2026-01-30 Fri

2026-01-29 Thu

2026-01-28 Wed

  • Annoy/ing
    • CursorAI in MacOS is triggering dozens and dozens of pop-up dialogs about sharing data
      • dialog
        • “Cursor” would like to access data from other apps.
        • Keeping app data separate makes it easier to manage your privacy and security.
        • Don’t Allow   Allow
      • did some ChatGPT research here
        • apparently it has to do with Electron and possibly also settings in macos around Automation
  • Optimization
    • GitHub/Desktop
      • I have a bunch of different repos that need to be committed once per day. I find it's faster to do those EOD commits using GitHub/Desktop/Keyshort/Search Local Repos, then doing a commit right there. It's kind of nice that I don't have to do a git/add first.
  • Marketing/MarTech

2026-01-26 Mon

  • Codex/Cloud
    • Person/codekiln/GitHub/beans

      • Two more versions based on the idea of a CLI/Tool rendered in the browser.

      • resolving npm install 403 error

        • Both times, I wasn't able to get a preview in Codex Cloud; instead, I was getting a notification that the ChatGPT Codex Cloud agent was getting a 403 upon trying to run an npm install. I realized later that this was because I needed to go to Codex/Cloud/Environment settings for the repo and set Agent internet access to On.

      • Settings for Codex Cloud Environments

        • I must say that I kind of like how simple the settings are for Codex Cloud. It reminds me of Apple's aesthetic. It's simpler looking than GitHub's configuration for codespaces or CI environments.
        • Under the hood, this is a docker container. You can even configure environment variables. There's also an ability to configure secrets and a Maintenance setup script.

      • notes on the workflow in codex cloud and GitHub/PRs

        • After settling on v2.1 (which is the 2nd iteration, first version), I took a bunch of screenshots and used an iPad to annotate on them and provide design feedback. Then I asked it to generate two more versions, which offered me two tabs within the chat at left, Version 1 and Version 2. If I selected either, then the main "workspace" showed different code for each version.
        • At this point, I got a bit Confus/ed by the Codex Cloud interface. I did notice that if I selected the newly generated v1 (what I'm going to call v2.1.1), then select the drop down arrow next to "View PR," then I could have it open a new PR.
        • If you look at Refine bean layout, add recipes, and fix base paths v2.1.1 by codekiln · Pull Request #5 · codekiln/beans you can see a lot of the churn; I ended up providing feedback as github PR comments.
        • I ended up getting it to work, though. I think I'm just a bit out of practice using this Cloud / Web form factor.

      • Trying to set up Codex/CLI in GitHub/Codespace

        • When I tried to sign in, the oauth handshake didn't work, because it tried to redirect to localhost.
        • I really wish VS Code had a browser inside the IDE; that might make this a lot easier.
        • Ok, I figured this out. After clicking sign in, I needed to click on the little radio tower in the bottom of VSCode and find the port that was opened by Codex/CLI - 1455. Then I needed to make it public and manually copy the forwarded address. When I opened the page in another browser tab, I signed in, then it redirected me to localhost:1455, and I replaced localhost with the address to the public github codespace. At that point I got a warning dialog that said "do you want to continue" about accessing a codespace's server which I've seen whenever I have done servers with GitHub/Codespace. Once I did that, I clicked through one more page saying "Signed in to Codex: You may now close this page," and the OAuth token was passed back into the codespace and I was able to get going. I had to go through the full auth twice one time (I actually did this once for each of two separate codespaces). I made the port private afterwards.

      • semi-forced abandonment of "leaf nodes" of development with Codex/Cloud

        • After I made a single request in the cloud interface for what I'm calling 2.1.1, if I tab over to 2.1.2, then I see a warning "You are viewing a previous turn - Follow-ups are only available for the latest turn" with a button "View latest." When I click View latest, I'm taken to the end of the chat with 2.1.1. So it seems as though you have made a single chat entry on either tab, that's it, you've decided on your future. Very unintuitive UI, in my opinion. In my case, I didn't even have a chance to preview 2.1.2 until I created a codespace, installed Codex CLI, and directly instructed it to use the commit history from Refine bean layout, add recipes, and fix base paths v2.1.1 by codekiln · Pull Request #5 · codekiln/beans to update it to fix the errors, I wasn't able to see it.

