dsh-gitbar-lite 0.0.0-stage → 0.1.1

package/docs/CONTRACT.md

@@ -0,0 +1,394 @@
+# dsh-gitbar-lite — implementation contract
+
+This document is the part of the design that is *not* visible from the outside:
+the wire between the two halves, where state lives, the invariants the code is
+built around, and — just as importantly — what has **not** been verified.
+
+It is written for whoever changes this plugin next, including a future me who has
+forgotten why the discard order is what it is.
+
+---
+
+## 1. What this is, and the shape of the split
+
+A GitHub Desktop-shaped Git panel that lives in DSH's built-in right sidebar.
+Three tab types, all session-scoped because that is what the right column is:
+
+| Tab | Kind | Address | Opened by |
+|---|---|---|---|
+| The panel | `gitbar` | page (`sidebar://gitbar`) | the guide card in the right column |
+| The trash | `gitbar-trash` | page (`sidebar://gitbar-trash`) | a button in the panel |
+| A diff | `gitbar-diff` | **resource** (`dsh-resource://dsh-gitbar-lite/diff/current`) | double-clicking a file, or a trash entry |
+
+The split between the halves is total: the browser half never touches the disk,
+never shells out, and never sees a path it did not get from the host.
+
+```
+right sidebar tab (browser)            host process
+──────────────────────────            ─────────────
+GitPanel / TrashBody / DiffBody
+        │
+        │  POST /api/dsh-gitbar-lite/rpc
+        │  { action, ...payload }            connection.fetch.register
+        ▼                                    → admit() (fence + cookie)
+   createRoute.fetch ─────────────────────►  → createHandlers[action]
+                                                   │
+                                                   ├─ git.js    ctx.subprocess
+                                                   ├─ repo.js   the operations
+                                                   └─ store.js  <DSH_HOME>/dsh-gitbar-lite
+```
+
+**Why one route and not one action per route.** `connection.fetch` matches on the
+exact pathname and method, and every call already carries a `sessionId` plus the
+few fields the action needs. One `POST` with an `action` field keeps the wire
+contract in one place — mirrored verbatim in the browser half — instead of
+spreading fifteen path literals across two languages.
+
+**Why it must be `connection.fetch` and never `webServer.register`.** The web
+server resolves `kind: 'exact'` routes *before* its `/api` prefix route, and
+everything that makes an `/api` request safe hangs off that prefix route:
+`connection.admit(req)` runs the Host/Origin fence and the browser-cookie check
+before `bridge()` dispatches. An exact route on the same path shadows all of it,
+turning this into an unauthenticated git surface reachable by anything that can
+guess the path. `test/index.test.mjs` and `test/check.mjs` both assert the
+carrier, and the check script proves its own assertion is sensitive by running
+the predicate against a deliberately broken copy of the source.
+
+---
+
+## 2. The wire contract
+
+`POST /api/dsh-gitbar-lite/rpc`, body `{ action, ...payload }`, `requestBody:
+'buffered'`.
+
+Every response is JSON with `cache-control: no-store`. Success is
+`{ ok: true, ...data }`. Failure is `{ ok: false, error: { code, message } }`
+**and** a meaningful HTTP status — both, on purpose, because the browser half
+reads the body and a non-2xx status must not be the only signal. The `message`
+the host sends is already user-facing (Chinese), and the browser half localises
+`code` when it recognises it, falling back to that message.
+
+### Actions
+
+| action | payload | returns |
+|---|---|---|
+| `status` | `sessionId` | the whole panel state (see below), or the picker's candidate list when nothing is bound |
+| `bind` | `sessionId`, `repoRoot: string \| null` | `{ repoRoot, bound }` or `{ cleared: true }` |
+| `repos` | `sessionId` | `{ candidates, scanned, scanTruncated, repoRoot, source }` |
+| `stage` | `sessionId`, `paths: string[]`, `staged: boolean` | fresh status |
+| `commit` | `sessionId`, `summary`, `description?` | `{ commit, ...status }` |
+| `commitAndPush` | same | `{ commit, push: { ok, ... } , ...status }` |
+| `push` | `sessionId` | `{ action: 'push', outcome, ...status }` |
+| `fetch` | `sessionId` | `{ action: 'fetch', outcome, ...status }` |
+| `discard` | `sessionId`, `paths: string[] \| null`, `hardDelete` | `{ discard, ...status }` |
+| `fileDiff` | `sessionId`, `path` | `{ diff }` |
+| `trashList` | `sessionId`, `all?` | `{ entries, total, showAll, repoRoot }` |
+| `trashDiff` | `id` | `{ entry, text }` |
+| `trashDelete` | `id` | `{ removed }` |
+| `trashClear` | `sessionId`, `all?` | `{ removed }` |
+| `prefs` | `sessionId`, `trashHardDelete?`, `trashShowAllRepos?` | `{ prefs }` |
+
+`paths: null` on `discard` means *every changed file*; an explicit array means
+exactly those paths. An array that matches nothing is an `E_BAD_REQUEST`, and a
+clean working tree is refused too, so "discard nothing" never writes an empty
+archive.
