Last Light: a self-hosted factory that works after dark on your GitHub
NearForm Last Light: webhook/cron triage + PR review, Architect→Executor→Reviewer builds, permission-scoped App tokens, and evals scored on your own repos — not someone else’s bench.
Source: Last Light — self-hosted AI software factory for GitHub
Last Light (NearForm; open nearform/lastlight, MIT) is a self-hosted after-dark software factory for a GitHub org: triage issues, review PRs, run security and health sweeps, and — on maintainer @mention — walk a full Architect → Executor → Reviewer line that opens a PR. The marketing line is the operating claim: the factory keeps working while you are away, on a production line you own.
The continuous spine is event-driven when you can expose a webhook, cron when you cannot. Issue triage every 15 min (cron) or on issue.opened; PR review every 30 min or on pr.opened; weekly health/security digests on Mondays; optional Slack digest with computed numbers, not model prose. Build cycles are opt-in expensive: Context → Guardrails (agent names the full test command into .git/lastlight-gate.sh, then a deterministic bash gate runs it once — READY only if exit 0) → Architect (read-only plan) → Executor (TDD) → Reviewer in a fresh context with up to two fix loops → Create PR. Approval gates can pause for a human from GitHub comment, Slack, or the dashboard. Every workflow phase runs under a downscoped GitHub App token (read / issues-write / review-write / repo-write) so a triage run literally cannot push code. Per-repo .lastlight/ overlays are read from the default branch only — a PR cannot reconfigure the agent reviewing it — and can only tighten operator bounds.
The trust move that matters for continuous factories: an evals harness that runs the same YAML workflows against a mocked GitHub, scored on cases mined from your merged PRs and resolved issues (add-case --pr / --issue), not SWE-bench cosplay. Model pick, prompt change, and calibration close against a blind held-out gate. Dashboard + OTEL traces make overnight spend and phase outcomes inspectable in the morning.
Why it matters for code review
Always-on review capacity is the scarce half of agent volume. Last Light treats PR review and stranded-review cron as first-class line jobs, keeps merge authority human via gates, and measures the factory on the operator’s own history. Continuous harnessed throughput looks like triage and review that never sleep — and FOMO when the queue was quiet overnight is often “no webhook, no cron,” not “Claude had a bad night.”