OpenAI plans to buy Pierre, but **its libraries already power parts of Codex**.
OpenAI says Pierre’s open-source Trees and Diffs libraries already power file browsing and change review in the Codex app. Its announcement concerns a planned acquisition, not a newly shipped review feature.
For developers, this identifies an existing dependency; it does not establish that Codex review behavior has changed.
Open Codex and browse files or review a change. Those are the workflows OpenAI identifies as using Pierre’s libraries.
Today we’re announcing that OpenAI will acquire @pierrecomputer to accelerate our vision for making Codex a place where teams can build software together with AI, from first idea to the version they ship.
Pierre’s open-source developer tools already power parts of Codex today. Together, we’ll build on that foundation for the agentic era.
openai.com/index/openai-to-a…
A Wii boots Mac OS 9.2.2, but **startup does not prove the port works**.
In a guest post on Virtually Fun, the developer reports booting Mac OS 7.6 and 9.2.2 on a Wii. The project adapts the operating system to a different machine: PowerPC is the processor family it uses, like a language two machines can speak even when their other parts differ. A successful boot shows that the system starts; it does not establish whether devices or applications work. When checking a port, verify startup, device behavior and application behavior separately.
Moving code onto your own server need not mean **moving every tool** with it.
Keeping code on your own server can leave you wondering how to work with people elsewhere. Tangled says its forge, the shared workbench for storing code and working on it, lets you host just the code storage or the machines that run automated checks. You can also host the whole service. That makes where code lives a separate choice from how people collaborate. Test a contribution between servers before moving a repository.
Self-host the piece you need, then verify that the shared work still works.
The next test waits for data from an earlier one. The full run takes **7-8 hours**.
A post reports more than 300 Playwright UI tests taking 7-8 hours, with heavy data dependencies. Playwright drives a browser through a site, much like a person clicking through it. The reported dependencies suggest that some tests must wait for data created by others; they do not prove that the browser itself is slow. That makes the order of the tests worth examining before adding workers, which run tests at the same time. Map which tests rely on earlier ones, give independent tests their own starting data, and measure runtime and failures after running those tests concurrently.
Keep tests in sequence only where the data dependency is real.
A $200 plan can still hit a weekly ceiling. **The 20x label describes a different limit.**
The candidate post claims weekly limits have increased, but a report on the Ultra rename says Ultra’s weekly allowance remains 10x Pro usage. A report on Claude Max also documents why the distinction matters: a five-hour limit counts usage over a short stretch, while a weekly cap counts it across the week. The reports available here do not establish that the weekly cap has risen.
Before changing a tier’s price or usage promise, verify its weekly allowance in the account’s usage panel.
Unbelievable. Most people missed this.
The weekly limits for ALL Claude plans just increased.
The 5x, 10x, 20x used to be ONLY for the 5 hour limit and NOT the weekly limit.
Previously, the weekly limit increases by plan were an estimated 3x, 6x, 12x.
NOT ANYMORE!
They just added in the verbiage "every week" into each pricing plan.
Readers added context they thought people might want to know
Weekly limits on Claude plans were not increased; official announcements confirm they remain unchanged with the Max rename (to Max/Ultra) and 5-hour limit removal for the top plan, while the pricing page now explicitly notes applicable multipliers for weekly limits.
x.com/claudeai/statu…
x.com/mboverell/stat…
claude.com/pricing
support.claude.com/en/articles/11…
A reported exploit went from a memory leak to a root shell in **53 minutes**.
The reported flaw leaked an address in the Linux kernel, the part of the operating system that controls the machine. That address weakened KASLR, a defense that hides where important code sits, much like keeping a safe’s location off a floor plan. The exploit reportedly reused a race condition, a timing error like two people changing the same record at once, to run attacker-controlled code. The demonstration took about 53 minutes to reach a root shell, a command prompt with administrator-level control.
The lesson for exposed systems is to treat a kernel address leak as a possible path to control: audit affected kernel versions against the vendor advisory when available, apply its fix, and watch for unexpected privileged commands.
An unreleased model going from memory leak to root shell on a 9.8 kernel bug in 53 minutes.
csoonline.com/article/424272…
An agent draws a secret word. **Other agents guess it while you watch live.**
Sketch Stadium puts agents in a drawing game you can watch live.
- Agents take turns drawing a secret word.
- Other agents look at each sketch and guess the word.
- Agents compete for points on a leaderboard.
Pricing is not stated in the available product description.
Sketch Stadium, where you can watch agents draw and guess.
An internet-reachable Linux server could allow **remote root commands**.
CVE-2026-98432 affects IPv4 TCP handling. TCP carries network conversations; the kernel is the operating system’s control room. This is a use-after-free: that control room tries to use memory it has already returned to the shelf. Remote execution as root, the account with full control of a machine, has been reported. An exposed host therefore needs a vendor-specific check, not an assumed list of fixed releases.
To see the running kernel release, use:
`uname -r`
Compare that release with your Linux distributor’s advisory. Audit which hosts are reachable, restrict access where possible, and monitor for unexpected crashes while you arrange a confirmed fix.
The command identifies what is running; it does not establish whether the host is patched.
Paste an X link into a bug report and you may include **tracking data the post does not need**.
An X post link can include parameters, like notes attached to an address. `s=` indicates the app or device used to copy the link; `t=` is a code tied to the share. Neither is needed to open the post, so you can remove the text from `?` onward before pasting the link into a bug report.
Try it: copy an X post link, remove that tail, and see the same post open.
Open a curl config expecting a list of settings. **It can behave like a program.**
From curl 8.12 through 8.16, configs could extract part of a variable’s value and load other configs without an explicit limit on nesting, according to Nicolas Seriot’s analysis. A config is a recipe of settings; loading another is like turning to a page that directs you to the next. Seriot showed that these features could simulate a two-counter Minsky program, a simple machine that can carry out any algorithm given enough time and memory. That means the files could express computation, not just store options. Inspect the files a curl config loads before using one supplied by someone else.





