@sabaiway/agent-workflow-kit weekly downloads statistics

<div align="center">

🧠 agent-workflow-kit

Durable, cross-agent memory & workflow for AI coding agents — the one command that installs it.

Run it once per machine, deploy it once per project — then every future session boots from a small, structured memory instead of re-reading your whole repo and re-deriving yesterday's decisions. Works with Claude Code, Codex, Cursor, and any agent that reads AGENTS.md.

Node ≥ 22 · dependency-free scripts · no telemetry in family code

One command to start:

npx @sabaiway/agent-workflow-kit@latest init

_{This installs the global skill — deploying into a project is a separate step (below).}

Works with any tool that reads AGENTS.md — Claude Code · Codex · Cursor · Devin Desktop (formerly Windsurf) · GitHub Copilot · Gemini CLI · Cline · Aider · and 20+ more.

Quick-jump: Install · What it deploys · How it works · Composition root

</div>

---

Part of the agent-workflow family. This package is the composition root + entry point: it delegates memory deployment to the substrate, injects the workflow methodology, and detects optional execution backends. This page is the kit's manual (install · commands · what it deploys) — for the whole-family story, start at the family front door.

---

❓ The problem

AI coding agents are stateless between sessions. Every new chat starts from zero:

── new session, no kit ───────────────────
    ▶ "continue the feature"
        ↓
    reads 18 files… greps ×6…
    re-infers the architecture…
    re-asks a decision you settled…
        ↓  (15k–40k tokens later)
    …finally starts working
──────────────────────────────────────────
→ re-derives what it knew yesterday, and
  re-introduces a bug you already fixed

No durable handover ⇒ drift between sessions, repeated mistakes, ballooning token cost.

---

⚡ Without vs. With

The kit gives the agent a small, structured memory it reads at the start of every session — instead of rebuilding context from source each time.

WITHOUT the kit · cold start, cost grows
  s1  ~30k tok  ██████████
  s2  ~28k tok  █████████    ← repeats a fixed bug
  s3  ~34k tok  ███████████  ← drift

WITH the kit · boots from memory, cost flat
  s1   ~4k tok  █
  s2   ~4k tok  █            ← no drift
  s3   ~5k tok  █            ← decisions kept

_{Illustrative/directional, not a measured guarantee — exact numbers scale with repo size. The point is the shape: cold re-reads that grow vs. a flat, cache-warm boot.}

	🚫 Without	✅ With agent-workflow-kit
Session boot	re-read source + grep to rebuild context	read a few small docs, ~constant
Boot cost	grows with repo, paid every session	flat; stable layer stays prompt-cache-warm
Cross-session memory	none	handover (where we left off)
Past decisions	re-litigated	decisions.md (ADRs) — settled once
Known bugs	re-discovered	known_issues.md — impact + workaround
Doc growth	unbounded sprawl	frontmatter caps + rolling changelog archive + one-file-per-ADR store
Drift	docs ≠ code over time	pre-commit gate keeps them honest
Cross-agent	re-explain the project to each tool	one AGENTS.md, read by 20+ agents

---

📦 What it deploys into your project

Invoking the skill inside a project creates a portable memory and its maintenance policy:

your-repo/
├── AGENTS.md              ← single entry point
├── CLAUDE.md → AGENTS.md  ← symlink, for Claude Code
└── docs/ai/
    ├── index.md           ← auto-generated navigator
    ├── handover.md        ← where we left off (read first)
    ├── active_plan.md     ← current task
    ├── agent_rules.md     ← session protocols + self-review
    ├── current_state.md   ← snapshot of the codebase now
    ├── architecture.md    ← layout rules pointer & extension points
    ├── technical_specification.md
    ├── decisions.md       ← ADRs — the HOT window (newest)
    ├── adr/               ← one file per archived ADR + log.md navigator
    ├── known_issues.md    ← bugs + workarounds
    ├── changelog.md       ← rolling, then archived
    ├── env_commands.md    ← daily commands
    ├── tech_reference.md  ← configs & patterns
    ├── pages/             ← one spec per page/route
    └── history/           ← changelog archive (HOT→WARM→COLD)
  + scripts/agent-workflow/ ← caps · index · archive (the kit folder)
  + pre-commit hook        ← keeps it all honest    (installed on a yes)

The Markdown memory is stack-agnostic; the kit folder scripts/agent-workflow/ + pre-commit hook are the Node path (dependency-free). They are offered on every project because the agent host runs them; a committer without Node on PATH gets a loud hook failure. On a project with a package.json, an upgrade places the caps and index check and the ADR rotator even after a no; the changelog and issue rotators and the hook still wait for a yes.…