Anthony Odie

@RichardSocher Congrats, noticed how many times you said “You’re absolutely right!” 😂 on @business

@EnoReyes @matanSF Thanks for fixing #1! Any update on #2?

@AnthonyOdie @matanSF Looking into this!

@EnoReyes Can you look into the hook execution + diff view in the CLI? I have 15+ hooks set up and execution clutters my screen. Any chance to make it more subtle like CC? The diff’s seem to never appear anymore by default (requires ctrl+o).

The Lumetric team are incredibly hard working AI-native builders that understood the direction all computer-based work was going. Looking forward to building an amazing native desktop experience for Droid with them!

@cognition 👀

Reply to this thread and fill out this form. If you're one of the first 100 to do so, you'll get a month of Max to try out SWE-1.6 Fast in Devin for Terminal and Windsurf: tally.so/r/b5KV40

Intelligence at 1000 tokens per second, right in your terminal. Now available with SWE-1.6 Fast, powered by @cerebras. We're giving the first 100 people who respond a free month of Max to try it out.

@mntruell @nikitabier can you allow X subscriptions to be used in Cursor so we can use Grok in @cursor_ai ? Grok in router logic could also be bedrock for opt-in data pipeline for RL and bolster the partnership.

Excited to partner with the SpaceX team to scale up Composer. A meaningful step on our path to build the best place to code with AI.

@matanSF @EnoReyes @EnoReyes I love the harness. Truly amazing for using open source LLM’s - can you use @OpenAI Codex models similar to how they allow OpenCode usage? Would be so huge. @thsottiaux

@matanSF Hey Matan and @EnoReyes , can you look into (1) when running droids in CLI on Windows / WSL, the interface becomes completely distorted for the last few versions (2) hook execution more subtle/not spam you in the CLI?

@bcherny Can you change the effort tuning on ultrathink so it is model dependent? Ultrathink is great to set Opus 4.6 to high but on Opus 4.7 why doesn't it set it to xhigh?

Separately, we’ve also heard reports of issues with Opus 4.7 in Claude Code. The team is working on those and we’ll share more as we roll out improvements over the coming days.

We’ve been looking into recent reports around Claude Code quality issues, and just published a post-mortem on what we found.

Wouldn’t it also be better from a cost + usage perspective for caching on the backend to have faster search tooling vs relying on shell commands for grep/find which may be slower/more token inefficient and lead to more compaction which invalidates caching? Let me know if I’m not thinking about this correctly. @trq212 cursor.com/blog/fast-rege…

If you’re not in auto mode, doesn’t this lead to decision fatigue? Why not just tweak the system prompting and make the native Grep/Glob more robust / unify it into a SearchTool? Not arguing, but I’ve just found the BashTool slower / less token-efficient for search. Would love to see the evals/data on testing Bash vs. native tooling (@cursor_ai improved tool capability & scaffolding vs. relying on shell for search). Evals show better performance/less context use. Are the dedicated Grep/GlobTools going to be deprecated? What was the initial logic with the dedicated tooling set if BashTool was “all you need”? How do you also even ensure rg is used over grep via BashTool? Does it respect permissions / gitignore? I just question that sometimes in the goal of simplicity, we maybe reduce quality but again would love to see the data on this. @bcherny @trq212

After 2.1.117, you may notice that Claude doesn't call its Grep or Glob Tool anymore. YES!!! It only took four months. It's faster than ever and it's all Bash. It's so much harder to take things away than to add them. Enjoy.

@eduferreyraok @dmwlff @noahzweben Most dedicated Grep tools are built on ripgrep including Claude Code’s

@AnthonyOdie @dmwlff @noahzweben Grep is not the fastest (maybe ripgrep), and I’ve seen posts about claude being buggy in regards globally and .env files

Can you look into auto mode + Hooks/Tools. CC has a great hook system, but if agents decides to use shell instead of dedicated tooling (Grep/Glob/Read/Edit/Write) where dedicated tools are better (hooks/permission checks/context efficiency/linters/formatters) this causes a series of issues and bypasses. I have noticed Opus 4.7 defaults using shell for nearly everything which auto mode seems to always approve which then given it's now in the context window reinforces the behavior and causes more aggressive changes that then needs to be fixed. Any way to add stronger enforcement of dedicated tools to Auto Mode?

@ohishi I created it for this, it’s just a few lines of promoting

Dogfooding Opus 4.7 the last few weeks, I've been feeling incredibly productive. Sharing a few tips to get more out of 4.7 🧵

Lovely! Can you look into auto mode + Hooks/Tools. CC has a great hook system, but if agents decides to use shell instead of dedicated tooling (Grep/Glob/Read/Edit/Write) where dedicated tools is better (hooks/permission checks/context efficiency/linters/formatters) this causes a series of issues and bypasses. I have noticed Opus 4.7 defaults using shell for nearly everything which auto mode seems to always approve which then given it's now in the context window reinforces the behavior and causes more aggressive changes that then needs to be fixed. Any way to add stronger enforcement of dedicated tools to Auto Mode? cc: @bcherny @ClaudeDevs @amorriscode

Opus 4.7 is a model I’ve loved working with in Claude Code. It’s more agentic and instruction following but also incredibly smart and creative. I think it takes a slight adjustment to get used to, but it's so good with auto mode.

@trq212 Can we get diffs on plans between turns? Would improve readability so much

New in Claude Code: /ultraplan Claude builds an implementation plan for you on the web. You can read it and edit it, then run the plan on the web or back in your terminal. Available now in preview for all users with CC on the web enabled.

@trq212 Can the team update the llms-full.txt for the platform.claude.com - it’s stuck to Jan 30

@trq212 @ejae_dev Can you guys update the llms-full.txt on platform? Stuck on Jan 30

@ejae_dev --effort max also works, docs are here: platform.claude.com/docs/en/build-… also I cant tell which of your responses are AI generated or not lol

A few end of week ships: You can now set effort to 'max' which reasons for longer and uses as many tokens as needed. This will spend your usage limits more quickly so you have to activate it per session. Hit /effort to try it.

@bcherny @ph_singer @0xPaulius It’s good where it is. Just leave a hint of how to configure it until users get used to it.

@ph_singer @0xPaulius Where would you ideally want to configure it?

claude is DUMB DUMB rn

explore repo + CST then restructure to utilize rules, lazy loading of nested CLAUDE.md files, checking git vs. latest updates. I’d assume better hierarchy of these files instead of monolithic root CLAUDE.md files would improve satisfaction + performance. Most people don’t know about nesting/lazy loading.

I want to make /init more useful- what do you think it should do to help setup Claude Code in a repo?

@jerryjliu0 @llama_index @learnwdaniel please serve OCR models for PDF! @cerebras

The fundamental issue with PDF parsing is that PDFs are designed for display purposes. The internal representation of data is outputting shapes at specific coordinates on the page (e.g. "render this string at coordinate (84, 720) with this font") each displayed character could be not contiguous at all, there could be no font mapping back to unicode so you have no idea what the character is. Any PDF parser needs to magically reconstruct this random sequence of display coordinate data into semantically meaningful text, tables, and more. VLMs do help (screenshot the page and read it), but besides collapsing the metadata they still struggle in terms of accuracy and cost. note: parsing Word/Pptx as text representations so typically a bit easier too read. Our entire company at @llama_index is laser-focused on PDF parsing so we've been really trying to understand all the nuances of doc formats, especially PDFs 🙂 more notes on this coming soon