Git is great as the durable layer, but what fiatpandas is pointing at is a bit different — real-time awareness that the agent is active right now, not a retrospective diff.
I've been working on this split in a side project (https://github.com/rillmd/rill — vault layer on top of Claude Code). Git still handles the durable side for free, since the CLI agent just writes files and commits normally. The live side comes from Claude Code hooks (UserPromptSubmit / Stop / PostToolUse) appending to a plain activity-log.md that the Electron GUI tails. Cruder than Google Docs cursor presence, but cheap — and the log itself is just another markdown file in the vault, same data model as everything else.
One trade-off: going the other direction and letting the frontend detect edits via FSEvents-style watching runs straight into the reconciliation issues the Zettlr subthread is about. Hooks on the agent side are less elegant but sidestep the whole class.
Respectfully, did you use an LLM to write this comment? You're responding to flatpandas here. Having "what flatpands is pointing at..." is similar to the output of an LLM if you were to link to that comment and get a generated summary.
I wrote it. English is my second language, so I reach for "what X is pointing at" constructions a lot — partly as a self-check that I've parsed the parent correctly. If any specific claim in there looks wrong or generated, happy to have it pointed out.
Typically we'd see the second person "you/your/you're" used and not the third person ("flatpands") here, since you (tarr1124) are directly responding to their comment, as if in conversation with them. ie: "what you're pointing at..."
Otherwise it reads like you're ignoring them and talking around them.
Yeah, fair. The third-person thing is an old habit from parsing parents before I reply, but you're right that in a direct reply it reads cold. Second person from here on.
I've been working on this split in a side project (https://github.com/rillmd/rill — vault layer on top of Claude Code). Git still handles the durable side for free, since the CLI agent just writes files and commits normally. The live side comes from Claude Code hooks (UserPromptSubmit / Stop / PostToolUse) appending to a plain activity-log.md that the Electron GUI tails. Cruder than Google Docs cursor presence, but cheap — and the log itself is just another markdown file in the vault, same data model as everything else.
One trade-off: going the other direction and letting the frontend detect edits via FSEvents-style watching runs straight into the reconciliation issues the Zettlr subthread is about. Hooks on the agent side are less elegant but sidestep the whole class.