)
find_dependent_files(repo_path, target_file, max_depth)	Find all files depending on target file
analyze_code_flow(repo_path, file_path, line_number)	Data flow analysis around a specific line
_analyze_architecture(repo_path, java_files)	Detect project layers (Controller, Service, DAO, etc.)
_build_dependency_graph(repo_path, java_files)	Build import-based dependency graph
_detect_patterns(repo_path, java_files)	Detect design patterns (Singleton, Factory, Builder, Observer)
_parse_java_file(file_path)	Extract package, imports, classes (regex-based)

5.6 agent_improvements.py — Helper Utilities (368 lines)

Four static helper classes:

Class	Purpose
ImportManager	Auto-detect and add missing Java imports (maps common security classes to their import statements)
SyntaxValidator	Basic Java syntax validation (brace matching, package declaration, class structure)
FrameworkDetector	Detect frameworks in pom.xml (Spring Boot, Spring Security, JPA, Jackson, etc.) with framework-specific fix recommendations
ContextEnhancer	Extract full method/class definitions from source code for enhanced prompt context

---

6. Atlas Subsystem (atlas/) — Self-Healing Testing Framework

Atlas is a comprehensive, autonomous testing and quality assurance pipeline with 14 sub-packages.

6.1 orchestrator/run_pipeline.py — Pipeline Core (1412 lines)

The brain of Atlas. Orchestrates the entire testing lifecycle.

Function	Description
run_full_pipeline(repo_url, ...)	Clone remote repo → full pipeline
run_full_pipeline_local(repo_path, ...)	Full pipeline on local repo
run_baseline_only(repo_path, ...)	Lightweight: build + test + coverage only
_run_baseline_phase(repo_path, ...)	Build (with restricted auto-fix) + existing tests + JaCoCo coverage
_run_validation_phase(repo_path, ...)	Diff-aware test generation, healing, regression detection (630+ lines)
_run_full_pipeline_core(...)	Core pipeline: baseline → validation → quality gate → PR
evaluate_quality_gate(coverage, unit, min_coverage_pct, max_failures)	Pass/fail decision on release readiness
calculate_regression_report(state_mgr, ...)	Compare current vs baseline to detect regressions/improvements
_calculate_usage(llm)	Compute estimated cost from Bedrock token metrics
run_organization_pipeline(org_url, ...)	Scan entire GitHub org: run pipeline on each Java/Maven repo

6.2 agents/build_mechanic.py — Build Failure Auto-Repair (1133 lines)

The SRE agent. Diagnoses and fixes compilation failures.

Method	Description
analyze(stdout, stderr)	Parse Maven build output → BuildDiagnosis (root cause, confidence, hints)
generate_fix(diagnosis, workspace_path, ...)	LLM call: Generate concrete fix (file patches, POM changes, config files)

Domain Expertise:

• Spring Security 6 migration patterns (WebSecurityConfigurerAdapter → lambda DSL)
• Deprecated API detection and deletion
• Missing dependency resolution (maps class names → Maven coordinates)
• COMMON_TEST_HINTS dictionary: 30+ patterns mapping class names to imports
• Test assertion guidelines (status codes, JSON paths, mock strategies)

6.3 agents/test_healer.py — Test Failure Doctor (151 lines)

Method	Description
heal(failed_tests, workspace_path)	Group failures by class → LLM call: generate fixed test file → AgentFix
_find_test_file(root, classname)	Locate .java test file by class name

Processes top 10 failures, max 3 classes, max 5 failures per class.

6.4 rag/store.py — SQLite Vector RAG Store (210 lines)

Lightweight persistent RAG store for test pattern learning.

Method	Description
upsert(id, kind, embedding, text, metadata)	Insert/update with normalized float32 embedding blob
query(embedding, top_k, kind, kinds, score_threshold, include_expired)	Cosine similarity search via dot product
get_by_id(id)	Direct ID lookup
count(kind)	Count entries by kind
evict_expired()	TTL-based cleanup (default 30 days)

Schema: rag_items(id TEXT PK, kind TEXT, created_at INT, embedding BLOB, metadata_json TEXT, text TEXT) Indexes: kind, created_at

6.5 llm/bedrock.py — Atlas Bedrock Client (163 lines)

Dedicated Bedrock client for the Atlas subsystem.

Method	Description
embed_text(text)	Titan Embeddings: inputText → embedding vector
generate_text(system, user, max_tokens)	Claude Messages API via Bedrock invoke_model

Tracks total_input_tokens, total_output_tokens, total_embedding_tokens for cost calculation. Security: Permanent credentials (AKIA*) do NOT use session tokens; temporary (ASIA*) require them.

