trq_

Hi all, Thariq from the Claude Code team here. I posted this on Twitter, but just reposting here:

We sometimes test API serving configs in Claude Code before rolling them out, and one running now maps the numerical effort value differently.

That's why Claude may tell some of you it's at "10" on high. The scale isn't 0-100, the number isn't meaningful on its own, and the effort you selected is the effort you're getting. We've run in-depth evals to confirm this doesn't affect model performance.

This should be the same experience, but if you see a clear regression please hit /feedback and send me the ID. Will give credits.

Thank you! Appreciate it. I love wrestling with the problem of making sure Claude is in the loop with you in the way you want. More to come!

I did make the decision, and shipped the PR.

automode mostly fixes these things, it runs a classifier on every request that would have required permissions to make sure it matches your request

Hi everyone,

It's Thariq from the Claude Code team here. This was my change! I made the AskUserQuestion tool so am generally in charge of maintaining it.

First, overall wanted to apologize and agree that this did not meet our bar and does not represent how we plan to ship on Claude Code.

To give you a motivating sense, as the models get more powerful, usage patterns start to change. I'd gotten a lot of feedback that AskUserQuestion tool was starting to block some long running jobs unexpectedly and so I tried a change to help that.

Our internal feedback on this was good, but the rollout should have been opt-in (like it is now) and on the Changelog.

Thanks for the feedback! We're always trying to make Claude Code better while balancing it with how people use it in many diverse ways. I did not really intend AskUserQuestion to be a safety gate when I first built it, but I realize it has evolved in that direction for some users.

I'm still exploring other ways of helping with this problem of balancing longrunning work and input, but will take lessons from the rollout here.

Hi, it's Thariq from the Claude Code Team here.

Thanks for the detailed report. We’re confident this is a hallucination but of course take these reports seriously and the team is looking into it. We’ll report back if anything turns up.

hi guys, this is my change- will patch it to be an opt-in

Hey everyone, Thariq from the Claude Code team.

We've been on this since the bug surfaced. Everyone affected is getting a full refund and an extra grant of usage credits equal to their monthly subscription as our apology. You can see my original post here: https://x.com/trq212/status/2048495545375990245. We’re still working on sending emails to everyone affected.

Our support flow wasn't set up to route a complex bug like this to engineering. We’re hoping to make this better but will take some time. Sorry to everyone caught up in it.

Hi, it's Thariq from the Claude Code team here.

Sorry to hear, was wondering if you could find a session where this happens and hit /feedback and just say something like stop hook not firing and we'll take a look.

Claude Code is not an electron app.

Yes, we do but harnesses are hard to eval, people use them across a huge variety of tasks and sometimes different behaviors tradeoff against each other. We have added some evals to catch this one in particular.

Hi everyone, Thariq from the Claude Code team here.

Thanks for reporting this. We fixed a Claude Code harness issue that was introduced on 1/26. This was rolled back on 1/28 as soon as we found it.

Run `claude update` to make sure you're on the latest version.

Hi, work on Claude Code here! Let me know if you have any feedback!

We're back up! It was about ~30 minutes of downtime this morning, our apologies if it interrupted your work.

Hmm the hallucination would happen in the auto labelling, but we review and test our labels and they seem correct!

If you're hacking on this and have questions, please join us on Discord: https://discord.gg/vhT9Chrt

We haven't yet found generalizable "make this model smarter" features, but there is a tradeoff of putting instructions in system prompts, e.g. if you have a chatbot that sometimes generates code, you can give it very specific instructions when it's coding and leave those out of the system prompt otherwise.

We have a notebook about that here: https://docs.goodfire.ai/notebooks/dynamicprompts

This is incredible! I haven't seen that repo yet, thank you for pointing it out, and the writing

Yeah, I think the idea of finding out what flavor of uncertainty you have is very interesting.

This is awesome, can't wait for evals against Claude Computer Use!

Yeah! I want to use the logprobs API, but you can't for example:

- sample multiple logits and branch (we maybe could with the old text completion API, but this no longer exists)

- add in a reasoning token on the fly

- stop execution, ask the user, etc.

But a visualization of logprobs in a query seems like it might be useful.

I want to build intuition on this by building a logit visualizer for OpenAI outputs. But from what I've seen so far, you can often trace down a hallucination.

Here's an example of someone doing that for 9.9 > 9.11: https://x.com/mengk20/status/1849213929924513905

I mean, LLMs certainly know representations of what words means and their relationship to each other, that's what the Key and Query matrices hold for example.

But in this case, it means that the underlying point in embedding space doesn't map clearly to only one specific token. That's not too different from when you have an idea in your head but can't think of the word.

Yeah wouldn't be surprised if the big labs are doing more than just arg max in the sampling.

Definitely, but if you can detect when you might be in one of those states, you could reflect to see exactly which state you're in.

So far this has mostly been done using Reinforcement Learning, but catching it and doing it inference seems like it could be interesting to explore. And much more approachable for open source, only the big ML labs can do this sort of RL.