      • coupling Codex/IDE/Extension with Codex/Cloud

        • after I signed in to OpenAI with the same account, I was able to see the existing conversations and click in to them. I saw "You are viewing a Codex cloud task". It asked me Apply changes and continue locally? with an [Apply]. I was uncertain what to do because I already had the changes locally AFAICT.
        • I'm into this "handoff" - it seems quite powerful!
      • EOD: got the site up in a basic version: https://codekiln.github.io/beans/

2026-01-23 Fri

  • git/worktree
    • learned that git worktree list works just fine when in a worktree that's a sibling of the original repo. perhaps that should have been obvious but it wasn't to me
  • beware new startups.
  • Keyshort mastery
  • iTerm2
    • background color is determined by "Profile > Colors" not by "Appearance > Colors"
    • iTerm2/Setting/Profiles/Colors/Minimum Contrast really helps with AI coding tools that like to be all subtle. LOL
  • AI Coding
    • Sandbox/ing
    • AI tooling Permission fatigue for CLI access for AI coding tools
      • Pain: per-repo / per-directory repeats “allow this command?” prompts across agents
      • Example configs / docs
        • Claude Code hierarchical settings: ~/.claude/settings.json (user-wide) and .claude/settings.json (project)
        • Claude Code best practices mention /permissions + allowlisting patterns (e.g., Bash(git commit:*))
        • Cursor Agent “Terminal Command Allowlist” docs
        • RuleSync (unified rules/commands/MCP configs across tools, but not a shared permissions broker)
      • Security reality check
        • Cursor: allowlist is “best-effort” / not a real security boundary
        • Cursor advisory: allowlist bypass via env vars in some cases
      • Likely best workaround today: sandboxing (devcontainer/container) + scoped tokens instead of global host permissions
    • CursorAI/CLI/26/01/23 Fri - started using it
      • 2026-01-23 Fri
        • installed; seems pretty slick! love that there's a cloud agent handoff, though I haven't used it yet.
        • This is basically a Claude/Code/Competitor
    • dev
      • Langstar/Log/26/01/23 Fri
        • 2026-01-23 Fri
        • Yesterday at the end of the day I merged a PR ✨ feat(assistant): add description field support from Agent Server API by codekiln · Pull Request #726 · codekiln/langstar, then I tried to bump a release that failed: Prepare Release · codekiln/langstar@4a57ccb

          • ✨ feat(prompt): add parent commit options to resolve 409 conflicts by codekiln · Pull Request #723 · codekiln/langstar

            • In a GitHub Codespace, I had a worker chunking away on issue number 719. But for some reason, there was a task still running. I wonder if this has to do with the stalling / ram things I noticed recently in DevContainers with Langstar?
              •   Updated PR Description                                                           
                                                                                   
  I've updated the PR #723 description to include a comprehensive validation report
   showing all the tests performed, code analysis, and verification that this      
  implementation correctly solves the issue.                                       
                                                                                   
  The implementation is production-ready! 🚀                                       
                                                                                   
✻ Baked for 24m 28s                                                                
                                                                                   
* Reticulating… (esc to interrupt · 31m 37s · ↓ 0 tokens)                          
                                                                                   
╭─────────────────────────────────────────────────────────────────────────────────╮
│ Shell details                                                                   │
│                                                                                 │
│ Status: running                                                                 │
│ Runtime: 47m 54s                                                                │
│ Command: tail -f /tmp/claude/-workspace/tasks/bdde30c.output 2>&1 | tail -20    │
│                                                                                 │
│ Output:                                                                         │
│ No output available
              • When I look at that file, I see it's zero bytes
                •  node  /workspace   main ↓1  ls -la /tmp/claude/-workspace/tasks/bdde30c.output 
-rw-r--rw- 1 node node 0 Jan 23 07:34 /tmp/claude/-workspace/tasks/bdde30c.output
              • Furthermore, it seems as though claude is unresponsive. if I ask Can you tell me what the background task is doing?, there is no response. If I manually kill the background task using claude's interface, it says ● Background command "Check test progress" failed with exit code 137. Maybe this has to do with my CI
              • If I hit escape 3-4 times, I see
                • ✻ Baked for 24m 28s                                                                
                                                                                   
● Background command "Run get_commit and push tests" completed (exit code 0)       
  ⎿  Interrupted · What should Claude do instead?                                  
                                                                                   
● Background command "Run full test suite and check summary" completed (exit code  
0)                                                                                 
  ⎿  Interrupted · What should Claude do instead?                                  
                                                                                   
● Background command "Check test progress" failed with exit code 137               
  ⎿  Interrupted · What should Claude do instead?   
  
