A plugin reload (or a disable/enable) tears the client fiber down while this
module's own service/store subscriptions are still live: unregistering each
built-in descriptor notifies the plugin's service, and every notification ran
`sync()` again. The `files` takeover is never a descriptor — its key is
`FILES_KIND` — so the drop loop released it every single time, and the block
below re-created it. Re-registering on the now-inactive context throws inside
`registerSlots` (`ctx.slots.inject` cannot create an effect on an inactive
context) *after* `tabs.register` already put the type into the host registry,
so that registration's disposer was lost with the exception.
The id then stayed taken for the rest of the page's life: every later attempt
was refused with
sidebarRight: tab type id "dsh-better-sidebar:files" is already registered
and the kind rendered the host's "nothing here can view this kind of content
yet" face until a page refresh.
- `sync()` skips the `files` takeover in its drop loop: the takeover is owned
by the editor-type switch, so an unrelated notification no longer puts its
host-side type through a tear-down/re-register window;
- a registration that fails after its type was registered releases the type
again (and whatever slot disposers it collected), so the same id stays
registrable and a retry can succeed;
- one descriptor that cannot register no longer aborts the whole sync, so the
remaining types are still brought up;
- teardown releases every registration even when one release throws.
Verified: tests/native-registration.spec.ts (new, 4 cases) is red 4/4 against
the unfixed code and green with the fix; tsc/eslint clean and the full unit
suite passes.
The rc.1 adapter first went out as 0.21.0-rc.1, but DSH 0.1.7-rc.1 is
upstream's release batch, so the plugin that targets it must not sit on a
prerelease channel. Re-cut it as v0.21.1: the version is now free of `-`,
so release.yml publishes it under npm's `latest` instead of `alpha`.
Version literals only — package.json, dsh.plugin.json and
SIDEBAR_SERVICE_VERSION (tests/service.spec.ts still pins the last two
together). No code change: the peer floor stays ^0.1.7-rc.1 and the
support line stays "DSH 0.1.7-rc.1 and later".
Docs follow the channel move: README/README_EN/guide drop the "alpha
channel" framing, the install command becomes @latest, and every older
DSH line now gets an explicitly pinned command in the version table
(@0.19.1 and so on) because `latest` no longer covers them. The rc.1
design record keeps its history and gains a §8 for the re-cut.
The table named a version spec per DSH line; it now carries the command a
reader can copy, one row per DSH generation, pinned to the exact version
(`@0.19.1`, not `@latest`) so a row cannot rot when `latest` moves to a
newer stable cut.
The command form is the one every era of this plugin documented itself —
`dsh plugin --profile web add dsh-better-sidebar@<spec>` — and DSH's CLI
has had `plugin --profile` since 0.1.0-rc.8, so the old rows are accurate
rather than invented.
Adds an explicit DSH-to-plugin table to the Installation section of both
READMEs, because "which version do I install" now has more than one right
answer: 0.21.0-rc.1 targets DSH 0.1.7-rc.1 and later, 0.19.1 is the cut for
0.1.6-alpha.2 and earlier, and the two 0.1.7 alphas have NO published plugin
at all (0.21.0-alpha.1 never reached npm), so those hosts must move DSH to
rc.1 first.
The peer column is included on purpose: it is a range, not an exact pin —
`^0.1.7-rc.1` already admits a later rc.2 or the stable 0.1.7, so upstream
shipping one does not require a new plugin version. Pinning the peer to an
exact 0.1.7-rc.1 would instead be actively harmful: DSH 0.1.7-rc.1's own
boot preflight silently disables any plugin row whose `@deepseek-ai/dsh*`
peers do not satisfy the running host.
`tests/primitives-exports.spec.ts` collected its file list with
`execFileSync('rg', …)`. ripgrep is installed on a developer machine but not
on the CI runners, so the whole spec failed to COLLECT there
(`spawnSync rg ENOENT`, which is why `ci` and `ci-windows` were red on this
branch while `pnpm test` was green locally). Replaced with a plain readdir
walk over `src/` and `tests/` — no external binary, and the scan is a few
hundred small files.
Verified by running the spec with a failing `rg` shim first on PATH: it
passes, which it cannot do while still calling ripgrep.
New design record `docs/plans/2026-09-23-dsh-0.1.7-rc.1-adaptation.md`
(method, the version-bump-only delta, the code-card break and the mermaid
regression it exposed, the over-yield recovery, the live 3080 upgrade with
its boot-log findings, and the evidence — including the note that the
mount lane first ran against a stale tarball because `e2e_resolve_tarball`
picks the newest existing one instead of packing).
AGENTS.md §3 gains the rc.1 rules: the boot-time peer preflight (a new
§3.4 item 12) with its four gotchas, the vendor hoist lesson, and the
code-card/toolbarLabels contract. Two claims are corrected while we are
here: the guide's four-entry description cap was attributed to rc.1 but
`GuideBody.tsx` and the host's terminal guide entry are byte-identical
across alpha.1 / alpha.2 / rc.1, and the icon surface is 25 named imports
plus a type-only `IconProps`, not 26.
README / README_EN / guide: version, badges, peer floor, the pinned-DSH
advice (`alpha` is now 0.1.7-alpha.2 while rc.1 rides `next`), the reduced
refusal list, and the correction that npm `latest` is 0.1.5-rc.3.