+
+### The status payload
+
+```
+{
+  workspacePath, workspaceTitle, source: 'binding' | 'detected' | 'none',
+  repoRoot, repoName, head: { oid, short, subject },
+  branch: { oid, head, upstream, ahead, behind, detached, unborn },
+  remotes: string[],
+  files: [{ path, origPath?, index, worktree, kind, untracked, conflicted,
+            staged, partial, added, deleted, binary, synthesized? }],
+  counts: { files, staged, partial, conflicted, additions, deletions },
+  primary: 'push' | 'publish' | 'fetch',
+  prefs: { trashHardDelete, trashShowAllRepos }
+}
+```
+
+`head` is the commit a discarded diff would be taken *against*, which is why the
+trash can say "against `a1b2c3d` *init*" and mean it.
+
+### Failure codes
+
+`E_BAD_REQUEST` · `E_UNKNOWN_ACTION` · `E_NO_WORKSPACE` · `E_NO_REPO` ·
+`E_NO_GIT` · `E_EMPTY_SUMMARY` · `E_NOTHING_STAGED` · `E_CONFLICT` ·
+`E_GIT_FAILED` · `E_NO_REMOTE` · `E_NO_UPSTREAM` · `E_AUTH` · `E_PUSH_REJECTED` ·
+`E_TRASH_NOT_FOUND` · `E_INTERNAL`.
+
+`E_AUTH` and `E_PUSH_REJECTED` exist so the panel can say the one useful
+sentence for each instead of pasting git's stderr. Both are derived by pattern
+matching git's output in `classifyGitFailure`, which reads **stdout as well as
+stderr** — `git commit` with nothing staged exits 1 with an empty stderr and
+"nothing to commit" on stdout, and a stderr-only matcher reports that as an
+unexplained internal failure.
+
+---
+
+## 3. Where state lives
+
+`<DSH_HOME>/dsh-gitbar-lite/`, falling back to `~/.dsh`. **Do not read
+`$DSH_HOME` alone**: the Host process usually does not have it set (the variable
+belongs to the shell DSH spawns), so a state file written under a bare
+`process.env.DSH_HOME` lands somewhere else or fails to resolve at all. This is
+the same trap `dsh-tailscale-remote` documents.
+
+```
+bindings.json   { version, bindings: { <canonical workspace path>: { repoRoot, workspacePath, updatedAt } } }
+prefs.json      { version, trashHardDelete, trashShowAllRepos }
+trash/index.json  { version, entries: [ ... ] }
+trash/<id>.diff   the archived diff, one file per discard operation
+```
+
+Every write goes through a temp file plus a rename, so a crash mid-write cannot
+truncate a file, and every read-modify-write is serialized per path — two tabs
+issuing overlapping mutations is the normal case here, not an edge case.
+
+**Why not the settings service or a storage domain.** A binding is keyed by an
+arbitrary absolute workspace path, and a `storageDomain` record key is restricted
+to `^[a-zA-Z0-9_-]+$` because it becomes a path segment; `C:\Users\…\<a folder
+with spaces>` cannot be one without hashing it, at which point the "native"
+option is doing less than the twelve lines here. A settings `Config` would put a
+growing map of machine paths into the user-visible settings form, which is worse.
+
+**The binding key** is canonical, separator-normalized, and case-folded on
+Windows (`workspaceKey`), so `C:\Foo\`, `C:\Foo` and `c:\foo` are one record. The
+original spelling is stored alongside for display.
+
+**A stale binding is dropped, not reported.** If the bound directory is gone, or
+is no longer a repository, `resolveBinding` clears the record and re-detects. A
+broken binding the user cannot clear from the UI is worse than re-deciding.
