How jasirdeen007/multi-agent-software-defect-prediction works
AgentTriage is a BugHub issue-label research pipeline: prepare and explore issue data, then train/evaluate defect-prediction baselines and audit labels with specialized agents and a judge. The README identifies data preparation and EDA as the current phase; the tree and source index also show baseline and auditing capabilities. Some relationships are explicit in excerpts or the README; unsampled subsystem internals are represented without inferred call chains.
Main entry points: orchestrator.py Audit runner · run_baseline.py Baseline runner · base_agent.py LLM agent client · data_agent.py Metadata agent
Components 23
Data preparation
- Dataset preprocessing
src/preprocessing - EDA pipeline
src/eda/run_eda_pipeline.py - Canonical dataset
data/processed
Prediction baseline
- Baseline runner
src/baseline/run_baseline.py - Dataset loading and splits
src/baseline/data_splitter.py - DeBERTa embeddings
src/baseline/embedding_extractor.py - XGBoost classifier
src/baseline/xgboost_classifier.py - Model evaluation
src/baseline/evaluator.py - SHAP explanations
src/baseline/shap_explainer.py - DeBERTa fine-tuning
src/baseline/finetune_deberta.py - XGBoost tuning
src/baseline/tune_xgboost.py
Label auditing
- Audit runner
src/agents/orchestrator.py - Policy agent
src/agents/policy_agent.py - Metadata agent
src/agents/data_agent.py - Pattern agent
src/agents/pattern_agent.py - Arbitration judge
src/agents/judge_agent.py - Audit evaluation
src/agents/evaluate_audit.py - Audited records
src/agents/storage.py
Shared audit components
- LLM agent client
src/agents/base_agent.py - Audit data schemas
src/agents/schemas.py
Other components
- BugHub PostgreSQL
- Researcher
- Groq LLM API
Connections 26
- BugHub PostgreSQL → Dataset preprocessing: provides source data
README.md - Dataset preprocessing → Canonical dataset: prepares canonical data
README.md - Canonical dataset → EDA pipeline: analyzed as Parquet
README.md - Researcher → Baseline runner: runs experiment
- Baseline runner → Dataset loading and splits: loads and splits
src/baseline/run_baseline.py - Baseline runner → DeBERTa embeddings: extracts embeddings
src/baseline/run_baseline.py - Baseline runner → XGBoost classifier: trains classifier
src/baseline/run_baseline.py - Baseline runner → Model evaluation: evaluates predictions
src/baseline/run_baseline.py - Baseline runner → SHAP explanations: explains predictions
src/baseline/run_baseline.py - DeBERTa fine-tuning → Dataset loading and splits: uses dataset splits
src/baseline/finetune_deberta.py - XGBoost tuning → Dataset loading and splits: uses dataset splits
src/baseline/tune_xgboost.py - XGBoost tuning → Model evaluation: uses evaluation
src/baseline/tune_xgboost.py - Researcher → Audit runner: runs audit
- Audit runner → Policy agent: dispatches verdict request
src/agents/orchestrator.py - Audit runner → Metadata agent: dispatches verdict request
src/agents/orchestrator.py - Audit runner → Pattern agent: dispatches verdict request
src/agents/orchestrator.py - Audit runner → Arbitration judge: arbitrates agent verdicts
src/agents/orchestrator.py - Policy agent → LLM agent client: inherits API client
src/agents/policy_agent.py - Metadata agent → LLM agent client: calls JSON generation
src/agents/data_agent.py - Arbitration judge → LLM agent client: calls JSON generation
src/agents/judge_agent.py - LLM agent client → Groq LLM API: requests completions
src/agents/base_agent.py - Audit runner → Audited records: saves audit records
src/agents/orchestrator.py - Audit evaluation → Audited records: evaluates audit output
- Audit runner → Audit data schemas: constructs audit record
src/agents/orchestrator.py - Metadata agent → Audit data schemas: returns verdict
src/agents/data_agent.py - Arbitration judge → Audit data schemas: returns judge verdict
src/agents/judge_agent.py