The alpha.1 handover refused 27 extensions so DSH's own document preview
could own them, but nine of those have NO host renderer at all: `avif` and
`ods` sit in `document/unviewable.ts`'s "known binary, no renderer" list
and get the unsupported empty state directly, while `xlsb` / `xlt` /
`xltx` / `xltm` / `ots` / `dot` / `dotx` fall through to the plain-text
body, fail the binary check and land on the same dead end.
Before that handover these files reached the plugin's `code` catch-all and
rendered the binary-download pane, so refusing them was a user-visible
regression we introduced ourselves — and rc.1 did not restore it. They are
claimed again (`fods` stays handed over: the host shows that flat XML as
plain text, which beats a download pane).
`canOpen` had zero test coverage, which is exactly why the regression went
unnoticed, so `tests/native-surface.spec.ts` now pins both directions:
every format the host renders must be refused, every format it cannot
render must be claimed. Reverting to the old list makes it fail with
`macro.xlsb: expected false to be true`.
Move the support line from 0.1.7-alpha.1 to 0.1.7-rc.1 (peer floor
`^0.1.7-rc.1`). 0.21.0-alpha.1 never reached npm, and once the code-card
labels below are fixed the same build crashes on alpha.1 (its DiffBlock
still reads the `labels.files` upstream deleted), so declaring only the
host we verify against is the honest floor — and from rc.1 on the host
enforces peers at boot anyway: a row whose `@deepseek-ai/dsh-*` peers do
not satisfy the runtime is silently disabled.
Dependencies: 28 devDependency pins and 14 peers move to 0.1.7-rc.1.
rc.1 also raised the vendored runtime (cordis 4.0.4, schemastery 3.18.4,
cordis-plugin-group 1.0.4, cordis-plugin-include 1.0.9) and publishes DSH
cross-references as exact versions, so bumping schemastery/cordis alone
left four unmet transitive peers — the three vendor packages are hoisted
into devDependencies, which is what makes `pnpm peers check` clean again.
One compile break: rc.1 added `CodeToolbarLabels` (codeLabel / wrapLabel /
unwrapLabel, all required), made DiffBlockLabels / ReadBlockLabels extend
it and deleted `DiffBlockLabels.files`. Fixing SideChatView then leaving
markdown on the legacy banner would render two different code-card styles
inside one side-chat view, so the plugin adopts the card everywhere:
`markdownTextProps()` now passes `labels.code.toolbarLabels` with the
host's own wording (codeBlock.title|wrap|unwrap, added to all 20
dictionaries; `sideChatBlockFiles` retires with the deleted footer).
Adopting the card surfaced a silent regression the suite caught: the new
toolbar shows `codeLabel` for languages the highlighter does not know, and
mermaid has no shiki grammar, so the info string no longer reaches the DOM
and the banner-text probe in mermaid.tsx matched nothing — mermaid fences
stopped being swapped for diagrams. Detection now matches the FENCE BODY
from the source that produced the render.
`tests/market-manifest.spec.ts` pins every DSH peer as a floating caret
range: an exact pin (the devDependency style) or a comparator without a
prerelease would fail rc.1's preflight and silently stop the plugin from
loading, and `pnpm peers check` cannot see that — it validates the
resolved dev tree, not the range shape the host reads.
Two things the Plugins list made obvious once the icon rendered beside the
built-ins.
The artwork was flat: one blue panel, one darker blue rail. The built-ins
each carry their own two-hue palette — the file header says so outright
("these glyphs carry their own brand colors and gradients instead of riding
currentColor") — and at the 30px a row renders at, a single-hue mark reads as
a plain block. Three candidates were rendered next to the real built-in
artwork at the real sizes (30 / 36 / 72, light and dark) before picking one: a
violet rail sat too close to the panel's deep blue to read as a second colour,
and a pale teal panel washed out. The shipped pair is blue for the panel
(gradient) and teal-green for the rail (gradient), which stays legible at row
size and matches the Subagent artwork's proven blue/green combination without
sharing its silhouette.
The description sold the plugin as someone else's sidebar. It now leads with
what it is — better — in one sentence rather than a colon and a feature list,
in both locales, and both `package.json` (the npm blurb and the list's
fallback label) and `dsh.plugin.json` say the same thing. The README never
framed it that way, so nothing there changed.
The artwork guard gained the two properties this decision rests on: the mark
must be colourised (two gradients, no currentColor) and must not use a
CSS-only paint, since the icon ships as an <img> and the search artwork's
foreignObject technique would stay empty there. Red/green verified — a flat
single-hue redraw fails it.
Verified on the live 3080 instance: the row renders the new artwork and
"更好用的右侧边栏——文件树、编辑器、文件变动、任务和侧边对话都收在这一栏里，每个会话互不干扰。"
with no VSCode framing anywhere in the copy.
0.1.7 reads `package.json#icon` plus `locale/<lang>.json` to illustrate and
label a plugin in the Plugins page and the `plugin_manager` list. The plugin
had neither, so it rendered as a bare row (and every other third-party entry
in this profile still does — ours is now the only one with a real mark).
