DiffXray
Zero-dependency local Git diff & GitHub PR X-ray MCP server (xray_git_diff, xray_patch_text, xray_github_pr) that maps untouched downstream callers, collapses cross-file code moves, and audits 12 deterministic silent-regression rules.
Documentation
GitHub diffs list changed files alphabetically. DiffXray shows what actually broke downstream.
Turn any local git diff, branch, or 40-file AI-generated Pull Request into an interactive Blast-Radius Map, collapsed code-move tour, and "Quietly Wrong" silent-bug audit — in under 1.5 seconds, with zero dependencies and zero API keys.
English · 简体中文 · Try Live Web Studio → · Architecture
Open Live Browser Studio (No Install) →
·
Paste any public GitHub PR URL (or open ?pr=owner/repo/123) — 100% client-side.
Why DiffXray?
Coding agents (Claude Code, Codex, Cursor, Antigravity) now generate 2,000-line, 30-file Pull Requests in minutes. Tests pass (exit 0), the diff compiles, and reviewers get stuck in cognitive overload because traditional diff viewers hide three things:
- Untouched Downstream Callers (The Hidden Blast Radius): When
verifySessionToken(token)changes toverifySessionToken(token, orgId), GitHub shows yousession.ts— it never shows the 3 untouched routes and background workers that still call it with 1 argument. - Mechanical Code Motion Mixed with Real Edits: Moving a 120-line function into a new module while changing 1 discount constant looks like 240 lines of red/green churn. DiffXray pairs cross-file moves via normalized shingle Jaccard matching, collapses pure motion, and highlights only the in-flight delta.
- "Quietly Wrong"
exit 0Regressions: Agents routinely make failing suites turn green by adding.skip, deletingexpect()assertions, wrapping errors in emptycatch {}blocks, droppingawait, widening types toany, or removingrequireAuthmiddleware.
Dogfood proof: On its very first run against our own PR #13, DiffXray's
silent-catchdetector caught a real JSON-RPC-32700parse-error swallowing bug in our MCP server before push.
Quick Start (30 Seconds)
1. Run Locally on Any Git Repo (Zero Install via npx)
Requires Node.js 20+. Zero third-party node_modules are downloaded.
# X-ray uncommitted changes (or current branch vs main) right in your terminal
npx github:unbridled-41/diffxray
# Generate & open the single-file offline Interactive HTML Studio
npx github:unbridled-41/diffxray main...HEAD --html --open
# X-ray any GitHub Pull Request directly by number, slug, or URL
npx github:unbridled-41/diffxray --pr 14
npx github:unbridled-41/diffxray --pr vercel-labs/skills#10 --html --open
# Pre-commit / CI gate: exit 1 if any high or critical silent hazard is introduced
npx github:unbridled-41/diffxray --staged --fail-on high
2. Install as an Agent Skill (Claude Code · Codex · Cursor · Antigravity · OpenCode)
Teach your coding agent to verify its own blast radius and eliminate silent regressions before finishing a task or opening a PR:
npx skills add unbridled-41/diffxray
Or install as a native Claude Code Plugin:
claude plugin install diffxray --marketplace unbridled-41/diffxray
3. Add Sticky PR Review Maps via GitHub Actions (5 Lines)
Create .github/workflows/diffxray.yml to automatically post and update a Guided Reading Order + Blast-Radius + Risk Radar map on every Pull Request (see it live on our own PRs):
name: DiffXray PR Review
on: [pull_request]
permissions:
contents: read
pull-requests: write
jobs:
xray:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
with:
fetch-depth: 0
- uses: unbridled-41/diffxray@main
Exportable Adaptive SVG PR Cards
Run npx github:unbridled-41/diffxray --svg diffxray-card.svg (or click Tab 4: PR Card & Export in the HTML Studio) to generate a prefers-color-scheme vector card for PR descriptions, architecture reviews, and Slack threads:
What the 12 "Quietly Wrong" Detectors Catch
