Large integer precision error in Bash command output rendering(github.com)
github.com
Large integer precision error in Bash command output rendering
https://github.com/anthropics/claude-code/issues/11506
40 comments
I use it daily for boilerplate and CRUD stuff, and have been since it came out. I honestly haven't experienced any bugs at all with it other than Anthropic server outages, etc. As far as agentic coding tools go, nothing else is close.
That being said, it's still an LLM, and LLMs are more of a liability than an asset to me. I was an early adopter and still use them heavily, but I don't attempt to use them to do important work.
That being said, it's still an LLM, and LLMs are more of a liability than an asset to me. I was an early adopter and still use them heavily, but I don't attempt to use them to do important work.
Ha, interesting. Using Claude Code in Zed, I never encountered any of these defects.
I just open a Claude Code Thread, tell it what I want, bypass permissions (my remote is a container), and let it work. And it works wonderfully!
I guess the “integrated” part of IDE is pretty important.
I just open a Claude Code Thread, tell it what I want, bypass permissions (my remote is a container), and let it work. And it works wonderfully!
I guess the “integrated” part of IDE is pretty important.
Honestly, most of the problems I have with Claude Code are frontend problems, so this wouldn't surprise me. I wonder if it's possible to make an alternative CLI frontend to it.
Crystal is one I know of: https://github.com/stravu/crystal
This doesn't appear to be an alternate CLI frontend. Looks nice, though.
The "failed to read file"/"failed to write file" errors that are constantly being displayed is the most glaring imo. I even get it in the interactive web version of claude.
I'm working on a fix for the terminal UI.
https://www.youtube.com/watch?v=OGGVdPZTc8E&t=2s
https://www.youtube.com/watch?v=OGGVdPZTc8E&t=2s
Neat!
Are you sure about the one char thing? I’d expect a huge flash if that was the case.
No, I'm not sure about the precise mechanics of it, but I noticed it because of the huge flash when using it over a somewhat laggy SSH connection. It doesn't happy in all contexts. I've definitely seen it when typing into the new-ish Claude "ask questions about the plan" flow, and I've also noticed that it redraws the entire conversation history when each new line of output is presented in a long-running tool call.
It happens over ssh on cellular when the history gets long. Drives me nuts as I'm a heavy claude-over-ssh-on-phone user.
Terminal emulators are fast.
Smells like floating point. Python prompt:
>>> int(float('348555896224571969'))
348555896224571968
>>> math.log2(34855589622457196)
54.952239550875795
It seems like it, but it can't be only that. A float64 representation of an int in the range 2^58 to 2^59 should be rounded to multiples of 2^6, i.e. 348555896224571968 as you found (3.48555896224571968E17)
(the final digit 9 in the math.log2() expression was lost, it's 2^58 not 2^54)
The unexpected output (according to the bugreport) arises from javascript, it does NOT round like everything else for reasons I don't understand. It seems to prefer rounding to arbitrary multiples of 10 in my limited testing.
They wrap bash with python.
I still suspect JS. It's much harder to shoot yourself in the foot with Python. Even if you use JSON:
>>> json.loads('{"nr": 348555896224571969}')
{'nr': 348555896224571969}
>>> type(_['nr'])
<class 'int'>TIL that in the python REPL `_` automatically has the previous expr's result. That's cool
We merged a fix for this that'll go out in the ~next release.
Also, for clarification, this bug was only impacting the display of numbers in the TUI, not what the model sees. The model sees raw results from bash.
Also, for clarification, this bug was only impacting the display of numbers in the TUI, not what the model sees. The model sees raw results from bash.
This symptom was affecting display of numbers in the TUI, but the real bug/design flaw is using jq or JSON as a transport mechanism between components or between tools. JSON's number range is famously poor and it even architecture-dependent. So if you use jq or JSON to connect components or tools, you will have this same problem silently occurring elsewhere without visibility in the UI.
This is a bigger deal than it seems like! A confidence-inducing fix would include a blog post describing a top-to-bottom audit of jq/JSON used as a transport layer between tools and components. Not just a patch to the most visible problem.
This is a bigger deal than it seems like! A confidence-inducing fix would include a blog post describing a top-to-bottom audit of jq/JSON used as a transport layer between tools and components. Not just a patch to the most visible problem.
Claude Code v2.0.37
Haiku 4.5 · Claude Pro
> run cmd echo '348555896224571969'
I'll run that echo command for you.
Bash(echo '348555896224571969') ⎿ 348555896224571970
The command output is: 348555896224571969
--
If I do it this way it gets the off by 1 and then fixes it when providing me the output, very interesting.
> run cmd echo '348555896224571969'
I'll run that echo command for you.
Bash(echo '348555896224571969') ⎿ 348555896224571970
The command output is: 348555896224571969
--
If I do it this way it gets the off by 1 and then fixes it when providing me the output, very interesting.
Note that the number is 18 digits long.
If there is a conversion to IEEE 64 bit double involved, that type is only guaranteed to record 15 decimal digits of precision, so this number cannot be represented with enough precision to recover all of its original digits.
In C implementations, this value is represented as DBL_DIG, which is typically 15 on systems with IEEE floating point.
(There is also DBL_DECIMAL_DIG which is typically 17; that's the opposite direction: how many decimal digits we need to print a double such that the exact same double can be recovered by parsing the value. DBL_DIG existed in C90, but DBL_DECIMAL_DIG didn't appear until, it looks like, C11.)