The icon follows the two icons DSH ships for this feature
(packages/experimental/*/icon.svg): a 36x36, fill-only geometric mark with a
blue gradient and no stroked outlines or text. It reads as a panel with a
docked right rail — the sidebar this plugin is. The first cut used two blocks
separated by a gap, but the rail shared the panel's light tone at the top, so
the split leaned on a 3px gap that shrinks away at list size; a darker rail
flush inside one panel survives any downscale.
`locale/en.json` is the discovery baseline English needs to be present for the
whole directory to be read — and `locale/zh.json` carries the Chinese title
and blurb. The manifest's own `description` still advertised the terminal and
the browser, which this plugin handed to DSH in 0.20.0; it is now user-visible
as the fallback label, so it was corrected too.
Every failure in this feature is silent: bad metadata only ever degrades to a
plain text row. `tests/plugin-meta.spec.ts` therefore pins both ends — the
manifest wiring and the packed tarball — plus the house style of the artwork,
and was red/green verified (removing `icon.svg` from `files` fails it twice,
once against the real tarball).
Verified on the live 3080 instance: `readPluginMeta` resolves the icon to a
`data:image/svg+xml` URI and the page renders it (`ourIconExact: 1`, no page
errors).
The one-time import of a pre-0.1.7 `settings.yaml` section never ran. It
started from the `ctx.inject(['settings'], …)` callback, which fires the
moment the settings SERVICE appears — while the loader is still mounting
rows, so `describe()` does not list this plugin's own entry yet, the lookup
returns undefined, and the import bailed on "no form" without a trace.
Upstream's own migration waits (`ctx.root.loader.await()`); this does the
same.
That made every early return of the import invisible, which is how it went
unnoticed: the callback simply produced nothing. It now returns a named
outcome ('imported' / 'no-profile-home' / 'no-form' / 'already-configured' /
'no-legacy-section' / 'rejected') and the caller logs it — anomalies at warn,
the rest at info. The `loader` mirror's `await` is optional so a composition
without the cordis loader still reaches the import instead of aborting the
whole settings face inside the inject callback; the smoke fakes model it.
Verified on the live 3080 instance: after the upgrade the settings form
reports the user's real values back (autoOpenJobs false, workspaceFence
false, the nine-entry tabsEnabled map), where before it reported pure
defaults. The mount lane stays 7/7.
DSH 0.1.7-alpha.1 is 1299 commits past 0.1.6-alpha.2 and rewrites four
contracts this plugin was built on. The peer floor moves to
^0.1.7-alpha.1 (semver's prerelease rule means ^0.1.6-alpha.2 can never
match a 0.1.7 prerelease); 0.1.6-alpha.2 and earlier stay on v0.19.1.
Released as 0.21.0-alpha.1 (dist-tag alpha).
Hard breaks:
- dsh-settings was rewritten: register(ns, schema) is gone, a form is
addressed by the plugin row's Loader entry id and read from the
module's exported Config. The preferences are merged into Config and
marked volatile — which requires DSH's own schemastery, because only
that build wraps a volatile field in the cosmokit reference the
loader's volatile commit walks. With the public package a settings
write reports success and revision 1 while the effective value never
changes.
- ui-primitives renamed its whole icon family (IconXxx14/16 ->
IconXxxRegular/Medium): 26 imported symbols, 17 files. The failing
mode is render-time, not load-time. Five call sites that relied on the
old 14px default now pass size={14} explicitly.
- Session format v4 rejects source.kind === 'plugin'; the side-chat
boundary injection now writes plugin:dsh-better-sidebar, the same kind
DSH's own v3->v4 migration produces, so historical threads survive.
Tool results moved to role 'tool' with top-level content/isError; all
four parsers accept both shapes.
- The client session snapshot dropped subagentsByParent/jobsBySession
and the on-demand subagent-catalog observation API. The Tasks page
reads projectionsBySession and a new jobs.list route.
Yields and convergence:
- The read-only previews go to DSH's documentpreview (Office->PDF, a
spreadsheet worker, zoom, per-directory refresh); editor.canOpen
refuses those 26 extensions. Markdown, HTML (with its sandbox
switches) and the editable code buffer stay.
- The host browser kind is desktop-profile-only now, so the web e2e
lanes expect no browser entry and assert its absence.
- Link takeover claims only urlTarget links and passes the rest to the
host, fixing a regression from the 0.1.6 round: the fallback pointed
at a 'browser' type the plugin no longer registers, so http links in
its own markdown did nothing.
Also fixed, found along the way:
- The file tree refreshes live (new directory watch over the plugin's
own socket), since the plugin owns the `files` kind.
- ctx.sessions.list has no `current` field at either tag; the three
readers were dead. They now use ctx.sidebarRight.mounted.
- package.json.files was missing lib/client-locale.js.
- Two genuinely dead call sites: sessions.openSubagent/open/
refreshSubagents do not exist on the current baseline, so the
topology click, the jump-back and the retry were all no-ops.
- The test harness window shim lacked addEventListener, which the 0.1.7
primitives call at module-evaluation time; and the e2e token exchange
needs Playwright's request stack, not Node's fetch.
Adds the per-version adaptation record (docs/plans/2026-09-21-…-adaptation.md),
rewrites AGENTS.md §3 for the new baseline (the two silent contract breaks, the
terminal/browser handoff, the new official plugin-manager surface, the three
ui-primitives breaks, the widened primitives bare-import set), and updates both
READMEs and the integration guide for the v0.20.0 feature set.
