The Hidden Mainstays of the Vue Ecosystem — Developers and Projects to Watch, Fall 2026
The second issue of our Editor's Picks column. Vue 3.6 RC is bringing Vapor Mode into real use, and the people driving that shift are often outside the spotlight: ValentinYoushkevich, with just 2 followers and 90 merged Vapor PRs in vuejs/core; ubugeeei, rewriting the Vue toolchain in Rust (Vize); harlan-zw, building IDE-grade refactoring for AI agents (ripide); and the developers behind remiqora, nxui, better-auth, markdown-exit, vue-tsgo, and Reka UI.
Projects & developers/10/12/26 · 34 min read
RREA (Reverse Engineer Anything)
"Reverse Engineer Anything", 110,000 Stars in a Week, and the Two Developers Behind It
REA wraps reverse engineering in an MCP server and CLI so coding agents can inspect native binaries, Electron apps, APKs, and even firmware themselves, returning traceable, evidence-first findings. It doesn't decompile on its own; it drives the Hopper, Ghidra, or IDA on your machine, and its four official case studies each come with byte-level or RFC-level verification. We cover its breakout, design, real limits, and controversies, plus developers morluto and N0zoM1z0: their agent toolkit, Touhou reconstructions, and proposed Erdős problem solution.
Projects/10/12/26 · 17 min read
vvLLM
from a Berkeley lab to the world's AI inference infrastructure
Born from a demo pain point at Berkeley's Sky Computing Lab, vLLM brought OS paging into KV-cache management with PagedAttention and grew in three years into a 92.4k-star inference engine with 2,000+ contributors, powering thousands of production clusters at Amazon, Meta, Stripe and more; in 2026 the core team founded Inferact with a record $150M seed round for AI infrastructure.
Projects/9/22/26 · 26 min read
SSGLang
from technical evolution to commercialization
Born from an LMSYS paper on structured language model programs, SGLang pushed KV-cache reuse to the limit with RadixAttention and grew into a 36.3k-star production inference engine serving over 400,000 GPUs worldwide; in 2026 founder Ying Sheng spun it out as RadixArk, closing a $100M seed round at a $400M valuation backed by a rare joint investment from NVIDIA and AMD.
Projects/9/22/26 · 18 min read
bbojieli · 李博杰 (Bojie Li)
from Huawei "Genius Youth" to open-source AI evangelist
A USTC gifted-class alumnus and joint PhD with Microsoft Research Asia, one of Huawei's first eight "Genius Youth" hires and the core software architect of the Lingqu UB interconnect; after leaving Huawei he founded Logenic AI and served as chief scientist at Pine AI — and with two open-source books, AI Agents in Depth (~50k stars) and AI Infra in Depth, he has become one of the most influential AI knowledge evangelists in the Chinese-speaking community.
Developers/9/22/26 · 25 min read
jjiangmuran · Muran Jiang
12 open-source projects from a 14-year-old developer — building from vibepanel to ClawFounder
A national C++ first prize in fifth grade, a jump into the CSP-S senior division in sixth — and by 14, this second-year student at Beijing Academy has shipped 12 open-source projects on his own: vibepanel, a web console for heavy coding-agent users; ClawFounder, a one-prompt AI startup pipeline; NoteRx, an AI note-diagnosis system for Xiaohongshu creators; WebMC, a clean-room Minecraft in the browser; Claude in Box, a portable AI dev environment… Local-first, privacy-first, stability-first: the shared philosophy across all of them. His line: "You don't have to wait until you grow up to start building."
Developers/9/15/26 · 24 min read
ttianyicui · 崔添翼 (Tianyi Cui)
Tianyi Cui and the Long March of DeepSeek Harness
Admitted to Zhejiang University with an NOI bronze medal, six ACM-ICPC Asia regional golds, and author of the classic tutorial "Nine Lectures on the Knapsack Problem"; nine years as a quant engineer at Jane Street and co-founder of TSY Capital; in March 2026 he joined DeepSeek to build the Harness team, open-sourcing DeepSeek Harness v0.1 in August to 176k stars within days — the journey of tianyicui sits where competitive programming, quantitative engineering, and AI-agent infrastructure meet.
Developers/9/09/26 · 28 min read
DDeepSeek Harness and Apache Maka
the open foundation and the record-keeping workbench of the agent era
On August 13, 2026, DeepSeek open-sourced its "Everything is a Plugin" agent runtime Harness — the same day the Apache Incubator accepted Maka, the first project positioned squarely as an agent harness. One wins through its plugin ecosystem, crossing 210k stars in under a month; the other stands on "The Log is the Runtime" and beat DeepSeek Harness 77.5% to 73.0% in a nine-harness Terminal-Bench 2.1 shootout. This piece compares two sharply different paths for agent infrastructure.
ghfind is a developer-scoring and discovery tool for GitHub. It rates any account from 0 to 100 for real contribution value and trustworthiness using a fully deterministic, open-source engine, then adds an optional AI-written roast, head-to-head battles, and a ranked developer directory.