Every rule is 100% deterministic (zero LLM hallucinations, <50ms execution, OASIS SARIF v2.1.0 compatible via --sarif):
| Rule ID | Severity | What It Catches |
|---|---|---|
secret-leak | Critical | Newly added AWS, GitHub, OpenAI/Anthropic/Stripe keys, Slack tokens, private keys, or DB passwords (masked in output). |
injection-risk | Critical | Interpolated SQL template/f-strings, eval / new Function, shell: true command execution, dangerouslySetInnerHTML. |
auth-guard-removed | Critical | Deleted authentication/authorization middleware or decorators (requireAuth, verifyToken, @UseGuards, @login_required) without replacement. |
breaking-signature | High | Exported function required-param increase, arity reduction, or async flip cross-referenced against untouched workspace callers. |
weakened-test | High | Added .skip / .only / @pytest.mark.skip or deleted expect() / assert calls without replacement assertions. |
silent-catch | High | Newly added empty or comment-only catch {} / except: pass blocks that swallow runtime failures. |
unawaited-async | High | Dropped await before async calls inside modified hunks, creating fire-and-forget race conditions. |
type-erosion | Medium | Newly injected : any, as any, @ts-ignore, # type: ignore, or unsafe {} escape hatches in production code. |
schema-drift | Medium | Prisma / SQL / ORM model files modified without a corresponding migration file in the diff. |
untested-core | Medium | >= 30 effective lines modified in core, entrypoint, or data layers with 0 test files updated. |
lockfile-drift | Low | Package manifest dependency blocks modified without updating the corresponding lockfile. |
debug-leftover | Low | Leftover debugger, console.log, pdb.set_trace(), breakpoint(), or dbg!() in non-CLI production modules. |
Multi-Language Support
Symbol extraction, signature-break diffing, and module import resolution work out-of-the-box across:
- TypeScript & JavaScript (
.ts,.tsx,.mts,.js,.jsx,.mjs,.cjs) - Python (
.py) - Go (
.go) - Rust (
.rs) - Java, Kotlin & C# (
.java,.kt,.cs) - SQL & Prisma Schemas (
.sql,.prisma)
MCP Server (Claude Desktop · Claude Code · Cursor · Codex)
DiffXray includes a built-in, zero-dependency JSON-RPC 2.0 STDIO Model Context Protocol server exposing xray_git_diff, xray_github_pr, and audit_silent_hazards:
{
"mcpServers": {
"diffxray": {
"command": "npx",
"args": ["-y", "github:unbridled-41/diffxray", "--mcp"]
}
}
}
How DiffXray Compares
| Capability | GitHub PR UI | SaaS PR Bots ($15–30/mo) | Static Repo Maps (floorplan) | DiffXray |
|---|---|---|---|---|
| Cost & License | Free | Paid per seat | MIT | Free · MIT |
| Works on uncommitted local edits | No | No | Static files only | Yes (< 1.5s) |
| Zero API keys / Zero telemetry | N/A | Cloud LLM required | LLM plugin | 100% deterministic & local |
| Untouched downstream caller Blast Radius | No | Partial | Static | Yes (interactive graph) |
| Cross-file moved code collapsing + delta | File rename only | No | N/A | Yes (shingle Jaccard) |
| Zero-install browser URL PR X-Ray | Alphabetical list | No | Static demo | Yes (?pr=owner/repo/123) |
Runtime node_modules dependencies | N/A | Heavy | 0 | 0 |
Zero-Dependency & Privacy Architecture
- Zero Supply-Chain Bloat:
package.jsonships with"dependencies": {}.scripts/verify-integrity.mjsenforces zero runtime dependencies and strict browser/Node ESM isomorphism on every CI run across Linux, macOS, and Windows. - 100% Private: Neither the CLI nor the offline HTML Studio ever transmits code or telemetry. Network calls happen only when you explicitly pass
--prto fetch a GitHub Pull Request.
Contributing & License
Contributions, new language symbol patterns, and low-noise risk detectors are welcome! See CONTRIBUTING.md and docs/ARCHITECTURE.md.
Released under the MIT License.