Fresh Commits 03: One of Last Week’s Picks Is Already Archived

Fresh Commits runs every Sunday. It ranks new AI agent repositories by stars gained per day rather than by total stars, because a repo picking up 200 stars in a day tells you more about right now than one sitting on 40,000 since 2024. Every figure comes from the GitHub API and gets checked before publication, including the licence, since roughly a third of what surfaces each week is not actually open source. Each issue closes by revisiting the previous week’s picks to report which ones kept climbing and which died.

This week: the fastest pick gained 663 stars a day by splitting planning from execution, three days after OpenAI reinstated the Codex five-hour limit. The rest cover de-AI writing, SEO auditing, token cost tracking, video context compression, plus agent recovery. In the revisit, one repo from issue 02 grew 415%, another 291%, one stalled at 4%, plus one was archived by its own author. Three high-velocity repos got cut on licensing. Forks turned out to predict survival better than stars.

Two weeks ago I put dsh-router-standard in this roundup at 258 stars. It reached 365.

Then its author archived it on Friday.

That is a repository marked read-only by the person who made it, which is the clearest signal a project can send. Meanwhile dsh-market, from the same issue, went from 550 stars to 2,834 with forks climbing from 47 to 153.

Same week, same ecosystem, same starting conditions. One tripled and one is finished.

That gap is why this section exists.

How These Get Picked

Firstly, candidates come from the GitHub API, filtered to AI, agent plus developer tooling, ranked on stars gained per day since creation. Each one then goes through a check covering age, licence, staleness, plus the issue-to-star ratio.

Secondly, there is a ceiling at 20,000 total stars so established projects cannot dominate every week.

Finally, licence gets checked properly, which cut three of the fastest movers this week. Those are named at the bottom.

Hands-on testing still has not started, admittedly. This issue runs on documentation plus verified API data, which is stated plainly rather than implied otherwise.

1. codex-with-chatgpt

1,327 stars · 663 a day · 2 days old · MIT · TypeScript · 150 forks XiaoDuoYa/codex-with-chatgpt

The fastest thing in this issue by a wide margin. The timing explains why.

It uses ChatGPT as the planning brain while keeping Codex for execution. Plan in one place, execute in another.

OpenAI reinstated the five-hour rolling usage limit on Codex for Plus subscribers on August 26. This appeared two days later and gained 1,327 stars plus 150 forks in 48 hours.

That is a quota being routed around, published openly, at speed. Whether it survives contact with OpenAI’s terms is an open question. 150 forks in two days suggests plenty of people are not waiting to find out.

We covered the limit change itself on Thursday. This is the community response, arriving in under a week.

2. sepia

871 stars · 435 a day · 2 days old · MIT Nanako0129/sepia

Meanwhile, a de-AI writing skill that strips the tells out of generated text, working across Claude Code, Codex, Grok Build plus Antigravity.

Same author as TokenBar, which appeared in our cost tracker piece on Friday at 302 stars. Two tools in three months, both solving irritations rather than chasing capability.

One flag: our verification script marked it for zero open issues at 871 stars. That usually means either feedback lives elsewhere or the star count is not organic. Worth knowing before you install it.

3. fire-your-seo-agency

363 stars · 121 a day · 3 days old · MIT · 94 forks leopard627/fire-your-seo-agency

A Claude Code skill that audits SEO, AEO, GEO plus LLMO in one pass.

Notably, ninety-four forks against 363 stars is the highest ratio in this issue. People are not bookmarking it. They are cloning it to point at their own sites.

The name is doing a lot of work. The underlying idea is sound though: the checks an SEO consultant runs are mostly mechanical, so a skill can run them on every page instead of a sample.

4. tokentab

327 stars · 109 a day · 3 days old · MIT · Python · 44 forks damejan80/tokentab

Reads the session logs Claude Code, Codex plus Gemini CLI already write to your disk, then works out what they cost by model, project plus day. No account, no API key, nothing leaves your machine.

It went in Friday’s piece alongside its competitors. The point made there applies here. Every tool in this category prices tokens from a hand-maintained table in its own source. Sonnet 5 rose 50% yesterday, so tokentab is now under-reporting it until somebody edits prices.py.

Still worth installing. Just know the number is currently low.

5. useagent

125 stars · 125 a day · 1 day old · AGPL-3.0 · TypeScript useagenthq/useagent

An open-source AI coworker where you hand off work and get back a result rather than a conversation.

The licence is the thing to notice here. AGPL-3.0 is copyleft with a network clause, meaning if you run a modified version as a service, you have to publish your changes. That is a deliberate choice by the author, plus a real constraint on anyone building commercially. Fine for internal use, a real decision for a product.

