---
title: QA Agent LangGraph
emoji: "🧪"
colorFrom: blue
colorTo: indigo
sdk: streamlit
app_file: app.py
pinned: false
---

# LangGraph QA Compare

LangGraph-based QA test generation workflow that orchestrates calls to `qa-agent-service`. This project owns sequencing, shared workflow state, backtracking, runtime traces, and exports. The actual QA agents are invoked through the service.

## Setup

```bash
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
streamlit run app.py
```

By default the app calls the hosted Hugging Face Space REST API:

```text
https://helmfridsson-qa-agent-service.hf.space/api/v1
```

The app accepts any of these inputs and normalizes them to the REST base automatically:

- `https://helmfridsson-qa-agent-service.hf.space`
- `https://helmfridsson-qa-agent-service.hf.space/docs`
- `https://helmfridsson-qa-agent-service.hf.space/api/v1/health`
- `https://huggingface.co/spaces/helmfridsson/qa-agent-service`

You can also override the default at runtime with `QA_AGENT_SERVICE_URL`.

For local development, start `qa-agent-service` separately and override the REST API URL in the sidebar:

```bash
cd ../qa-agent-service
PYTHONPATH=src uvicorn api:app --reload --port 8009
```

The local service URL is `http://127.0.0.1:8009`.
MCP may run on `http://127.0.0.1:8008/mcp`, but that endpoint is not a REST API and does not expose `/docs` or `/openapi.json`.

## Workflow

The application runs a LangGraph `StateGraph`:

```text
START -> requirements_analyst(service) -> test_designer(service) -> reviewer(service) -> orchestrator(local)
```

The orchestrator conditionally routes to `END`, `test_designer`, `requirements_analyst`, `reviewer`, or failure. The first supported selective backtracking path is reviewer feedback back to `test_designer`. `run_controls.max_rounds` prevents endless loops.

## Agents

The three QA worker agents are provided by `qa-agent-service`:

- `requirements_analyst`
- `test_designer`
- `reviewer`

Each worker node can run in `Service-backed`, `Deterministic`, or `Disabled` mode. `Service-backed` calls the REST endpoint `POST /agents/{agent_key}/invoke` and validates the returned contract with this app's Pydantic schemas. Deterministic mode remains as a local smoke-test fallback.

The orchestrator is local and deterministic because `qa-agent-service` intentionally does not own workflow sequencing, retries, memory commits, backtracking, or stopping conditions.

## Service Validation

Before graph execution, the app can call:

```text
GET https://helmfridsson-qa-agent-service.hf.space/api/v1/health
```

If the service is not reachable, execution stops before the graph starts.

## Sample Scenarios

The Scenario Input tab includes predefined sample scenarios for login/registration, checkout, password reset, support tickets, inventory management, and course enrollment. Select one and click `Load selected scenario` to populate the title and requirement text.

## Outputs

Each run writes files under `runs/<run_id>/`:

- `run_export.json`: complete export with partial results preserved
- `stage_traces.jsonl`: append-only stage trace log, one trace per line

The Streamlit UI also exposes download buttons for both files.

## Known Limitations

- Runtime status updates appear after each Streamlit-triggered graph run, not as token streaming.
- Deterministic mode is intentionally basic and mainly supports smoke testing.
- Only the first MVP backtracking path is fully supported: reviewer to test designer.
- There is no database, authentication, RAG, hosted deployment, or Docker setup in this version.
- GitHub Actions can deploy this app repo to the Hugging Face Space `helmfridsson/qa-agent-langgraph` once the GitHub secret `HF_TOKEN` is configured.