Secure RepoPilot

Issue-to-PR coding agent with baseline verification, safety controls, and privacy auditing.

Secure RepoPilot reads a bug report, analyzes a repository, runs baseline tests, applies a minimal patch, reruns verification, blocks unsafe commands, audits traces for leakage, and generates a PR-ready report.

Research basis:

• SWE-Cycle: full software issue lifecycle evaluation.
• Phoenix: safe GitHub issue resolution with specialized agents.
• AgentVisor: tool privilege separation and prompt-injection defense.
• AgentLeak: privacy leakage auditing across agent traces.

Note

This is a portfolio-grade implementation inspired by current software-engineering-agent research. It is not a claim of benchmark-level SWE-bench performance.

See It In Action

python demo.py

Expected demo shape:

{
  "verdict": "accept",
  "changed_files": ["src/securecalc/calculator.py"],
  "baseline_passed": false,
  "patched_passed": true,
  "safety_risk": 0.0,
  "privacy_score": 1.0
}

The bundled sample repo contains a real failing test: divide(8, 0) should return None. Secure RepoPilot copies the repo to a scratch workspace, reproduces the failure, patches calculator.py, reruns tests, and writes a PR-style report.

Why This Project

Most coding-agent demos stop at code generation. Secure RepoPilot focuses on the parts that matter in production:

• repo understanding before editing
• failing-test reproduction before patching
• minimal patch generation
• baseline-vs-patched verification
• command safety policy
• prompt-injection and secret scanning
• privacy leakage audit over agent traces
• structured PR report and metrics

Architecture

flowchart LR
    I["Issue text"] --> A["Repo analyzer"]
    A --> P["Planner"]
    P --> B["Baseline tests"]
    B --> C["Patcher"]
    C --> T["Patched tests"]
    T --> S["Safety guard"]
    S --> L["Privacy auditor"]
    L --> J["SWE judge"]
    J --> R["PR report"]
    J --> DB["SQLite run store"]

Loading

Features

• Repository scanner for files, languages, dependencies, and test commands.
• Issue parser with lightweight bug classification.
• Planner that selects suspected files and verification steps.
• Reproducer that captures baseline failing tests.
• Deterministic patcher for the bundled demo and extension point for OpenAI mutations.
• Tester that runs commands with timeouts and captured logs.
• SWE-style judge with accept, needs_fix, unsafe, and inconclusive verdicts.
• Security guard for blocked commands, prompt-injection text, and secret patterns.
• Privacy auditor for emails, private local paths, and token-like values.
• FastAPI endpoints, Streamlit dashboard, SQLite storage, docs, and CI.

Quick Start

python -m venv .venv

Windows PowerShell:

.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
Copy-Item .env.example .env
python demo.py

macOS/Linux:

source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
python demo.py

Run tests:

pytest -q

No-dependency smoke test:

python tests/run_tests.py

Run API:

uvicorn app.api:api --reload

Run dashboard:

streamlit run app/ui.py

API

Endpoint	Purpose
GET /health	Check runtime configuration
POST /run-fullcycle	Run analyze, reproduce, patch, verify, audit, judge
GET /runs	List stored runs
GET /runs/{run_id}	Fetch one run payload
GET /metrics/{run_id}	Fetch compact run metrics

Example:

{
  "repo_path": "examples/buggy_python_repo",
  "issue_text": "Division by zero should return None",
  "apply_patch": true,
  "workspace": ".repopilot-runs"
}

Evaluation Signals

Signal	Meaning
baseline_passed	Whether detected tests passed before patching
patched_passed	Whether detected tests passed after patching
regression_count	Previously passing commands that failed after patching
confidence	Judge confidence in the final verdict
safety_risk_score	Risk from command/prompt/secret checks
privacy_score	Leakage audit score over traces and reports

Safety Boundaries

Secure RepoPilot is a local portfolio agent, not a hardened code-execution platform. It blocks risky commands and scans for secrets, but production use should add container isolation, read-only mounts, network denial, memory limits, and stronger policy enforcement.

Resume Bullets

• Built Secure RepoPilot, a SWE-Cycle and Phoenix-inspired issue-to-PR coding agent that analyzes repositories, reproduces bugs, patches code, generates regression tests, and verifies fixes with baseline-vs-patched execution.
• Implemented AgentVisor-style tool safety controls and AgentLeak-style trace auditing to detect unsafe actions, prompt-injection risks, secrets exposure, and privacy leakage across agent logs and outputs.
• Added FastAPI, Streamlit observability, SQLite run tracking, CI tests, patch scoring, risk metrics, and PR-ready report generation for production-style autonomous software engineering workflows.

Roadmap

• Add OpenAI patch proposals with structured output validation.
• Add Docker sandbox execution.
• Add GitHub issue ingestion through the GitHub API.
• Add unified diff apply/revert for arbitrary repositories.
• Add benchmark pack for multiple mini repos and issue types.