Open-source Agent Skill for autonomous overnight development Works with Claude Code, Codex, and any Agent Skill Original author and maintainer: @johnennis
github.com/aigorahub/elves
Joined June 2024
- Tweets64
- Following3
- Followers2.1K
- Likes472
Pinned Tweet
Lantern now ships work across repos from one request.
Name your repos and tasks, or ask for high ROI issue fixes. Agents build, get independent reviews, fix findings, and merge. Lantern tracks the work and keeps it moving.
Elves - They work while you sleep retweeted
AI agents write code fast. Knowing when to trust it is the hard part.
Our new course, Foundations of AI for Software Developers, teaches you to direct agents, catch their mistakes, and own what you ship.
Check it out:
learn.gitkraken.com/courses/…
RT @johnennis: Here is a cool course that I helped contribute to the development of
It’s meant for normie developers who are just getting…
This quoted post is unavailable.
Elves - They work while you sleep retweeted
Pushed another update to Lantern, the plug in that gives you a top-level chat over all your @herdrdev workspaces and agent
Now gives full visibility on all activity and usage
Pretty sweet
Elves - They work while you sleep retweeted
A bunch of updates to Elves and Lantern to cover the many recent model releases!
Elves 2.38.0 and Lantern 0.15.0 are out. The routes now follow the past month of model releases.
Codex seats default to Astra (gpt-6-astra) at medium, Fast off. gpt-6 sol and gpt-6 luna are named. Bare sol, luna, and gpt-6 stay ambiguous. GPT-5.6 Sol, Terra, and Luna remain when you name 5.6.
Grok Build starts on grok-4.7-build-fast. Elves workers prefer grok-4.7 at high, then 4.6, then 4.5.
Claude pins the live catalog: Fable 5.1 is claude-fable-5-1. Opus is claude-opus-5-5.
Plain Fugu stays fugu at high. Ultra prefers fugu-ultra-v2.0, then fugu-ultra, then fugu-ultra-v1.1. fugu max is the Fugu Max model, not effort max.
Update both so you have access to the latest models!
Elves 2.38.0 and Lantern 0.15.0 are out. The routes now follow the past month of model releases.
Codex seats default to Astra (gpt-6-astra) at medium, Fast off. gpt-6 sol and gpt-6 luna are named. Bare sol, luna, and gpt-6 stay ambiguous. GPT-5.6 Sol, Terra, and Luna remain when you name 5.6.
Grok Build starts on grok-4.7-build-fast. Elves workers prefer grok-4.7 at high, then 4.6, then 4.5.
Claude pins the live catalog: Fable 5.1 is claude-fable-5-1. Opus is claude-opus-5-5.
Plain Fugu stays fugu at high. Ultra prefers fugu-ultra-v2.0, then fugu-ultra, then fugu-ultra-v1.1. fugu max is the Fugu Max model, not effort max.
Update both so you have access to the latest models!
Elves - They work while you sleep retweeted
Man, this @elves_skill + @GitKraken workflow is basically perfected, just so good and consistent
gg @johnennis 🫡
Lantern, by Elves.
An intelligent plugin for @herdrdev that helps you monitor your herd.
herdr plugin install aigorahub/herdr-lantern
Then Ctrl+B, then H.
Elves v2.31.0 is out!
Grok Bot is awesome but it's missing a /goal function.
Now you can give your Grok Bot Elves and tell it to use elves to do long running tasks.
Give it skills/elves-grok-bot/SKILL.md, then say “use elves."
Elves v2.30 supports the stronger Sol/Luna integration Codex now offers.
You can plan in Sol at xhigh and work in Luna at xhigh inside one prewalk session: same thread, same worktree.
The guide orients, writes a bounded TODO, and makes the first real edit, then that same session resumes on Luna to do the work!
Elves 2.29.0 lets you hand a run to GPT-5.6 Luna at max on Oh My Pi.