Also retires the support text the terminal deletion made false: the install
flow is one step again (no build-script approval), the FAQ's two node-pty rows
collapse into one pointing at DSH's own terminal, the platform note no longer
promises prebuilt native binaries (there are no native deps left), and the
installers drop the -Repair mode and their allowBuilds writes along with
pnpm-workspace.yaml's now-empty allowBuilds block. Removes the dead
browserNoSandboxWarning locale key (its setting went with the plugin's browser)
and tightens theme.spec's capsule guard so a literal colour fails it.
The e2e resolver now refuses a PATH that is not on the baseline line:
this machine's PATH is 0.1.6-alpha.1, and running the mount lane against
it fails at 'the bundle was not registered' with a zero exit from the CLI —
a misleading dead end worth one explicit error instead.
DSH 0.1.6-alpha.2 ships a right-Sidebar terminal and a right-Sidebar browser
of its own. Both were things this plugin had built from scratch, so keeping
them meant two implementations of one surface plus a plugin that shadows a
host type the product is actively developing.
Terminal: the plugin's whole PTY stack goes — the UI-tab manager, the
agent-owned registry, the lazy node-pty loader, the eight model-facing
terminal_* tools, the xterm view and its font/link/scrollback helpers, and
the cross-session 'pinned terminal' rail that existed only to keep a terminal
alive across a session switch. The host's terminal kind becomes the only
provider, which also retires the node-pty dependency, its pnpm allowBuilds
entry, and the installer's -Repair mode (issue #140's whole class of
'node-pty failed to load' support).
Model-facing capability note: the plugin used to inject terminal_* tools
(off by default). Those are the model's only cross-call persistent terminal,
and DSH's own equivalent (@deepseek-ai/dsh-tool-terminal) is not mounted by
any shipped bundle, so removing them leaves the model with the one-shot
bash/pwsh tools only. That is the accepted cost of not maintaining a second
PTY surface; the README documents how to enable the upstream toolset.
Browser: the plugin's browser tab type is dropped, so the host's own browser
kind owns embedded pages. The DOM-level link takeover STAYS (it is the only
thing that can route an https link to the system browser by protocol, which
the host does not do) but now delegates to the host's browser tab via
ctx.sidebarRight.openTab('browser', …) with a window.open fallback, and the
two settings that only ever configured the plugin's own iframe (sandbox
escape hatch, loopback allowlist) go with it — along with the host-side
browser.probe route they fed.
Cards: the empty-pane welcome cards adopt the host's guide-capsule recipe
(0.5px l4 hairline, 24px radius, layer-1 fill, 56px floor, 14px/20px padding)
so the plugin's 'pick what opens here' surface reads as the same design as
DSH's own. A theme.spec guard pins the geometry and the token-only paint —
including that the upstream-only --dsw-alias-bg-l1/-l2 spellings (defined
nowhere, so their declarations are silently dropped upstream) never appear
here.
Also drops the dead state the removed features left behind (SidebarTab.pin,
nextTerminal, agentWaits, bottomOpenedOnce, the agent-tab helpers, the
pinned-tab memo key), re-points every test fixture that used them at
surviving equivalents, and prunes the specs of the deleted features.
The peer floor had to move: semver's prerelease rule needs a comparator on the
same [major,minor,patch] tuple, so `^0.1.5-rc.1` can never match any 0.1.6
prerelease. The floor becomes `^0.1.6-alpha.2` (which does admit 0.1.6 alpha/rc
prereleases and excludes 0.1.5), and `engines.dsh` follows. This is a deliberate
support-line break, matching how v0.19.0 dropped 0.1.2: 0.1.5-rc hosts stay on
0.19.1.
Three upstream changes actually reach this plugin and are adapted here:
- `IconSendOutline16` no longer exists in ui-primitives (only
`IconSendOutline14`).
- `TerminalBlockLabels` gained a required `noExitCode` pill text.
- `ConnectionIndicator` dropped `reconnectLabel`: the outage label is now the
one static label that already names the retry action, so the label copy moves
there, matching the host's own '连接异常，点击立即重连'.
ui-primitives still declares no `dependencies` while its bundle bare-imports
more than it used to: `diff` and `simple-icons` were previously satisfied only
transitively and had to be hoisted into devDependencies, exactly like the
anser/shiki/katex group before them. `@deepseek-ai/dsh-code-runtime` is gone
upstream (the package no longer exists) and leaves the dev tree.
The e2e fallback npx spec is pinned rather than version-less: an unpinned
`npx @deepseek-ai/dsh` resolves npm's `latest`, which is still the 0.1.5 line
this plugin no longer supports.
DSH 0.1.6-alpha.2 made two plugin-facing contracts stricter, and both broke
this plugin silently:
- `SidebarRightGuideEntry.id` is now required and unique per provider. The
native tab registrations sent guide entries without one, so the second kind
collided on `undefined` and `SidebarRightTabRegistry.register` threw
'duplicate guide entry id'. The throw happens inside `ctx.inject`'s callback
body, which cordis swallows, so the whole native surface stayed empty with no
symptom beyond a missing guide row.
- `conversation.chat.turnTail` changed from a `chain` to an additive `list`.
