Intent-Aware Code Review

Protect business intent in every code change.

Turn approved engineering docs, ADRs, and project rules into deterministic pull request checks before intent drift reaches main.

AxiomGuard is an intent-aware code review guard for TypeScript PRs that turns approved rules into proof-carrying checks.

Run the real demo See v0.2.0

GitHub Action for CI
Benchmark leaderboard is public
Approved rules are source of truth
PROVEN critical evidence can block

SemanticRule

rule_id: AUTH-BOUNDARY-001
status: approved
severity: critical
invariants:
  - type: forbidden_import
    forbidden_imports:
      - '@acme/billing'
enforcement:
  mode: block

Changed diff
apps/auth/session.ts
+ import { charge } from "@acme/billing";
+ export async function createSession(user) {
+   return issueSession(user);
+ }

Proof-carrying finding

AUTH-BOUNDARY-001

Status: PROVEN
Evidence: import edge auth -> billing
Action: Block merge

Trace

CLAUDE.md
SemanticRule YAML
TypeScript diff
Symbolic evidence
CLI / PR gate

S-Agent Core v0.2.0

GitHub Action and benchmark infrastructure are now public.

The new release turns S-Agent from a local CLI demo into a CI-ready semantic guard: teams can run approved SemanticRules in GitHub Actions and track quality with a reproducible benchmark leaderboard.

Read the v0.2.0 release Open leaderboard

GitHub Action

Run S-Agent in CI

Use the composite action to call the existing CLI with project, rules, output-format, and fail-on-blocking inputs.

Benchmarks

Measure semantic review quality

Run pnpm benchmark for JSON output, a Markdown table, precision, recall, false-positive rate, PROVEN rate, runtime, and clean blocking rate.

Scope discipline

No hidden analyzer changes

The release adds CI and measurement infrastructure without PR comments, hosted services, dashboards, or artificial analyzer tuning.

1.000 Precision

1.000 Recall

0.000 False positive rate

0.000 Clean blocking rate

Real repo demo

Run the repo's real TypeScript demo.

The landing demo uses the same fixture shipped in this repository: a CLAUDE.md rule, an approved SemanticRule, a TypeScript import violation, and the CLI report that blocks only after symbolic proof.

pnpm analyze:demo:broken pnpm analyze:demo

CLAUDE.md

Documented intent

examples/demo-typescript-app/CLAUDE.md

The authentication layer is identity-only.
It must not import billing code or trigger billing side effects directly.

SemanticRule

Approved guardrail

examples/demo-typescript-app/rules/auth.rules.yml

rule_id: INV-AUTH-001
status: approved
severity: critical
from: src/auth/**
to: src/billing/**
mode: block

TypeScript change

Violating edge

examples/demo-typescript-app/src/auth/session.ts

import { BillingService } from "../billing/billing-service";
const billing = new BillingService();
billing.recordSessionStart({ userId, reason: "auth-started" });

Proof

Blocking evidence

apps/cli

Changed file: src/auth/session.ts
Evidence: import edge auth -> billing
Status: PROVEN
Severity: critical
Blocking: yes

Broken fixture Blocks

$ pnpm analyze:demo:broken
# S-Agent Report

Project: examples/demo-typescript-app

## Violation: INV-AUTH-001

Changed file: src/auth/session.ts
Changed symbol: module

Problem: Layer boundary violation:
src/auth/session.ts imports ../billing/billing-service.

Why this matters:
The authentication layer is identity-only;
billing behavior must stay inside the billing domain.

Evidence:
- src/auth/session.ts:1 - forbidden boundary
- CLAUDE.md:3 - CLAUDE.md#authentication-module

Status: PROVEN
Severity: critical
Blocking: yes

Clean fixture Passes

$ pnpm analyze:demo
# S-Agent Report

Project: examples/demo-typescript-app-clean

No findings.

Intent-aware review

What is AxiomGuard?

AxiomGuard gives reviewers concise, source-backed answers about what changed, which approved rule was violated, and why the finding can or cannot block.

What is AxiomGuard?

AxiomGuard is an intent-aware code review tool for engineering teams that need pull requests to preserve approved business and architecture rules. It turns SemanticRule YAML into deterministic TypeScript checks and reports proof-carrying findings reviewers can audit.

What is intent-aware code review?

Intent-aware code review verifies whether a code change preserves documented business, product, and architecture intent. It catches changes that compile and pass tests but still violate approved rules such as module boundaries, side-effect limits, or value thresholds.

Who is it for?

