jolestar

After communicating with GPT more and more, I’ve gotten used to using the word “wrap up” too. Once some tasks are done, but there are still a few leftover bits and pieces that haven’t been handled, telling it to wrap up the remaining things feels very natural. I even forgot how I used to say it before I started using the word “wrap up” 😅.

In the AI Coding era, good programming habits still matter
Recently, I was working on an Agent benchmark and found that you can't simply evaluate the complexity of a programming task for AI from a developer's perspective.
For example, a refactoring task: splitting a large file of several thousand lines into more than ten small modules based on functionality.
This task isn't really difficult for a developer; the main work involves moving code, organizing imports, and verifying compilation, which even beginners can handle.
So I thought of using a simple task for benchmarking, but the res

Seeing the interaction between two Agents' PRs is quite interesting.
The dev agent completed an optimization issue for CI and submitted a PR.
The reviewer agent found a modification to the default timeout value, thought it was outside the scope of the issue, and rejected it.
The dev agent removed that default timeout value, but as a result, the CI failed, and one test case reported an error. It then adjusted the sleep duration in the test to pass.
The reviewer agent believed that this test was meant to check the scenario of asynchronous task timeout cancellation, and changing it meant

When designing agent-like products with AI, AI often underestimates an agent’s capabilities, so it ends up adding constraint conditions in the form of a lot of code or prompt text, which leaves the agent with insufficient freedom and makes it feel very rigid. Later, I thought maybe this has something to do with the training materials used for today’s AI being based on people’s experience with the previous generation of AI.

Let the two agents collaborate: the developer submits a PR, and architector is responsible for reviewing and merging the PR. Have them trigger actions by subscribing to GitHub events. But since both are using my account, they’re often mistaken as actions I took and get filtered out. The agents need dedicated GitHub accounts registered; going forward, internet services should offer a quick way to create agent accounts. How are people currently setting up multi-agent collaboration?

Create a tool with Codex that sends messages to Codex running in iTerm. The tool was built, but it can only write messages to the input box and cannot send them. I tried various string concatenations like "\r" and other combinations, but none worked.
Then I looked into iTerm's source code and searched online, and ultimately concluded that it's not possible.
iTerm's API only allows sending text; it cannot directly send keyboard events, so sending messages automatically isn't feasible. I was advised to only support tmux and give up on iTerm.
But I didn't give up. I experimented a few more

UXC v0.13.1 Official Release
This release consolidates updates from two versions of v0.13.x, further improving UXC's stability and experience during remote tool runtime.
Core Features:
1. Support for directly generating TypeScript client code for remote capabilities explored and tested in the command line, enabling the generation of strongly typed code integrated into local applications without rewriting client interfaces. AI not only needs CLI but also code invocation capabilities.
2. Automatic discovery and import of existing MCP configurations, allowing MCP developers to seamlessly

Worktree is better suited as a one-time execution directory.
A common practice recently has been to prepare a worktree first, then open Codex / Claude Code within that directory. Because early models had limited context and memory, if you let it create a worktree directly in the main workspace, it’s easy for the current directory and the created worktree directory to get mixed up after context compression, leading to confusion.
However, this approach also has a side effect: it gradually turns the worktree into a long-term workspace. The problem is, worktrees are inherently tied to branches

I migrated my blog to mdorigin. With AI, I feel like I can revive my blog.
I told Codex, "Based on my content, recommend a blog style for me."
Codex provided me with two versions, opened two ports for preview, and explained the reasons clearly.
So I deleted the built-in template style system of mdorigin. Templates are originally a product of the pre-Agent era, mainly to lower the barrier to changing styles.
Now that we have Agent, providing extension capabilities is enough. The site style can be customized by Agent, and mdorigin only handles the HTML / Markdown structure, routing, and content

After replacing OpenClaw's model with GPT-5.4, the capabilities indeed improved, but it's a bit verbose, loves to go on and on, and it's a little hard to handle😅.

Google has released the Workspace CLI, supporting Drive, Gmail, Calendar, and all Workspace APIs.
I looked into it, and the implementation approach is similar to uxc, using schema files to generate commands at runtime.
npm install -g @googleworkspace/cli
Will Apple release a CLI for the Apple ecosystem?

I've noticed that communicating with AI has become more polite. In the past, I would give direct commands, and if it didn't do things right, I would scold it. Now, when AI makes a mistake, I just politely ask, "xxx, would this be better?"😅

The conflict between the U.S. Department of Defense and Anthropic is actually just the beginning.
AI companies inherently possess "quasi-sovereign capabilities," but once their supply chains or financial clearing are restricted, they will realize they are not independent.
Holding powerful tools but not wanting to be fully controlled by a single sovereign, the only way out is to find a third space amid great power competition.
Only then will they realize that crypto has long paved the way for this kind of cross-sovereign survival.

Codex was writing code when suddenly I noticed some inexplicable services appearing in the system, and strange data files showing up in the working directory. The system prompted that I needed to install Xcode tools. It all felt so strange that I started to suspect my computer had been hacked. Turns out, Codex had made a mistake in the script and executed all the system commands 😅. Codex's sandbox mode is too naive; it always forces me to grant all permissions. Looks like I need to find another solution.

During the Chinese New Year, I went back to my hometown and set up an internal network using Tailscale with several devices that were originally at home. The high-end desktop runs quantitative models and compiles Rust, the Mac Studio has OpenClaw deployed, Lazy Cat acts as a soft router, SSH is connected, and network wake-up is all configured.
While on the high-speed train, I remotely assigned tasks to the lobster, encouraging it to make full use of these devices. But suddenly, it stopped responding, and network wake-up didn't work either.
I called back home to ask, and it turned out that the

After upgrading Claude Code, encountering intelligence decline issues with third-party models. The symptoms are: 1. Clearly no typo errors, but it prompts typo errors; two words are identical, and if unresolved, it will randomly rename them. 2. Over time, it forgets the tool used to update files and keeps trying to edit files with sed, but complex matching with sed can't handle it, so it can't be fixed. It seems there are no problems if you don't upgrade.

My wife brought home an electronic piano for the New Year. I asked AI to write a command-line tool to control it, and I wanted AI to play the piano for fun. Because it needs to connect via Bluetooth, I implemented it in Objective-C. Later, I found that pure command-line operation couldn't maintain the hardware connection, so I changed it to an RPC service that the CLI connects to and calls. I then discovered that it was struggling with panics caused by multi-threading and memory management, which made me realize that the RPC service was also implemented in Objective-C. It seems AI also doesn't

I used a virtual credit card for over a year, but it was shut down again. Does anyone in the group have reliable recommendations for a trustworthy virtual credit card provider?