I used Fable to rewrite 65kLoC of Go in Rust. It cost $400
iurii.net
iurii.net
I'm sure there are other caveats. Also cost of the more or less straightforward Bun port was $165000 for 500KLoC.
If not an intermediate representation, my first question would be: how to split the work. You cannot just prompt full rewrite of 60K LoC, you cannot do it module by module — modules do translate 1:1, like in my case golang packages did not matched Rust crates. Bun did file by file.
With hierarchical state machines in the middle, I did it almost in one shot.
Agents are actively destroy QA gates in many ways, usually by cheating ("the test is buggy, not my changes" — changes the test), or just rot QA slowly by writing buggy overcomplicated tests
What works for me 10/10 is fuzzing and my own constant usage. For this project specifically (text editor), I asked LLM to create human-like fuzzing session, it sends keystrokes like: "the user is searching for a file, editing, <ordering a lizard>, saves changes".
On top of it, I run https://mutants.rs/ which is kind of tests fuzzing. It flips random switches in the app itself, and if tests are silent - they missed a bug.
The downside of this, is I usually find bugs after 1-2 hours of running. I use local Qwen to babysit these sessions, to make initial investigation, a repro case, and file a ticket.
It is probably because I read tons and tons of LLM output.
I absorbed so many different models at this point :)
Thank you for the kind words.
*be me over eager
(I ported first 80% practically in one shot, planning and then leaving Fable overnight to orchestrate). Then I added a bunch of features, so at the end I ported more like 150% of the original code, I added tree-sitter, and a bunch of syntaxes highlighters. At the end with all that, price went up to ~$650
Yes, I completely forgot to mention, this is exactly my case. Rune is a TUI editor, so I feeded the same terminal sequences to the old and new apps.
It didn't translate 1:1 (I ported core editor first, there were side panels, and different chrome elements) so I instructed LLM to use ttyd (tty -> browser render), Fable then could open both apps with playwright, make and compare screenshots.
To rephrase, one critical component is to establish a feedback loop for the model. This new generation of models: Opus 5, Fable, GLM-5.2, even Qwen3.8-27B can self-correct, provided they know whether they are progressing or not.
A month ago, especially smaller model would fall into a rabbit hole it dug for itself and would never recover. This generation can sometimes run tens of hours without losing track.
I still wouldn't trust a model after 70% context window, but the progress is noticeable.
A TUI editor’s real output is the bytes stored on disk, while rendering can look identical the saved files might diverge (encoding, line endings, trailing new lines etc). Did you manage to diff that?
Totally agree on the overnight roadmap runs I have the same experience here. The agents have to know how to self-correct and if it’s progressing, otherwise it’s failing!
The way rune works with files minimizes chances of silent corruption. I keep original byte blobs immutable, separately there is a journal (kinda WAL) of positional deltas (inserts and deletes).
So I only need to validate that blob + deltas = snapshot.
Disk IO is encapsulated through VFS, and writes are atomics (write to a temp file, then rename).
Separate virtual rendering buffer is built on top of that. Rune, like Obsidian, renders markdown preview inline, so the same chunk could be rendered as "Header" as well as `## Header` when under the cursor.
All of that makes it quite easy to work with text. The core function is to translate offset in a byte array to line and column and back, which is pure math and relatively easy to test.
Another trick that helped a lot is to use sqlite extensively: blobs, deltas, vfs, redo and undo history graph are all sqlite tables.
I noticed that LLMs make stupid decisions when it comes to data structures, but they understand CRUD and SQL, so I turned all Rune's internals into dumb CRUD.
If the experiment was "spend 400 bucks to see if it'll work" then that's awesome, and fun, and a cool use of AI. It's impressive that AI can do that.
If it was to make something useful ... has it?
$400 are subsidized into the subscription, and this was mainly an experiment to prove the theory about data conversion step. I call it a success and I use rune editor daily.
To me, running multiple agents is not very different from managing multiple teams — I won't be able to keep up with the changes by reading the code.
I may make certain architectural decision, and I need to act based on some signals.
The simplest example is clusters of bugs are signaling that certain modules are dirty. Sometimes I read a plan and understand that the agent is trying to workaround some auwful engineering.
My friend once sent me a snippet, maybe 10 lines of C+++, asking "can you spot the UB?".
So I'm staring at these 10 lines, I KNOW there is an UB. I wasn't able to find it without a hint.
My all-time favourite example is (again, my memory, I may be a bit wrong):
for (int i = 0; i < (size_t)limit; ++i) {
}
while(true)
But somehow compiler has decided that i <= limit is always true.
In my experience it mostly comes down to the harness (or lack of) that you use. Something like 'superpowers' can be pretty verbose and hash out things for a long time (and use a decent amount of tokens) but the output is pretty decent. The better instructions and the more brainstorming you do initially the better - as the subagents encounter fewer issues.