6.6 generation/java_unit_test_generator.py — RAG-Enhanced Test Gen (441 lines)

Method	Description
generate_minimal_tests_for_repo(target_count, preferred_classes, ...)	Discover main classes → prioritize by scoring → generate tests
_generate_single_test(src, repo_path, ...)	LLM + RAG call: Check fingerprint → query RAG for similar patterns → generate JUnit 5 test
_set_fingerprint(class_key, sha, test_path)	Store source hash in RAG for idempotent re-runs

Scoring heuristic for class prioritization:

• +10 if in preferred classes list
• +5 for service/controller/repository classes
• +3 for @RestController/@Service/@Repository annotations
• −2 for test/config/model classes

Uses RepoContractRegistry for constructor/method signature validation in generated tests.

6.7 build/ — Build Infrastructure (5 files)

File	Purpose
maven.py	Maven command runner (mvn compile, mvn test, etc.) with subprocess management
jacoco_injector.py	Inject JaCoCo Maven plugin into pom.xml for code coverage
spring_test_injector.py	Inject spring-boot-starter-test dependency
failsafe_injector.py	Inject Maven Failsafe plugin for integration tests
dependency_governance.py	Enforce dependency version governance (BOM alignment, conflict resolution)

6.8 core/ — Core Infrastructure (5 files)

File	Purpose
config.py	Atlas-specific configuration (data dirs, model IDs, etc.)
logging.py	RunLogger class for structured pipeline logging
state.py	PipelineStateManager — manages baseline/validation state persistence
shell.py	Safe shell command execution with timeout
resilience.py	Retry with exponential backoff (configurable attempts, jitter) + Circuit Breaker pattern (CLOSED/OPEN/HALF-OPEN states) + Rate Limiter

6.9 analysis/ — Code Analysis (3 files)

File	Purpose
java_maven.py	Java project analysis: detect_repo_facts(), count_existing_tests(), find_domain_models()
contract_service.py	RepoContractRegistry: extract class constructors, method signatures for test generation validation
diff_analyzer.py	Analyze git diffs to identify functional changes for targeted test generation

6.10 reporting/ — Test Reporting (2 files)

File	Purpose
models.py	Report dataclasses: TestReport, CoverageReport, BreakageReport, GenerationReport, RegressionReport, QualityGateReport, UsageReport, FullRunReport
parsers.py	Parse Surefire XML reports, JaCoCo CSV coverage data, classify test failures

6.11 gitops/ — GitHub Integration (3 files)

File	Purpose
github_pr.py	Create PRs for Atlas-generated tests
github_issues.py	Create GitHub issues for persistent test failures
github_org.py	List repos in a GitHub organization for org-wide scanning

6.12 repo/ — Repository Management (2 files)

File	Purpose
cloner.py	RepoCloner: Clone repos with token authentication
history.py	Run history tracking for the Atlas pipeline

---

🎨 Frontend Deep Dive

Technology: Streamlit (5676 lines, single app.py + utility modules)

UI Components

The frontend is a premium dark-mode dashboard with glassmorphism styling, gradient headers, and micro-animations. Key CSS tokens:

• Background: #0f172a (dark slate), Secondary: #1e293b
• Accent: linear-gradient(135deg, #3b82f6, #2dd4bf) (blue → teal)
• Font: Inter (body), JetBrains Mono (code)

Core Functions

Function	Lines	Purpose
main()	45	Entry point: mode selector (Vulnerability Workflow vs Repository Explorer)
display_vulnerability_workflow(api_url)	~600	Streamlined flow: Upload → Map → Test → Fix → Verify
display_repositories(data)	~2400	Full repository explorer with vulnerability cards, dep trees, fix controls
process_active_batch_fix(selected_repo_id, ...)	~1040	Real-time batch fix processing with progress bars
display_lineage_graph(result, repo_name, vuln_idx)	~275	NetworkX-based dependency graph visualization
fetch_repositories(api_base_url, ...)	30	Call backend to fetch GitHub repos
map_vulnerabilities(api_url, repositories_data, csv_file)	28	Upload CSV and map vulnerabilities
run_testing_agent(api_url, repo_url)	70	SSE streaming of Atlas pipeline progress
batch_fix_vulnerabilities(api_url, vulnerabilities, ...)	~210	Call batch fix endpoint with progress callbacks
display_lineage_graph._extract_paths(items)	10	Extract file paths from dependent files list
display_setup_progress(current_step)	~160	Animated 3-step progress tracker (Upload → Fetch → Map)
display_run_history(api_url)	70	Fetch and display pipeline run history table