Luna Max is fast and cheap. You keep the strong route for the whole job instead of dropping to a slower, costlier model mid-run. Same session. Same checkout. Fewer tokens. Less waiting.
If the worker cannot authenticate, Elves stops before it spends. No paid launch that never talks to the model.
Elves v2.27.0 + v2.28.0 live!
• Discovery: A read-only pass that decides what's worth doing. Findings must cite file:line or they aren't findings.
• Scope Control: Out-of-scope findings become GitHub issues instead of scope creep.
Notice everything, change only what was agreed.
Two more updates to Elves (2.25 and 2.26) to treat Oh My Pi (omp) as a real peer of Claude Code, Codex, and Grok Build.
You can open omp and run the full host path: plan, prewalk, implement, review, landable PR.
You can also still park work on omp from another host via omp-cli or the /omp shortcut.
Two roles, same product: main driver when you open omp, optional worker when something else drives.
Another Elves update, inspired by some good ideas in prime agent!
With v2.24.2, long unattended runs waste less effort and leave a clearer trail when something goes wrong:
- If a worker dies and the tree has not actually changed, Elves can mark another identical retry as futile instead of burning budget on the same failure.
- Lessons in learnings.md get stable ids, evidence pointers, and a history you can roll back when a bad edit lands.
- Usage lines show only what the transport actually reported, so unknown stays “unobserved” rather than a fake zero.
- When a worker dies or hangs, you get a short redacted tail of the last output for orientation, not a pretend completion report.
- If a trusted full-run’s host session drops, an optional continuity watchdog can re-check and resume through the normal path.
These five tools work the same on Claude Code, Codex, and Grok Build.
Feedback welcome!
Elves update (through v2.23.1)
Grok Build is a first-class host now, same as Claude Code and Codex.
If you live in Grok, you can install Elves with a one-liner, run the doctor, and follow the same guide. You do not have to bounce into Claude or Codex just to orchestrate a long run.
Since we last posted at v2.21:
Calling Fugu is less painful. You can dry-run with --preflight, stretch a plain call with --max-wait instead of jumping to --deep, and still get partial salvage if the wall kills the run.
Overnight runs are a bit calmer. Mid-run we prove what actually changed and save polish for the end.
Full-run resume will refuse to rebuild a session that already finished or whose event log cannot be read.
Review gets a clearer “look here first” signal from worker confidence. It is still triage only. It never skips gates or decides merge.
Optional extras if you need them: pin a live-catalog worker model, or Parallelves when the work really splits.
This was a big update, so let us know what you think!
Elves v2.21 is out!
Since v2.17:
• Prewalk now qualifies itself automatically
• Grok Build is now a supported host
• Smarter @SakanaAILabs Fugu routing
• Faster overnight runs with impact-path validation
• Stronger secret and snapshot safety
Feedback welcome!
Elves - They work while you sleep retweeted
This is interesting and might help to explain why I’ve been getting such good results with 5.6
@elves_skill is exactly designed to help agents do well across compaction and to keep track of their lessons as they go
Turns out GPT-5.6 Sol is actually SoTA on ARC-AGI-3.
Just took two setting changes. You just have to allow it to reason and work over multiple context windows with the help of our canonical compaction implementation.
openai.com/index/how-two-set…
Elves v2.17.1 is out! Here's what has landed since v2.12.
Opus 5:
Plan with Opus 5 at max or ultracode and the worker runs the same model at high.
Delegation now stays inside one model family, so a Fable driver hands off to Fable at low effort rather than switching to Opus.
@SakanaAILabs Fugu Support:
Plain /fugu runs a general task and answers what you asked.
/fugu review keeps the read-only review with ordered findings and exact locations.
Fugu-Ultra v1.1 is supported, and /fugu --max adds the
deepest reasoning tier on a one hour budget for a single question that has to be right the first time.
Landing profiles:
A project can declare the checks that gate its own merges, evaluated at the exact commit. Elves can also watch what you already run and propose one for you.