Post

Heavy user here. One thing: when subagents run with run_in_background=True, file reads silently fail — the file_path gets dropped. Causes infinite retry loops. Took us days to isolate. Would love a fix or at least an error instead of silent failure.

long shot! can you do anything about making a double-pro subscription level a reality?

i rarely use subagents because i can't see their thinking and can't monitor if they're on the right track. :(

Second this, /clear command is the biggest one for me

I think most make sense? The more the web experience matches my TUI experience, the better. If some are more difficult than others, I think state management is first need, followed by custom commands for skills.

Asking this question publicly is hugely powerful. A problem with most corporate systems is getting basic feedback into the development team. Developers (particularly tool devs) have a strong tendency to be receptive to these inputs, but layers of insulation get put between them and users.

I just got stung by this. `@` isn't filename completion. It is file inclusion. If you `@` complete the name of a 5GB file, but tell Claude not to read the file because it's too big, it crashes Claude because it tries to put the entire file in the context.

When you try to complete, Claude Code used to cause iTerm2 to suggest JS functions and variables, but now the screen just flashes. Both completely break completion, which means Claude accounts for 80% of all my copy and pasting.

Thanks!, I'm annoyed with myself that I didn't think of doing this. In my defence, I've been using Claude Code (on and off) since it didn't know about Claude Code, and I got out of the habit of asking it how to fix itself.

👋

curious what's guiding the team, is there a big vision or just sifting through user feedback?

I have a tmux command to take fork the session for the current tmux pane (open a new pane, runs `claude resume {session_id} --fork-session`) And the advantage of it being a tmux command is that you can run it while the current session is doing things without interrupting

It is so fun! Hard to ignore when I'd really been enjoying using it for the last year, very fun to work on a product I use all the time.

mac / iterm2, idk what other info is relevant

/me rubs hands together

There's a dangerous bug in how CC checks its allowlist for commands. If the user permits 'safecmd' unprompted, CC will run 'safecmd && rm -rf' and 'rm -rf && safecmd' unprompted. I saw this bug a couple days ago. This report links to some of the many dupe reports: github.com/anthropics/c...

Considering Claude is trained on open-source software without compensating the authors, when will Claude Code be open source?

the message swapping thing sounds fascinating in a horrible way - like the system is trying to show two states at once and can't decide which is real

it JUST happened so this time it was a compaction and three queued messages that did it

Congratulations Tom! Claude Code is a fantastic product

noted

yup I am using ghostty without tmux (somehow I like the simplicity of it) -- would definitely be appreciative if a fix for copying softwraps without hard newlines lands 👾

I currently have a weird workaround workflow of copying claude code output into a text file so I can remove the newlines before pasting it into another terminal. with opencode i can just copy and paste directly, but not sure what tui magic they use

Following on, it's not built: Test-directory (3 new/modified files): - app/api/simulate-verify/route.ts — ... - app/verify-button.tsx — "Simulate Verify" button I'm going to have to create a skill. In fairness, I've worked with junior developers who take this approach.

Hey Tom, just hit a pretty serious problem. I've also been working heavily with a TDD approach this week. Claude was failing on a particular feature when I ran it, but the tests passed. A quick skim of the code, and saw that the E2E test case was injecting a local storage value to force it!

The big one for me is when a framework/library changes a pattern, and the model isn't aware. For example, Nextjs changed their middleware pattern to "proxy". Any auth issues I hit, Claude will think the problem is that middleware.ts is missing, and struggle to see past that.

Coming back to this: an easier win could be supporting MathJax rendering in claude.ai/code. Terminal-based claude code sessions can start /remote-control and render in the browser (or just use the browser directly)