PixelsTech (@PixelstechNet) header
PixelsTech (@PixelstechNet) profile picture

A place for programmers 程序员自留地

Joined December 2011
useful tips
Firefox can take a full-page screenshot with no add-on: right-click an empty part of the page, Take Screenshot, Save full page. Or Ctrl+Shift+S. Chrome hides its version in DevTools. For a cleaner PDF, the print settings to change: pagesaver.ai/blog/save-webpa…
7
Replying to @xinxianhub
It's narrower still: only sorted sets past the listpack threshold change, since small ones never used the skiplist. And the 43% is the ordered index alone, measured on the PR branch before merge. Worth checking how many of your sets are big enough first.
1
1
4
Replying to @MizoChris
One catch: the NVIDIA App won't offer 621.35 as an update, so anyone waiting for a prompt keeps the writes. It's on Nvidia's driver download and support pages. Rolling back to 621.01 works too.
46
Replying to @nator1_hrj
Exactly. Only code that opted into the simple protocol is exposed, so a quick check is to grep for QueryExecModeSimpleProtocol, and for default_query_exec_mode=simple_protocol in connection strings. Any hit wants v5.12.0.
1
4
GeForce hotfix 621.35 stops NVDisplay.Container.exe rewriting nvlog.nvlgstg while the PC idles, the 621.14/621.24 bug users measured at up to 30 GB an hour. The NVIDIA App won’t prompt you for it: get it from Nvidia’s driver download or support pages, or roll back to 621.01.
68
GeForce 621.14 and 621.24 users report NVDisplay.Container.exe rewriting nvlog.nvlgstg while the PC idles: 515 MB in 5 minutes on one, about 30 GB an hour on another. Check in resmon's Disk tab. Users who rolled back to 621.01 say it stopped. Nvidia hasn't responded.
47
GPT-6.2 drops reasoning items when your app edits earlier conversation history: shortening an old tool result, rewriting a developer message, or replacing old turns with a summary. Expect more tokens and possibly worse answers. Append instead, or use /responses/compact.
5
Replying to @ZackShutt
Worth knowing before you wait for it: 621.35 doesn’t show up as an update in the NVIDIA App. You have to grab it from Nvidia’s driver download or support pages. Rolling back to 621.01 also stops the writes.
9
1,565
Replying to @nator1_hrj
Right, once statements are joined into one string, a trailing -- comment swallows whatever follows on that line. The extended protocol avoids it because each batched query goes over as its own Parse/Bind/Execute, never as concatenated text.
1
21
Go code using pgx's simple protocol with SendBatch needs v5.12.0 or later. Before it, a batched query ending in a -- comment with no newline let an attacker-controlled string on the next query's first line run as SQL. The default extended protocol is unaffected. #golang
1
13
gemini-4-neon is GA on the Gemini API's paid tier at $3/$15 per 1M input/output tokens, flat across its 1M context. Not a drop-in swap: temperature, top_p and top_k should go, thinking defaults to high, minimal is rejected, and explicit caching isn't supported.
1
30
PyPI now rejects sdists packed as V7-style tar and wheels whose ZIP comment could be read as ZIP structures. Build backends already produce valid archives; if you pack by hand with GNU tar, pass --format=pax. Under review next: rejecting GNU-format tar sdists entirely.
11
On Windows, an unprivileged process can watch C:\ with ReadDirectoryChangesW and see paths of files touched by every user. TU Graz tracked Firefox site visits that way with 97.8% accuracy. Microsoft calls it by design. On Linux, the inotify fix (CVE-2025-68788) is only partial.
138
Solid 2.0 is stable. createResource, batch, startTransition, on, produce, createComputed and createMutable are gone: a computation can now return a Promise or async iterator. SolidStart is retired; server functions are in core. Packages are ESM-only, Node 22.12+.
11
jQuery 4.1.0 and 3.8.0 both fix a ReDoS in selector parsing: a long chain of hex escapes ending in an invalid character could hang the page. 3.8.0 is the last planned feature release of 3.x; from here that branch gets only critical security and bug fixes.
9
Nov 30: claude-sonnet-4-5 requests fail. Dec 11: gpt-5, gpt-5-mini, gpt-5-nano and o3 stop. Dec 1: gemini-2.5-pro and 2.5-flash don't error at all, they route to 3.8 Flash and 3.5 Flash-Lite and bill at those rates. Grep your configs. pixelstech.net/article/17947…
3
127
gpt-6.2-astra is in the API at $8/$40 per 1M input/output tokens, cached input $0.40: 20% under GPT-6 Astra, per The Decoder. gpt-6.2-sol keeps GPT-6.1 Sol's $2/$10. Both ship with tighter safeguards on higher-risk security help, so vulnerability research may hit more refusals.
28
Replying to @lassiecoder
The gap is auto mode: its classifier reviews the script, not each call, so it can't see a path computed at run time. For anything you never want touched, an explicit deny rule is the safer bet, since those are checked on every call.
1
7
Claude Code's experimental Code Mode has Claude write JavaScript that calls its tools. Deny/ask rules and hooks still see each call, but auto mode's classifier reviews only the script, so it can't see a path computed at run time. To log scripts: a PreToolUse hook on REPL.
1
1
52