Three preview (render) bugs where DSH's chat-optimized markdown primitives
leak into the file-preview scroll world:
- CodeBlock banner: DSH's bannerWrap is position:sticky (top:0, z-index:6) so a
block keeps its copy action reachable inside a per-message scroll container.
In the preview every fence shares one .editorMd scroll container, so the
banner pins to the viewport top until the whole block scrolls past and reads
as a floating header. Neutralize stickiness for non-mermaid fences in the
preview (mermaid chrome replaces the banner and is excluded via
data-mermaid-processed).
- TOC popover z-index: .tocBar was z-index:3, below the code banner's 6, so the
outline panel painted beneath a sticky fence header. Raise the bar to 7.
- TOC dismissal: only the toggle re-click closed the popover. Add a document
pointerdown listener that closes on an outside press (blank preview space),
exempting the button and panel so the toggle still flips.
Adds md-toc unit coverage for outside-click dismiss and toggle-preserving
behavior; records the deviations in the TOC design doc.
shift fix) — but as a BARE declaration. Engines without clip support
(Safari < 16, older Chromium) drop an unsupported declaration at parse
time, which leaves the host `overflow: visible`: no clipping at all,
and the historical "collapsed panels extend the document's scrollable
area, whole page draggable" regression returns on those engines.
Declare `overflow: hidden` BEFORE `overflow: clip`: clip-aware engines
apply the later clip (identical behavior to #453), while old engines
fall back to hidden and keep clipping (with its pre-clip
scrollability — the same state they had before #453, strictly better
than visible). The cascade order is the contract.
Also extends the rule comment (the fix guards the whole
scroll-into-view class — focus moving into off-viewport regions,
nested iframes/workbenches claiming focus while loading — not just the
reveal case) and adds tests/panel-host-css.spec.ts guarding the
cascade (hidden first, clip last), the fixed viewport layer, and that
the degraded rule never reintroduces overflow. Contract documented in
the integration guide §12 (root anchor bullet).
- pins: CI mount-gate CLI, dsh.plugin.json engines floor, and the
@deepseek-ai/* peer floor / devDependencies baseline move to
0.1.2-alpha.5 (dsh-client-locale trails at 0.1.2-alpha.3 — no such
release upstream; ^0.1.2-alpha.3 already accepts the alpha.5 runtime)
- code: migrate the alpha.4 Session.events removal to snapshotEvents()
(sidechat transcript live reads, fork inheritance, jobs.output replay,
subagent activity) and drop the seedLength meta field the host removed
from CreateAgentOptions.meta
- gates: pnpm peers check clean, typecheck + 1137 unit tests green,
real-host mount smoke 14/14 on npm-published 0.1.2-alpha.5
getEntriesByType widens to PerformanceEntry; transferSize/
decodedBodySize live on PerformanceResourceTiming (the e2e specs sit
outside tsconfig.build.json, so the bundler never saw this).
The comparison table (3-run medians on the same machine against a real
0.1.2-alpha.2 host): mount latency 377→273ms (-28%), the host's
aggregated client bundle ships 203KB less over the wire (-14%,
matching the 19 dictionaries' gzip weight) and 640KB less decoded
(their source weight), boot settings.get 4→2, the bottom-drag width
leak 504px→0. Frame pacing was already at 8/10ms p50/p95 — the drag
fix removed the host-layout snap, not a frame drop. Structural wins
that don't show in a single sweep (git spawns per poll tick, fold
passes per poll, transcript row re-renders, per-keystroke doc
stringification) are listed with their per-event arithmetic.
The bottom-push rule anchored on #root :has(> [data-slot=conversation]):
a :has() match cache invalidates on every #root subtree mutation, and
the streaming chat mutates that subtree constantly — every layout-push
frame (the bottom drag writes --dsh-sidebar-height per frame) paid a
whole-document selector re-evaluation. The shell's locator (which
already holds the measured column node for the ResizeObserver) now
tags it with [data-dsh-center-col]; the rule evaluates once per node.
The tag retargets with the ref on boot swap / HMR node replacement and
is removed on unmount; the drag lane pins exactly-one-tagged-node and
its identity as the conversation slot wrapper's parent. The dead
[data-pane] branch (renamed away in alpha hosts) is gone.
TextEditor: the CodeMirror updateListener re-stringified the WHOLE
document into React state on every keystroke (O(docLength) per key,
re-rendering the shell per key). The listener now only flips the dirty
flag; the preview's draft is snapshotted from the live view when
entering preview (or on a content swap), and save() already reads the
live document.
The core client bundle statically imported every third-language
dictionary (ja/de/fr/…/zh-HK/MO/TW, ~640KB of source — roughly half
the bundle's weight) for a code path most sessions never enter: the
sidebar's own t() only consults zh/en plus the better-locale override
store, and the dicts' ONLY consumer is that store's register() call.
They now ride the existing lazy-chunk machinery as lib/client-locale.js
(fetched from /sidebar/bundle only when better-locale is actually
installed; a generation counter keeps the async registration moot when
the integration effect re-runs or disposes mid-flight). Until the chunk
lands, t() keeps the zh/en chain; the store's own revision bump on
register re-renders the chrome once the dicts arrive.
Core client.js: 1.46MB → 822KB (-44%). The chunk-boundary key-set
check moves to the existing runtime test (every shipped dict stays
key-set-equal to zh).
