repo-engineer

repo-engineer is an autonomous repository engineer that acts as a verified tool-using coding agent. Given a small repository and an engineering task, it inspects the repo, plans, edits files through validated tools, runs the test suite, observes failures, replans, and only declares success when the verification suite is green. The core design rule is that the planner proposes, the runtime disposes: the LLM or scripted policy only emits tool-call proposals, while a deterministic state machine validates every call against a schema, checks permissions, jails every path inside the workspace, enforces step and replan budgets, and trusts nothing except the test suite. The system addresses common failure modes of AI agent demos, such as malformed tool calls, arbitrary shell access, and unverified claims, through schema validation, a no-shell tool with allowlisted command prefixes, and a VERIFY state that re-runs the suite. It includes two benchmark modes: a deterministic mode using a ScriptedPlanner that runs four software-engineering tasks in isolated workspace copies, and a real LLM mode using LLMPlanner with gemini-2.5-flash. Results are recorded separately in results/benchmark_results.json and results/llm_benchmark_results.json. The project is implemented in Python with no third-party runtime dependencies and includes a technical report and tests.