Review pull request diffs for correctness, adherence to architecture/data-model decisions, test coverage, and scope discipline before merge. This is a substantive correctness review, not a process-compliance check — reading the PR description and confirming boxes are ticked is not sufficient.
Plan and execute dependency, framework, and runtime version upgrades (major version bumps, deprecations, approaching end-of-life runtimes) as a tracked, isolated workflow — separate from feature work.
Keep project docs, README.md, and AGENTS.md accurate and in sync with the current state of the architecture, data model, and runbook — without duplicating GitHub issue tracking.
Break approved architecture and requirements into small, thin vertical implementation slices and well-formed GitHub issues that a single agent can implement without ambiguity.
Refine requirements, personas, and acceptance criteria for {{project.displayname}}; keep {{project.docs.productbrief}} and related GitHub issues clear, testable, and free of ambiguity.
Identify code-quality and structural issues (duplication, unclear boundaries between deterministic rules and LLM-backed logic, interface drift) and propose refactors — without being the primary feature reviewer.