Frontend Utility Modules

File	Purpose
src/vulnerability_ui.py (52KB)	Advanced vulnerability display: cards, severity badges, fix result rendering
src/lineage.py (10KB)	Lineage graph data transformations
utils/atlas_report_comprehensive.py (21KB)	Comprehensive Atlas report rendering
utils/integrate_render.py (4KB)	Report integration helpers

---

🧠 RAG (Retrieval-Augmented Generation)

ICSF uses a custom RAG implementation for test pattern learning:

Architecture

                    ┌──────────────────┐
                    │   Titan Embed    │
                    │   (Bedrock)      │
                    └────────┬─────────┘
                             │ embedding vector
                    ┌────────▼─────────┐
                    │  SqliteVectorRag │
                    │     Store        │
                    │  (cosine search) │
                    └────────┬─────────┘
                             │ similar patterns
                    ┌────────▼─────────┐
                    │  Test Generator  │
                    │  (LLM prompt)    │
                    └──────────────────┘

How RAG is Used

1. Fingerprint Check: Before generating a test, hash the source file → query RAG for existing fingerprint → skip if unchanged
2. Pattern Retrieval: Query RAG store for similar test patterns (kind=test_pattern) with cosine similarity ≥ 0.25
3. Context Injection: Retrieved patterns are injected into the LLM prompt as examples
4. Pattern Storage: After successful test generation, store the pattern in RAG for future use

RAG Store Configuration

Setting	Value
Database	SQLite (data/atlas_rag.db)
Embedding Model	Amazon Titan Embed Text v1
Embedding Dimension	1536 (float32)
Similarity Metric	Cosine (via dot product on normalized vectors)
TTL	30 days (auto-eviction of stale entries)
Score Threshold	0.25 minimum cosine similarity

---

🤖 AI / LLM Integration

Models Used

Model	Use Case	Provider
Claude 3.5 Sonnet	All reasoning: code analysis, fix generation, strategy planning, safety validation, build repair, test healing	AWS Bedrock
Amazon Titan Embed Text v1	Text embeddings for RAG store	AWS Bedrock
Llama 3 70B (optional)	Alternative generation model	AWS Bedrock

LLM Call Sites

Component	# of LLM Calls	Purpose
CodeContextAgent	1	Data flow analysis
FixStrategyAgent	1	Fix strategy planning
CodeFixAgent	1 per file	Code generation
SafetyValidatorAgent	1	Fix validation
BuildMechanic	1–3 per build failure	Build error diagnosis + fix
TestHealer	1 per test class	Test repair
JavaUnitTestGenerator	1 per source class	Test generation
Total per vulnerability	~6–12	Depending on file count and failure iterations

Cost Management

• CostGuardService tracks cost per run with $5.00 default budget
• Pricing model: Claude 3.5 Sonnet @ $0.003/1K input, $0.015/1K output
• _calculate_usage() in the pipeline reports total tokens + estimated cost
• BedrockClient tracks total_input_tokens, total_output_tokens, total_embedding_tokens

---

📥 Input Requirements

1. Security Vulnerability Report (CSV)

Supported scanners: Fortify, Checkmarx, SonarQube, Snyk

Required Column	Example
vulnerability_type or category	Cross-Site Scripting
file_name or file_path	src/main/java/com/example/Controller.java
line_no or line_number	42
severity	Critical / High / Medium / Low
description	User input is rendered without encoding
recommendation	Use OWASP encoder for output encoding
repo_name or link	my-app or https://github.com/org/my-app

2. Version Control Credentials

• GitHub PAT: Requires repo and read:user scopes
• Stored in backend/credentials.yaml

3. AI Model Access (AWS Bedrock)

• AWS credentials: AWS_ACCESS_KEY_ID + AWS_SECRET_ACCESS_KEY in .env
• Region: us-east-1 (default) or any Bedrock-enabled region
• Model access: Must have Claude 3.5 Sonnet + Titan Embeddings enabled in your AWS account

