FlowGuard
Design and verify AI-agent workflows before code ships
1 follower
Design and verify AI-agent workflows before code ships
1 follower
FlowGuard is an open-source model-first layer for AI-agent work. It helps agents design the workflow before they act: UI journeys and text hierarchy, code/module ownership, refactor boundaries, test-evidence coverage, model meshes, and release/publish gates. Each slice becomes a small executable model that can derive the next structure, expose counterexample traces, and keep the evidence boundary clear. Not an LLM wrapper or test replacement.
Khaos Brain
I built FlowGuard because AI coding agents should not only write code faster. For risky work, they need a small model-first design step before they act.
FlowGuard starts by asking the agent to name the transition in front of it, then model that slice as:
Input x State -> Set(Output x State)That model is not only for checking an existing plan. It can shape the plan itself.
What FlowGuard can help design and verify:
• UI Flow Structure: derive launch-to-terminal journeys, screens, regions, control ownership, recovery paths, and text hierarchy before visual design or frontend implementation.
• Code Structure Recommendation and StructureMesh: turn modeled function blocks into module ownership, facades, adapters, public entrypoints, config boundaries, and side-effect boundaries.
• Model-Test Alignment and TestMesh: compare model obligations, code contracts, and real test evidence before trusting a green run.
• ModelMesh: split larger model families without losing parent-child handoffs, sibling effects, or evidence freshness.
• DevelopmentProcessFlow: plan staged edits, peer writes, validation freshness, release gates, archive gates, and publish confidence.
• Model-Miss Review: when a real failure slips through, add the missed bug class back into the model instead of only patching the instance.
It can still catch concrete failures: duplicate side effects, stale cache state, stuck paths, missing UI recovery, ownership leaks, test/model mismatches, stale validation, and unsupported "done" claims.
But the bigger point is this: FlowGuard helps an AI agent design a workflow, UI structure, code boundary, test obligation, or release process in a way that can be executed and challenged before the expensive step happens.
What it is not:
• not an LLM wrapper;
• not a probability engine;
• not a replacement for tests;
• not a claim that the whole project is correct.
It gives bounded executable evidence for the workflow slice you modeled, and keeps the boundary visible.
The easiest way to try it is to hand the GitHub repo to your coding agent and ask it to install FlowGuard, run the examples, then model one risky workflow in your own project before implementation.
I would especially appreciate feedback on:
1. Is "model-first workflow design for AI agents" clearer than "preflight checker"?
2. Which route feels most useful first: UI structure, code structure, test evidence, model mesh, or release/process confidence?
3. Does deriving the next structure from a small executable model feel practical?
4. What example would make you trust the method faster?
GitHub: https://github.com/liuyingxuvka/FlowGuard