DevPilot : Autonomous AI Software Engineering Agent
Link to open source: https://github.com/nikhil-mutreja/Buildathon
DevPilot is an autonomous AI Software Engineer that automates end-to-end coding workflows. Powered by LangGraph and the Swytchcode Execution Authority, DevPilot inspects real repositories, diagnoses root causes via AST analysis, generates code fixes with unified diffs, runs automated tests as a safety gate, creates Git branches, commits verified changes, opens Pull Requests, tracks work in Jira, and broadcasts alerts on Slack.
User Request → Understand → Plan → Inspect Repository → Analyze Code
→ Generate Fix → Run Tests → Create Branch → Commit → Push
→ Open PR → Jira Tracking → Slack Notification → Result
| Problem Statement | DevPilot Implementation |
|---|---|
| Understand Development Tasks | NLP intent parser extracts task type, target issue, and required tools from natural language. |
| Work with Code Repositories | Reads real code via github.content.get, generates patches, commits via Git, opens PRs via github.pull.create. |
| Run Tests | Executes pytest as a safety gate — blocks PR creation if tests fail. |
| Track Issues | Retrieves and triages GitHub issues, creates Jira tasks via jira.api.issue.create. |
| Assist Development Teams | Posts Slack alerts via slack.chat.postmessage.create with PR links and Jira tickets. |
- Real Code Operations: Reads real source files, writes real patches, runs real
pytest, creates real Git branches with actual commit SHAs. - Test Safety Gate: Blocks PR creation when tests fail — never pushes broken code.
- Conditional Agent: Different requests trigger different workflow paths. Investigate-only skips code changes; Jira/Slack only when explicitly requested.
- AST-Based Defect Scanner: Detects CWE-681 (float precision), CWE-89 (SQL injection), CWE-532 (credential logging), CWE-775 (resource leaks), PEP-484 (mutable defaults).
- End-to-End Traceability: Links Issue → Patch → Tests → Branch → Commit → PR → Jira → Slack.
- Premium UI: IDE-like workspace with execution timeline, code diff viewer, test results, PR review cards, and audit trail.
- Dual Execution Modes: Real mode (live Git + Swytchcode) and Demo mode (offline simulation).
| Provider | Tool ID | Role |
|---|---|---|
| GitHub | github.issue.get1 |
Retrieve open issues |
| GitHub | github.content.get |
Read source code files |
| GitHub | github.content.update |
Commit code patches |
| GitHub | github.pull.create |
Open Pull Requests |
| Jira | jira.api.issue.create |
Create tracked tasks |
| Slack | slack.chat.postmessage.create |
Send team alerts |
- Agent: LangGraph (StateGraph, conditional routing)
- Execution: Swytchcode Runtime
- UI: Streamlit 1.64
- API: FastAPI + SSE (Server-Sent Events)
- Validation: Pydantic v2
- Testing: Pytest 9.1
- Language: Python 3.12
cd /home/nikhil-mutreja/buildathon
source .venv/bin/activate
pip install -r requirements.txt
streamlit run app/ui/streamlit_app.py
uvicorn app.api.server:app --host 0.0.0.0 --port 8000
| Repository | Defects | Tests |
|---|---|---|
real_test_repo |
Greeter edge-case (empty/whitespace name) | ✅ 3 tests |
ecommerce_service |
CWE-681, CWE-89, CWE-775 | ✅ Tests included |
auth_microservice |
CWE-532, PEP-484 | ✅ Tests included |
realtime_stream_service |
CWE-775 (connection pool) | — |
pytest -v
23 tests covering:
- Intent parsing and tool selection
- Severity triage classification
- Code patch generation and unified diff
- Full autonomous workflow (Issue → Fix → Tests → PR → Jira → Slack)
- Real-mode end-to-end lifecycle
- Safety gate (test failure blocks PR)
- Investigate-only mode
- Multi-repository isolation
