Skip to main content

Architecture Decisions

These records capture the load-bearing decisions behind the agent peer-review workflow: why GitHub itself is the datastore, how labels and native reviewer routing divide responsibility, why the codebase is one core library behind thin adapters, how concurrent panel review avoids collisions, how review context gets assembled and bounded, and how pi.dev integration ships as its own package, and how untrusted pull-request content is treated as data rather than instructions. Each one explains what was decided and why, grounded in the code that actually shipped, and together they supersede the design specs that shaped the project during its construction.