Design
Slipway is a small governance control plane for local AI-assisted development. It does not replace an AI coding tool, project tracker, or Git. It makes agent work legible by binding every governed change to a lifecycle, a current authority file, and evidence that can be inspected after the session ends.
For the full design document, see Design Philosophy. This page summarizes the concepts that matter most when deciding whether to use Slipway.
Core Principles
Section titled “Core Principles”| Principle | Meaning |
|---|---|
| Local-first | The repo contains active state and audit records. |
| One authority | change.yaml owns current lifecycle state; logs explain how it changed. |
| Bounded autonomy | Agents can advance work, but gates, blockers, review, and done-ready proof stay visible. |
| Adapter thinness | Host files route to the CLI instead of becoming independent workflow engines. |
| Artifact traceability | Intent, research, requirements, decisions, tasks, evidence, review, and assurance stay connected. |
| Fresh verification | Completion is valid only when current evidence proves the current worktree state. |
Why Evidence Is Re-Derived
Section titled “Why Evidence Is Re-Derived”Slipway does not trust a verdict just because a file says pass. It derives
freshness from authoritative inputs such as the current bundle, task plan,
execution run version, selected review set, the terminal ship-verification
suite run, and runtime task evidence.
That is why stale proof fails closed. Recovery is to rerun the owning stage, reviewer, or task evidence path, not to restamp files.
Why Worktrees Matter
Section titled “Why Worktrees Matter”Governed changes often bind to dedicated worktrees. The current worktree is the
behavioral surface. A root checkout, old branch comparison, or archived bundle
can help with review, but it is not a substitute for fresh status, validate,
and next output from the owning worktree.
Why Adapters Are Thin
Section titled “Why Adapters Are Thin”Generated Claude, Codex, Copilot, Cursor, Kilo, Kiro, OpenCode, Pi, Qwen, and Windsurf files help an AI tool find the right command or skill. They do not own lifecycle semantics. If a generated adapter and the CLI disagree, refresh the adapter and trust the current CLI.