Hey HN, I’m Ali.

I built Decision Guardian to solve a problem I call "institutional amnesia." Senior engineers build complex systems (like highly-tuned DB connection pools or sensitive auth middleware) and document the "why" in wikis or ADRs. But inevitably, a new developer modifies that code without reading the docs, and things break. While tools like CODEOWNERS are great for telling you who needs to review a file, I needed something that automatically tells the developer why the review matters before they even ask.

What it is: Decision Guardian is an open-source tool that shifts teams from passive documentation to active enforcement. You write your architectural decisions in a standard Markdown file (e.g., .decispher/decisions.md) and map them to file paths using glob patterns or advanced JSON rules. When a developer modifies those protected files, it automatically analyzes the diff and drops a structured comment into the Pull Request explaining the context, warnings, and checklists.

I just released v1.1, which includes a major internal SOLID refactor. The core engine is now 100% decoupled from GitHub (using ISCMProvider and ILogger interfaces), allowing it to ship as a standalone 430KB CLI (npx decision-guardian check) that works locally, in pre-commit hooks, or on GitLab/Jenkins.

Under the hood (How it works): I wanted this to be fast and secure enough to run on massive enterprise repositories, which required a few fun technical implementations:

O(1) Pattern Matching: Naively matching N decisions against M files in a 3,000+ file PR is painfully slow. The core engine compiles all file patterns into a Prefix Trie (an Aho-Corasick variant), turning candidate lookup into an O(1) operation relative to the number of decisions.

ReDoS-Protected VM Sandbox: For complex rules, you can use Regex to scan diff contents (e.g., detecting hardcoded credentials). Because user-supplied regex can cause catastrophic backtracking, all regex evaluation runs inside an isolated Node VM sandbox with a hard 5-second timeout and pre-flight safe-regex checks.

Progressive Comment Truncation: GitHub limits PR comments to 65,536 characters. To prevent failures on massive PRs, the tool uses a 6-layer progressive truncation algorithm that intelligently summarizes details before falling back to a hard truncation.

Privacy & Telemetry: I know HN cares deeply about telemetry. v1.1 introduces anonymous usage metrics to help me understand performance (e.g., node version, duration, match counts). However, I designed it with strict privacy constraints:

Zero PII: It never collects source code, file contents, repo names, branches, or paths.

Runtime Blocklist: The privacy module physically validates every payload against a blocklist before sending. If any blocked field is detected, it throws an error and rejects the payload.

Open Source Backend: It is fire-and-forget (never slows down the tool), entirely opt-out (DG_TELEMETRY=0), and the Cloudflare Worker backend code is fully open-source in the repo.

Links:

Repo: https://github.com/DecispherHQ/decision-guardian

Docs/Architecture: https://decision-guardian.decispher.com

I also included 5 production-ready templates (security, database, api, etc.) to help teams get started instantly.

I’d love to hear your thoughts on the architecture, the Trie implementation, or how your teams currently handle enforcing ADRs during code reviews!

Ask HN: How do I use AI as a tool if some answers are objectivity incorrect?

Will AI Replace You? – Roast My Career

KeyWalking: Pattern Based Passwords (2015)

Agent-md/session-commit: Update your AGENTS.md after each session

Show HN: MQTT with Runtime Protobuf and Service Discovery (Zig)

Ask HN: Anthropic has stood its ground. What excuse is left for other companies?

Open Source Brain Stimulation: TDCS (Transcranial Direct Current Stimulation)

Marianne Bachmeier

We Will Not Be Divided

Stop Putting Secrets in .env Files

Struere: Lovable for AI Agents

GW of datacenter demand queues up for UK grid access

QED in Context: An Observation Study of Proof Assistant Users [pdf]

Japan's Rapidus lands $1.7B to chase 2nm chip production by 2027

The whole economy pays the Amazon tax

Oak Ridge spawns institute to curb AI datacenter power surge

Mash – graphical remote browser shell

Zclaw – The 888 KiB Assistant

Block, the parent of Square and Cash App, is laying off over 4k people

There Are No Psychopaths

Banks can Prevent Bank Runs by operating more like Bond Funds

Jsontap – Reactive access to incrementally parsed JSON for Python

Show HN: Decision Guardian – Auto-surface architectural context on PRs and CLI

FCC approves Charter Communications' $34.5B deal to buy Cox

Shift from passive documentation to active enforcement

Show HN: Accept.md now supports SvelteKit – return Markdown from any page

I stopped using JSON for MQTT and use Zig to develop gRPC-like communication

Raided by the Police – Investigating Nintendo, Sega, & Devkit Arrests [video]

Sam Altman Says OpenAI Is Working on Pentagon Deal

Autokey Wayland – fork of AutoKey a desktop automation app with Wayland support

Ask HN: Do you enforce Prettier in all Next.js projects?

Comments