The chain's `select` made one entry render and let this plugin REPLACE the
host's produced-files row; a list cannot express that, and the registration
now throws 'list slot ... requires options.id'. Upstream chose the list
deliberately ('An additive list lets both plugins contribute without either
knowing the other's data or rendering'), so the plugin's row can only ever
duplicate the host's deliverables row -- which itself opens files through the
host's openFile, i.e. into this plugin's editor. The contribution is removed
and the row is the host's alone.
The file-opening half survives as `sidebar-file.ts` (the changes tab and the
editor host still use it), and `resolveSidebarPath` moves to the `paths.ts`
path-helper module beside its siblings. `registerNativeSurface` grows an
optional `reportFailure` so a future contract break reaches the visible
diagnostic strip instead of a silently empty guide, and one bad descriptor can
no longer take the remaining registrations down with it.
Copilot review: the openTab JSDoc claimed createTab descriptors ignore
title/path/id, but that is only the bottom-workbench landing (the
factory mints the tab wholesale there). On the native right Sidebar the
seed fields are forwarded as navigation params (createTab values as
defaults, seed wins) and only the id is host-minted. State both
landings so consumers do not drop path seeds.
The native openTab branch rerouted EVERY path seed into openResource (a
dsh-resource://file/... open) without looking at seed.type, so any
component tab opened with a path landed in the editor (the only
dsh-resource://file/** claimant) and its registered component never
mounted. This also contradicted the guide, whose TabDescriptor/openTab
contract and WebDocsView example document path-seed component tabs.
Route by type instead: only the editor kind (the only registration with
dsh-resource://file/** patterns) turns its path seeds into resource
opens; every other kind keeps the page open and carries the path in the
navigation params, which NativeTabRecords.ensure already merges back
onto the synthetic record's tab.path for the component. The bottom
workbench path is untouched, editor opens behave exactly as before, and
onOpen still receives the synthetic tab with the path seed.
main went red on the merge commit (run 34579312780) with a real assertion:
`tests/agent-pty.spec.ts:413` expected `skipWait` to report 1 registered wait
and got 0. The cause is the test's own preparation, not the registry: an
interactive shell ECHOES what is written to it, so a terminal opened with a
command that contains the wait needle already has that needle in its
transcript by the time `waitFor` runs — `waitFor`'s documented fast path then
resolves the wait immediately and `skipWait` finds nothing to skip.
POSIX /bin/sh does not echo a command written to a non-tty, so this had been
latent for as long as the suite existed; PowerShell does echo it, which is why
it only ever failed on Windows, and why it took a Windows run to expose it
rather than a code change.
The three affected cases (skip / waiting-snapshot / wait-notification) now open
a BARE shell and wait for the shell to be up rather than for a printed marker,
via a new `waitForShellReady` helper that watches for output of any kind. That
removes both halves of the hazard: no needle in the command, and no dependency
on which words a given shell prints at startup. Nothing about what the cases
assert changed — they are about the registry's wait bookkeeping, which needs a
live terminal and no particular output.
Verified: tests/agent-pty.spec.ts 33 passed (and the three cases now finish in
milliseconds instead of waiting out a transcript poll), full suite 127 files /
1343 passed / 9 skipped.
The hairline outline I added to the colored tab glyphs reads as a thick stroke
in the live instance, so it goes: the glyphs stay exactly as the icon set draws
them, tinted only through the wrapper's token. What the feedback actually asked
for was the terminal SIZE, and that is all that remains — it renders a step
down (0.85x, floor 10px) because the VSCodicon terminal is a wide filled
rectangle that read heavier than its neighbours.
tests/builtins.spec.ts keeps pinning both readings (tasks = layers, terminal =
the shrunk VscTerminal with no style override), so the assertion now also
guards against an outline creeping back in.
The `--maxWorkers=2` fix from the previous commit was falsified by real CI:
run 34574847774 still ended with `[vitest-pool]: Worker forks emitted error /
Worker exited unexpectedly` while reporting 126 of 127 files passed and zero
FAIL lines. The worker that died was the one running
tests/native-surface.spec.ts — it produced NO results at all, which is the
same shape as the two earlier crashes (125/126 and 123/124: N-1 files
reported plus a silent worker exit at teardown).
So the consistent signature across all four crashes is a worker-recycling
race, not a test failure. `test:windows` now runs the whole suite in ONE fork
(`--maxWorkers=1`), which removes recycling entirely; the cost is wall-clock
(14 s -> 41 s locally, and the Windows step stays well inside its 30-minute
budget) and the signal is unchanged. Retries are still deliberately not used —
they would hide a real regression instead of removing the race.
The CI comment now records both Windows failure classes with their evidence,
so the next person can tell what this flag is for.
Two glyphs read wrong in the live instance:
- Tasks shows subagent sessions AND the host's background jobs, so a checklist
glyph described a page that does not exist. `VscLayers` (sheets stacked, i.e.
work running behind the conversation) says what the page is; the amber token
stays, so it still reads as live activity.