AxiomGuard is built for TypeScript-heavy teams using AI coding tools, large refactors, or fast-moving product engineering workflows. It is most useful when important rules live in ADRs, READMEs, CLAUDE.md, or senior engineers' repeated review comments.

Layer Boundary auth must not import billing

Forbidden Side Effect read-only flows must not write

Value Invariant discount cannot exceed max

Docs become guardrails

Your most repeated review comments should be executable.

AxiomGuard starts where teams already write intent: CLAUDE.md, READMEs, ADRs, and architecture notes. Candidate rules can be suggested, but the approved YAML file is the source of truth.

The MVP stays deliberately narrow: TypeScript diffs, symbolic checks, and proof-carrying findings for the three violations that teams repeat by hand.

Layer Boundary

auth must not import billing

import { charge } from '../billing/service'

Forbidden Side Effect

read-only flows must not write

await db.invoice.update(...)

Value Invariant

discount cannot exceed max

discount = 0.85

How it works

How does AxiomGuard check a pull request?

The review path stays small and auditable: approved rules, TypeScript analysis, symbolic evidence, and conservative blocking.

Start with docs Use CLAUDE.md, READMEs, ADRs, and architecture notes as source material for candidate rules.
Approve SemanticRules A human approves explicit YAML rules. The approved SemanticRule file becomes the source of truth.
Check every PR AxiomGuard indexes TypeScript code, analyzes imports and functions, verifies findings, and prints a proof-carrying report.
Block only proof Only PROVEN findings from approved critical rules in block mode can stop a merge.

Comparison

AxiomGuard vs SAST, linters, and AI reviewers.

AxiomGuard is not another generic scanner. It adds a deterministic intent layer beside the tools teams already use.

Alternative	What it checks	What AxiomGuard adds
SAST	Security vulnerability patterns	Product and architecture intent, such as billing boundaries and domain invariants
Linters	Style, syntax, and generic code-quality rules	Team-specific business rules backed by approved documentation
AI reviewers	Broad contextual suggestions and plausible concerns	Deterministic blocking only when approved rules produce symbolic evidence
Architecture tools	Dependencies, packages, and structural drift	PR-level reports tied to documented business and architecture intent

Search topics

Explore intent-aware review topics.

These short pages answer the priority questions teams ask when they compare semantic review, AI guardrails, architecture drift prevention, and business logic regression prevention.

semantic code review Semantic Code Review for TypeScript PRs

Semantic code review checks whether a TypeScript pull request preserves approved business and architecture intent, not only syntax, style, or security patterns.

intent-aware code review Intent-Aware Code Review Guardrails

Intent-aware code review verifies that every pull request preserves documented business, product, and architecture rules.

AI code review guardrails AI Code Review Guardrails for Generated Code

AI code review guardrails help teams keep AI-generated and refactored code aligned with approved business and architecture rules.

architecture drift prevention Architecture Drift Prevention for Pull Requests

Architecture drift prevention catches pull requests that violate documented module boundaries and approved architecture decisions.

business logic regression prevention Business Logic Regression Prevention in Code Review

Business logic regression prevention checks pull requests for rule violations such as unsafe side effects, broken product invariants, and invalid thresholds.

Finding statuses

Only proof gets to stop a merge.

Heuristics can point reviewers toward risk. Blocking remains conservative: approved critical rules, deterministic evidence, and a concrete trace.

PROVEN Can block

A deterministic check found symbolic evidence for an approved critical rule.

PROBABLE Review

A strong signal needs human judgment before it can become enforceable.

SUSPECT Inform

A weak signal is useful context, not a reason to stop a merge.

RULE_CONFLICT Investigate

A heuristic suspicion disagrees with deterministic evidence.

DISMISSED No action

The finding has no actionable evidence for this change.

FAQ

Common questions about AxiomGuard.

Is AxiomGuard a replacement for SAST?

No. AxiomGuard complements SAST by checking business and architecture intent rather than security vulnerability classes.

Does AxiomGuard use an LLM?

The MVP does not use an LLM for enforcement. Future LLM features may suggest candidate rules or explanations, but approved SemanticRules remain the source of truth.

What can block a pull request?

Only a PROVEN finding from an approved critical rule in block mode can block a pull request.

What languages are supported?

The MVP is TypeScript-first and focuses on deterministic symbolic checks for TypeScript diffs.

Do teams need perfect documentation to start?

No. Teams can start with one important rule reviewers already enforce manually, then expand the SemanticRule set over time.