dsh-sidebar-git-graph

A right-sidebar page for DSH (DeepSeek Harness) that draws the current session's repository as a branches-and-merges commit graph — VS Code style: colored lanes, bezier curves where branches split and merge, ref chips, per-commit details — with a working-tree Changes area above it.

Read-only. It never runs checkout, commit, reset, rebase, fetch, push, add or restore. The only git commands it runs are rev-parse, status, for-each-ref, log and diff.

The Git graph page: colored lanes, ref chips, commit rows

The screenshot above predates the Changes area and shows the commit graph alone.

Features

Working-tree changes, VS Code style — a Changes and a Staged Changes section above the graph (plus Merge Changes during a merge), each row carrying its status letter (M/A/D/R/C/U) and its added/deleted line counts. Click a file to expand its diff inline; the sections fold and remember it.
Branch/merge relationships at a glance — one lane per concurrent line of history, curved connectors where a branch forks or merges back, thicker dots on merge commits, and the HEAD lane emphasized.
Ref chips per commit — local branch / remote branch / tag, with the current branch filled in the theme's brand color and the theme's matching on-brand foreground.
One workspace, several repositories — a workspace often holds more than one repository (this project's own checkout has a nested one under vendor/@zeng/). The header then grows a picker listing every repository it found; the choice is remembered per workspace. A workspace that is itself only a subdirectory of a repository works too — the enclosing repository is offered as well. Discovery never leaves the workspace: it scans downwards, bounded by depth, directory count, repository count and a time budget.
Header at a glance — repository name, current branch, ↑ahead ↓behind, dirty-file count.
Commit search — highlights matches and steps through them (the graph stays intact; it never re-lays-out a filtered list).
Details on click — full sha, author + email, absolute time, parents, participating refs.
Virtualized list — a 2000-commit window still renders only the visible rows; nothing is requested while the page is not the visible tab.
Honest error states — not a git repository / no git on PATH / no workspace / trust fence refused / git failed each get their own message instead of a blank page.
Zero dependencies, no build step — the host half is plain ESM, the browser half is a single hand-written client-bundle file. What you read in lib/ is what runs.

Install

dsh plugin --profile web add github:AFunDog/dsh-sidebar-git-graph

Then restart dsh web (the host half adds a route) and hard-refresh the browser. Open the right sidebar and pick Git graph — from the + menu when dsh-better-sidebar is installed, or from the sidebar's guide page on a bare DSH (the + control is not drawn while that pane already holds the guide tab). The page follows the session you are viewing: switch sessions and it re-reads that session's workspace.

Pin a version, or install from a local clone

# pin a tag (recommended: lockfiles record the resolved commit)
dsh plugin --profile web add github:AFunDog/dsh-sidebar-git-graph#v0.2.0

# from a local clone (absolute path only — relative paths are rejected by spec parsing)
dsh plugin --profile web add /path/to/dsh-sidebar-git-graph

With or without dsh-better-sidebar

The page works both ways, and picks automatically:

Your setup	What happens
dsh-better-sidebar installed	The page registers through its public `ctx.betterSidebar.registerTab` service: it shows up in the `+` menu, gets a card in Settings → Side cards (with an on/off switch and a per-plugin setting for the commit window), and appears under Git graph.
Bare DSH	The page registers with DSH's own sidebar tab type (`ctx.sidebarRightTabs` + the `sidebar.right.pane.tab` slot) — same column, same `+` menu, minus the settings card.

The service contract is re-stated locally (only registerTab / openTab are used) and is activated with ctx.inject, never a static inject: a statically declared service that is absent would park this plugin's fiber, which shows up as "works after a hot reload, gone after a restart".

How it works

lib/index.js      host half  — POST /dsh-sidebar-git-graph/api, git execution, trust fence
lib/git-read.js   host half  — pure: argv builders + parsers (importable by plain node)
lib/changes.js    host half  — pure: working-tree argv builders + parsers
lib/repos.js      host half  — "how many repositories are in this workspace" discovery
lib/workspace.js  host half  — sessionId → working directory, workspace allow-list, fence
lib/client.js     browser half — tab registration, lane layout, SVG rendering (single file)
test/             zero-dependency tests (node scripts/test.mjs runs them all)

The route dispatches on the request envelope's method: graph (commit DAG), changes (working-tree status) and diff (one file's patch).