- The terminal was the widest glyph in the set and the only bare fill in a
strip of the plugin's own 1.5px-stroked icons, which made it look like a
heavy foreign block. It now renders a step down (0.85x the strip size, floor
10px) and every colored tab glyph gets a hairline outline of its own ink
(`stroke: currentColor` + `non-scaling-stroke`, so the line stays one device
pixel across the sets' mixed 16- and 24-unit viewBoxes; `paint-order: stroke`
centres it on the fill edge instead of fattening the shape).
tests/builtins.spec.ts pins both readings by glyph identity (tasks = layers,
terminal = the shrunk, outlined VscTerminal), so a future swap has to argue
with a failing assertion rather than with a comment.
Gates: typecheck / lint green, pnpm test 127 files / 1343 passed / 9 skipped.
- docs/plans/2026-09-11-file-icons-and-tab-glyphs.md: the design record —
what the two stacked PRs did, the measured upstream facts that made their
colored implementation redundant (rc.2's `FileTypeIcon`/`CodeFileIcon` with
the host's own palette, the classifier never returning a folder, the native
tab definition having no icon field while its title slot IS the chip), the
three merge conflicts and how each was resolved, what was deleted, the
catch-all semantics change, and the verification evidence.
- docs/plans/2026-09-10-dsh-0.1.5-rc.2-adaptation.md: cross-link it from the
change table.
- docs/external-plugin-guide.md: the file-icon fallback chain now ends at the
host's artwork, with the catch-all caveat spelled out; §12 loses its "one
exemption" paragraph entirely — the plugin draws no color of its own any
more (file icons are host pixels, tab glyphs ride tokens).
- AGENTS.md §5: replaces the icon-theme exemption with the two new rules
(file-icon chain + catch-all hazard, and where the colored tab glyphs live
plus the fact that the native chip glyph is drawn by the plugin inside the
host's title slot).
- README.md / README_EN.md: the v0.19.1 entry gains the icon section
(#611 / #531 / #594) with the dataset deletion and the new public API.
Issue #611 asked for file-type icons like the official explorer's, and #429
built a colored theme for them. DSH 0.1.5-rc.2 makes most of that work
unnecessary: `@deepseek-ai/dsh-client-ui-primitives` now exports
`FileTypeIcon` / `CodeFileIcon` with the host's own 48-type full-color code
artwork plus category-colored sheet glyphs (and its classifier behind them).
So the plugin renders the host's icons instead of shipping a color table, and
color becomes the only behaviour rather than an opt-in theme.
File icons (src/client/file-icons.tsx)
- `builtinFileIcon(path, size)` → `<FileTypeIcon path size />`; folders use the
host's folder category. The plugin's own per-extension VSCodicon map is gone.
- The public `registerFileIcon` API from #531 is unchanged and still outranks
the built-ins, so an external plugin can take over any row. One semantic
consequence, now documented and tested: because the host classifier covers
every path, a registered catch-all (`exts: []`) claims rows it used to lose
to the built-in map. An unregistered catch-all (the old icon theme's shape)
is unaffected.
- Nothing is lazy any more: the artwork lives in a platform module the client
already holds, so `lib/client-file-icons.js` and its 563-entry dataset (255
kB) are deleted outright — the core bundle grew by 10 kB (0.85 → 0.86 MB).
- `src/client/file-icon-theme.ts`, the `fileIconTheme` pref (prefs-shared /
config / client prefs), the settings-page select row and the six
`settingsFileIconTheme*` keys in all 20 dictionaries are removed with it.
Tab glyphs (src/client/builtins/tab-icons.tsx)
- Every built-in tab type now declares a colored glyph: files (the host's
folder artwork), changes/diff (commit), tasks (checklist), side chat
(discussion), terminal (shell), browser (globe).
- The color comes from a `--dsw-alias-*` token through a wrapper class, never
from a literal — the skin contract (§12) holds and `tests/theme.spec.ts`
enforces both halves (no literals in the module, every color declaration in
the stylesheet resolves to a token).
Native right Sidebar chips (src/client/native/tab-adapter.tsx)
- The host's tab definition has no icon field, but the
`sidebar.right.pane.tab.title` slot IS the chip's content, so the plugin
draws the glyph itself: an editor tab with a path shows that file's icon,
every other tab shows its descriptor's icon — the same `descriptor.icon` the
workbench strip and the guide capsules already render, so the three cannot
drift. The glyph is `aria-hidden`, which keeps the chip's accessible name
exactly the title the mount lane matches on.
Gates: typecheck / lint / check:consumer-types green; pnpm test 127 files /
1342 passed / 9 skipped (3 new chip cases, 1 rewritten built-in case).
Brings in the theme plumbing around #531's API — the `fileIconTheme` pref
thread, the settings row, the 20 dictionaries' keys and the theme loader's
test — so the working tree is a complete, green starting point before the
colored implementation is replaced.
Conflicts (three; resolved by taking this line's side where it had moved on):
- src/client/service.ts: the import block keeps `baseName, extOf` (the
`isNarrowWidth` import the theme branch needed belongs to its own
fresh-session seeding, which main has since rewritten) and the feature-list
doc comment keeps main's "v0.19.0 REMOVED 'floatWindows'" note while
adopting the sharper `fileIcons` wording (matched by exts / names /
folderNames).
- tests/plugin-shape.spec.ts and tests/prefs.spec.ts: main DELETED the store
seeding tests these files used to hold and rewrote the schema assertions
(the `SIDEBAR_PREFS_DEFAULTS` destructuring), so this line's side wins —
the theme branch's re-added blocks describe behaviour main no longer has.
