Most people recoiled from that shift. The dominant reaction was that intermediate edits are scratch space, not team-facing artifacts. Developers use that space to think, try dead ends, rewrite history, and clean up secrets, confusion, and half-baked ideas before publishing something others should have to read. Turning that into durable shared history felt less like better collaboration and more like keystroke-level surveillance with extra security risk. Several people also rejected the claim that pull requests are where design collaboration should happen in the first place. In healthy teams, design already happens in docs, chat, and direct conversation. PRs exist to inspect a proposed change, not to replay the entire search process that led there.
Where the conversation got more interesting was around AI. A few commenters said the strongest use case is not human collaboration at all. It is
provenance, auditability, and training data for agent-heavy workflows. If code increasingly arrives through long chats with models, teams may want the prompts, referenced files, and model outputs preserved so they can answer review questions, reproduce decisions, or satisfy compliance demands later. That framing made DeltaDB look less like a Git replacement and more like infrastructure for capturing the human-agent work trace. Even supporters usually wanted hard limits though. Local-only capture, optional sharing, and some way to summarize or promote useful checkpoints beat dumping every keystroke into permanent team memory.
A second theme was that this is not obviously a new version-control problem. Some argued Git already supports high-frequency history if teams make smaller atomic commits, preserve branch history, or use tools like `
--first-parent` to separate polished milestones from exploratory work. Others pointed to
Jujutsu and to Google’s long-running fine-grained internal history as evidence that the idea itself is not new. The unresolved question is not whether you can store more history. It is whether humans, not just models, actually benefit from raw edit streams more than from curated semantic checkpoints. On that point, most people thought Zed had not made the case.