How sanju234-san/mini-devin works
Mini-Devin turns a GitHub issue into a reviewed pull request through an autonomous multi-agent workflow, with human approval of the plan before code changes. The backend also provides live run visibility, retrieval, sandbox verification, artifact and metrics APIs, and evaluation support. The README documents the product workflow and major capabilities; sampled routing code confirms approval and retry stages, but the graph builder and most API, agent, and integration implementations were not inspected, so detailed internal calls are not asserted.
Main entry points: graph.py Agent orchestration · live_stream.py Live run stream · retrieval_agent.py Retrieval agent · planner_agent.py Plan authoring
Components 26
Agent workflow
- Agent orchestration
backend/app/agents/graph.py - Workflow routing
backend/app/agents/routing.py - Run and plan schemas
backend/app/agents/schemas.py - Input and output guardrails
backend/app/guardrails - Issue triage
backend/app/agents/triage_agent.py - Retrieval agent
backend/app/agents/retrieval_agent.py - Plan authoring
backend/app/agents/planner_agent.py - Code generation
backend/app/agents/coder_agent.py - Debug and revise
backend/app/agents/debugger_agent.py - Change review
backend/app/agents/reviewer_agent.py - Pull request creation
backend/app/agents/pr_agent.py
Knowledge and models
- Index and vector search
backend/app/retrieval - Text embeddings
backend/app/ml_models/embedding_model.py - Triage classifier
backend/app/ml_models/triage_classifier.py - Model selection
backend/app/agents/model_router.py - LLM provider fallback
backend/app/integrations/llm_provider_client.py
Backend services
- REST API
backend/app/api - Isolated verification
backend/app/sandbox - GitHub integration
backend/app/integrations/github_client.py - Live run stream
backend/app/websocket/live_stream.py
Run experience
- React dashboard
frontend/src/app/page.tsx - Run socket hook
frontend/src/hooks/useAgentRunSocket.ts - Chainlit console
chainlit_console/app.py
Evaluation
- Benchmark evaluation
evaluation/swebench_lite/run_eval.py
Other components
- GitHub issue
- Plan approver
Connections 22
- GitHub issue → Agent orchestration: starts run
- Agent orchestration → Input and output guardrails: checks input
- Agent orchestration → Issue triage: routes triage
- Agent orchestration → Retrieval agent: routes retrieval
- Agent orchestration → Plan authoring: routes planning
- Plan authoring → Plan approver: requests approval
- Plan approver → Agent orchestration: approves plan
- Agent orchestration → Code generation: routes coding
- Agent orchestration → Isolated verification: verifies live
- Agent orchestration → Debug and revise: revises failures
- Agent orchestration → Change review: routes review
- Agent orchestration → Pull request creation: creates pull request
- Triage classifier → Issue triage: classifies
- Retrieval agent → Index and vector search: queries index
- Text embeddings → Index and vector search: embeds chunks
backend/app/ml_models/embedding_model.py - Model selection → LLM provider fallback: selects provider
backend/app/agents/model_router.py - Agent orchestration → REST API: serves runs
- Agent orchestration → Live run stream: emits run events
- React dashboard → Run socket hook: subscribes
frontend/src/hooks/useAgentRunSocket.ts - Run socket hook → Live run stream: receives events
- Chainlit console → Live run stream: renders stream
chainlit_console/app.py - Benchmark evaluation → Agent orchestration: evaluates runs