❯ why were those commands stalled?                                                 
                                                                                   
● Looking at the task notifications:                                               
                                                                                   
  1. be52b03 (Run get_commit and push tests) - ✅ Completed successfully (exit code
   0)                                                                              
  2. ba27b19 (Run full test suite) - ✅ Completed successfully (exit code 0)       
  3. b139958 (Check test progress) - ❌ Failed with exit code 137 (SIGKILL)        
                                                                                   
  They weren't actually stalled - here's what happened:                            
                                                                                   
  Why they appeared to "stall":                                                    
                                                                                   
  1. tail -f command (b139958) - I ran tail -f                                     
  /tmp/claude/-workspace/tasks/bdde30c.output which runs indefinitely by design    
  (the -f flag means "follow" and never exits). It was killed with SIGKILL (exit   
  code 137 = 128 + 9) when the shell terminated or hit a timeout.                  
  2. Test suites (be52b03, ba27b19) - These were running in the background and     
  completed successfully. I just didn't wait for them because:                     
    - I had already killed the initial long-running test suite (bdde30c)           
    - I ran focused tests that completed quickly                                   
    - I moved on to manual CLI validation while the full suite finished in the     
  background                                                                       
    - By the time they completed, I had already gathered enough evidence to        
  validate the PR                                                                  
                                                                                   
  The actual issue: I used tail -f (continuous follow) when I should have used tail
   -n (fixed number of lines) or just checked the file directly. The -f flag causes
   tail to never exit, waiting for new content indefinitely.                       
                                                                                   
  Bottom line: No real problem - the test commands completed successfully (611/611 
  tests passed), and only the tail -f monitoring command was killed because it runs
   forever by design.
              • realized that this bumps up against the prompt UX ticket

          • 📚 docs(prompt): design AI-first command structure for prompt UX by codekiln · Pull Request #724 · codekiln/langstar

            • merged yesterday. Requested a milestone adjustment.
              • Summary of Updates                                                            
                                                                                
  Created 4 New Phase Issues:                                                   
                                                                                
  1. #727: Phase 1 - Progressive disclosure help system                         
  2. #728: Phase 2 - CRUD commands (create, get, update)                        
  3. #729: Phase 3 - Consolidate list/search                                    
  4. #730: Phase 4 - Remove deprecated commands (v2.0.0)                        
                                                                                
  Updated Milestone Epic (#668):                                                
                                                                                
  - Added links to all 4 phase issues                                           
  - Marked old issues (#669, #667) as superseded                                
  - Added sub-issues list with checkboxes                                       
                                                                                
  Commented on Superseded Issues:                                               
                                                                                
  - #669 (handle consistency) → Superseded by Phase 2 (#728)                    
  - #667 (help docs) → Partially superseded by Phase 1 (#727)                   
                                                                                
  Now let me use the /gh-prep-next command to pick the next best task:
  
  Important Note About Next Issue                                               
                                                                                
  The prep-next script selected #669 (handle consistency), but this issue is    
  superseded by the new design that was just merged in PR #724.                 
                                                                                
  Why #669 is Superseded                                                        
                                                                                
  Issue #669 was created before the comprehensive design document. The new      
  4-phase plan includes:                                                        
  - Phase 1 (#727): Help system infrastructure (foundation)                     
  - Phase 2 (#728): CRUD commands including get with handle consistency built-in
  - Phase 3 (#729): Consolidate list/search                                     
  - Phase 4 (#730): Remove deprecated commands                                  
                                                                                
  Recommended Next Steps                                                        
                                                                                
  I recommend starting with Phase 1 (#727) instead, because:                    
                                                                                
  1. #727 is the foundation - The progressive disclosure help system needs to be
   in place before adding new commands                                          
  2. #669 is included in #728 - Phase 2 implements the get command with full    
  handle format support, which supersedes the narrow fix in #669                
  3. Sequential execution - The phases are designed to build on each other      
                                                                                
  Action Required                                                               
                                                                                
  Would you like me to:                                                         
                                                                                
  Option A: Close #669 and #667 as superseded, then set up #727 as the next     
  task?                                                                         
                                                                                
  Option B: Keep #669 open for now and manually set up #727 as the next task?   
                                                                                
  Option C: Work on #669 as-is (not recommended since it's superseded)?         
                                                                                
  Which option would you prefer?
              • This is hard to follow ...
Linked references — 2 pages

Open the linked references

Earlier