flow

Decision 0067 · Accepted

Flow boundaries: plugin, starters,
scripts, host

start grew into the assumed front door: it waits for a green deployment and always opens discovery. Since 0062, most users meet Flow as a global plugin in a repository that never touched a starter.

Status
Accepted
Built from:
da88c48

Status: Accepted (2026-11-03). Spec: 066. Amends 0022 and 0024 (explore runs beside start only when the request describes a product), 0026 (flow no longer waits on a missing brief), and 0051 (design is offered after the initial commit, not the first deployment). Builds on 0062.

Context

start grew into the assumed front door: it waits for a green deployment and always opens discovery. Since 0062, most users meet Flow as a global plugin in a repository that never touched a starter.

Decision

  1. Four owners.
    • Flow plugin: intent, judgment, specs, delivery principles, and reusable skills. Its value never depends on a starter, start, Vercel, or the whole delivery sequence.
    • Starters: optional, opinionated technology foundations.
    • Project scripts: the authoritative verification commands.
    • Native host: sessions, worktrees, delegation, and built-in review.
  2. start ends at the initial commit. It pushes where authorized and reports governance, handing over the one push command when the guard blocks a default-branch push. A deployment is reported if the host shows one, or reported as not connected, never awaited here; "deploy it" connects a host.
  3. explore beside start is conditional. It runs when the request describes a product; otherwise start offers it once.
  4. qa stays in the commit gate. It measured under ten seconds on both starters (warm store, install scripts skipped, small suites; limits), and it is the check that runs the template's code after renaming, beyond identity_scan's text scan.
  5. Mission and principles carry this. The mission opens with velocity with intent; principles 4 and 5 state composable skills, what-not-how, and proportional evidence.
  6. Critical thinking is selective. It applies where a wrong answer is costly: direction changes, consequential uncertainty, and what a person must understand before it ships. It is not a phase.

Alternatives

  • Drop qa for a lighter gate, as the audit proposed. Rejected on evidence: it would save seconds and lose template-integrity coverage.
  • Remove main-branch CI work in the starters. Rejected: runs take about a minute on any event, so the saving is negligible and the post-merge signal is lost.
  • Fold start into flow. Rejected: 0009 keeps alignment and scaffolding apart.

Consequences

  • Time to start now ends at the initial commit, which measures initialization, not hosting.
  • On GitHub with the plugin, the guard stops start's initial push to the default branch, as it always has; the person pushes once. A guard exception is a separate Human-tier change.
  • A new project can sit undeployed until "deploy it" or ship; that is delivery, and the report says so.
  • Starters vendor pinned Flow skills (0045), which load beside a global plugin. Worth revisiting; not settled here.

Jump to, or try a question