Skip to content
godic97Public

About

An AI code guardrail and validation framework that strictly constrains and controls rogue AI coding agents and flawed codebases through logic and code.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

10 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 

Repository files navigation

Gingoa logo: a robot head wearing the golden headband from Journey to the West

Gingoa

An AI code guardrail and validation framework that constrains and checks AI coding agents through logic and code, not through prompts.

The name

Gingoa (긴고아) is the Korean reading of 緊箍兒, the golden headband in the Chinese novel Journey to the West.

In the story, Sun Wukong, the Monkey King, is the strongest of the pilgrims escorting the monk Tang Sanzang to India: fast, clever, and quick to act on his own judgement. The bodhisattva Guanyin gives Sanzang a golden headband and a spell to go with it. Once Wukong puts the band on, he cannot take it off, and when he goes too far Sanzang recites the Tightening Spell (緊箍咒, 긴고주) and the band closes until he stops. The band takes nothing away from Wukong's strength; it is what lets Sanzang travel with that strength safely. At the end of the journey, when Wukong becomes a Buddha, the band disappears by itself.

An AI coding agent is in Wukong's place: capable, fast, and able to do real damage when it acts on a wrong judgement. Gingoa is the headband:

  • It is worn, not argued with. The checks sit in the harness around the agent, the way the band sits on Wukong's head, and do not depend on the agent agreeing to them.
  • It leaves the agent its strength. Gingoa does not make the agent do less; it makes what the agent does visible and checkable.
  • The person holds the spell. Decisions that matter, such as unmasking personal data or accepting a test gap, stay with the user.

Why

An AI coding agent reads your data, writes your code and tells you the result is fine. Instructions in a prompt can ask it to be careful; they cannot make it so, and they cannot prove that it was. Gingoa puts the checks where the agent cannot talk its way past them:

  • Enforced by the harness. Each check runs as a hook or a local program around the agent, so the verdict does not depend on the model's own account of what it did.
  • Measured, not asserted. A check produces evidence you can read, such as a score, a list of surviving bugs, or the exact values that were withheld.
  • Local. Checks run on your machine and send nothing anywhere.
  • Honest about limits. These are guardrails against mistakes and overreach by a cooperating agent. They are not a sandbox against an agent that sets out to defeat them; each component documents what it does not cover.

Components

Each component is its own repository, released and versioned on its own. Gingoa links them; it does not copy their code.

Component What it controls How Repository
deid-guard What personal data reaches the model Data files are read through a profile card, then a de-identified copy; every row the model reads is scrubbed against known values and detectors; unmasking a column needs the user's approval in a local dialog. Korean and English. godic97/deid-guard
mutation-gate Whether the agent's tests actually catch bugs When the agent writes or runs tests, a hook has it mutation-test the code under test with the language's own tool (Stryker for JS/TS, mutmut for Python, cargo-mutants for Rust, gremlins for Go, PIT for Java/Kotlin, Stryker.NET for C#, Stryker4s for Scala), add assertions until the surviving mutants die, and report the score. Other languages use mutants the agent writes itself, run against the project's own tests. A missing tool is installed only after the user agrees. An optional end-of-turn report covers the lines the agent changed. godic97/mutation-gate
claim-check Whether the agent's "tests pass", "build succeeds" and "lint is clean" are true At the end of each turn, a hook finds those claims in the agent's answer (English and Korean) and checks each against the last matching command in the session transcript: its recorded exit code or summary line, and whether code changed after it. The verdict (backed, contradicted, or unverified) is shown to the user; the turn is never blocked. godic97/claim-check

Install

The components are Claude Code plugins. This repository is a plugin marketplace that lists them all; add it once, then install the ones you need from your shell:

claude plugin marketplace add godic97/Gingoa
claude plugin install deid-guard@gingoa
claude plugin install mutation-gate@gingoa
claude plugin install claim-check@gingoa

Inside Claude Code, the same works as /plugin marketplace add godic97/Gingoa followed by /plugin install …. The marketplace points at each component's own repository, so you get the same release either way; each repository also works as a marketplace on its own.

Each repository's README lists its requirements, configuration and limits.

License

MIT. See LICENSE. Each component carries its own license in its repository.

About

An AI code guardrail and validation framework that strictly constrains and controls rogue AI coding agents and flawed codebases through logic and code.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors