Label-cascade software factory: plan → CI → review → squash without a happy-path human

dmitriy-yefremov/software-factory is a fork-and-own template for a continuous factory: human writes specs/NNNN-name.md, opens a PR, labels state:queued, and a scheduler drives planner → plan-reviewer → implementer → CI gate → code-reviewer → auto-merger. Every stage is a label swap; REQUEST_CHANGES and CI-red get bounded retries (up to two per stage); anything that cannot proceed parks at state:needs-attention. The happy path squash-merges with no human in the loop — orchestration lives in nine GitHub Actions workflows plus Claude role briefs, not in a single overnight chat session.

The load-bearing idea is separating product code from factory machinery. You stamp placeholders, wire two GitHub Apps and secrets, rewrite CLAUDE.md / principles as the constitution, then file specs; the skeleton stays CI-green on clone so the toolchain is proven before agents touch it. Auto-merge is gated by guardrails after rebase, not by the implementer’s self-report. That is harnessed throughput: modest increments, independent review roles, and a merge queue that only advances when labels and checks agree.

Why it matters for code review

Continuous factories compound when the merge train is a state machine with evidence, not a model that stayed awake. Peak Claude mastery still wins for novel intent and constitution design; overnight and always-on win when each verified stage hardens the next. FOMO that your backlog sat idle while someone else’s label cascade drained specs is the psychological half — answer it with parked needs-attention and CI-gated squash, not with an unattended burst onto main.