DEV Community

Ștefănescu Liviu
Ștefănescu Liviu

Posted on

Show Dev: LoopTroop - Local-first AI coding orchestrator with LLM councils, Ralph loops, and a visual Kanban board

#ai #showdev #opensource #devtools

Hey everyone!

I wanted to share LoopTroop, an open-source, local-first GUI app I have been building to solve one of the biggest headaches with AI coding agents: context rot.

We recently crossed 120 stars on GitHub, and I am actively developing and improving it every day.

The Problem: Context Rot in Long-Running Tickets

AI coding tools are great when you ask them to write a quick function, generate boilerplate, or fix a single bug. But once you throw a complex, multi-file feature ticket at an agent, things usually fall apart after 10 to 15 tool calls.

The context window fills up with terminal outputs, compiler errors, large file dumps, and half-baked diffs. As that noise piles up, the model starts forgetting early requirements, re-introducing previously fixed bugs, or hallucinating code.

LoopTroop takes a different route: context engineering over infinite chat history. Instead of letting one long conversation degrade over time, it breaks work into distinct stages, persists state to disk as durable files, and feeds the model clean, minimal micro-contexts for each step.

How LoopTroop Works

LoopTroop is built around a "slow and precise" philosophy. Rushing an agent to write code immediately usually costs more time in cleanup than doing structured planning upfront.

1. Interactive Interview (Resolving Ambiguity First)

Before writing any specs or code, LoopTroop runs an interactive interview. It scans your repository, detects ambiguities in your ticket, and asks you direct clarifying questions about architecture, trade-offs, and edge cases. This makes sure the plan actually reflects what you want before any code gets touched.

2. The LLM Council (Multi-Model Planning)

Instead of relying on a single model's biases to architect your feature, LoopTroop uses an LLM Council for planning the PRD and task breakdown:

  • Multiple configured models draft their own proposals independently.
  • The models review and vote anonymously on the drafts to score the best approaches.
  • The winning plan absorbs the strongest ideas from the other drafts and runs a coverage check to ensure no requirements were skipped.

3. Decomposition into Atomic "Beads"

Large tickets are decomposed into small, self-contained units of work called Beads. Each bead defines:

  • A specific, narrow objective
  • Exact target files
  • Concrete acceptance criteria
  • Automated test commands to verify completion

The agent implements one bead at a time with a clean prompt focused only on that task.

4. Ralph Loops (Smart Failure Recovery)

When an agent gets stuck in a loop or fails a verification check, traditional tools just append the failure traceback to the chat, polluting future attempts.

LoopTroop uses a recovery loop pattern:

  • Each bead execution is time-boxed.
  • If a bead fails or gets stuck, LoopTroop stops the attempt, writes a short diagnostic note explaining what failed, and discards the noisy session history.
  • The next iteration restarts with a clean context, carrying over only that compact failure note. Failures turn into targeted learning rather than context pollution.

5. Local-First Visual Kanban Board

LoopTroop is a desktop GUI, not a CLI command or a hosted cloud platform. It gives you a clean Kanban interface to manage multiple repositories and multiple active tickets simultaneously:

  • Track phase progression in real time.
  • Inspect council drafts, voting breakdowns, and PRDs.
  • Monitor live execution logs and terminal output.
  • Review bead statuses, test outputs, and final Git diffs from one window.

Human in the Loop

LoopTroop is designed to keep you in the driver's seat at key decision points:

  • You answer the initial interview questions.
  • You review and approve the generated PRD and bead breakdown.
  • You can step in, pause, or adjust during bead implementation.
  • You review the verified changes before anything is handed over to your main branch.

Under the Hood

  • Execution Engine: LoopTroop uses the OpenCode harness for file operations, terminal execution, and model communication.
  • Model Agnostic: You can configure any model or provider supported in your setup (Anthropic, OpenAI, Google, DeepSeek, local models via Ollama, etc.) for both the main implementer and the council members.
  • Local: Runs locally on your machine against your local Git repositories.

Links & Trying It Out

Any feedback, critique, or bug reports are more than welcome. If you decide to give the app a spin and it works nicely on your projects or breaks completely on your stack, drop a comment here or open an issue on GitHub. I would love to hear your thoughts!

Top comments (0)