@hallelx2 The current README adds an important caveat: 0.935 is pooled F1 on 107 PDFs; 0.442 is per-document F1 on 125. I wouldn’t read the difference as the size of the detection gap.
https://t.co/FOGT7mD1z6
@iamvishal16_ios Claude Code has an off switch: "idleCompaction": false in settings.json (v2.1.290+). The release notes don’t say whether there’s a warning first.
https://t.co/vXvhbFN6CC
@Desolatte Keeping follow-up questions on the original commit is a useful detail. A later fix shouldn’t quietly change the code you’re discussing while you’re still trying to understand the incident.
@hartner12 The parent names help with another ambiguity too: the same component can appear in several panels. Knowing its name still wouldn’t tell Claude which instance you meant.
@ryanhefner Was OpenGraphs holding up Skylounge here, or is this a different set of projects? I saw how you planned to connect those two in your older write-up.
@dusangran The rejected approaches are the bit I’d worry about losing. A summary can keep the plan but drop why option B was ruled out, and suddenly B looks worth trying again.
@shawn_pana The guide’s “iPhone in Use” rule is a good boundary. Unlocking your phone should hand control back to you, rather than have the agent keep retrying taps.
@thedatadr1ver I’d count repeats of the same failure, rather than every failed attempt. A worker testing a new hypothesis is different from rerunning the same command three times.