Without a harness you could try a direct port (prompt: 'convert x to y, don't bother me') and if the languages are roughly compatible you could get a seemingly working port much quicker but likely with major hidden issues. A comprehensive testsuite is obviously a must.
What is the scale? Because I'm pretty sure it's impossible to one shot 65k LoC with "good luck, make no mistakes" prompt.
I also did this within a subscription. I counted the number of tokens afterwards and calculated the cost as if I'm paying per token. $400 is of course arbitrary, but it's a ballpark number, bun was $165000.
> In my experience it mostly comes down to the harness (or lack of) that you use.
Yep, it is.
Tokens per task is a good proxy measure of skills, 'superpowers' or any other.
Either a skill gets you the thing more efficiently (less tokens), or you don't need to redo the result afterwards (less tokens). I benchmark all my skills that way.
Models need less and less steering at this point, especially frontier ones.
* How many lines was the result? Considering how much err != nil and line splitting needs to be done in Go, did you at least reach 30k SLoC?
* How was sloc counted here? includes comments, blank lines or not? (something like cloc will give a good answer).
* Performance characteristics of resulting rust, was there an improvement? It maybe appealing to say Rust is Always faster than Go.
65k LoC of Go without comments resulted in roughly 60k LoC of Rust witout comments (code column of the cloc tool).
The error handling is not so different between Rust and Go, in both cases I cannot panic to avoid the data loss. So it boils down to if (failure) return something for graceful degradation. And generally errors in my case (a text editor) are rare, only disk IO, which is encapsulated in one VFS module, everything else, like non-closed brackets in code is expected behavior.
The biggest differences were in third party libraries, UI, markdown parsing — completely different API and paradigms.
> Performance characteristics of resulting rust
I haven't measured. I don't think there's any significant difference between Go and Rust if app doesn't do allocations on a critical path. The reason I started this project was mainly to experiment (now I use similar approach to refactor much bigger legacy code base), and tree-sitter support is better Rust so it seemed like a good fit.
It would help to clarify these things in the post. Your software is atypical of common Go software in the wild. The purpose of Go -> Rust rewrite would be usually the efficiency of rust.
Although, in the agentic environment, a benefit of not having GC at all definitely helps.
Another point is, Fable is reasonably cheap if you don't allow it to read or write.
There is a way to control tool invocations at the harness level when writing skills or agents.
Example: ~/.claude/agents/critic.md
---
name: critic
description: >
Plan Critic. Reviews an implementation plan. Use before the implementation.
tools: Read, Agent
---
allowed-tools: Read Grep
disallowed-tools: WebSearch Glob
In my experience, you can get good results if Fable does't write code itself, only spawn subagents.
I can run Fable for 10 hours, and it would output 50k tokens and read 300k (30% of the context window). The resulting code is okay-ish. I would rarely merge LLM-produced code first try without an adversary review.
LLM-powered rewrites and huge refactors are better done using 1 additional step "convert the code to <something> that represents it best".
The simplest example is, for a CRUD app it can be swagger description. The more complex behaviour exhibit the app, the more raw information should be provided.
Like ontologies, "A is a child of B" model can derive and enforce that "B is a parent of A", and so on.
On top of that, I write that Fable is reasonably cheap if one uses it solely for agent orchestration.
1. Extract the data representation
Ask the LLM to represent your code as any combination of:
graphs
ontologies
hierarchical state machines
UML process charts
constraints
math formulae3 Convert representations back to code
My gut feeling is this is doing way too much, and it would've figured it out.
The whole point of this experiment was to try and make the rewrite as cheap as possible.
To be fair, I haven't try exactly "rewrite in Rust" prompt, but I cannot imagine for this to work.
Giving the benefit of doubt, we all might be writing a bit like claude nowadays.
If that is the case, I'd recommend reviewing the content before publishing to see if it sounds like a LLM.
Or if you are trying to create "better" AI slop and think that is enough to say the text is human-written, don't do that, just say it was AI-generated or assisted.
Probably I need some fresh air and a good fiction book.
I haven't played with GLM 5.3, only with 5.2, so I cannot say for sure.
GLM is at the level of Opus. Fable is something different entirely. It is capable of tracing the data flows of the app, I even tried it in a huge PHP codebase, it works.
PHP is a weird beast because it allows something like
$v = 'SomeClass' + 'Controller'
... // and later
new($v)
And another distinct feature of Fable — it is an amazing orchestrator. I prohibit it basically read and write, and it operates a swarm of haiku and sonnet.
Languages have conventions and best practices that aren’t portable.
We could port COBOL to c and do a crap ton of goto/jmps and make a
BEGIN MESSMy point is that one can translate the data flow to another language.
You could imagine any program as input -> [blackbox] -> output. For example, same pixels rendered on the screen provided identical keyboard input.
I propose a way to decompose the blackbox.
I did because I have a huge legacy code base, a distributed monolith, a few millions lines of code. Ideally, I want to get rid of it.
At this point, I know a recipe to break the monolith, so I finally could eat the elephant piece by piece.