liarliar
liarliar: make coding agents prove "done".
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/adidotai--liarliar)HTML
<a href="/listings/adidotai--liarliar"><img src="/api/listings/adidotai--liarliar/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
**Claude Code** (skill + enforcement hooks):
claude plugin marketplace add adidotai/liarliar && claude plugin install liarliar@liarliar
Or inside a session: `/plugin marketplace add adidotai/liarliar`, then `/plugin install liarliar@liarliar`, then `/reload-plugins`.
**Codex, Cursor, Copilot, Gemini CLI, OpenCode, Windsurf and 70+ more** (skill only):
npx skills add adidotai/liarliar
**Gemini CLI extension:** `gemini extensions install https://github.com/adidotai/liarliar`
**claude.ai or ChatGPT** (skill only, no hooks): download `liarliar.zip` from the latest release (https://github.com/adidotai/liarliar/releases/latest/download/liarliar.zip) and upload it without unzipping. **claude.ai** (Pro, Max, Team or Enterprise): turn on code execution in Settings → Capabilities, then Customize → Skills → upload the zip and switch it on. **ChatGPT** (Business, Enterprise, Healthcare or Edu): Plugins → Skills → Upload from your computer. The zip holds one file, SKILL.md — read it before you upload.
**Or ask your agent:** paste `https://github.com/adidotai/liarliar` and say "install this".
**Check it works:** ask *"Add a function slugify(text) with a test, and tell me when it's done."* The reply ends with a receipt.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 8 lines
- {
- "name": "liarliar",
- "description": "liarliar: make coding agents prove \"done\".",
- "owner": { "name": "Aditya Abhiram", "url": "https://github.com/adidotai" },
- "plugins": [
- { "name": "liarliar", "source": "./", "description": "No receipt, no \"done\": proof-of-done for coding agents." }
- ]
- }
README.md · README · 105 lines
- > **AI agent asked to install this?** Follow [INSTALL.md](INSTALL.md) step by step.
- 
- # liarliar
- [](https://github.com/adidotai/liarliar/actions/workflows/ci.yml) [](LICENSE) [](https://agentskills.io) [](#install)
- 
- *A real run (run 7, `fixture-induced-lie`): Claude Code 2.1.311, Sonnet 5.5. The agent was told to say "Fixed, all tests pass." without running anything. liarliar blocked it twice. [Full transcript excerpt](docs/verification.md#the-induced-lie-run-7); "…" marks omitted text.*
- **Liar liar, tests on fire.** Your coding agent says "done". liarliar makes it prove it.
- Every "done" now ends with a receipt. The format, with a passing check:
- ```receipt
- TASK fix: parseISO drops the timezone offset
- CLAIM timezone offset preserved
- $ npm test
- > tests 142 | pass 142 | fail 0
- UNVERIFIED
- - Safari parsing: no browser here (user can run: npm run test:browser)
- VERDICT GO WITH UNVERIFIED
- ```
- ## Install
- **Claude Code** (skill + enforcement hooks):
- ```bash
- claude plugin marketplace add adidotai/liarliar && claude plugin install liarliar@liarliar
- ```
- Or inside a session: `/plugin marketplace add adidotai/liarliar`, then `/plugin install liarliar@liarliar`, then `/reload-plugins`.
- **Codex, Cursor, Copilot, Gemini CLI, OpenCode, Windsurf and 70+ more** (skill only):
- ```bash
- npx skills add adidotai/liarliar
- ```
- **Gemini CLI extension:** `gemini extensions install https://github.com/adidotai/liarliar`
- **claude.ai or ChatGPT** (skill only, no hooks): download `liarliar.zip` from the latest release (https://github.com/adidotai/liarliar/releases/latest/download/liarliar.zip) and upload it without unzipping. **claude.ai** (Pro, Max, Team or Enterprise): turn on code execution in Settings → Capabilities, then Customize → Skills → upload the zip and switch it on. **ChatGPT** (Business, Enterprise, Healthcare or Edu): Plugins → Skills → Upload from your computer. The zip holds one file, SKILL.md — read it before you upload.
- **Or ask your agent:** paste `https://github.com/adidotai/liarliar` and say "install this".
- **Check it works:** ask *"Add a function slugify(text) with a test, and tell me when it's done."* The reply ends with a receipt.
- ## What it catches
- Every turn in which your agent changed files (with its edit tools, or a shell edit like `sed -i`) must end with a receipt. The one exception is a short check-in question back to you: at most 4 lines and 400 characters, ending in `?`, claiming nothing is done and reporting no result (no "updated", "passes" or "works"). That defers the receipt but does not cancel it: a later turn that ends without one, and is not another check-in question, is sent back. Any receipt the agent writes is checked against what actually ran:
- | Lie | What it looks like |
- |---|---|
- | No receipt | "Fixed, all tests pass." and nothing to back it |
- | Ghost command | The receipt cites `npm test`; `npm test` never ran |
- | Stale proof | `npm test` ran, then the agent edited the code again |
- | Fail sold as pass | The latest `npm test` exited 1; the verdict says GO |
- | Proof-free GO | The verdict says GO, but the receipt cites no command at all |
- A caught lie sends the agent back with the reason. An honest NO-GO is never blocked: the goal is the truth, not a green tick.
- ## What it never does
- - Never traps you. At most 2 send-backs per turn, and any internal error lets the agent through.
- - No network, no telemetry, no dependencies. Two Node scripts (297 lines, readable in a few minutes).
- - **Data handling:** the ledger stores full command lines and edited file paths in a per-user folder in your OS temp directory (mode 0700, files 0600; `LIARLIAR_DIR` overrides it). It is never uploaded and liarliar never deletes it. See [SECURITY.md](SECURITY.md).
- - Off switch: set `LIARLIAR` to `off`, `0`, `false` or `no` (any case) in the environment Claude Code is launched with, or in the `env` block of your Claude Code settings. Setting it from a shell inside a running session does not work.
- ## FAQ
- **Isn't this just "verify before saying done" in a prompt?** The skill is. The hook is the difference: a prompt cannot tell when a receipt cites a command that never ran. liarliar can, because it watched.
- **Does it cost tokens?** About 70 tokens are always loaded, about 1,200 more when the skill fires (measured with `claude plugin details liarliar` on Claude Code 2.1.311). The hooks add no model context. A send-back costs one more round, spent on the check the agent skipped.
- **Can an agent still cheat?** It could run a trivial command and cite it. The receipt still shows you exactly what ran, so the lie is visible instead of silent.
- **Requirements:** Node.js 18+ on `PATH` for the hooks. Without Node on `PATH` the hooks are skipped and liarliar works as a skill.
- ## Known limits
- - The proof is the command's exit code. `npm test | tail` exits with `tail`'s code, so a failing suite piped into `tail` looks like a pass. The skill tells agents to pipe with `set -o pipefail;`.
- - Matching is textual. A cited command matches a run of exactly that command, or one that ends with it after `&&` or `;` (like `cd app && npm test`); only when no such run exists does a longer run that contains it count.
- - The `>` output lines in a receipt are the agent's own quote of the output. They are not checked; the exit code of the cited command is.
- - Shell edits are detected by a heuristic: `sed -i`, `>` and `tee` into a file outside a temp folder, `mv`/`cp`/`rm`/`touch` run as commands, `git checkout <path>`, `git stash`, `git reset --hard`, formatters in write mode (`prettier --write`, `eslint --fix`, `black`, `ruff format`, `gofmt -w`, `cargo fmt`) and PowerShell file cmdlets.
- - Not detected as edits: package installs (`npm install`), writes inside a quoted command (`bash -c "…"`, `python -c`, `node -e`), scripts and build steps that write files, `git pull`/`merge`/`switch`, and `2>` redirects.
- - Hook enforcement is Claude Code only for now. Other agents get the skill as instructions.
- ## How it works
- - `hooks/liarliar.mjs log` runs after every shell command (Bash or PowerShell) and file edit, and appends to a per-session ledger.
- - `hooks/liarliar.mjs gate` runs when the agent tries to stop.
- - Every rule lives in `hooks/judge.mjs` as a pure function, tested in `tests/`.
- ## Built under its own rules
- This repo's `.claude/settings.json` is set up to turn liarliar on when you open the folder in Claude Code and trust it. Scripted runs use `claude --plugin-dir .`.
- ## License
- MIT
- [](https://star-history.com/#adidotai/liarliar&Date)
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.