perk

perk is a Pi-native workflow for plan-driven software engineering.

It combines a Python CLI that manages the session exterior, a TypeScript Pi extension that governs the session interior, and an issue backend for durable workflow state, GitHub by default. The workflow is organized around planning, implementation, code review, and learning.

Workflow stages and surfaces can use perk’s built-in defaults or pluggable third-party extensions, such as plannonator for plan review.

perk is built for work that may be interrupted, resumed elsewhere, superseded, or moved across machines. Resumability, human feedback, and remote execution are not special cases; they are part of the core model.

perk is inspired by erk, formerly maintained by the team at Dagster.

What perk is

perk's unit of work is a plan: a reviewed, durable description of a change, written before files are edited and saved as canonical workflow state. The core spine is plan -> save -> implement -> submit -> address -> land -> learn; address only happens when review feedback needs a response. Longer-running objectives can emit bounded plans into the same spine.

The implementation is split across two planes:

• exterior — the Python perk CLI, which scaffolds repos, manages worktrees, mints run ids, and launches primed pi sessions;
• interior — a TypeScript Pi extension, which owns in-session stage transitions, tool gates, context, and warm /... commands.

Both planes read the language-neutral shared/ registry and contracts directly. There is no codegen layer to drift.

The operating model is deliberately simple: canonical state lives in GitHub by default (Linear is supported), .perk/workflow/ is a local cache, and session state is throwaway. Cold doors launch fresh context from a shell (perk implement, perk plan resume); warm doors keep context inside a session (/submit, /land, /learn). implement and address can dispatch to CI; review, merge, and learning stay local.

Quickstart

From the git repo you want to wire:

uv tool install perk
perk init
perk doctor
perk plan

perk init is idempotent and safe to re-run. perk doctor reports health; perk doctor --fix repairs managed drift.

You need a git repo plus git, gh, node >= 22, pi, and uv. ast-grep is preferred for structural search and reported when absent, but it is advisory. GitHub auth is verified and reported by init; it becomes required when you drive real plans, PRs, and merges.

For the guided first run, follow Get started with perk.

Documentation

Operator docs live under docs/user-docs/:

• Tutorials — learn by doing; start with Get started with perk.
• How-to guides — recipes for specific jobs: resume a plan, address review feedback, switch to Linear, attach a skill, dispatch to CI.
• Reference — CLI commands, in-session commands and tools, objectives, configuration, providers, backends, and schemas.
• Explanation — the mental model: plans, planes, state tiers, doors, and remote maturity.

Internal research, design notes, and durable learnings live under docs/.

Layout

• perk/ — Python CLI, the session exterior.
• extension/ — TypeScript Pi extension, the session interior.
• shared/ — stage registry and cross-plane contracts.
• docs/ — operator docs, research, design notes, and learnings.

Develop

Two pinned toolchains:

• Python — uv (3.13, pinned in .python-version), ruff (lint/format), ty (types).
• TypeScript — npm (Node ≥ 22, .npmrc), Biome (lint/format), tsc (types).

With just:

just setup
just install-cli
just fmt
just lint
just typecheck
just test
just ci

just setup runs uv sync, npm install, just hooks, and just install-cli. The hook is managed by prek and runs ruff check on staged Python. Re-run just hooks after a fresh clone if hooks are missing.

Without just: uv run … for Python (uv run perk init, uv run pytest, uv run ruff check perk tests, uv run ty check) and npm run … for TypeScript (npm run lint, npm run typecheck).