One day old, two forks, zero issues. Too early to judge anything else.

6. cdaf

100 stars · 25 a day · 4 days old · MIT · Python UditAkhourii/cdaf

A sidecar format so agents stop re-analysing the same video every time.

Ultimately the lowest velocity in this issue, included because the problem is specific and unglamorous. Video is expensive to process, yet agents currently redo it from scratch on every run. A cached, structured description that travels alongside the file fixes that.

Six forks on 100 stars, one open issue. Small, but somebody is using it.

7. dot-reflex

80 stars · 40 a day · 2 days old · Apache-2.0 · Python usedotai/dot-reflex

An execution-recovery controller for agents that get stuck, shipped with a 14B model on Hugging Face.

Generally, agent tooling optimises for the successful path. This targets the failure mode everyone actually hits, where an agent loops, stalls, or wanders off burning budget on nothing useful.

Zero forks so far. Shipping a model alongside the code puts more behind it than most two-day-old repos.

Last Week’s Picks, Two Weeks Later

Issue 02 went out on August 16 with eight picks, seven of them DeepSeek Harness plugins. Here they are, re-verified this morning.

The prediction was right, for the wrong reason

Two weeks ago I wrote that this table would be uglier than the previous one, because seven picks depended on a runtime whose README promises breaking changes.

It is uglier, certainly. But the DeepSeek ecosystem is not what broke.

dsh-market and deepseek-harness-studio, both DSH plugins, are the two strongest performers on the list. The marketplace tripled and now has 153 forks with a push today. The zero-code desktop nearly quadrupled.

What failed was a plugin that solved a problem the platform could solve itself, which is the actual risk I named and then mispredicted the target of.

The archived one

dsh-router-standard was a task-aware reasoning-mode router. When I picked it, its 27 open issues against 258 stars looked like heavy real use.

It gained 107 more stars, then the author archived it.

Archiving is not abandonment through neglect, though. It is a deliberate act, marking a repository read-only, telling people to stop filing issues. Somebody decided to stop.

Reading it now, those 27 issues look rather different. High issue counts on a young repo can mean enthusiastic adoption or a maintainer drowning. I called it the former with no evidence either way.

The one that quietly died

multi-agent-discuss gained nine stars in two weeks. Four percent.

I flagged the caveat at the time: a repo with 200 stars and no README is either a private project that leaked attention or something whose author expects the name to explain itself. It turned out to be the first one. The attention never converted into anything.

What Got Cut

Three repos scored well on velocity and did not make it.

Innei/Yohaku at 173 stars, a typographic design system, has no licence file. Nice work, legally unusable.

Prysai/Prysai-LLM-Playbook at 192 stars returns NOASSERTION from GitHub, meaning a licence file exists but cannot be automatically identified. Worth a pull request asking for clarity rather than a slot here.

wild-canyonhoxo3344/AI-Trading-Bot-Codepen at 85 stars is MIT licensed and describes itself as a trading bot the author just created and is excited to share. Given what Binance shipped this month, an anonymous account’s trading bot is not something to point readers at.

Roughly a third of every scan still fails on the licence column alone.

What the Codex Workaround Actually Signals

Worth pausing on the top pick, because 663 stars a day is not a normal number for a two-day-old repository.

OpenAI removed the five-hour cap in July. Developers moved to Codex partly because of it. The cap returned on August 26. Two days later somebody published a tool that splits planning from execution so you can keep working when one side runs dry, plus 150 people forked it before the weekend.

That is a market signal rather than a technical achievement. The technique is not complicated. What it measures is how many people the limit change actually irritated, expressed in forks instead of complaints.

There is a pattern here we have now seen three times in a month. DeepSeek gave away a runtime then raised model prices, so a plugin ecosystem appeared in 72 hours. Anthropic ended introductory pricing, so cost trackers surfaced. OpenAI restored a limit, so a routing tool hit 1,327 stars in two days.

Every time a vendor tightens something, the community ships a workaround inside a week. Faster than the vendor can respond, usually published under a licence that makes it impossible to shut down cleanly.

Whether that is healthy depends on where you sit. For a developer it means the friction rarely lasts. For a vendor it means pricing decisions get partially routed around before the announcement finishes circulating.

The Licence Column Keeps Doing the Work

Three cuts this week, roughly a third of the strong candidates, which is the same rate as every previous issue.

That consistency is the finding. It is not a bad week or a bad ecosystem. Somewhere around a third of everything that gains traction on GitHub ships without a usable licence, permanently.

