Jim Cheung

Sunday, May 03, 2026

when using Noj

java -jar noj-<version>-clay.jar

there're can use global config ~/.config/scicloj-clay/config.edn:

{:base-source-path "clay_notebooks"
 :browse false
 :port 1971
 :host "127.0.0.1"}

more config options: https://scicloj.github.io/clay/#configuration

for :format [:quarto :html], need installl the quarto cli first: https://quarto.org/docs/get-started/

Tuesday, May 05, 2026

looks like Google Chrome silently installs a 4 GB AI model on your device without consent.

I checked my ~/Library/Application\ Support/Google/Chrome/OnDeviceHeadSuggestModel/ it's a weights.bin in a subfolder

I then disabled optimization-guide-on-device-model and prompt-api-for-gemini-nano in chrome://flags

not a fan of using skill / agent markdown files, however I can see the benefits of keeping some:

may steal some from here:

Thursday, May 07, 2026

I'm following this Sandboxing opencode on macOS

it's simple, just sandbox-exec

using these two with opencode for daily workflow, they're quite minimal (oh-my-openagent and superpowers are more heavy)

in general, these are the steps for a new project:

  1. git init
  2. openspec init --tools opencode
  3. open openspec
  4. /init to create AGENTS.md
  5. start with /opsx-explorer

Friday, May 08, 2026

looking for java agent framework, mainly these three:

then I found this talk Comparing Agentic AI Frameworks for Java by Timo Salm / Sandra Ahlgrimm @ Spring I/O 2026

I think the conclusion is Embabel

Sunday, May 10, 2026

been using opencode for few days, I like the openspec workflow so far, but I'm not sure about keeping the archive around, as important feedbacks already been marked to AGENTS.md time to time, not much use of the archives (maybe just not yet)

session here I mean one openspec workflow (explore -> propose -> apply -> verify -> archive)

after few sessions, I felt I wish we had more discussion before implementation. so I think for new project it's better spend the first session to generate architecture / features / roadmap markdown files, or just put them all in AGENTS.md

human in the loop (HITL)

I also think it's important to keep human in the loop. I've done only few sessions, but already overruled couple design decisions that agent proposed.

when agent stuck, they don't look into github issues. for example when using Embabel library, it gets stuck on this and this issues. I resolved both in copilot using opus by giving it the issue link. again, HITL

when I noticed agent looked into jar files to verify implementation, I started to place source code of some libraries and marked in AGENTS.md told them to look into there instead. I think it really helps and will keep this practice.

I'm using glm-5.1 and kimi-2.6 for oh-my-opencode-slim agents, even though it will be faster using smaller models for some agents, I don't see speed is an issue right now.

Thursday, May 14, 2026

I really like the opencode with openspec workflow, I start with /opsx-explore for almost everything

recently I discovered plannotator, mainly for annotate md files, still trying to setup to do code review on it

models I changed to glm-5.1 and gemma4:31b, I found kimi-2.6 is overthinking a lot

Blog Archive
    Newer Entries
  • N/A