Suggested capstone

Build an agentic maintenance system for a real repository. It should accept a well-specified issue, create an isolated workspace, research and plan the change, implement it, run repository checks, perform an independent AI review, and open a pull request for human approval.

• A repository with one-command setup and a clear architecture map.
• A concise AGENTS.md or equivalent repository guide.
• One reusable agent skill and one MCP integration.
• Deterministic hooks, tests, and CI gates.
• A threat model and least-privilege permission design.
• A durable issue-to-PR background workflow with retries.
• Model gateway telemetry for cost, latency, routing, and failures.
• A fixed evaluation suite and a small results dashboard or report.
• A five-to-ten-minute demo and a retrospective describing failures and improvements.

• A fresh clone can be set up and validated with one documented command.
• At least ten fixed evaluation tasks have baseline and final results.
• One interrupted background job demonstrably resumes or fails cleanly.
• Generated changes cannot bypass tests, security checks, or human merge approval.
• Every run is traceable to its issue, spec, prompts/context, tool calls, outputs, costs, and final review decision.