Next commit removes `fileIconTheme` again along with the 563-entry dataset and
its lazy chunk: DSH 0.1.5-rc.2 exports the official `FileTypeIcon` artwork, so
the plugin renders the host's icons instead of shipping its own color table.
Brings in the plugin-side file-icon feature (issue #611): the public
`registerFileIcon` API plus the resolver chain the tree, the editor tab and
the external-plugin surface all read.
Conflicts (three, all comment/list unions — resolved by union, verified green):
- src/client/service.ts: main's "v0.19.0 REMOVED 'floatWindows'" note stays
(the capability is gone and must not come back); the feature list gains
'fileIcons'.
- src/client/Sidebar.tsx: one comment above `tabIconOf`; both sides' wording
is kept (the editor-with-path special case and the descriptor fallback).
- docs/external-plugin-guide.md: the `features` list comment keeps main's
"additive only, the one exception being floatWindows" wording plus the new
'fileIcons' entry.
Notes on what this merge does NOT bring in: #594 (feat/file-icon-theme) is
stacked on this branch and lands next, but its colored-theme implementation
(563-entry react-icons dataset in a 255 kB lazy chunk plus a `fileIconTheme`
switch) is being replaced — DSH 0.1.5-rc.2 exports the official
`FileTypeIcon`/`CodeFileIcon` artwork from ui-primitives, so the plugin does
not need to ship its own color dataset.
Gates on the merge result: pnpm typecheck clean, pnpm test 127 files /
1331 passed / 9 skipped.
ci-windows is green twice on 5451f5c — the PR-triggered run 34561189320 and a
follow-up workflow_dispatch run 34561457067, both 126 files / 1318 passed /
3 skipped with zero `Unhandled Errors` and zero `Worker exited unexpectedly`
(the window that motivated this work saw two of the latter). plugin-mount is
green twice as well: `added 521 packages in 24s` with no ERESOLVE lines
against 3062 in the historical OOM run, mount smoke 7 passed, aggregate lane
passed.
Also records the falsification this PR went through — the first cut still
failed on real CI in two places (the third 5000 ms default victim,
tests/git.spec.ts:85 at 9607 ms, and `--legacy-peer-deps` breaking the CLI
install with ERR_MODULE_NOT_FOUND) — and states plainly that two green runs
do not prove the worker-death root cause, only that it stopped reproducing
once concurrency was bounded and the timeout budgets were corrected.
The first run of this PR against real Windows and real npm found two things
the local macOS pass could not:
1. ci-windows lost a THIRD file to vitest's 5000 ms default —
`tests/git.spec.ts:85` ("caps status entries at the truncation limit")
took 9607 ms building a pathological untracked set with real git calls.
Three different files (git / powershell / agent-pty), one root cause, so
per-file timeouts were whack-a-mole: vitest.config.ts now sets a global
`testTimeout: 15_000`. The two heaviest suites keep their own 30 s budgets
and a genuinely hung case still fails.
2. `--legacy-peer-deps` made the mount lane's CLI install broken instead of
robust: it skips exactly the peers a global install has to supply, and
`@deepseek-ai/cordis-plugin-group` — required at boot by
`@deepseek-ai/dsh-app-boot` — is a peer, not a dependency, so the CLI died
with ERR_MODULE_NOT_FOUND (`Cannot find package
'@deepseek-ai/cordis-plugin-group'`) before the smoke could start. The
flag is removed; the pin to a fully-published rc.2 (no ETARGET) plus the
4 GB heap (no resolver OOM) is the whole fix.
Verified locally: a plain `npm install -g @deepseek-ai/dsh@0.1.5-rc.2` into a
scratch prefix (with NODE_OPTIONS=--max-old-space-size=4096) resolves the peer
(`dsh/node_modules/@deepseek-ai/cordis-plugin-group` present), that CLI reports
0.1.5-rc.2, and `DSH_CMD` pointed at it runs the mount lane green — 7 passed.
`pnpm test` stays green with the new timeout: 126 files / 1312 passed / 9 skipped.
Docs (plan §C, AGENTS.md §2, both READMEs' v0.19.1 entries) record the rejected
middle step and why, so nobody re-adds it.
- docs/plans/2026-09-10-dsh-0.1.5-rc.2-adaptation.md: the adaptation plan —
the upstream rc.1 -> rc.2 delta per affected area, why there is no
plugin-side code change, and the CI failure taxonomy (A real Windows
timeouts, B the two silent worker deaths, C the mount lane's npm
ETARGET/OOM) with the evidence and the fix for each.
- AGENTS.md §2/§3: CI pins rc.2, the mount lane's npm flags and their reason
(floating transitives + staged prerelease publishing), the Windows lane's
bounded pool and why, and the primitives-dependencies note extended to rc.2.
- README.md / README_EN.md: support badge (v0.19.1, verified on rc.2), the
install section's stable-release note (rc.1 hosts need no upgrade — the peer
floor is unchanged) and a v0.19.1 changelog entry covering the two CI fixes.
- docs/external-plugin-guide.md: baseline line -> v0.19.1 / DSH 0.1.5-rc.1+,
verified on rc.2.
