HIMANSHU.KUMAR
← Selected work01 / 03

AI developer tool

Loop Engineering

A workflow that helps coding agents plan, test, and finish long software tasks.

Role
Open-source contributor · derivative extensions
Date
Apr–Jul 2026
Stack
TypeScript · Node.js · MCP · Git
Links
Repository ↗
Loop Engineering system map showing scheduling, sub-agents, worktrees, skills, and persistent task state connected in one workflow.
Loop Engineering system map connecting scheduling, sub-agents, worktrees, skills, and persistent state.

Case in one minute

Problem
Long tasks lose context and test evidence can go stale
Solution
A bounded workflow tied to the current Git workspace
What I built
Resume state, code discovery, verification, and checker gates
Result
Workspace changes require fresh proof before completion

01 / Problem

Problem

Long coding tasks can lose context—and a passing test becomes unreliable after the workspace changes.

An agent needs enough saved state to resume useful work, while current files, diffs, and fresh checks must remain the source of truth.

02 / Solution

Solution

Carry each task through an explicit workflow and bind its verification to the Git workspace that was checked.

The task record stays small and inspectable. If files change after verification, the workflow marks that evidence stale and sends the task back for a fresh check.

03 / How it works

How it works

The complete path from input to a finished, inspectable result.

Loop Engineering workflow
Loop Engineering closed-loop workflow showing task clarification, planning, implementation, fresh checks, independent verification, delivery, and the verify-again return path.

  1. Start the task

    Record the objective and the evidence needed to finish.

  2. Save the plan

    Keep the next steps small enough to inspect and resume.

  3. Implement

    Change the real workspace and record the files that were touched.

  4. Run fresh checks

    Attach real command output to the current Git workspace.

    Files changed after the check? Verification expires and the task returns here.

  5. Independent check

    A separate checker reviews the result and its evidence.

  6. Finish

    Completion is recorded only while the evidence is still current.

04 / What I built

What I built

The concrete parts I designed, implemented, and tested.

  1. Bounded task state and resume

    I added a task record that preserves the objective, plan, changed files, verification, and checker result without treating saved context as repository truth.

  2. Workspace freshness enforcement

    I tied successful verification to the Git workspace so tracked or untracked changes invalidate old proof before checking or finishing.

  3. Codebase discovery workflow

    I connected Codebase Memory-first discovery with targeted local confirmation, keeping the full graph outside the prompt while exact files settle decisions.

  4. Maker and checker separation

    I kept implementation and review as distinct stages so completion depends on explicit, independently examined evidence.

05 / Results

Results

What the finished system demonstrates through working behavior, tests, and project artifacts.

  • The CLI records an implementation before it accepts verification evidence.
  • Workspace changes after verification prevent a clean finish until the check is refreshed.
  • The checker stage is kept separate from the implementation stage.
  • The derivative provenance and the exact additions are documented in the repository.
Loop Engineering architecture showing codebase discovery, task state, implementation, verification, checker review, and the stale-evidence return path.
Expanded system view from the repository, including the path that returns changed work to verification.