The review bottleneck moved
Diffle is a public Apache-2.0 project for reviewing Git changes in a browser. It can compare a worktree, branch, commit, revision range, or GitHub pull request. The interface uses Vim-style movement, line and file comments, file search, full-file views, and optional language-server navigation.[1]
The project keeps draft comments under the Git directory instead of the worktree. It can print those comments as a prompt for an agent. It can also add them to a pending GitHub review, but it leaves final submission on GitHub to the reviewer.[1]
A fast review tool should reduce hand travel without reducing the reviewer's field of view.
Changed text is only the first witness
A diff shows the edit. It does not automatically show every caller, an unchanged branch in the same file, or the type contract behind a symbol. Diffle's language-server keys address that gap. A reviewer can request a definition, type definition, references, file symbols, repository symbols, or hover information without leaving the review route.[1]
Those actions map to standard Language Server Protocol requests. The specification defines go-to-definition as a request for the definition location at a document position. It defines references as a project-wide lookup for the symbol at that position. Hover asks the server for information at the same position.[3]
The protocol does not make the answer true. The server may still have stale or partial project state. The gain is navigation speed. Reviewers can widen the sightline before they accept a local-looking change.
Local does not mean inert
Diffle says language servers run through a shell inside the repository and can read anything the current user can read. The --no-lsp option starts none. That warning belongs beside the feature, especially when the repository is untrusted or the language-server command came from configuration.[1]
The GitHub route has a separate authority boundary. Pull-request loading needs a token. Review export works only when the comparison matches the committed pull-request diff. Unpushed commits, worktree changes, and stale threads are excluded from that path.[1]
The artifact exists, with a narrow check
Release v0.2.4 includes a Linux x64 archive. We downloaded that exact archive, matched its SHA-256 digest to the GitHub release record, inspected its three members, and ran the bundled binary with --help. The command listed worktree, pull request, commit, range-diff, local host, random port, language-server, and no-language-server options.[2]
That check proves the published binary starts on this Linux host and exposes the documented command routes. We did not open its browser server, give it a repository, start a language server, provide a GitHub token, or submit a review.
Keep a human sampling lane
The Hacker News discussion split around the current review problem. One commenter wanted a local way to turn line comments into an agent prompt without giving the review tool access to a remote forge. Another described reviewing through an agent procedure. A reply proposed manual review of a random commit sample to measure what that procedure misses.[4]
That is the useful operating rule. Faster navigation and review automation can share one bench. Keep a manual sample, record misses, and compare their cost against the time saved. Throughput is not a review result.