Emmanuel Galanos

icymi - @LGUS TVs spy on your network, nearby WiFi, pictures, and audio. They should be banned from import. youtu.be/6IFVTcM28KA

No idea who the owner for Dots is on OpenAI/Codex team, but the unnecessarily restrictive system prompt is killing the product usability for no good reason. If a PA responded like this, you would fire them.

It may be the least interesting model capability overhang to use models to sort out 20+ years worth of old drives & folders, but it's so satisfying

I wonder what happens to the Opus outputs from Grok bot users allowing their data to improve the model? As Anthropic's normal terms don't allow for distillation...

Claude desktop Fedora rpm via existing Claude Code cli dnf repo would be awesome. Pretty please @amorriscode @felixrieseberg ??

Astra is incredibly capable, but it only takes less than a day for a new codebase to become complete slop. It still lacks taste & judgment around architecture, file org, semantics of interfaces, etc even though it can fix any of these with prompting. Default is chaos.

PSA: Tailscale users Do NOT run `tailscale set --ssh` as suggested in setup guides. It's not needed to ssh into your boxes. Setting it makes their coordination server an auth provider (i.e. auths root logins). A legit auth/keys convenience feature, but not worth the risk IMO.

Opus 4.7's kink is adding meta-narrative/commentary to every document edit

Very very strange. Claude Code /effort max warns about the conversation being cached for the current effort level. That doesn't make sense if it is just reasoning effort per turn. Re-reasoning over previous turns? Different KV cache fidelity? Different model weights?

I've had to switch Claude Code back to Opus 4.6. I can't trust Opus 4.7 to work reliably even if peak intelligence is probably higher. I doubt this is something fixable by the harness.

Why does the Codex team not want to be paid? Business plan $25/month with no Business Pro plans. Meanwhile Claude offers Team Premium seats $150/month providing more usage. I'm now stuck using Claude for the rest of the week. Please help biz users pay for Codex @thsottiaux

Codex default sandbox is broken on Linux: Ran git diff 2a2d1810efcd15fed8417dcfb011a5887bfb6aaa --stat && echo '---DIFF---' && git diff │ 2a2d1810efcd15fed8417dcfb011a5887bfb6aaa └ bwrap: Can't mkdir /home/devuser/.codex/memories: No such file or directory @reach_vb

Claude Code v2.1.71 adds /clear to the conversation messages which then confuses Opus CC: @trq212 a quick one to fix

Bad experimental design leads to incorrect conclusions. > Anthropic learning: "Users don't want high speed + large context" > Actual learning: "Users aren't willing to pay $225/M token for high speed + large context"

Claude Code TUI has overly dumbed down the tool calls display - Read() does not even show the filename. Verbose mode can show it, but is too much detail. Not a good change. CC @trq212

Claude Code's new persistent auto memory directory feature is worse than useless for worktree users: ${HOME}/.claude/projects/${worktree-dir}/memory/MEMORY.md Claude will now write memories that will never get used. Sigh.