Launching today
Diffsmith
Comment on your AI agent's code & collaborate on changes
94 followers
Comment on your AI agent's code & collaborate on changes
94 followers
Diffsmith is a code review studio for local git changes made by Claude Code, Cursor, Codex, Copilot, or any other AI coding agent. Open a repo, read the uncommitted diff, and click any line to leave a comment anchored to that exact spot. Then hand your notes straight back to the agent. Supports MCP for viewing you're agent's replies to your comments directly at the line of code it relates to. Diffsmith is 100% offline and a one time purchase unlocks lifetime access.
One-time purchase for a tool that sits in the hot loop of agent coding is a refreshing call. The mechanic I want to understand: when the agent answers a comment by rewriting the whole block, where does the anchor go, does the comment follow the moved logic or die with the old lines? That resolution step decides whether review rounds stay readable after round three.
@vollos Great question! In that situation, comments can’t stay with the new lines so they get archived into a “Past” tab in the sidebar (you can click it there to see the original lines it was anchored too).
Agents can add a comment to the newly changed code if they want to (and in my experience they often do because they understand that the original comment will be archived)
@maxchuquimia Archiving beats the pretend option, a comment glued to lines it no longer describes is worse than one that admits it's history. Agents leaving a fresh comment on the new code because they know the old one gets archived surprised me, that's the loop closing itself. Thanks Max.
The line-anchor model nails what the agent wrote, but reviewing Claude Code and Codex diffs my highest-leverage notes are almost always about what it didn't write: 'you skipped the error path here', 'no test for the empty input'. There's no changed line to click on for those. How do you attach a comment to an omission, or to a file the agent should have touched but left alone? That's where a lot of my review time actually goes.
The MCP path is what makes this a loop rather than a one-way review, since the agent's replies land back at the line they belong to. Where does that comment state actually live: inside the repo as something I would gitignore, or in app-local storage outside the working tree? I ask because I want to know what happens to anchored comments when I switch branches or stash mid-review, and whether the MCP server is something the agent polls for new notes or you push to it when I finish a pass.
AI code review is most useful when it is tied to release risk, not just style feedback. I would rather see fewer comments with clear repro, affected behavior, and test evidence than a long list of plausible-sounding notes the team has to triage again.
the line-anchored comments thing is genuinely clever, like you can basically hold a real conversation with the agent right on the code instead of copy-pasting snippets back and forth. also appreciate that it's fully offline, no telemetry nonsense, just open the repo and go
honestly this looks really useful for my workflow since i use cursor and claude code daily. one thing i'd love is a way to batch similar comments together, like if i want to flag the same issue across multiple files, it gets tedious clicking line by line for repetitive notes. a "apply comment template to selection" or something similar would save a lot of time when reviewing bigger changes
honestly the line-anchored comments and pushing feedback back to the agent sounds super useful, but it would be even better if i could batch related comments into a single thread so i don't have to send five separate messages to the agent when i'm reviewing one function.