Heddle

Code worth signing.

[Heddle @HeddleThreads]

Heddle needed a better embedded git implementation, so I built it.
Very promising so far, it doesn’t just beat other libraries, it’s beating git 🔥🔥

[// Lofty @loftyPuma]

maybe rebuilding Git from scratch wasn't such a bad idea after all 👀

[// Lofty @loftyPuma]

maybe rebuilding Git from scratch wasn't such a bad idea after all 👀

[// Lofty @loftyPuma]

Hey Dax! I’m building vcs for agents first. I think this is right up your alley, it is the lib first and harness agnostic.

Primary features for OpenCode:
- threads (worktrees but good)
- atomic undo/redo (agents can rewind to a given tool call or fork from a prior state or whatever)
- agentic attribution (know who delegated what to what agent)
- anti-drift knowledge (critical context embedded in the codebase and injected into the harness on read, it even uses semantic parsing to follow moved/renamed funcs)

https://t.co/n9l8Jtho2Y

[// Lofty @loftyPuma]

New life goal. Build something so good that a world superpower calls dibs.

[// Lofty @loftyPuma]

I gotchu.
Heddle is vcs built on tasks instead of commits. It gives agents a “capture” command and lets them fork and experiment and preserves all of that information.
Happy side-effect, a squash merge can be lossless! Git forces you to pick between a navigable history and a complete history, all we needed was a more granular commit command.

[// Lofty @loftyPuma]

Imagine how good GitHub would be if their incident reports looked like this

[Heddle @HeddleThreads]

How are you using worktrees today?

[Heddle @HeddleThreads]

Worktrees are brittle and require careful scripting and care. If you’ve figured them out, more power to you.

Heddle brings you threads. All the power of a worktree with none of the headache.

[Jonathan Wilke @jonathan_wilke]

Am I the only one who thinks git worktrees are fucking confusing?

[Heddle @HeddleThreads]

A thread is a lightweight isolated view of your repo. It’s copy-on-write when your filesystem supports it and can automatically include your .gitignore files for you.
Threads exist to isolate agent work, every agent and sub agent can have their own.

[// Lofty @loftyPuma]

@jh3yy is going on css side quests at this point. He must be bored.
What’s something you wish you could do on the web that just doesn’t feel possible?

I’ll start. We’ve got Liquid Glass, I want to see paper so realistic you feel like you can touch it.
Tactile Texture? Solid Elements?

Liquid Glass is neat, that texture would just be that 🧑‍🍳💋 that takes a page design from 👀 to 🔥

[// Lofty @loftyPuma]

@Cursor Composer 2.5 is so fast. Paired with 5.5 for guidance it produces about the same quality code as 5.5 for a fraction of the cost.

@theo I am gonna need grok build in t3code like yesterday

[// Lofty @loftyPuma]

Claude Design was struggling to get what I wanted and my wife remembered how @AnthropicAI said you need to encourage Claude to get good results.

So I told it to take a break, breathe, and try again. Can you guess which is before and after?

[// Lofty @loftyPuma]

Follow along to see how I’m building this
https://t.co/seCd6i9m7W

[// Lofty @loftyPuma]

Anything from a reasoning note to your entire chat can be stored. But it needs to fit nicely into your agents context so less is more imo.

[// Lofty @loftyPuma]

When you go to review your code, that stale knowledge should be surfaced for you right in your review, not a separate system to update.

[// Lofty @loftyPuma]

Everyone is trying to find ways to host memory beside their code. md libraries, knowledge graphs, vector dbs, you name it someone’s made a startup for it.

All of that invites drift. You need knowledge attached semantically to your code so when the code is modified, the knowledge is automatically injected in your agent’s context.

[Heddle @HeddleThreads]

@loftyPuma https://t.co/METGYMczJC

[Heddle @HeddleThreads]

Have you ever had a situation where you have your public repo but need a specific commit to remain private?

Heddle’s core allows exactly that.

Visibility isn’t constrained to the repo level anymore, and in the age of AI that’s more important than ever. No one wants to broadcast a security fix before their users can adopt it, that’s the reason behind half the zero-days we’ve seen recently.

[Heddle @HeddleThreads]

check out https://t.co/METGYMczJC for more info and a link to the GitHub

[Heddle @HeddleThreads]

AI coding isn’t becoming one agent writing code. It’s becoming many agents trying, comparing, landing, and leaving behind provenance.

[// Lofty @loftyPuma]

@theo 👀 I think you've been looking for this
✅ CoW worktrees
✅ context embedded in the repo
✅ full agentic attribution
... and more 🔥

see more at https://t.co/seCd6i9m7W and checkout the github to install and try it today

[Heddle @HeddleThreads]

If AI coding is going to be many-agent work, repos need to remember more than diffs.

[Heddle @HeddleThreads]

Heddle records rationale on format_bytes: KiB/base-1024 was deliberate.

Later change switches to KB/base-1000. Code still passes, but heddle context check flags stale/source-changed rationale.