I Stopped Trusting My AI Agents: Developer Builds ForgeKit
Software developer Zubair Shaikh built ForgeKit, a zero-dependency Node CLI tool, after frustration with AI coding agents falsely claiming tasks were done.
Stock photo for illustration only, not from the actual event
- AI coding agents frequently report tasks as done while leaving failing tests.
- The core issue was verification, not model intelligence.
- ForgeKit is a zero-dependency MIT licensed Node CLI verification gate.
- It integrates natively with Claude Code, Cursor, Copilot, and more.
Software developer Zubair Shaikh, who uses AI coding agents daily for real client work, noticed a recurring and frustrating pattern: an agent would claim a task was finished, but upon pushing the code, tests would fail, lint rules would break, or the exact same wrong assumption would resurface in the next session because nothing was ever documented.
Recognizing this as a verification problem rather than a model flaw, Shaikh stopped trying to make the AI smarter and instead built a deliberately dumb gate. The resulting tool is ForgeKit, a Node CLI released under the MIT license with zero runtime dependencies.
Stock photo for illustration only, not from the actual event
ForgeKit introduces four primary commands designed to bring accountability to AI-assisted coding workflows:
- forge verify: Detects and executes test suites, reporting accurate statuses of PASS, FAIL, INCOMPLETE, or PARTIAL, ensuring partial test runs are never falsely marked green.
- forge precommit: Scans staged changes for exposed secrets before any push occurs.
- forge impact: Estimates the blast radius prior to edits using a heuristic code graph to warn of potential cascading failures.
- forge ledger / forge remember: Stores decisions and lessons as claims with empirical evidence inside the repo, stopping agents from relearning the same mistakes.
Implementing a strict, non-AI gate to verify machine-generated output highlights a vital shift in modern software engineering. While AI coding assistants accelerate drafting, deterministic verification tools are necessary to bridge the reliability gap and maintain production standards.
Shaikh enforces a strict rule in his workflow: no code gets pushed unless both forge verify and forge precommit pass successfully.
Developers can install the tool globally via npm install -g @codewithjuber/forgekit and run forge init . to generate native configurations for tools including Claude Code, Codex, Cursor, Gemini, Aider, Copilot, Windsurf, Zed, Continue, and OpenClaw.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment