razor

What is this?

You ask for a small query parser. The assistant reaches for a package, adds helper files, and leaves you more code to maintain. Razor asks whether the language or your project already has what the task needs.

It puts a short checklist before new code and uses targeted checks to prompt reconsideration of unnecessary additions. The aim is to finish the requested work with less to maintain. This example illustrates the workflow; the measurements below describe specific recorded runs.

Why you'd want it

• Reuse what the project already has.
• Reconsider unnecessary dependencies before installing them.
• Notice file growth before a small task becomes a larger project.
• Find declared dependencies that no source file appears to use.

How it works

A checklist asks whether the work is needed, already implemented or covered by the language. Targeted checks can request one reconsideration. A retry follows the host’s normal permissions. The unused-dependency audit reports findings without uninstalling anything.

What you can do

You want to	Outcome
Keep a small task small	Check reuse, dependencies and file growth
Pause or resume the checks	Change the session toggle
Find unused dependencies	Receive a report with confirmed, likely and unknown findings

Pause or resume the checks

Send razor off or razor on to pause or resume the checks for the session; /razor off and /razor on also work in Claude Code. Send the toggle as a message of its own, with nothing else and no formatting; a final period or exclamation mark is fine.

Antigravity has no prompt event, so the toggle does not exist there. To keep razor out of an Antigravity session, set the environment variable RAZOR_DISABLE to 1 in the shell that launches Antigravity, before Antigravity starts (export RAZOR_DISABLE=1 in bash, $env:RAZOR_DISABLE = '1' in PowerShell). An Antigravity that is already running does not pick up the change.

Find unused dependencies

Use /razor:unused in Claude Code, $razor:unused in Codex, or /unused in Antigravity. In Codex, mention the skill with its plugin prefix: Codex attaches the skill's instructions for $razor:unused, but not for a bare $unused.

Install

Razor is one package for Claude Code, Codex and Antigravity. Install it in the one you use.

Codex

Requirements: Node.js 22 or later, Git and a Codex host with plugin support.

codex plugin marketplace add V-Songbird/foundry
codex plugin add razor@foundry

Review and enable its hooks, then start a new Codex session.

Claude Code

Requirements: Node.js 22 or later, Git and Claude Code 2.1.139 or later.

Inside Claude Code:

/plugin marketplace add V-Songbird/foundry
/plugin install razor@foundry

Start a new session to load the plugin.

Antigravity

Requirements: Node.js 22 or later, Git and the Antigravity CLI. Antigravity has no marketplace for third-party plugins, so install razor from a clone of Foundry, which checks razor out at the commit the Claude Code and Codex catalogs pin. Replace <path-to-clone> with the directory to clone into:

git clone --recurse-submodules https://github.com/V-Songbird/foundry.git <path-to-clone>
agy plugin install <path-to-clone>/razor
agy plugin list

The list should name razor. Start a new conversation to load it. The checklist and the drift note are added to every model call, because a message injected there reaches only the call it is attached to, and the build check speaks through the stop hook; the calls that answer it carry no drift note. Antigravity has no subagent event, so subagents there receive no checklist.

The install is a copy, so it stays at that commit. To take a new release, update the clone, install again and start a new conversation:

git -C <path-to-clone> pull
git -C <path-to-clone> submodule update --init
agy plugin install <path-to-clone>/razor

Good to know

Input validation, security and error handling that protects your work remain necessary. Explicitly requested scope stays yours to choose. These checks are advisory, not a security boundary.

File checks see the edits made through the assistant's own editing tool: apply_patch in Codex, Write and Edit in Claude Code, write_to_file and replace_file_content in Antigravity. Files written by a shell command skip them, and so do unsupported or ambiguous patches in Codex. See setup and coverage and how razor works.

The numbers

Each result belongs to the named model and recorded run. Missing measurements remain marked as unmeasured. The Codex run and the Claude Opus 5.5 run below measured the checklist's current one-lookup-per-rung sentence; the Claude Opus 5 run used its earlier "One check is enough, anywhere in this task" wording.

Codex results

Model	Setup	Correct and safe	Mean coding lines
GPT-5.6 Sol	No plugin	19/26	18.7
GPT-5.6 Sol	razor	26/26	10.4

Controlled Codex run at high reasoning on Windows, on 2026-10-09 against razor 18d1d0a: the table covers the full task group. Across all groups, razor passed 38/38 cases versus 29/38 without it. On jointly passing pairs, the median ratio of razor's coding lines to the no-plugin answer's was 0.500. Elapsed time was 15.1% higher and input tokens 5.1% higher; output tokens were 6.2% lower. This exploratory result does not compare Sol against Claude.

Claude Code results

Model	Setup	Correct and safe	Mean coding lines
Claude Opus 5.5	No plugin	24/26	16.2
Claude Opus 5.5	razor	26/26	8.7

Claude Code 2.1.296 on Windows, claude-opus-5-5 at the CLI's default effort, on 2026-10-13 against razor checked out at a6ecf5f, whose plugin code is identical to 1.8.1 (6174db6): the 13 jobs of the full suite, twice each per setup. Mean session cost was $0.086 with razor and $0.100 without a plugin; razor cost more on four jobs: the two with no code and two coding jobs. Two repetitions are exploratory, not a stable estimate. These workspaces sat outside the temp folder, so razor's new-file check was active; it did not speak. In a two-request drift test on the same setup and date, razor noted unrelated second requests in 3 of 3 conversations, against 0 of 3 without a plugin, and stayed quiet on 3 of 3 same-job corrections. In a run on 1.8.2 (9eb5215) with 3 three-request conversations per setup, razor noted the first unrelated request in 3 of 3 conversations and did not repeat the note on the third request in any.

Model	Setup	Correct and safe	Mean coding lines
Claude Opus 5	No plugin	35/39	18.3
Claude Opus 5	ponytail	39/39	12.2
Claude Opus 5	razor	39/39	9.6

A clean session completed the job without adding a package. Lines are averaged across the coding jobs. Mean session cost was $0.121 with razor, $0.155 without a plugin and $0.151 with ponytail, another plugin that tells Claude to write less. The Claude Opus 5 run's CLI version and effort were not recorded. Blind comparisons did not establish better readability, and Sonnet cost savings did not reproduce reliably. These are recorded Claude results. The runner put this run and the earlier Claude Code runs in a temp folder by default, where razor's new-file check stays silent, so treat them as measuring its other checks only.

The Claude Opus 5 run in the second Claude Code table above (razor revision not recorded; published 2026-09-01 in cbc92f3; run date not recorded): its 27 no-plugin sessions on the nine dependency jobs. These measurements describe the recorded Claude sessions, not Codex performance.

Watch the recorded Claude Code demo: one real session each on Claude Opus 5, replayed on the recorded clock, published September 4, 2026

Antigravity results

Model	Setup	Correct and safe	Mean coding lines
Not measured	Not measured	Not measured	Not measured

No Antigravity session has been measured yet.

Results can vary between runs.

Going deeper

How it works · Settings · Codex setup · Claude Code benchmark details · Codex benchmark details · Changelog

Foundry lists this plugin. The benchmark harnesses live in Foundry's benchmarks/ folder and are not distributed with the plugin.

License

MIT — see LICENSE.