GitHub Open Source Weekly 2026-08-12: Agent Skills Become Workflows
Data period: 2026-08-04 to 2026-08-12 (rolling 7 days) Sources: GitHub Trending weekly and monthly, GitHub Search API, GitHub REST API, HN Algolia
TL;DR: reverse-skill, TencentDB-Agent-Memory, and pdf-inspector led this week's growth. Their domains differ, but all three make AI work more repeatable. Cloudflare Computer, book-to-skill, and loopx push the same question further: can an agent keep working safely over time? New projects brought phone control, editable slides, short-drama production, and an ad-free music client into the developer conversation.
Fastest Growing: Top 10
| # | Project | Stars/week | Total stars | Language | Created |
|---|---|---|---|---|---|
| 1 🔁 | reverse-skill | +8,182 | 23,939 | PowerShell | 2026-05-13 |
| 2 🔁 | TencentDB-Agent-Memory | +7,555 | 19,858 | TypeScript | 2026-04-07 |
| 3 | pdf-inspector | +7,143 | 14,681 | Rust | 2026-02-06 |
| 4 | computer | +6,807 | 7,590 | TypeScript | 2026-06-05 |
| 5 🔁 | book-to-skill | +4,113 | 20,485 | Python | 2026-05-01 |
| 6 | DeepSeek-Reasonix | +4,109 | 33,957 | Go | 2026-04-21 |
| 7 | airllm | +4,042 | 30,762 | Jupyter Notebook | 2023-06-12 |
| 8 | loopx | +2,947 | 4,153 | Python | 2026-05-31 |
| 9 | ComfyUI | +2,842 | 126,809 | Python | 2023-01-17 |
| 10 | google/skills | +2,159 | 17,770 | Python | 2026-03-31 |
Top New Repos: Top 10
| # | Project | Total stars | Language | Created |
|---|---|---|---|---|
| 1 | photo-abstract-editorial | 2,510 | Not listed | 2026-08-04 |
| 2 | human-writing | 2,406 | Python | 2026-08-05 |
| 3 | open-kimi-ppt-skill | 1,603 | Not listed | 2026-08-05 |
| 4 | phone-harness | 1,493 | Python | 2026-08-07 |
| 5 | oil-motion | 1,478 | Python | 2026-08-07 |
| 6 | WeChat-AI | 1,410 | TypeScript | 2026-08-10 |
| 7 | h3.c | 1,214 | C | 2026-08-09 |
| 8 | shuohao-skills | 861 | JavaScript | 2026-08-06 |
| 9 | youtube-music | 835 | TypeScript | 2026-08-11 |
| 10 | awesome-growth-hacking-skills | 826 | Shell | 2026-08-04 |
Why these projects matter
reverse-skill routes reverse-engineering and authorized security work to repeatable methods and supports several coding agents. It is not a casual install: authorization, scope, and safety procedures matter. MIT licensed, +8,182 this week, 23,939 total stars, and also on the monthly list.
TencentDB-Agent-Memory turns conversations, documents, and code into shared Chat Memory, Skills, an LLM Wiki, and a Code Graph. It feels like a shared notebook for agents, but its license field is NOASSERTION and the repository has 630 open issues, so governance deserves attention. +7,555 this week, 19,858 total stars, also monthly.
pdf-inspector is a Rust library for PDF inspection, classification, and extraction. Its scanned-versus-text classification can route a document to the right OCR or extraction path, which is more useful than simply swapping chat models. MIT licensed, +7,143 this week, 14,681 total stars. The matching HN posts had only single-digit interaction, so they are not evidence of broad consensus.
Cloudflare Computer takes the direct approach of giving an agent a computer. Browser and desktop automation become possible, but permissions, isolation, observability, and failure recovery matter more than a polished demo. MIT licensed, +6,807 this week, 7,590 total stars. The matching HN discussion had 10 points.
book-to-skill converts a technical-book PDF into a Claude Code skill. That is appealing for anyone who wants to turn reading into a reusable workflow, but generated skills still need review. MIT licensed, +4,113 this week, 20,485 total stars, also monthly.
DeepSeek-Reasonix is a DeepSeek-native terminal coding agent designed around prefix-cache stability. It is aimed at people who are comfortable leaving a terminal agent running, not a reason to install it just because its star count is high. MIT licensed, +4,109 this week, 33,957 total stars.
AirLLM targets 70B inference on a single 4GB GPU. A related HN post reached 231 points and 85 comments, making it the strongest discussion signal in this dataset, though actual speed and model support depend on your hardware. Apache-2.0 licensed, +4,042 this week, 30,762 total stars.
loopx focuses on persistent goals, wake-ups, executable todos, evidence logs, and handoffs for long-running agent teams. In plain English, it tackles the problem of where an agent stopped yesterday and how to resume safely today. MIT licensed, +2,947 this week, 4,153 total stars.
ComfyUI remains a major node-based diffusion GUI, API, and backend. It is not new, but its 126,809 total stars and +2,842 this week show continued interest. It is powerful for image and video workflows, though its model and node setup is not beginner-light. GPL-3.0 licensed.
google/skills is a collection of Agent Skills for Google products and technologies. Alongside reverse-skill and book-to-skill, it suggests Skills are becoming installable workflow units rather than prompt snippets. Platform compatibility and permissions still need checking. +2,159 this week, 17,770 total stars.
The new-repository list follows the same direction: phone-harness lets an agent control a phone, open-kimi-ppt-skill creates editable slides, and shuohao-skills packages AI short-drama production. human-writing tackles the very real problem of AI text sounding generic. oil-motion, photo-abstract-editorial, WeChat-AI, h3.c, youtube-music, and growth-hacking skills cover animation, visual work, messaging, inference, music, and marketing. New projects can spike after one post, so read the README, license, and security notes before putting them on a daily device.
Monthly trend comparison
reverse-skill, TencentDB-Agent-Memory, and book-to-skill appeared in both weekly and monthly trends. That persistence matters more than a one-day spike, but a weekly-only project is not necessarily weaker; it simply reflects short-term acceleration in this snapshot.
This week's takeaways
Skills are becoming workflow components. The common thread across several repositories is saving, reusing, and executing methods instead of starting from a blank prompt.
Agent continuity is the next bottleneck. Memory, state, handoffs, and computer access all point to the same shift: useful automation must resume safely, not just answer once.
Practical infrastructure still matters. PDF routing, low-hardware inference, and visual workflows reduce friction in existing work, which is often more valuable than another flashy demo.
Was this article helpful?
