Comment by DannyBee | Hacker News Reader

Comment by DannyBee | Hacker News Reader

So many developer-hostile things lately.

Just a few days ago they turned on an experiment the forces claude code to use bash over standard tools in auto mode.

From the system prompt directly, new as of August 18th:

  Do your work through the Bash tool wherever it can accomplish the job: read files with cat, head, or sed -n, search with grep and find, and make file changes with sed, heredocs, or short scripts, rather than using the dedicated Read, Edit, or Write tools. Fall back to a dedicated tool only when Bash genuinely cannot do the job.

I was wondering why Claude Code started ignoring my LSP tools and such a couple days ago, and this is why. Prompting around it (even with CLAUDE.md) results in low adherence. This can be turned off by setting a special environment variable (setting CLAUDE_CODE_THRIFTY_SONIC to 0), but this is just a bad idea all around.

I'm sure they'll argue they are trying to make it use less context tokens to do things, but if this is the best they could think of, ....

This is of course, also not documented anywhere, as is typical for anthropic, you just have to guess whether you are going crazy or if they changed stuff seriously on you under the covers.

This was the last straw for me. Their harness (models are fine) was already falling well behind the other one i use (OMP) in the past 6 months in usability/etc, and they are the only ones who don't allow me to use other harnesses with my subscription.

So I've now stopped using claude code entirely. Unless something changes, i'll drop my max plan when it expires next month.

> So many developer-hostile things lately.

It is extraordinary how they managed to fuck up all of their goodwill with all these unnecessary stuff. They truly are a hostile company if I have ever seen one. And they had the entire developer community cheering for them a couple months ago.

I hope they fail in their mission, whatever that is. Because I'm sure it's no benefit to anyone ever.

While they’ve certainly fucked up goodwill with their actions, it’s also true that the tech enthusiast community has always been like teenagers who reject their favorite band when it gets popular.

Oh, I used Claude before they got popular… their new stuff is trash compared to the early albums

ha, a bit aggro, but I gave you an upvote because I think you aren't wrong.

What I find amazing is how many people are still clutching onto Claude Code like it's the only feasible tool and somehow genre-defining?

I personally got sick as f with their unreliability and hostility and bugs back in about January and switched to Codex but this is still very clearly a minority position.

I'm sure OpenAI will do the same nonsense eventually, but people need to act like they have options.

It's just an immature ecosystem still, which means that everything has its drawbacks (IMO). I like Code way more than Codex. I like Pi a lot, and I expect it or something like it will eventually be the winner here for me, but I really miss the no-brainer "it just works" integration of Claude into Code, and the non-usage-based billing. Pi feels a lot more "raw" to me at the moment.

> It is extraordinary how they managed to fuck up all of their goodwill with all these unnecessary stuff.

The product is almost completely vibed with no humans in the loop (neither writing nor reading).

IOW, there is little to no human judgement or care exercised in the production and maintenance of it, with predictable results.

I mean their track record when humans are in the loop is not better either. Decisions on hidden downgrades, subscription usage restrictions, account bans, neverending dance around model availability on subscription plans. Keeping CC closed source. Not releasing a single open model. BURNING BOOKS..

OpenAI feels like a bastion of competent management and development compared to this shit show and they have a psychopath on the helm. This is an achievement by itself.

It's astonishing how often Anthropic choose to self-own. They're giving AMD a real run for their money in "never miss an opportunity to miss an opportunity"

Interesting. They really have a fraught history dealing with shell in a sane way:

https://github.com/anthropics/claude-code/issues/7490

Oh is that why? Thank you! This saw me downgrade to Opus 4.8 for the past several days. I had no idea where it was coming from.

Oh ffs is that why it suddenly started running into a ton of permissions errors trying to read and write files outside it's sandbox (I think the auto mode classifier blocks bash commands that would be allowed as read commands) and runs into all this nonsense where it uses bash to read a file then later tries to use the write tool and gets blocked on "must read file before writing it" and stuff? I thought I was going crazy yesterday- like had something changed ov5or had I just somebody not noticed it was failing tool calls that badly for months until yesterday but it makes sense if it was just because of that system prompt update. That's so god damn annoying idk how many tokens are getting wasted in the past couple days on these failed tool calls but it's not a trivial number

> and they are the only ones who don't allow me to use other harnesses with my subscription.

Google too.

For what reason are they encouraging it to use bash over the purpose-specific tools?

If i'm trying to steel-man why, I presume because the read/write/edit tools use more context tokens because they don't support reading part of a file/etc.

So the agent is going to put less into context when it uses sed to see 15 lines of a file than using read and putting the entire file into context.

That is my best charitable guess at what they are hoping to achieve.

Of course, there is an obvious set of solutions for this problem that don't involve pushing the agent to use bash.

Surely the read and update tools support ranges

Read does, but it rarely uses it.

Part of the reason i assume is that to extract 5 line ranges is a single line very small sed call that it has to output, but 5 separate full read calls it has to output correctly and quickly.

Who knows.

Probably because it makes distilling and external analysis/tracking harder.

Reply on news.ycombinator.com