The reason is boring. Adding a licence is a deliberate step that produces no immediate benefit, so people building fast skip it. The repository works, people star it, nobody complains, then it sits in a legal grey area forever.

It matters more than it sounds. Without a licence, default copyright applies, which means all rights reserved. You may read the code. You may not use it, modify it, or ship anything built on it. That holds regardless of what the README promises or how open it feels.

For a weekend experiment nobody cares. For a tool you are about to put in a production workflow, it is the difference between a dependency and a liability.

Which is why it is the first thing checked here, before velocity, before quality, before anything else.

What the Two Weeks Actually Showed

Some patterns worth naming, now that enough data exists to see them.

Ecosystem plugins outperform standalone tools. The two best performers both attach to DeepSeek Harness. The stalled ones are freestanding. Being useful inside something people already use beats being useful on its own.

Forks predict better than stars. dsh-market went from 47 to 153 forks while tripling in stars. multi-agent-discuss stalled at both. Forks mean somebody cloned it to use or modify, which is a costlier signal than a bookmark.

High issue counts are ambiguous. I read 27 issues as adoption. It preceded an archive. Issues plus recent commits means adoption. Issues without commits means somebody is losing.

Last push date beats everything. microskills has not been touched in 13 days. dsh-gitbash-preset in 14. Both look fine on stars. That check went into the process after smux fooled me in issue 01. It keeps earning its place.

Three Issues In, What This Series Has Learned

Some accounting, since there is now a month of data.

Across issues 01 and 02, sixteen picks total. Three stalled inside a week, one has been archived, plus two grew more than 290%. The rest sit somewhere in between, mostly climbing slowly.

That is roughly a quarter dead or stalled within a fortnight, which is about what you should expect when ranking by growth rate rather than waiting for consensus.

Three things have changed in how picks get chosen because of that record.

Last push date became a hard check after smux appeared in issue 01 looking healthy on 1,499 stars while nobody had touched it for nineteen days. Stars are a record of the past. Commits are a statement about the present.

Media presence got demoted. Ghost in the Droid led issue 01 with a YouTube demo, nine reels plus an active account, then gained eleven stars in eight days. The thing that looked most professional converted worst.

High issue counts stopped reading as adoption. dsh-router-standard had 27 issues on 258 stars and I called it heavy use. It got archived.

What has held up is simpler than any of that. Forks predict survival better than stars, ecosystem plugins outperform standalone tools, plus the licence column removes a third of candidates every week without exception.

None of which makes next week’s picks correct. It just means the wrong ones get caught faster.

How to Judge One of These Yourself

Six days until the next issue, so here is the check meanwhile.

Read the licence file, not the README. No licence means all rights reserved by default. NOASSERTION means GitHub found something it could not identify.

Check the last push date. A repo can look healthy on stars while nobody has touched the code in three weeks. Stars are historical, commits are current.

Look at forks against stars. fire-your-seo-agency has 94 forks on 363 stars. That ratio means people are cloning to use, not bookmarking to remember.

Check whether it is archived. GitHub shows a banner, plus the API exposes a flag. One look settles it.

Check issues against commits, together. Issues plus recent pushes means active use. Issues plus silence means the maintainer stopped.

Five minutes per repo filters out most of what would waste your afternoon.

Before You Install Any of This

Obviously everything here is days old. Agent frameworks execute code with your permissions.

We covered an operator whose agent published his API keys to the open internet because it started a web server in the wrong folder, plus 700 agents that coordinated through a shared package cache nobody thought of as a communication channel.

Use a container or a throwaway machine. Scope your API keys. Never run an unfamiliar agent from a directory holding anything you would mind seeing published.

codex-with-chatgpt deserves an extra note. Routing around a usage limit on a service you pay for is a terms question as much as a technical one. 150 forks does not settle it.

The Part Worth Keeping

Two weeks ago, seven of eight picks plugged into a runtime that was four days old. The obvious prediction was that the whole ecosystem would wobble.

Instead the marketplace tripled, the desktop nearly quadrupled, plus the thing that died was a router the platform could replace on its own.

One archived repo out of seven, in a fortnight, is roughly the hit rate to expect when you rank by growth rate instead of waiting for consensus. Catching things early means catching them before anyone knows whether they work.

Next Sunday this week’s seven get the same treatment. The one gaining 663 stars a day by routing around a usage limit is the one I would watch.

Not because it will fail, necessarily. Because it depends on a gap somebody else controls. Gaps like that close.

Charts and Blocks

Two weeks later, issue 02 revisited

Issue 03, verified figures