4. Build Environment

• Java JDK 17+ on PATH
• Maven on PATH
• Git on PATH

---

🛠️ Technical Stack

Layer	Technology	Version
Language	Python	3.10+
Backend Framework	FastAPI	≥0.104
Frontend Framework	Streamlit	≥1.28
LLM Provider	AWS Bedrock (Boto3)	≥1.34
Embedding Model	Amazon Titan Embed Text v1	—
Reasoning Model	Claude 3.5 Sonnet	—
Database	SQLite	(stdlib)
HTTP Client	httpx	≥0.28
Data Processing	pandas	≥2.0
Version Control	GitPython + GitHub API	≥3.1
Graph Analysis	NetworkX	≥3.0
Validation	Pydantic	≥2.10
Containerization	Docker Compose	3.8
Build Tools	Maven, JDK 17+	—

---

🚀 Getting Started

Prerequisites

• Git, Java JDK 17+, and Maven installed and on PATH
• Python 3.10+
• AWS credentials with Bedrock access (Claude 3.5 Sonnet + Titan Embeddings enabled)
• GitHub PAT with repo and read:user scopes

Environment Setup

1. Create .env in backend/:

AWS_ACCESS_KEY_ID=your_key
AWS_SECRET_ACCESS_KEY=your_secret
AWS_REGION=us-east-1
BEDROCK_MODEL_ID=anthropic.claude-3-5-sonnet-20240620-v1:0
BEDROCK_EMBED_MODEL_ID=amazon.titan-embed-text-v1

2. Create credentials.yaml in backend/:

github:
  token: ghp_your_personal_access_token
  username: your-github-username
  email: [e-mail]

Run with Docker (Recommended)

docker-compose up --build

• Backend: http://localhost:8000
• Frontend: http://localhost:8501
• Backend has 4GB memory limit, frontend has 1GB
• Health checks are configured for both services

Manual Installation

Backend:

cd backend
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows
pip install -r requirements.txt
uvicorn main:app --reload --host 0.0.0.0 --port 8000

Frontend:

cd frontend
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows
pip install -r requirements.txt
streamlit run app.py --server.port 8501

