What problem does it solve? Auditing a real software project before implementation requires collecting trustworthy repository evidence, challenging internal analysis with an independent reviewer, and turning verified findings into bounded implementation tasks — a process that is error-prone and ungoverned when done ad hoc. ## Core Features & Use Cases - Deterministic Evidence Collection: Gathers git state, files, languages, tests, CI config, and churn into repo-evidence.json bound to an exact commit SHA. - Independent External Review: Sends a secret-redacted review packet to Codex CLI (read-only sandbox) or the OpenAI API, then reconciles every finding as AGREE/PARTIAL/DISAGREE/NEW/UNVERIFIED. - Governed Task Decomposition and Dispatch: Labels analysis claims FACT/INFERENCE/UNKNOWN, decomposes accepted findings into an acyclic task DAG, and dispatches to Devin behind a run-level circuit breaker with cost budgets. - Use Case: Before refactoring a legacy service, run the full pipeline to produce an evidence-grounded audit, an independent critique, a reconciled findings list, and a Devin-ready task plan — all tracked by a 14-state state machine. ## Quick Start Ask the agent to run the project-review-orchestrator against the current repository to collect evidence, obtain an independent review, and produce a dispatchable task plan.