+
+---
+
+## 4. The invariant: a tick *is* the index
+
+This is the claim that makes the plugin safe next to a terminal, and the claim a
+pathspec-limited commit would quietly break.
+
+- Ticking a file runs `git add -A -- :(literal)<path>`.
+- Unticking runs `git restore --staged -- :(literal)<path>`.
+- A file reads as ticked when **all** of its change is in the index
+  (`index !== '.'`), and as *indeterminate* when the index and the worktree are
+  both dirty (`MM`) — one tick cannot represent half a file, so the box shows a
+  dash and either click restates the file as a whole.
+- Commit runs a plain `git commit -F -` with **no pathspec**. What the panel
+  shows, what `git status` shows, and what gets committed cannot drift apart.
+- Untracked and `.gitignore`d files are never ticked: they are not in the index
+  and cannot be.
+
+`:(literal)` matters: without it, discarding a file named `docs` would also reach
+`docs/anything`, and a file named `a[1].txt` would be treated as a pattern.
+
+The two user-visible consequences, stated plainly: the plugin *does* mutate your
+index (that is the point — GitHub Desktop does the same), and a file you staged
+by hand shows up ticked, because it is.
+
+---
+
+## 5. Discarding, in order, and why
+
+```
+1. read the full status                      ── decide the target set
+2. refuse if any target is conflicted        ── E_CONFLICT, one sentence
+3. build the archive                         ── BEFORE anything is destroyed
+     tracked:   git diff --no-color --find-renames HEAD -- <literal paths>
+                (+ git diff --numstat … in the same round, for the per-file table)
+     untracked: synthesizeNewFileDiff per path, from the file's own bytes
+4. cap it on a line boundary                 ── limitDiff, records droppedLines
+5. write the trash entry + blob              ── unless hardDelete
+6. restore
+     tracked:   git restore --source=HEAD --staged --worktree -- <literal paths>
+     untracked: delete the file, then prune directories the discard emptied
+```
+
+Steps 3 and 6 are not interchangeable. `git restore` destroys the original, so
+the archive has to exist first; that is the whole reason the plugin has a trash.
+
+**Why tracked and untracked are restored by different means.** A mixed
+pathspec — one untracked path alongside tracked ones — makes the *entire*
+`git restore` fail with `pathspec … did not match any file(s) known to git` and
+restores nothing (verified on git 2.54). Splitting them is not tidiness; it is
+correctness.
+
+**Why untracked paths are deleted rather than restored.** They are not in HEAD,
+so `git restore --source=HEAD` cannot reach them. Their content is captured in
+step 3 or lost, which is exactly why the synthesized diff exists.
+
+**One entry per discard operation, not one per file.** Discarding a single file
+archives one blob with one file in it; "Discard all" archives **one blob holding
+every file's diff**, with `scope: 'all'`. That is what was asked for, and it also
+means a whole-tree discard is one write rather than N.
+
+**Conflicts are refused, not resolved.** Discarding a conflicted path would mean
+`git restore` against HEAD in the middle of a merge, which can disturb the merge
+state. The panel says one sentence instead, and `test/e2e.mjs` S12 proves both
+the single-file and the discard-all paths refuse.
+
+**`hardDelete` is a per-operation choice, remembered.** The confirmation box's
+"delete for good" starts unticked for a first-time user (the safe answer keeps a
+copy) and is persisted after each discard, so it mirrors the user's last
+decision rather than latching.
+
+---
+
+## 6. The trash
+
+An entry records, so that a diff can still be read months later:
+
+```
+{ id, at, repoRoot, repoName, workspacePath, sessionId,
+  scope: 'file' | 'all',
+  branch: { head, detached },
+  base: { oid, short, subject } | null,     ← the commit the diff is against
+  files: [{ path, origPath?, kind, added, deleted, binary, synthesized?, oversized? }],
+  bytes, truncated, droppedLines, blob }
+```
+
+`base` is HEAD at the moment of the discard. That is the answer to "which version
+is this a diff against" — and it is recorded rather than inferred, because HEAD
+moves.
+
+The list is filtered to the currently bound repository by default; `total` still
+reports the whole index, and a toggle shows everything. Clearing with the filter
+on clears only the visible repository's entries.
+
+Reading an entry whose `.diff` was deleted by hand still lists it: the metadata
+is worth keeping even when the blob is gone, and the read reports the blob as
+missing rather than pretending the entry vanished.
