ARCHIVE
The English edition of my blog.
In my custom browser Omoikane, I've corrected the default 8 px margin on the <body>, implemented declarative Shadow DOM, and fixed Unicode character direction for vertical text. These seemingly minor web‑compatibility issues actually affect the cascade, the HTML parser, and the rendering pipeline.
Added a Judgment Layer to the Visual Store that records, for each image, the decisions made before opening the image as well as the outcomes verified by the Vision LLM or a human. Notes on how to integrate with Jev and on migrating to v3.
To jev-mcp, which invokes Jev from Codex, we added an evaluation JSONL for Japanese coding judgments, an evaluation CLI, telemetry that does not record inputs, structured error handling, and decision recipes.
We built jev-mcp, which uses Codex’s TypeSafe AI decision model Jev. The design isolates classification and relevance scoring, includes four MCP tools, supports batch processing, validates responses, and summarizes the current scope of Codex Skills and verification.
I added Pointer Lock and ElementInternals to my custom browser, Omoikane. This summarizes fixes for X11 input duplication, iframe realms, form state restoration, performance degradation caused by unnecessary DOM traversal, and a garbage‑collection reference leak in Boa’s Array.from in the Rust‑based JavaScript engine, along with the necessary corrections.
I have added the Fullscreen API and TextEncoderStream/TextDecoderStream to my custom browser, Omoikane. I will address synchronization between the top layer and the native window, chunk boundaries, backpressure, decoder state that spans garbage collection, and the remaining issues identified in Web Platform Tests.
Progress on cubase_mcp, which controls Cubase from MCP. This includes investigations of the Mixer Bank and DirectAccess for obtaining the track list, points where we're stuck during hardware testing, and the input guard, Track Probe, and I/O Probe that were merged on September 19.
We have released a compatible simulator lqs (local queue service) for testing SQS locally. It supports an HTTP API that can be accessed via the AWS SDK, as well as Standard/FIFO queues, visibility timeouts, long polling, DLQs, and other features.
I recalled the Labelflash that came with the Gateway PC I used back when Windows Vista was around, and when I looked it up, I realized there are hardly any opportunities to use it anymore.
We will clarify the meaning of encountering “Error running remote compact task: {"detail":"Bad Request"}” in Codex’s remote compact, and differentiate it from context_length_exceeded and 413 Payload Too Large.