The 22nd dev log covering 54 commits — inpaint editor mask tools, LLM-intent-driven model reference injection, and allowlist-based login gating
Three takes on where agentic coding is heading — Fable 5 loop engineering, GLM 5.2 as a usable open-weight model in Claude Code, and managing skills as deployable artifacts
A tour of four open-source speech projects that run on your own machine — Whispree for voice-to-prompt, VoxCPM and Voicebox for TTS and cloning, and SpeechFeedback for Korean pronunciation correction
Comparing Upscayl and UpscalerJS — the same ESRGAN-family super-resolution tech delivered two opposite ways, one as an Electron desktop app and one as a JavaScript library
A dev log on fixing the live gate marker so the feed shows the in-progress pipeline stage correctly over SSE, plus kicking off Typst-based pptx export and Firebase auth
A thematic read of the science Gemini medical AI quantum computing and edge model announcements Google Research made at IO 2026 framed as a research to product pipeline
A short dev log on fixing OpenAI generated image size mismatches and preferring location-matched tone references in the generation path
A deep dive into AI slide generation tooling comparing native editable pptx engines against slides-as-code frameworks like Marp Slidev and Typst
Reading the Claude Opus 4.8 coding-agent gains and honesty improvements alongside the 2028 AI leadership paper released around the same time to map the two axes of frontier competition
Twenty-three commits over four days closing out the build phase — production LLM-call telemetry with token usage and logCtx, per-project state isolation under workspace refresh, the inline past-gate rewind that lets users edit any prior decision from chat without leaving the session, and the final pass to unify the revision-request marked-card visual across every canvas tab