- Comment-only: `DSH 0.1.5-rc.1` -> `0.1.5-rc.1+` where the note describes a
behaviour that arrived in rc.1 and is unchanged in rc.2 (native guide
description rendering, in src/client/native/index.ts, src/client/service.ts
and the three specs that assert it).
Two independent, evidence-backed defects in the 2026-09-09/10 CI window
(60 completed runs: 26 success / 17 failure; ci-windows|Test red 8 times,
plugin-mount|Install dsh CLI red 4 times).
1. Windows: six of the eight red ci-windows runs were genuine 5000 ms
vitest timeouts in specs that really start a process —
`tests/agent-pty.spec.ts` (`:71` / `:106`; a PowerShell + ConPTY pair per
terminal) and `tests/install-powershell.spec.ts` (`:105`; a cold
`powershell.exe` start measured 12.1 s). vitest.config.ts sets no
testTimeout, so both suites ran on the 5000 ms default, and
`waitForTranscript`'s own poll budget was *also* 5000 ms — the inner wait
consumed the whole outer budget, which is a structurally guaranteed
timeout, not a flake. Both suites now declare a 30 s budget (describe-level
options, verified working on vitest 4) and the inner poll drops to 15 s so
it stays strictly below the outer bound.
2. Windows: `ci-windows`'s Test step runs `pnpm test:windows`
(`vitest run --maxWorkers=2`, the script added with the baseline commit)
instead of `pnpm test`. Five specs spawn real processes, node-pty's ConPTY
helper and a cold powershell.exe are both console-bound, and an unbounded
fork pool on a 2-core runner is the amplifier behind both the timeouts
above and the two silent worker deaths (`Worker exited unexpectedly` with
0 FAIL and every file reported passing — one of them on a commit whose
sibling run was green). Local `pnpm test` semantics are unchanged; only the
Windows lane is bounded.
The worker-death mode itself is NOT proven fixed here: it does not
reproduce on macOS and the two occurrences have no same-commit rerun. This
is the evidence-backed part; the real Windows result decides whether more
is needed, and retries are deliberately not used to paper over it.
3. plugin-mount: `npm install -g @deepseek-ai/dsh@<ver>` never really pinned
anything — the pin's transitives are floating `^` ranges, so while upstream
publishes a prerelease in stages (rc.2 went out package by package between
14:43 and 14:57 UTC on 09-10) npm resolved a mixed rc.1/rc.2 peer graph:
3062 ERESOLVE lines then `Ineffective mark-compacts near heap limit`
(exit 134), or a hard ETARGET on a sibling that appeared 5.5 minutes (resp.
25 seconds) later. The same mechanism hit the alpha line on 09-09. The lane
now pins an already-complete rc.2 and passes `--legacy-peer-deps` plus
`NODE_OPTIONS=--max-old-space-size=4096` to decouple the install from the
upstream publish window; a mismatch would not be silent because the mount
smoke in the same lane boots the real shell.
The stale comment claiming rc.1 is both `latest` and `next` is corrected —
it documented a reality that had already changed.
The pinned baseline moves to DSH 0.1.5-rc.2 (npm `next`; `latest` is still
rc.1) while the plugin's peer floor stays ^0.1.5-rc.1, so rc.1 hosts need no
DSH upgrade to run this version. Nothing in the rc.1 -> rc.2 delta touches
this plugin: of 300 changed upstream files, the overwhelming majority are
single-line `version` bumps in per-package package.json files, and the only
real code edits are the message-feedback dialog, deliverables card CSS and
the CodeFileIcon SVG artwork split (`code-file-icon-artwork.ts`). There are
zero changes under packages/api|host|session|agent and zero type-surface
changes beyond one line in message-feedback/types.ts, so the adaptation is
the pin itself — devDependencies, the CI mount lane and
SIDEBAR_SERVICE_VERSION — with no plugin-side code change and no runtime
compatibility layer.
- package.json: devDependency pins 0.1.5-rc.1 -> 0.1.5-rc.2 (including
dsh-client-locale, which is still published on the DSH tuple); peers keep
the ^0.1.5-rc.1 floor; version -> 0.19.1
- dsh.plugin.json: version -> 0.19.1; engines.dsh keeps >=0.1.5-rc.1 (same
supported range as the peer floor — the two must not fork)
- pnpm-lock.yaml: lockfile realigned to rc.2
- src/client/service.ts: SIDEBAR_SERVICE_VERSION -> 0.19.1 (tests/service.spec.ts
asserts it equals package.json's version)
Verified locally on macOS arm64 against the rc.2 tree: pnpm install clean,
`pnpm peers check` reports no peer dependency issues, typecheck / lint /
check:consumer-types green, `pnpm test` 126 files / 1312 passed / 9 skipped.
The real-mount smoke (scripts/e2e-mount.sh with DSH_CMD pinned to rc.2) and
the aggregate double-mount lane both pass — evidence recorded in
docs/plans/2026-09-10-dsh-0.1.5-rc.2-adaptation.md.
Conflict: src/client/sidebar/use-host-feeds.ts
- ours: import { mirrorAgentWaits, ... } from '../state.ts'
- theirs: same line + new import { isNarrowWidth } from '../breakpoints.ts'
Resolution keeps both imports; the two sides' semantic changes are in
disjoint regions (ours: tab-disabled wait mirroring; theirs:
activateTasksPage native-landing) and both survive intact.
ymac