Show HN: WTF > Auto-check what your coding agent changed

A new open-source tool called WTF (agent-wtf) lets developers auto-audit changes made by coding agents such as Claude Code, Cursor, Copilot, Codex, and Aider by inspecting the resulting Git diff rather than the agent itself. WTF compresses mechanical churn from lockfiles, minified bundles, and snapshots, reporting that 3,812 changed lines reduced to 94 meaningful lines to review (97.5% compressed), and its `wtf verify` command runs tests to produce an independently verified local execution receipt. The tool runs locally with no account and no AI required via `npx agent-wtf`, and `wtf init-agent` configures AGENTS.md, CLAUDE.md, .cursorrules, and .github/copilot-instructions.md so agents self-audit before declaring completion.

Any agent. Any Git repo. Local. No account. No AI required. WTF works with any coding agent Claude Code, Cursor, Copilot, Codex, Aider because it inspects the resulting software change, not the agent. npx agent-wtf bash $ npx agent-wtf WTF — what just happened? 2 files changed · +7 / -5 VERIFIED ○ Tests not yet run · Run wtf verify to validate tests npm test PAY ATTENTION 1. TESTS Test skipped or disabled test/charge.test.js:9 test.skip 'handles VIP coupon cap calculation', = { ALSO ⚠ 1 skipped test ⚠ 1 debug statement console.log Review surface: 12 lines to review across 2 files Then: bash $ npx agent-wtf verify WTF — what just happened? 2 files changed · +6 / -5 VERIFIED ✓ tests 254ms Review surface: 11 lines to review across 2 files Agent finishes ↓ WTF ↓ Evidence ↓ Agent fixes ↓ WTF verify ↓ Human gets receipt Coding agents can generate more code in two minutes than you can review in an afternoon. When an agent claims: "Done Refactored the billing module and all tests pass." — humans are left with three questions: 1. What actually happened? 2. Did it actually work, or did the agent just say it did? 3. Where do I actually need to look? WTF gives you the answers in under a second. Most agent diffs are dominated by lockfiles, minified bundles, snapshots, and generated boilerplate. WTF separates mechanical churn from code that actually deserves human attention: 3,812 changed lines ↓ WTF ↓ 94 meaningful lines to review 97.5% compressed You review what matters. WTF accounts for the rest. | Standard Tooling | What Happens with Coding Agents | How WTF Solves It | |---|---|---| | git status | Lists modified files, but treats a 2,000-line lockfile the same as an auth timeout modification. | Separates signal from churn : Classifies mechanical lines vs. meaningful lines deserving human review. | | git diff | Floods your terminal with generated boilerplate, snapshots, and minified bundles. | Focuses human attention : Automatically highlights high-risk patterns auth, DB migrations, env vars, debug leftovers . | | Agent Claims | Believes the agent when it claims "refactored billing module and all tests pass" . | Verifies independently : Flags skipped or disabled tests test.skip and generates an unforgeable local execution receipt wtf verify . | No installation required: npx agent-wtf Or install globally: npm install -g agent-wtf | Command | What it does | |---|---| | wtf | See what changed and what needs attention in your working tree ~50ms . | | wtf init-agent | Automatically configure your repository for autonomous agent self-auditing. | | wtf verify | Discover and run your tests/builds to produce an independently verified receipt. | | wtf show | View exact diff snippets and line evidence for every finding. | | wtf --json | Machine-readable evidence schema wtf/0.1 for coding agents. | Run this once in any repository: npx agent-wtf init-agent This automatically configures your repository's agent rules AGENTS.md , CLAUDE.md , .cursorrules , and .github/copilot-instructions.md . From that moment on, whenever Claude Code, Cursor, Copilot, or Cline works in your repo, the agent autonomously: 1. Runs WTF before declaring completion. 2. Catches shortcuts : Detects its own skipped tests test.skip , debug leftovers console.log , and schema risks. 3. Executes tests : Runs wtf verify to independently validate your test suite. 4. Hands you proof : Attaches the unforgeable verification receipt directly to its final reply before you review. To view the markdown template without modifying files, pass npx agent-wtf init-agent --print . WTF is designed to inspect machine-generated changes, so it treats repository content as untrusted input. - No code uploads : Zero code or diffs ever leave your machine. - No telemetry : Works completely offline with zero tracking or background pings. - No account or API key : No signup, no LLM tokens, no monthly bill. - No required AI model : Fast, local deterministic analysis. - No shell-based Git commands : Direct binary spawning shell: false with baseline Git configuration overrides. - Repository filesystem containment : Enforces realpath containment to prevent symlinks from escaping the repository. - Terminal control-sequence sanitization : Strips ANSI cursor escapes, OSC sequences, and Unicode Bidi controls. - Zero runtime npm dependencies : Pure ESM package with 0 runtime dependencies, reducing third-party supply-chain exposure. Normal wtf analysis does not intentionally execute project code. wtf verify is different: it runs your project’s verification commands locally with your user permissions and is not sandboxed. Only use it on code you trust to execute. See SECURITY.md https://github.com/LinusInnovator/wtf/blob/main/SECURITY.md for details. WTF is an evidence ledger, not an oracle. We strictly avoid fabricated confidence scores e.g. "87% safe" or "clean code guarantee" . Instead, WTF categorizes facts into four strict evidence tiers: - REPORTED : What something claims happened e.g., an agent summary . - OBSERVED : What WTF directly confirmed in the Git diff e.g., session timeout altered, .env introduced . - VERIFIED : What WTF independently executed and validated e.g., test runner exited code 0 . - UNKNOWN : What available evidence cannot prove e.g., tests exist but have not been run . WTF does not claim to catch every bug or replace human judgment. It eliminates the blind spots between what the machine claimed and what the machine actually did. git clone https://github.com/LinusInnovator/wtf.git cd wtf npm install npm run build npm test npm run gauntlet Built by @LinusInnovator https://github.com/LinusInnovator . Explored in depth at Great Delights https://great.delights.pro/ai-patterns .