+
+---
+
+## 7. The right-sidebar tab identities
+
+```
+id:   dsh-gitbar-lite/git     kind: gitbar         page   keepMounted, one guide card (order 30)
+id:   dsh-gitbar-lite/trash   kind: gitbar-trash   page
+id:   dsh-gitbar-lite/diff    kind: gitbar-diff    resource, patterns: dsh-resource://dsh-gitbar-lite/diff/**
+```
+
+`definition.id` is **also the slot key** the body and title register under. Two
+consequences the code leans on: a rename in one place without the other silently
+renders nothing, and `test/client.test.mjs` asserts every slot key is one of the
+three ids.
+
+**Why the diff tab is a resource with a fixed address.** Re-opening a resource
+address reveals the tab that is already open — the host looks the content id up
+across the whole layout before planning a new tab — and the `params` payload
+re-navigates that tab (each open bumps `navigation.revision`). A page kind would
+reuse only within the active dock pane, so a tab the user had floated would get a
+second copy. The fixed address *is* the "one diff tab, ever" requirement, and the
+diff's own chip is a title seat because it has to read the live
+`navigation.params` to name the file currently on screen — a title captured at
+open time never changes.
+
+**Why `keepMounted` on the panel and not the trash.** The panel holds the tick
+state that mirrors the index and a half-written Summary; peeking at a diff must
+not lose either. The trash holds only what it re-reads on every visit, so
+retaining it would buy nothing.
+
+**The panel re-reads on `tab.visible`.** Coming back to the panel refreshes
+quietly, because the working tree may have changed under it — an agent in the
+same session is a normal author of those changes.
+
+---
+
+## 8. Configuration
+
+Every field is volatile, which is what lets the settings page edit them without a
+restart — and is why `cordis.patch.yml` deliberately carries **no** `config`
+block: values written into the profile's own patch layer must not be shadowed by
+this package on the next upgrade.
+
+| field | default | why |
+|---|---|---|
+| `gitPath` | `''` | Empty means "whatever is on PATH". Set it when git is somewhere unusual. |
+| `stateDir` | `''` | Empty means `<DSH_HOME>/dsh-gitbar-lite`. |
+| `timeoutMs` | `30000` | Per-command. A push to a slow remote may need more. |
+| `maxDiffBytes` | `524288` | Ceiling on one archived diff. A discarding user is having a bad day already; a 200 MB generated file must not become a 200 MB write that then fails. |
+| `maxFileBytes` | `2097152` | Larger untracked files are listed but their content is not read, archived, or diffed. |
+| `trashLimit` | `200` | Entries kept; the oldest are dropped and their blobs deleted. |
+| `scanDepth` | `3` | Directory depth the picker scans below the workspace root. |
+
+---
+
+## 9. Verified, and not
+
+### Verified
+
+| what | how |
+|---|---|
+| porcelain v2 parsing, including a rename, a staged delete, a conflict, a non-UTF-8-hostile path, and a path containing a space | `test/parse.test.mjs`, 47 cases; the conflict fixture pins the nine-field `u` record |
+| diff assembly, truncation on line boundaries, per-file splitting | `test/diff.test.mjs` |
+| state files, the binding key, corruption tolerance, write serialization | `test/store.test.mjs` |
+| the route carrier, single registration under a hostile service order, error classification | `test/index.test.mjs` |
+| every registration identity, the wire literals, the dictionaries, and the hook-free components' rendered trees | `test/client.test.mjs` |
+| the whole host stack against **real git**, mutation by mutation | `test/e2e.mjs`, 230 assertions over 19 sections that build their own fixtures |
+| the manifest, the patch layer, the tarball's contents, the two copies of the wire contract, the dictionaries, the workflow | `test/check.mjs` |
+| that the client bundle would not abort the Host boot | `scripts/check-client-bundle.mjs` |
+
+The e2e suite is deliberately hostile about §4: after ticking one file it reads
+`git status --porcelain` back and asserts git agrees, which is the only way to
+show the index claim rather than assume it.
+
+### Not verified
+
+- **`ctx.subprocess` is doubled in the tests.** The shipped service lives inside
+  DSH's own bundle and is not importable from a plugin repository, so
+  `test/helpers/subprocess.mjs` reproduces its documented shapes (the spawn spec,
+  `collected.stdout.readFrom`, `done`) on top of `node:child_process`. It does
+  **not** reproduce the Windows Job containment or the environment scrubbing —
+  those belong to the real service. The first live run in a real DSH is what
+  closes that gap; see `test/live.mjs`.