The browser half reads the session's working directory and asks the host for a snapshot.
The host resolves which workspace belongs to that session (session service → persisted workspace table → workspace registry → newest session directory), refuses paths that are neither registered nor the session's own cwd, then picks the repository to draw: the one the browser named (if it passes the fence and rev-parse agrees it is a repository root), else the one containing the workspace, else the first one it finds inside the workspace. It runs the four read-only commands with a sanitized environment, a 10 s timeout and a 48 MiB output cap, and answers with { repo, repos[], selection, refs, commits[{ sha, parents, author, email, time, subject, refs }] }.
The browser half lays the commits out into lanes with one pass over --topo-order history and renders SVG bezier connectors under a virtualized row list.

The layout function and the geometry builder are pure and exposed on the plugin's internals so the test suite can drive them without a browser.

Settings

Setting	Where	Default	Meaning
Commits loaded per read	Settings → Side cards → Git graph → feature settings (dsh-better-sidebar only)	`400`	100–2000. Bigger is more complete and slower on first read.
Repository scan depth	same place	`5`	0–8. How deep to look for repositories to offer in the picker; `0` means "no scan — only the repository the working directory is in". Lower it if a huge workspace reports an incomplete list.
History scope	page header	all branches	All branches or current branch only.
Commit search	page header	—	Highlights matches and steps through them.

Limitations

Read-only by design. No staging, no commit, no checkout, no push. To change the working tree, use another tool.
Untracked files carry no +/− counts. git diff --no-index compares exactly two paths, so per-file counts would mean one process per file — hundreds of them in a repository that has not been added yet. The section header counts what it can and leaves the rest blank rather than printing a 0 that would read as "nothing changed". A file changed on both sides shows up in both sections, which is what the two sections are for.
A nested repository is one non-clickable row. git does not descend into a repository inside a repository (even with -uall), so the outer repository sees a single directory and there is no diff to show for it. The page says so on the row instead of doing nothing when clicked.
At most 3000 rows, and at most 512 KiB of patch per file; both are reported in the page rather than silently cut.
The window is bounded (up to 2000 commits per read, default 400); older history is reported as truncated rather than silently dropped.
Repository discovery is bounded, and it scans the filesystem. Depth, directory count, repository count and elapsed time are all capped; hitting a cap is reported in the page instead of being hidden, and lowering the scan depth is the fix. Directories that never hold repositories of interest (node_modules, target, .venv, build output, …) are not descended into — though each of them is still stat'd once, so a repository that happens to be named build is still found.
The picker only offers what is reachable from the workspace — the workspace itself, something inside it, or the repository that encloses it. The host refuses any other path even if the browser asks for it, and re-checks with rev-parse that the path really is a repository root.
Search does not filter the graph — filtering would break lane continuity, which is the whole point of the page; matches are highlighted and stepped through instead.
Lane colors are derived from DSH theme tokens with color-mix(). On a browser without color-mix(), lanes fall back to the four theme state colors in rotation.
Persisted DSH internals are a fallback only — reading storages/workspace.json and sessions/ is the last resort when the session/workspace services are unavailable.
Verified on DSH 0.1.7-rc.2 with dsh-better-sidebar 0.21.1; the native fallback path is covered by the same tests but has had less real-world use.
The npm name dsh-git-graph belongs to a different plugin by another author (enoughpower/dsh-git-graph). This one is @zeng/dsh-sidebar-git-graph and is currently distributed from GitHub only.

Development

No build step: edit lib/, restart dsh web for host-half changes, hard-refresh for browser-half changes (a changed client bundle is picked up with a reload — its rev is re-derived from the file's mtime).

npm test        # = node scripts/test.mjs: node --check on every lib file, then every test

git must be on PATH for the route test. scripts/test.mjs discovers test/*.test.cjs, so a new test file cannot be forgotten in CI (listing files in the workflow once already meant three new tests ran nowhere).

