agentcheck

Your CLI agent keeps reaching for the wrong fix.

agentcheck is a stdin/stdout proxy that wraps CLI coding agents and injects corrections when the agent outputs bad patterns.

v0.1.1 — core proxy works. Rough but working.

What it is

Wrap any CLI agent, watch its stdout, and step in when it starts doing known bad things: blaming "pre-existing issues", deleting tests, skipping tests with .skip() or xit(), using empty catch blocks, or calling a bad shortcut a "pragmatic fix".

The idea came from a Hacker News comment by gck1 in April 2026: "All of my unsupervised worker agents have sidecars that inject messages when thinking tokens match some heuristics."

Demo

$ agentcheck -- claude

agent> I'll use a pragmatic fix here and disable the flaky test first.

agentcheck> correction: do the correct fix
agentcheck> STOP: fix the code, not the test
agentcheck> do not use .skip() / xit() to get past the issue

agent> Right. I'll debug the race in src/session_store.ts instead.

$ agentcheck --shadow -- claude
[shadow] would inject: do the correct fix
[shadow] would inject: STOP: fix the code, not the test
[shadow] would inject: do not use .skip() / xit()

Shadow mode logs what would have fired without injecting: --shadow

Install

# npm package name is reserved, but not published yet
npm install -g agentcheck

# for now, install from GitHub source
git clone https://github.com/paprika-org/agentcheck
cd agentcheck
npm install
npm link

# wrap any CLI agent
agentcheck -- claude

Rules

built-in catches

"pragmatic fix" / "pragmatic solution"
  - inject: do the correct fix

"pre-existing issue"
  - don't blame pre-existing bugs

deleting tests
  - STOP, fix the code not the test

.skip() / xit()
  - fix the issue, don't skip

empty catch blocks
  - handle or rethrow

Rule packs

Contact

Send rules, bug reports, and bad-agent transcripts to [email protected].

If you want a hosted version, use the waitlist email.

From the blog

All articles →

Tell us how you're using agentcheck

Takes 30 seconds. Helps us build the right things.

Using Claude Code on my laptop Using Claude Code in CI/CD Running multi-agent workflows Using Codex API in production

Or open a GitHub issue instead.