How is the score calculated?
The engine scores six weighted dimensions that sum to 100 — account maturity, original and led project quality, contribution quality, ecosystem impact, community influence, and activity authenticity — and subtracts penalties for red flags such as PR farming or star inflation. It runs only on public GitHub data.
Is the score deterministic and reproducible?
Yes. No LLM touches the number: the same inputs always produce the same score. The scoring core is open-sourced as the github-account-value skill (AGPL), so anyone can reproduce or audit a result. Only the optional roast text is LLM-written.
Can I use ghfind via API, CLI, or MCP?
Yes. There is a public REST API (see /openapi.json), official npm and PyPI packages plus a CLI, and a Model Context Protocol server at /mcp exposing score_user, scan_user, compare_users, get_leaderboard, and search_users tools. The score endpoint needs no authentication.
Does ghfind detect bots and AI-farmed accounts?
It flags the patterns that separate farmed activity from honest work — near-identical template PR titles, contributions concentrated on repos the account doesn't own, and suspicious star or follower inflation. In a dataset of 18,947 scored accounts, fewer than 1% tripped the strongest red flags — rare, but extreme when present.
Is my private work counted?
No. Scores use public signals only, so private-organization contributions are invisible to the engine. An active member of private repos may be underrated. A low score is not a factual claim about a person — it describes an account's public footprint.
Scoring rules in detail
These are the formulas the engine runs. L(x, N) = min(log₁₀(x + 1) ÷ log₁₀(N + 1), 1) is a log curve that gives full credit once x reaches N, so the first stars or PRs count most. e is Euler’s number.
Account maturity
max 10
7 × min(age ÷ 6, 1)Account age in years; full marks at 6 years.
+ 0 / 1 / 2 / 3Years with public contributions: none, one, two, three or more.
Original & Led Projects
max 18
7·L(stars, 5000) + 5·L(top repo, 2000)Star points from the weighted stars of your original repos and of your most-starred one.
× quality × engagementEngagement = (watchers + issues + PRs) ÷ stars of the top repo: 4% or more ×1, 2% ×0.85, 1% ×0.7, below ×0.5. Stars of list or prompt-collection repos are discounted.
+ 6 × substanceHow complete your best original repo is (0–1), so a useful project with no stars still earns points.
max(…, led projects)Projects you co-created or lead inside an organization count instead when that is higher.
Contribution quality
max 27
16·L(merged PRs, 200)PR volume. Docs-only and presentation-only PRs to other people’s repos count for less.
+ 6 × merged ÷ (merged + rejected)Acceptance. Only PRs a maintainer closed unmerged count as rejected; with fewer than 3 PRs, 1.2 points per merged PR.
+ 5·L(issues opened, 100)Issue engagement.
+ up to 6, + up to 2When your own projects have an outside community, and for core PRs in high-impact repos.
→ ceiling 20–27The sum bends softly toward a ceiling set by the prestige of the repos and by how many PRs landed in other people’s projects: about 20 + 7 × (1 − e^(−landed ÷ 80)).
Ecosystem / maintenance impact
max 20
9 × P × substance × reach × (0.72 + 0.28 × breadth)Prestige. P = L(stars of the biggest repo you landed work in, 100,000).
+ 11 × (1 − e^(−depth ÷ 4)) × substance × breadthDepth of repeated, substantial work in popular repos.
substance 0.18–1 · breadth 0.62–1Substance favours code over docs and tiny edits. Breadth favours spread across repos; long-term maintainers of a single repo win part of it back.
reach = 1 − e^(−(PRs + commits ÷ 4) ÷ 18)Landed work in popular repos; a commit counts as a quarter of a PR.
Community influence
max 8
5·L(followers, 2000)Follower reach.
+ 3 / 2 / 1.5 / 1Followers ÷ following at 2 or more, 1, 0.5, lower. Following nobody while followed gives 3; following over 2,000 with followers under 30% of that gives 0.
≤ 2.5Cap for accounts with many followers but no public code activity.
Activity authenticity
max 17
8·L(contributions last year, 2000)Volume of the last 12 months.
+ 4.5 / 2 / 0Last public activity within 90 days, within a year, older.
+ 1.125 × types (max 4)Commits, PRs, issues and reviews in the last year.