Checking the theme tokens against a real DSH install

Colour tokens are the one place where a typo fails silently: var(--typo, fallback) just falls back, so a wrong name shows up as "the colours look wrong" rather than as an error. This plugin shipped exactly that bug once — the current-branch chip used --dsw-alias-label-inverse, which does not exist, so its text inherited the theme's body colour while its background was --dsw-alias-brand-primary. In the light theme those two resolve to the same value: the chip was 1.00:1 contrast in both themes (white on white in dark mode). The fix is --dsw-alias-label-primary-foreground, the token DSH itself pairs with button-primary-fill (= brand-primary): 18.9:1 and 18.1:1.

test/style-tokens.test.cjs has a built-in snapshot of every --dsw-alias-* token and fails if the CSS references anything outside it. To re-check that snapshot against the real theme:

ZGG_THEME_FILE=<dsh>/node_modules/@deepseek-ai/dsh-client-ui-theme/lib/client.js \
  node test/style-tokens.test.cjs

Notes for implementers

Traps that cost real time here, in the order they bite:

String.split(sep, limit) discards the remainder — it does not fold it into the last element. 'a file with spaces.txt'.split(' ', 11)[10] is 'a'. Every porcelain v2 record ends with a path that may contain spaces, so indexing by field number truncates it silently. Split the fixed fields off and join the rest back.
The u (unmerged) record is not shaped like 1 or 2 — it has 11 fields (u <XY> <sub> <m1> <m2> <m3> <mW> <h1> <h2> <h3> <path>). Parsing it as a 9-field 1 record drops conflict files entirely, which is silent data loss that only appears during an actual merge. Produce a real conflict and dump the bytes instead of trusting the docs.
git diff --no-index exits 1 when the files differ, and stderr is empty in that case. Without special-casing it, untracked files can never open a diff and there is no error message to explain why.
Pass both paths for a rename. With only the new path, -M does not detect the rename and you get "new file mode" plus the whole file as additions (measured: 1 added line became 41).
Do not inspect -z output through a PowerShell pipeline. PowerShell splits on newlines and stringifies; -z output has no newlines, so the whole thing collapses into one element and the NUL boundaries become easy to misread. Use Node with encoding: 'buffer' and split on NUL.
--literal-pathspecs is mandatory. Without it a browser-supplied path is parsed as pathspec syntax, and a value like :(top)* widens "read one file" into "read the whole tree".
Nested repositories have no diff. git does not descend into a repository inside a repository, even with -uall, so the outer repository reports one directory row and --no-index on it fails with Could not access '<dir>/null'.
git diff on an unmerged path emits a combined diff — @@@ hunk headers with two-character prefixes (++, +, + ). A unified-diff reader cannot parse it and will return zero lines, i.e. an empty diff with no error. Compare such paths against HEAD instead; that is a normal unified diff and shows the file exactly as it is on disk.
A conflicted file and an untracked file share the status letter U — so the client must not decide "use --no-index" from the letter. Getting it wrong renders the whole file as additions, again with no error. The host states untracked explicitly per row.
method is not part of the old contract. Before this release the host ignored it and always answered with commit-graph data. Because the browser half reloads on every page load while the host half needs a restart, an older host is a normal state to be in — so validate the response shape, or you will render "no changes" from a payload that never mentioned any.
Do not sanitize the git environment by disabling system config. Setting GIT_CONFIG_NOSYSTEM=1 also drops core.autocrlf, which Git for Windows sets in its system config. A repository with LF in the index and CRLF in the working tree then reports every line as changed — +1 −0 became +88 −87, a one-line addition drawn as a whole-file rewrite. Setting GIT_CONFIG_SYSTEM to an empty string does the same thing (measured), so "redirect it instead" is not a safe variant either. Set neither and let git find its own config; whitelist the environment to keep credentials out. Verify by diffing your counts against git diff --numstat on the same tree.

License

MIT — see LICENSE.

中文说明

AFunDog/dsh-sidebar-git-graph