+- **No headless-DOM render test.** The panel, the trash and the diff are driven
+  by hooks, and the three components with no hooks in them (the row, the
+  confirmation, the picker) *are* asserted on their element trees. What is not
+  covered is real layout: whether the panel actually fills the column, whether a
+  diff row's background is the token it claims. A browser harness with a
+  hand-written React shim would be needed, and the two bugs that motivate one in
+  other plugins in this family were both DOM-measurement bugs — this plugin
+  measures nothing and augments no host DOM, so the harness was judged not worth
+  its fixture-fidelity cost. That judgement is worth revisiting if the UI grows
+  anything hover-dependent or portalled.
+- **What a 401 from `/api` does and does not prove.** `connection.admit()` — the
+  Host/Origin fence and the browser-cookie check — runs on the `/api` prefix
+  route *before* dispatch, so an unauthenticated request to a path nobody serves
+  answers 401 as well. `test/live.mjs` therefore reports 401 as *inconclusive*
+  rather than as "mounted"; only a 200 (or, when authenticated, a 404) is
+  evidence about whether this plugin is actually registered.
+- **`openTabIn` / `openResourceIn` are outside the published type contract.**
+  They are prototype methods on the sidebar-right controller and are called by
+  shipped plugins (including `dsh-deepseek-webchat`), but three source comments
+  label them "not part of `ISidebarRight`". This is the first thing to re-check
+  on a DSH upgrade.
+- **`@deepseek-ai/dsh-client-ui-dockkit` was not available to read**, so the tab
+  strip's height and the pane's padding could not be confirmed. The styles
+  therefore add no top padding and claim only `height: 100%; min-height: 0`, so
+  they fill whatever the kit provides instead of guessing.
+- **Git's path encoding on this machine.** `core.quotepath=false` plus `-z`
+  produced UTF-8 paths in every fixture that was checked (including a non-ASCII
+  name), and the parsers split on NUL rather than on quoting. A repository whose
+  paths git emits in another encoding is untested.
+
+---
+
+## 10. Known limitations
+
+- One repository per workspace, one workspace per binding. No submodules, no
+  worktrees, no multiple remotes beyond "prefer `origin`, else the first".
+- Push and Fetch always act on a single remote. There is no branch switching, no
+  history list, and no stash — that was the agreed scope.
+- `git fetch --prune` is what Fetch does, matching GitHub Desktop. It deletes
+  remote-tracking refs whose branch the remote no longer has.
+- Commit hooks run. A hook that blocks or hangs is the user's own hook; the child
+  is killed at `graceMs` only if the command times out, and `timeoutMs` is the
+  knob for that.
+- The trash keeps the diffs, not the files. Recovery means applying the archived
+  diff by hand; the plugin does not offer a restore button, by agreement.
+- Binary content is not archived — the diff says `Binary files … differ`, which
+  is honest but is not a backup.
+- An untracked file over `maxFileBytes` is deleted with only a stub in the
+  archive. The confirmation box does not yet warn about that specifically.

package/icon.svg

@@ -0,0 +1,15 @@
+<svg width="24" height="24" viewBox="0 0 16 16" fill="none" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Git bar">
+  <title>Git bar</title>
+  <!--
+    The branch mark the panel itself draws, at the plugin-list size.
+    Same 16-grid geometry as the host's IconBranchOutline artwork, so the icon in
+    the plugin list and the icon on the tab chip are recognisably one mark;
+    stroke-width is raised because this one is rendered at 24px rather than 16.
+  -->
+  <g stroke="#4F6BED" stroke-width="1.35" stroke-linecap="round" stroke-linejoin="round">
+    <path d="M1.01503 8.0001L5.6964 8.0001C6.41913 8.0001 6.78049 8.0001 7.12115 7.91951C7.4232 7.84804 7.71233 7.73014 7.97821 7.57C8.27809 7.38939 8.5364 7.13669 9.05303 6.63129L11.3281 4.40564" />
+    <path d="M1.01221 7.9999L5.6964 7.9999C6.41913 7.9999 6.78049 7.9999 7.12115 8.08049C7.4232 8.15196 7.71233 8.26986 7.97821 8.43C8.27809 8.61061 8.5364 8.86331 9.05303 9.36871L11.3281 11.5944" />
+    <circle cx="12.4502" cy="3.3079" r="1.56962" />
+    <circle cx="12.4502" cy="12.6921" r="1.56962" />
+  </g>
+</svg>