GitHub stacked pull requests turn a large agent-generated feature into a dependency-ordered sequence—data, API, application wiring, then UX—so each layer has a bounded reviewer and its own CI evidence. Teams read the stack top-down for intent, review bottom-up for correctness, and use local rebase and sync commands when a lower layer changes, preserving signed commits and propagating fixes upward.
Cloudflare's Codex supplies the standards behind a similar review system. More than 60 RFCs are reduced into structured SHOULD and MUST statements with stable identifiers, owners, and separate approved and enforced states. In four months the code reviewer flagged nearly 230,000 deviations and withheld approval about 16,000 times; mechanical rules move into fast linters while model-based review stays available in CI and locally.
Schalk Neethling's 18-package Calavera release shows why delivery itself needs product design. Four prereleases uncovered repository permissions, mismatched tarball paths, packed provenance metadata, and trusted-publishing bootstrap problems. Exact-version checks, dry runs, inspected archives, signed provenance, retry-safe publishing, and revocation of the one-use bootstrap token converted those failures into reusable release contracts. Cheap code increases throughput, but reviewable units, governed rules, observable failure states, and reversible release steps determine whether that throughput remains maintainable.
Two browser-level guardrails complete the picture. Remy Sharp captured a file-drop interaction with a tiny inline handler before 600 KB compressed of modules loaded, queuing the user's intent even on GPRS. Microsoft's experimental Network Efficiency Guardrails use Document Policy and the Reporting API in Edge to surface images over 200 KB, data URLs over 100 KB, and uncompressed text resources during real browsing.