Slipway Flow

Slipway: the software factory, as built

Drawn after the final diagram in Web Dev Cody's "Stop Writing Code. Build a Software Factory Instead." The team works at the inputs and at the review; the factory does the rest and improves itself through pull requests.

Scheduled (US Eastern) sweeps, opt-in per repo: performance 02:00 · security 02:30 · cleanup 03:00 · docs 03:30 shepherd every 30 min · knowledge sync 15 min · retro 04:00 · digest 05:00 · report Mon 08:00 Anthropic model API, via the proxy Slipway core · VM 106 slipwayd · Postgres · egress proxy · tokens stay here context context pack, cited [KB-n] Slack thread or issue text attachments, mockups knowledge (git) slipway/knowledge/ all apps .slipway/knowledge/ per app indexed every 15 min skills (git) slipway/skills/ stages, review .slipway/skills/ conventions triggers issues · PR comments · Slack Claude Code (MCP) · cron workflow stages plan questions ⇄ you implement docs + knowledge records decisions verify fresh-sandbox evidence review panel parallel reviewers risk rules + AI, higher wins PR shepherd keeps PRs up to date, fixes conflicts, re-checks, merges when policy allows sandbox · VM 107 isolated (gVisor), one container per stage, destroyed after Claude Code, headless repo + .task (brief, context) no tokens, no production data internet only via the proxy retro + evals nightly: repeated review findings, your corrections, failures → one skill or knowledge PR with an eval case; the score must not drop Pull Request overview · risk · screenshots architecture · security performance · external impact tests · AI review context used labels risk:low/medium/high check slipway/gates merged → outcome → shadow precision report opens Slack a thread per request, tied to a repo watched channels → digest PRs GitHub issues (slipway:ready) · PR comments Claude Code slipway_* tools on the laptop questions, PR links and nudges go back to the same thread or issue feedback: @slipway → address-feedback · merged skill/knowledge PRs team you now; Bryan and Leon later each as engineer, product or designer per request users the people the apps are for external systems APIs and services the apps call humans in the loop: inputs prompts, answers, approvals, "merge" humans in the loop: review read the PR, comment, merge needs → insights facts → systems/ files

Click any box, circle or workflow stage for what it does and how. Links like #shepherd open straight to one.

request and resultscheduled and sandboxin git, changed by PRfeedback into the factory