For Claude Code, Codex, Pi, Muse Code, and Grok Build
Smartypants listens to the conversation and the working tree. It draws the system the way a staff engineer would on a whiteboard, remembers what you intended, and tells you the moment the code leaves it.
Click any box, arrow, boundary, or layer to read about it. Drag boxes, frames, and layers. Double-click a box to go deeper. Scroll to zoom.
One convention everywhere, so every diagram reads the same way.
Top to bottom is the stack, from who calls to where data rests.
Left to right is the journey inside a layer, as long as it needs to be.
Dashed frames are network and trust boundaries — public internet, edge, cluster namespaces, private subnets — taken from your Kubernetes, Helm, compose, and Terraform files.
Every box has a plain name and a one-line description. Click it for what, why, deep-dive notes, flows, drift, and the intent recorded about it.
Arrows point the way data moves; a reply is its own arrow back to the caller. Solid is data, dashed is control.
Turn it on in an existing repo and it builds the whole diagram in the background: services, dependencies, routes, tables, topics, then the compose, Kubernetes, Helm, and Terraform that wire them, including ingress, network policies, namespaces, and subnets.
When the agent finishes a turn, it looks at what actually changed in git — dirty files, new commits, untracked files — and asks one batched question per file. Code that leaves the design is flagged in red; new services and infra changes redraw the diagram.
If you change a decision in chat, the design evolves and the old choice is remembered as excluded. If the code contradicts a decision, constraint, or boundary, that is drift.
“Go deeper on the aggregator”, a double-click, or smartypants deeper expands a system into components, a component into modules, and a module into the data model, algorithm, and failure notes an interviewer would ask about.
IntentCode keeps what you meant as keyed atoms — N redirect.latency p99<50ms, D link-store dynamodb>cassandra — grouped by part, never a transcript. A whole design interview fits in about 200 tokens.
Two models, two jobs: Jev selects, Claude writes. A free local pass drops noise, Typesafe Jev decides in about 200 ms what a turn is worth and which part it touches, and Claude, through your own Claude Code sign-in, writes the diagram update or drift note only when there is work. Greetings, “run the tests”, and doc edits cost nothing, and your agent never waits.
npm install -D @logan-robbins/smartypants
npx smartypants init # hooks, config, and catch-up if code exists
# Claude writes the diagram through your Claude Code sign-in: no key needed
export TYPESAFE_API_KEY=... # optional: Jev picks in ~200 ms (console.typesafe.ai/keys)
npx smartypants serve # open the printed URL# Claude Code
/plugin marketplace add logan-robbins/smartypants
/plugin install smartypants@smartypants
# Codex
codex plugin marketplace add logan-robbins/smartypants
codex plugin add smartypants@smartypants
# Pi
pi install git:github.com/logan-robbins/smartypantsCommands: deeper, catchup, review, intent, drift, mermaid, stats, reset. Everything is inert without smartypants.config.json.
Live runs: Typesafe Jev selecting and Claude writing, on labeled agent changes to a real repo and held-out design turns.
| Jev + Claude | Claude alone | |
|---|---|---|
| End-of-turn drift review, 22 labeled turns | 22/22, none missed, no false alarms, 7.3 s | 22/22, 9.7 s |
| “Worth remembering?” on held-out turns | none worth keeping dropped | 34/34 (local rules alone: 23/34) |
| Per-turn decision | ~0.2 s, a fraction of a cent | ~2.9 s, ~1.5¢ |
| Catch-up of an existing repo | real service names, flows, and network boundaries from code plus compose, k8s, Helm, Terraform; 36–115 s and about a cent on 10 open-source repos | |
Details and raw runs: results/.