WHO
Tech Leads / Staff Engineers at 5–50 person teams using Cursor / Claude Code / Codex for backend, who put architecture rules in Markdown but still fix layer and ORM violations on every PR.
ConstraintGate · Waitlist
Turn architecture rules into CI that coding agents cannot ignore
Declare YAML/TOML invariants — layer boundaries, forbidden imports, required ORM — then fail the PR with SARIF plus an agent-readable repair brief. Built for Node (Express/Nest) and Python (Flask/FastAPI). Not another JVM-only ArchUnit or heavy orchestrator.
Market-validation page · Free waitlist · No fake user counts
Problem
Constraint decay on every PR
Structural constraints pile up and LLM coding-agent assertion pass rates drop ~30pp (constraint decay). Markdown rules in AGENTS.md have no force — PRs keep fixing layer and ORM/DB violations.
Before
- Rules live in AGENTS.md / Cursor rules; agents can ignore them
- PRs keep fixing layer and ORM/DB violations
- ArchUnit is JVM-heavy; eslint DIY; evals don’t gate structure
After
- YAML/TOML invariants; CI must pass
- Fail fast with SARIF locating the break
- Lightweight Node + Python gate + agent repair brief
Solution
What ConstraintGate does
An executable architecture gate for teams already shipping backend with coding agents — YAML invariants to CI red + SARIF + repair brief, not another AGENTS.md.
PROBLEM
Architecture rules in Markdown are hints, not gates. Constraint decay leaves PRs as repair yards for layer and ORM/DB misuse.
SOLUTION
Declare YAML/TOML invariants → GitHub Action/CLI on PR → CI red + SARIF + agent-readable repair brief (optional webhook into the next agent loop).
RESULT
Architecture rules become CI that coding agents cannot ignore — not another skipped Markdown file.
Features
Features
Marketing points from the product hypothesis — for demand validation, not a formal spec promise.
YAML/TOML architecture invariants
Declare layer dependencies, forbidden/allowed imports, required ORM/DB access boundaries — turn Markdown intent into machine-checkable rules.
GitHub Action / CLI on PR
Static checks on every pull request; violations fail CI — agents and humans cannot “skip the MD.”
SARIF + human report
SARIF and a human-readable report for GitHub Code Scanning / PR checks with violation locations.
Agent-readable repair brief
Failure output you can feed back into Cursor / Claude Code / Codex; optional webhook into the agent loop.
Node + Python MVP path
Express/Nest and Flask/FastAPI common layouts first; flat $49–$149/mo hypothesis — not JVM-only, not metered by eval score.
How it works
How it works
A three-step loop matching the waitlist-stage product hypothesis.
- 1
Declare invariants in YAML/TOML
Layer boundaries, forbidden imports, required ORM/DB access — aligned to Express/Nest/Flask/FastAPI layouts.
- 2
Run on every PR
GitHub Action or CLI; violations → CI red + SARIF for Code Scanning/checks.
- 3
Feed the agent a repair brief
Human report + agent-readable repair task; optional webhook into the next Cursor/Claude Code/Codex loop.
Use cases
Who it’s for
If these situations sound familiar, join the waitlist to help us validate.
AGENTS.md says Clean Architecture — agents still cross-layer import
You need a PR CI red light, not another reviewer eyeballing imports.
Forbid domain→DB; force repository/ORM
When data-layer violations dominate failures, invariants should stop them.
Nest/Express module boundaries rotting
Tired of DIY dependency-cruiser/eslint — want YAML OOTB + agent repair brief.
FastAPI/Flask teams generating routes with Cursor
Rules must enter CI, and failure briefs must return to the agent loop.
Evaluated ArchUnit (JVM) or archfit
You want a Node/Python lightweight gate + flat pricing — not a heavy orchestrator.
Public signals
Public signals from the industry
Public reporting and industry discussion only — not customer testimonials or invented metrics.
- arXivConstraint Decay: capable configs drop ~30pp assertion pass rate; data-layer ~45% of logic failures.
- Hacker News~287 points, 197 comments — practitioners argue for linters / ArchUnit / CI over Markdown constraints.
- ArchUnitOfficial userguide for JVM architecture tests — the pattern teams want beyond Markdown.
- alexei-led/archfitDeterministic arch check + agent-readable repair task direction on GitHub.
- DEV.toFrom prompts to invariants — re-architecting with ArchUnit and LLMs.
FAQ
FAQ
How is ConstraintGate different from ArchUnit?
ArchUnit is strong on the JVM. We start with Node/Python, YAML invariants, and an agent-readable repair brief — not a pile of test classes.
How is it different from eslint / dependency-cruiser / nx?
They can coexist. We focus on architecture invariants → CI + SARIF + a repair brief for the agent loop, so Markdown doesn’t drift.
How is it different from Braintrust / LangSmith?
Those lean behavioral/eval. We are a structural gate (imports, layers, ORM boundaries) that fails PR CI — not prompt scores.
Why not only AGENTS.md / Cursor rules?
Markdown is a hint. ConstraintGate is fail-able CI — coding agents cannot ignore it.
What’s in MVP scope?
Node (Express/Nest) + Python (Flask/FastAPI) common layouts. Other languages are not a day-one promise.
When is early access?
Waitlist invites go out in batches by email. No fake launch date.
What’s the expected price?
Hypothesis band $49–$149/mo flat; final pricing arrives with launch email.
Waitlist
Join the waitlist
Leave your email for ConstraintGate early access and launch notes. We will not invent a ship date.
Thank you — you’re on the ConstraintGate waitlist.
We’ll email early-access invites and launch updates in batches. Unsubscribe anytime.