dsh-diff-review

Codex-style change review for DeepSeek Harness. Whenever the agent edits files, a pill above the composer reports N files changed +M −K for that turn; hovering it floats the file list up, and the right sidebar shows each file's syntax-highlighted diff with a one-click revert.

English | 中文

• A pill above the composer: centered over the input box (not in a sidebar), counting only the files the current turn touched and resetting at each new turn.
• One pill, nothing else: N files changed +M −K — no separate "Review" or "Refresh" buttons. Hover opens the list, a file row opens the review page, and polling refreshes every 2 seconds.
• Hover only: moving into the dock floats the list up from above it and turns the caret up (▴); moving away fades it out. Clicking the pill itself changes nothing. The list is absolutely positioned, so it never pushes the composer down; in a short window its max height shrinks to the space above and the rest scrolls.
• Per-turn counts: edits to the same file within one turn are measured against that turn's starting content, for both the counters and the diff — not just the last edit.
• Counters bump: a small springy lift when a number changes (prefers-reduced-motion disables it).
• The list hugs its content: as wide as the longest file name, never a stretched empty right side; only a genuinely long path clamps at the composer width with an ellipsis.
• Names left, counts right: names start at one left edge; the two counter columns are right-aligned columns.
• Above the native docks: the pill sits above the native todo and queue docks, which keep their own appearance untouched.
• Right-sidebar review page: a tab strip along the top (browser-like tabs, scrollable sideways) and the whole area below is the selected file's diff — no side-by-side panes.
• Syntax highlighting: the language comes from the file extension (dart, cpp, py, js, mjs, …) and colors come from the client's own shiki highlighter and theme tokens; a not-yet-loaded grammar renders plain until it loads, then repaints.
• Scroll performance: diff rows are grouped 60 at a time into content-visibility: auto boxes, so off-screen groups skip layout and paint; highlighting and element construction are memoized per payload. A 2000-row diff scrolls at p50 16.7 ms / p95 17.7 ms per frame (containing every row individually instead degrades to p50 20.6 ms).
• Conservative revert: writes back only while the file still matches what was recorded, and reports a conflict instead of overwriting blindly.
• Only the agent's own writes: recorded from edit / write tool results (recognized by result shape, so Code-Mode sub-dispatches count too); > redirection or sed -i inside bash is not tracked.

The list that floats up on hover:

The review page: tabs on top, the whole diff below, colored by Dart grammar.

Install

dsh plugin --profile web add dsh-diff-review

Restart the app afterwards (the plugin list is read at startup). In the desktop app, use your own profile name instead of web.

Use

Just let the agent edit files. Once the pill appears during a turn:

• hover it to float the file list up, move away to dismiss (hover only; clicking the pill does nothing);
• click a file name to open the review page on that file in the right sidebar;
• switch files with the tabs at the top, and use "Revert all" to put this turn's changes back.

To remove it again, delete dsh-diff-review from the dsh.profile.bundles list in your profile's package.json and restart.

API

All three routes are behind the browser session (the page calls them same-origin, so the cookie rides along):

Method	Path	Purpose
GET	/api/diff-review.changes?sessionId=	this turn's changed files + summed line counts
GET	/api/diff-review.diff?sessionId=&path=	one file's diff rows + both full sides (for whole-file highlighting)
POST	/api/diff-review.revert	{ sessionId, paths? }; omit paths for the whole turn

How it works

• index.js (Host): records before/after from a tools/post-execute listener, groups changes by turn from session/event turn/start, and serves the three routes above. The diff is computed in-house with an LCS walk, so it needs no git and works outside repositories.
• client.js (browser): the conversation.input.dock pill plus the sidebar.right.pane.tab review page, with a zh/en dictionary.
• test/host.test.mjs: Host-side unit tests — node --test test/host.test.mjs.
• UI regressions (each needs BASE=… TOKEN=… pointing at an instance with the plugin loaded):
  • test/review.mjs: review-page layout + syntax colors (stubbed data);
  • test/popover.mjs: hover open, upward placement, fade in/out, caret direction, click does nothing, composer stays put;
  • test/order.mjs: the pill sits above the native todo dock;
  • test/shot.mjs: dock screenshots and the name→counter gap;
  • test/width.mjs: the list hugs its content (ONE=1 single row, LONG=1 long path, WIN=760,800 narrow window);
  • test/counter.mjs: counter right-alignment + the bump animation (getAnimations() proves it runs);
  • test/multi-edit.mjs: three edits to one file in one turn — asserts cumulative counts and the full diff;
  • test/perf.mjs: scroll frame times over a 2000-row diff (and the single-pill shape);
  • test/dart.mjs / test/e2e.mjs: end-to-end runs with a real turn.

Known limits

• Records live in process memory: after a restart, older chats have no history diff until they run a new turn.
• Each session keeps the most recent 12 turns.
• Diffs are truncated past 4000 lines per side.
• Deleted files do not go through edit/write, so they are not shown as changes.
• Client-side dsh.client changes (new slots, bundle shape) need an app restart; editing client.js content does not hot-reload an already-open page (client-hmr only covers runtime-rebuilt bundles; installed plugins need a restart).

License

MIT