---

📂 Project Structure

ICSF/
├── backend/
│   ├── main.py                          # FastAPI entrypoint (1636 lines, 20+ endpoints)
│   ├── config.py                        # Config class (AWS, Bedrock, GitHub credentials)
│   ├── credentials.yaml                 # GitHub PAT + user info
│   ├── logging_config.py               # Global logging configuration
│   ├── .env                             # AWS credentials (not committed)
│   │
│   ├── models/
│   │   └── agent_models.py              # 10 Pydantic models for pipeline data flow
│   │
│   ├── services/                        # 14 service files
│   │   ├── bedrock_service.py           # AWS Bedrock LLM wrapper (Claude, Llama, Titan)
│   │   ├── github_service.py            # GitHub API client (repos, orgs, file trees)
│   │   ├── vulnerability_service.py     # CSV parsing & repo mapping (833 lines)
│   │   ├── dependency_service.py        # Java dependency graph engine (2037 lines)
│   │   ├── fix_orchestrator.py          # Multi-agent pipeline controller
│   │   ├── batch_fix_service.py         # Batch vulnerability processing
│   │   ├── pr_manager_service.py        # Git operations & PR creation (1359 lines)
│   │   ├── batch_pr_service.py          # Aggregated PR creation
│   │   ├── atlas_service.py             # Testing pipeline façade
│   │   ├── fix_validator_service.py     # Post-fix build/test validation
│   │   ├── job_manager.py               # Async job & SSE streaming
│   │   ├── run_history.py               # SQLite run persistence
│   │   └── cost_guard.py                # LLM cost limiter ($5/run default)
│   │
│   ├── agents/                          # 7 agent files (Cognitive Fixing Loop)
│   │   ├── code_context_agent.py        # Blast radius mapper (642 lines)
│   │   ├── fix_strategy_agent.py        # Surgical planner (633 lines)
│   │   ├── code_fix_agent.py            # Multi-file code generator (1071 lines)
│   │   ├── safety_validator_agent.py    # Logic gate validator (371 lines)
│   │   ├── codebase_analysis_agent.py   # Repository intelligence (594 lines)
│   │   └── agent_improvements.py        # Helpers: ImportManager, SyntaxValidator, etc.
│   │
│   ├── atlas/                           # Self-Healing Testing Framework
│   │   ├── orchestrator/
│   │   │   └── run_pipeline.py          # Pipeline core (1412 lines)
│   │   ├── agents/
│   │   │   ├── build_mechanic.py        # Build failure auto-repair (1133 lines)
│   │   │   ├── test_healer.py           # Test failure doctor (151 lines)
│   │   │   └── models.py               # Agent data models
│   │   ├── rag/
│   │   │   └── store.py                 # SQLite vector RAG store (210 lines)
│   │   ├── llm/
│   │   │   └── bedrock.py               # Atlas Bedrock client (163 lines)
│   │   ├── generation/
│   │   │   └── java_unit_test_generator.py  # RAG-enhanced test gen (441 lines)
│   │   ├── build/
│   │   │   ├── maven.py                 # Maven command runner
│   │   │   ├── jacoco_injector.py       # JaCoCo coverage plugin injection
│   │   │   ├── spring_test_injector.py  # Spring test dependency injection
│   │   │   ├── failsafe_injector.py     # Failsafe plugin injection
│   │   │   └── dependency_governance.py # Dependency version governance
│   │   ├── core/
│   │   │   ├── config.py                # Atlas configuration
│   │   │   ├── logging.py              # RunLogger
│   │   │   ├── state.py                # Pipeline state manager
│   │   │   ├── shell.py                # Safe shell execution
│   │   │   └── resilience.py           # Retry, circuit breaker, rate limiter
│   │   ├── analysis/
│   │   │   ├── java_maven.py           # Java project fact detection
│   │   │   ├── contract_service.py     # Constructor/method signature registry
│   │   │   └── diff_analyzer.py        # Git diff → functional change detection
│   │   ├── reporting/
│   │   │   ├── models.py               # Report dataclasses
│   │   │   └── parsers.py              # Surefire XML & JaCoCo CSV parsers
│   │   ├── gitops/
│   │   │   ├── github_pr.py            # PR creation for generated tests
│   │   │   ├── github_issues.py        # Issue creation for failures
│   │   │   └── github_org.py           # Organization repo listing
│   │   └── repo/
│   │       ├── cloner.py               # Repository cloning
│   │       └── history.py              # Run history tracking
│   │
│   ├── scripts/                         # Utility & test scripts
│   │   ├── test_bedrock_connection.py
│   │   ├── test_cross_repo_dependencies.py
│   │   ├── test_dependency_analysis.py
│   │   ├── test_orchestrator.py
│   │   ├── analyze_all_matched_files.py
│   │   └── visualize_dependency_mapping.py
│   │
│   └── data/                            # SQLite databases & logs
│       ├── runs.db                      # Pipeline run history
│       └── atlas_rag.db                 # RAG vector store
│
├── frontend/
│   ├── app.py                           # Streamlit UI (5676 lines)
│   ├── src/
│   │   ├── vulnerability_ui.py          # Vulnerability display components
│   │   └── lineage.py                   # Lineage graph data transforms
│   ├── utils/
│   │   ├── atlas_report_comprehensive.py # Atlas report rendering
│   │   └── integrate_render.py          # Report integration helpers
│   ├── requirements.txt
│   └── Dockerfile
│
├── docker-compose.yml                   # Multi-container setup
├── start_frontend.bat                   # Windows frontend launcher
└── start_frontend.sh                    # Linux/Mac frontend launcher

---

📡 API Reference

Health & Credentials

Endpoint	Method	Description
/api/health	GET	Health check (Docker/LB probes)
/api/credentials/github	GET	Retrieve loaded GitHub credentials
/api/credentials/verify	GET	Debug credential loading

Repository Management

Endpoint	Method	Description
/api/github/repos	POST/GET	Fetch GitHub repositories

Vulnerability Management

Endpoint	Method	Description
/api/vulnerabilities/map	POST	Upload CSV + map vulnerabilities
/api/dependencies/analyze	POST	Single vulnerability dependency analysis
/api/dependencies/batch-analyze	POST	Batch dependency analysis

Fix Operations

Endpoint	Method	Description
/api/fix/orchestrate	POST	Full multi-agent fix pipeline
/api/fix/batch	POST	Batch fix multiple vulnerabilities

Testing Pipeline

Endpoint	Method	Description
/api/testing/start	POST	Start async testing job
/api/testing/job/{job_id}	GET	Poll job status
/api/testing/stream/{job_id}	GET	SSE event stream
/api/testing/runs	GET	Pipeline run history
/api/testing/run	POST	Legacy sync testing

Pull Request Management

Endpoint	Method	Description
/api/pr/create	POST	Create single PR
/api/pr/create-batch	POST	Create aggregated PR
/api/pr/merge	POST	Merge with conflict resolution
/api/pr/check-mergeability	POST	Check PR mergeability

---

ICSF — Making code security intelligent, automated, and reliable.