What is deccy? deccy is a decision management framework we developed, from first principles, to perform software engineering with AI. It is a pre-alpha version of the system we are building to reshape the way engineering is done. We have made it in three parts: deccy-cli provides a CLI tool which encodes most of the framework deccy-agent is a set of custom agents that we use deccy-viewer visualises the decisions made and their relationships The primary motivation behind deccy is to separate high-level decision-making (the key contributions of human engineers) from execution (which AI excels at), with emphasis on having high traceability of decisions and precision in making changes. It’s an explicit format in which humans and agents share context on the decisions that have gone into shaping the codebase, as opposed to opaque memory layers, or searching through (potentially stale) documentation and session logs. A codebase is some sequence of decisions, that are related in a vaguely hierarchical way You start with high-level decisions (I want to make a todo app) Gradually, you refine those decisions (What features should this app have? What platforms should it support? What language/framework should I use?) You refine them further (How does the “mark complete” feature work?) And further (When I mark an item as complete, what happens to the item?) Then you translate those into implementation decisions (schemas, data structures) And finally into actual implementation (code) With deccy, we model the decisions made in a codebase with a directed acyclic graph: Everything is a decision, from broad feature requests, to data schemas and even implementation code Each decision can depend on several decisions that precede it Each decision can influence several decisions Every decision is justified with a rationale Every decision is evaluated against the decisions it depends on Evaluation is a core part of deccy: when making a decision, we implicitly ask questions like “Does this make sense? Does it go against other decisions? Does it satisfy my intent?”. As such, we designed deccy to require that every decision is evaluated against every parent that it depends on. flowchart LR classDef decision fill:#ffffff,stroke:#1f2937,stroke-width:1.5px,color:#111827 classDef evaluation fill:#f5f3ff,stroke:#7c5cff,stroke-width:1.5px,stroke-dasharray:5 3,color:#3b2b7a PRE["Decision prerequisite"]:::decision EVAL("Evaluation"):::evaluation DEP["Decision dependent"]:::decision PRE --> EVAL PRE --> DEP EVAL -.->|evaluates| DEP So far, this has helped us make, and later change, many decisions without diving into the code, but has also made it easier to do a deep dive if needed. We have also observed (albeit anecdotally) that coding agents are much better at understanding what changes we want to make (and how to make them) with less steering when working on projects that are modelled with deccy. Using deccy has also allowed us to be more agile than with the usual PLAN.md approach to communicating decisions. Admittedly, deccy is not for every project. Often, a one-shot/few-shot product with the latest Fable/Astra model (or whatever tops the benchmarks when you read this) is good enough for a demo or even a MVP. The benefits we have observed so far are also primarily intangible (like finding deccy less frustrating to work with). We’re targeting long-term projects, where intent can be lost/forgotten, and our objective is to make collaborating with AI on these projects a better experience for developers and teams. deccy is our first step in that direction. How do I use deccy? First, it helps to install deccy-cli. Then, you’ll have to decide what types of decisions matter to you. For deccy-cli, we went with high level text decisions for intent, refined them with models written with EARS and Quint, and wrote the implementation with Rust. You’ll also want to choose how to evaluate these decisions - we chose simple non-contradiction checks for text and EARS decisions, checks for Quint models and tests for Rust code. flowchart TD classDef decision fill:#ffffff,stroke:#1f2937,stroke-width:1.5px,color:#111827 classDef evaluation fill:#f5f3ff,stroke:#7c5cff,stroke-width:1.5px,stroke-dasharray:5 3,color:#3b2b7a classDef heading fill:none,stroke:none,font-size:20px,font-weight:700,color:#1c1917 %% Layer headings are nodes rather than subgraph titles, so edges route around them. %% Node bodies are a two-column row: a bold "content:" label and a left-aligned content column. subgraph intent[" "] h1["Intent — plain text"]:::heading nzrnrlky["C-DRAFTS content:`deccy drafts` is one verb: invoked bare, it lists the registry's records... "]:::decision end subgraph ears[" "] h2["Model — EARS"]:::heading qspsyqlp["EARS content:When invoked bare, the deccy CLI shall list every record in the draft registry... "]:::decision wtnnxtql("gate: EARS content:1. License, applied once per prerequisite in depends_on: no element of the... "):::evaluation end subgraph rust[" "] h3["Implementation — Rust"]:::heading kpkplmqy["M-verb-drafts content:src/verb_drafts.rs "]:::decision rvpsuzys("gate: M-verb-drafts content:tests/verb_drafts/Cargo.toml tests/verb_drafts/Cargo.lock tests/verb_drafts/src/lib.rs "):::evaluation end nzrnrlky --> qspsyqlp qspsyqlp --> kpkplmqy nzrnrlky --> wtnnxtql wtnnxtql -.