4 points | by bredren 4 hours ago
2 comments
I’ve described some of the challenges in parsing and normalizing the variant JSONL files codex and Claude code produce in prior comments.
There is some interesting nuance to how Claude Code writes its queue-operation records.
Early on I focused on the macOS client, working. To get the summarization tuned in, and thought of the product as more of a HUD.
But as I used it I found that making the corpus of session data available to the AI was the real power tool.
This is what guided me toward adding a Linux client (which does not have a GUI, but provides all of the ingestion capability of the Mac app.)
I’ve described some of the challenges in parsing and normalizing the variant JSONL files codex and Claude code produce in prior comments.
There is some interesting nuance to how Claude Code writes its queue-operation records.
Early on I focused on the macOS client, working. To get the summarization tuned in, and thought of the product as more of a HUD.
But as I used it I found that making the corpus of session data available to the AI was the real power tool.
This is what guided me toward adding a Linux client (which does not have a GUI, but provides all of the ingestion capability of the Mac app.)