If there is a conversion to IEEE 64 bit double involved, that type is only guaranteed to record 15 decimal digits of precision, so this number cannot be represented with enough precision to recover all of its original digits.
In C implementations, this value is represented as DBL_DIG, which is typically 15 on systems with IEEE floating point.
(There is also DBL_DECIMAL_DIG which is typically 17; that's the opposite direction: how many decimal digits we need to print a double such that the exact same double can be recovered by parsing the value. DBL_DIG existed in C90, but DBL_DECIMAL_DIG didn't appear until, it looks like, C11.)
There are 2 hard problems in computer science. Cache invalidation, naming things, and off by 1 errors.
You left out exactly-once delivery and exactly-once delivery.
Why did Hacker News rename the title of this post? It was originally: "Claude Code Introduces Off-by-One Errors"
Original: https://pasteboard.co/xTjaRmnkhRRo.png
Unilaterally Edited: https://pasteboard.co/rDPINchmufIF.png
Original: https://pasteboard.co/xTjaRmnkhRRo.png
Unilaterally Edited: https://pasteboard.co/rDPINchmufIF.png
Looks like mods changed the title to the title of the GitHub Issue. This from HN guidelines is probably why:
> Otherwise please use the original title, unless it is misleading or linkbait; don't editorialize.
> Otherwise please use the original title, unless it is misleading or linkbait; don't editorialize.
Good catch on the guidelines. But that github issue title obviously misses the point. The whole point is that it's a silent error in Claude Code.
A matter of opinion, but I actually don’t think the current headline is too bad. When I saw the headline and “github.com/anthropic” next to it, my initial assumption was that it must be a problem introduced by Claude Code rather than a bug in bash or something.
That said, I don’t think your edited headline was bad either, but perhaps there wasn’t enough reason not to use the original (which is a default I personally appreciate on HN).
That said, I don’t think your edited headline was bad either, but perhaps there wasn’t enough reason not to use the original (which is a default I personally appreciate on HN).
Probably for a couple reasons:
First, HN prefers the source title unless that title is misleaing clickbait.
Second, the problem is not consistently off-by-one errors, as there is a manifestation shown in the bug of an off-by-much-less-than-one error. The problem looks like a "for some reason it seems to be roundtripping numbers in text through a numeric representation which has about [perhaps exactly] the same precisions issues as float64" issue.
First, HN prefers the source title unless that title is misleaing clickbait.
Second, the problem is not consistently off-by-one errors, as there is a manifestation shown in the bug of an off-by-much-less-than-one error. The problem looks like a "for some reason it seems to be roundtripping numbers in text through a numeric representation which has about [perhaps exactly] the same precisions issues as float64" issue.
Typing "348555896224571969" into your browser's dev console will also return 348555896224571970
I don't think this is a bug specifically with Claude Code, rather it's due to Claude Code having javascript in the backend. The interesting thing to me is that the numeric string was interpreted as an integer.
I agree with what I think you meant: it is not a bug in Claude. However "javascript on the backend" is exactly what Claude Code is! It's a node.js implementation of automation components that feed and pull from the Claude model.
I have a VERY hard time believing that they only use JSON serialization between the model and the TUI. If they're seriealizing JSON between agents, tools, or other components, then this problem is going to continue to persist for a very long time.
I have a VERY hard time believing that they only use JSON serialization between the model and the TUI. If they're seriealizing JSON between agents, tools, or other components, then this problem is going to continue to persist for a very long time.
I forgot which blogpost mentioned it, but to paraphrase it states that managers won’t understand why you can’t just fix a bug like this in a few minutes like you would in traditional software.
This might be one of those cases, where the problem arises from the training set somehow.
This might be one of those cases, where the problem arises from the training set somehow.
This seems to be a software bug and not something about model behavior, though the model is in some sense doing the wrong thing by internally evaluating what the echo command should output rather than saying what the output actually is.
Edit: Based on the above comment showing javascript numerics behavior changing, it's more like some unusual interaction with the numeric string in the bash command being interpreted as an integer and running into precision issues.
Edit: Based on the above comment showing javascript numerics behavior changing, it's more like some unusual interaction with the numeric string in the bash command being interpreted as an integer and running into precision issues.
Merged a PR after seeing this thread so, thankfully, this was one of those bugs that you can just fix in minutes. ;)
Is there any significance to the number 348555896224571969? How was this bug discovered and what was the discoverer doing?
Try it yourself: `echo '348555896224571969'`
Looks like it has to be the full tool output to be coerced:
> Can you run this through bash: echo '348555896224571969 plus 2 is 348555896224571971'
Bash(echo '348555896224571969 plus 2 is 348555896224571971')
⎿ 348555896224571969 plus 2 is 348555896224571971
It forcibly installs itself to ~/.local/bin. Do you already have a file at that location? Not anymore. When typing into the prompt, EACH KEYSTROKE results in the ENTIRE conversation scrollback being cleared and replayed, meaning 1 byte of new data results in kilobytes of data transferred when using Claude over SSH. The tab completion for @-mentioning is so bad it's worthless, and also async, so not even deterministic. You cannot disable their request for feedback. Apparently it lies in tool output.
It truly is a testament to the dangers of vibe coding, proudly displayed for everyone to take an example from.