-> qspsyqlp qspsyqlp --> rvpsuzys rvpsuzys -.-> kpkplmqy We chose one intermediate “layer”, but depending on factors like scale, complexity or stakes of your project, you may choose more (or fewer) layers. Crucially, each decision should be a sufficiently small logical step (by your chosen standard) from the decisions it depends on. deccy-agent is our set of helper agents for managing decision graphs. It comes with the steps to turn a proposed feature or change into a modification of the decision graph. The workflow using deccy-agent looks something like this: You tell the deccy-plan agent about your intended change The planner agent reads the graph, determines the right starting location(s) for your change You review and refine the plan, then pass it to the deccy-orchestrator agent The orchestrator kicks off the workflow: It creates skeletons for any new decisions and evaluations Then, it starts subagents which draft the evaluations and decisions, before a final subagent judge assesses quality Any problems are surfaced for you to deliberate on, before the drafts are merged into the decision graph You review the changes and, once approved, the orchestrator closes the round To view and navigate the graph more easily, use deccy-viewer, a VSCode extension which displays the decision graph and provides a search feature. Why deccy? We find software engineering with AI is broken. Producing slop has become too easy, but producing good software has ironically become more difficult. More on that here. Instead of producing more software that’s (barely) tolerable, why shouldn’t we direct the power of AI into crafting great software? Why not Spec-Driven Development? deccy is very similar to SDD! The key difference is that specs are broken up into unit decisions, with relationships between decisions more explicitly mapped. Historically, this required a lot of upfront work which was rarely justified. Now with AI, most of that work is cheap, and it’s much more justified to make these mappings once and early, than to rediscover them every time. Will this become obsolete when models improve? Probably not. While we initially made this to overcome the shortfalls of LLM agents, we’ve discovered that we humans are really the problem. As most software engineers and project managers are aware of, humans are notoriously poor at defining their demands clearly. This problem is amplified by AI, and the practice of deploying before we have understood the demand. deccy provides a sensible pacing, the grill-me effect by design and a way to identify and correct incorrect decisions with minimal blast radius. This is cool! How can I get involved? Thanks! The quickest way to get involved is to use deccy - we want to explore what deccy can and can’t do - and send your feedback to us at hello@edawise.com. Or send us an email to discuss other ways we could work together, we’d love to hear from you! This sucks Please let us know why at hello@edawise.com! We know that deccy has a long way to go, and we’re working hard to improve it, so all feedback is very welcome. I don’t know how to use this for my specific use case Drop us an email at hello@edawise.com! We’d love to help
AI NEWS JOURNAL
Hacker News Show HNAI NEWS / JOURNAL
Show HN: Deccy – software engineering as decision management
Hi HN,We built deccy out of frustration from working with coding agents.Although we love the speed that AI has brought to our work, we also want to have control over what we build. While it's not impossible to keep up with the changes AI makes, it is significantly more difficult when compared to writing code ourselves.We realised that what we cared about most was not controlling every line of the implementation, but preserving our intent (we especially hated when agents would wrongly guess what we wanted). We iterated through planning and spec-driven methods, but found that both writing our own plans/specifications and reading AI-generated versions was tedious and exhausting.Eventually, we decided to split the specifications up into the decisions that make them up, and get agents to do the boring work of tracking and bookkeeping. deccy is the first realisation of this idea. We're already working on the next iteration, but we wanted to share the mental model we developed.There's a long way to go from here, and we'd really love any feedback :) Comments URL: https://news.ycombinator.com/item?id=50483376 Points: 1 # Comments: 0
번역 서비스가 일시 중지되어 원문을 표시합니다.