SideChat transcript: every 2s poll rebuilt ALL row objects, so every
row (markdown parse + DOMPurify included) re-rendered even when the
log had not changed. transcriptRows now takes the previous poll's
rows and re-adopts the position-aligned rows whose content is
identical (reuseRows) — only the changed tail rebuilds. The view
feeds the memoized array back through a ref; unit tests pin the
identity contract (unchanged rows keep object identity, changed rows
and everything past the change rebuild, tool-result landings never
re-adopt).
FileTree: expanded/revealed membership was an includes() scan per
rendered row (O(rows × expanded) per render on deep trees); one Set
per render keeps it O(1) per row.
Shell: buildNewTabOptions was inlined in the two Workbench JSX
sites — two fresh arrays per render (drag state, resizes) re-rendered
every LeafView's + affordance. One shared memo now.
DiffPane resize / TerminalView ResizeObserver: pointer and resize
streams fired a setState / glyph-measuring fit() per event; both ride
per-frame batching now (frame-batcher in the core, a local rAF gate
in the lazy terminal chunk). SessionLens sizes each row's content
with a Blob (a UTF-8 encode of the whole edit) — that moved from
per-row-per-render into one memoized pass per changed fold.
Boot: loadPrefs and loadExternalDisable each fetched the settings
document, serially, before the first paint — and the second had no
timeout, so a stalled wire kept the sidebar unmounted forever. Both
decisions now ride ONE fetch (loadBootDecision) under the existing
2s race.
GitLens: the 2s silent poll re-listed worktrees on every tick (a
second git process per tick for a list that almost never changes, and
a fresh worktrees array re-render each time). Silent ticks now fetch
only the selected checkout's status; every 15th tick (~30s) re-lists
and re-runs auto-selection, so an agent-created linked checkout is
still picked up mid-session.
ChangesTab: extractFileOps parses every accumulated tool/call (up to
4000 events); the poll ran it once for the badge and the render ran
it AGAIN per tick, and the fresh array defeated the downstream memo.
The fold now runs once per poll and its result rides a ref shared by
badge and render (reset on session switch).
A measurement harness, not a pass/fail gate (beyond the mount sanity
it shares with the other lanes): an init-script probe samples the
navigationStart → host-attach latency, a longtask collector runs
through the full built-in tab sweep, sidebar-shaped resource entries
report transfer/decoded/duration, and a frame sampler records the
bottom-strip drag's pacing (p50/p95/max) plus the width-leak guard.
Results print as single-line PERF_JSON records so a shell run can tee
and grep them; before/after comparisons land in
docs/plans/2026-08-31-perf-optimization.md. Baseline (main @ d70db8f,
3 runs): mount ≈ 377ms, 2 longtasks ≈ 161ms, drag p95 = 10ms,
width leak = 504px.
The bottom drag applied the panel's persisted width preference to
--dsh-sidebar-width regardless of panelOpen (only the idle push effect
gated it through layoutPushSize). With the right panel closed, the
first drag frame jumped the variable 0 → ~400px: #root lost 400px
instantly, the host frame's viewport dropped across its 1024px
auto-collapse breakpoint, and the NATIVE left sidebar snapped to its
56px rail — then snapped back on release. The bottom strip is the only
drag reachable with the panel closed, which is why only it showed.
applyDrag now rides the same gate as the push effect for the PUSHED
width; the panel-side DOM writes (panelRef/bottomRef/lastDragSize)
keep the raw width — the bottom panel's right edge derives from it.
The e2e regression pins the pose: panel closed + bottom open, drag the
strip, and every sampled frame must keep the pushed width at 0, the
host grid template, the left column width and #root's right edge
unchanged. Verified red on the old code (leaked 504px).
Replace the pnpm-workspace.yaml overrides with the plain-repo pattern:
install every peer the dev tree touches (dsh-attachment, dsh-code-runtime,
dsh-scope, dsh-session-projection, dsh-system-prompt, dsh-user-approval,
dsh-util-time) as direct devDependencies pinned to the baseline, same
as the thirteen already there. Root-level hoisting makes pnpm resolve
the peers natively (peers check: zero unmet, no global forcing), and
the list extends the existing bump workflow instead of introducing a
new mechanism.
Verified: typecheck clean, 107 test files / 1137 tests green, and the
real-host mount smoke passes 14/14 against @deepseek-ai/dsh@0.1.2-alpha.3.
Upstream declares its dsh-* siblings (dsh-attachment included) as
peerDependencies provided by the host bundle tree, so the host-side
mount is unaffected. The stale copies only surface in this repo's dev
tree, which installs a subset of the peers directly — making the
overrides a standing shim to extend on each baseline bump, not a
temporary patch awaiting an upstream fix.
Point-by-point audit of all 117 commits between alpha.2 and alpha.3:
every host contract this plugin relies on (token auth, slash RPC,
MarkdownText labels, the event-stream/persistence APIs behind
sidechat.events, SettingsNamespaceInput, SUBAGENT_DESCRIPTOR_VERSION
(still 3), dsh-client-store, the profile loader, the node-pty pin) is
unchanged, so no plugin code needed rewiring. alpha.3's breaking
changes (required BeginSubmissionInput.mode, the renamed
attachment-invalid subagent error, the removed SQLite persistence
backend, the identity-gated projection change feed) do not touch this
plugin.
What did need adapting was the dependency tree: the alpha.3 npm
artifacts ship with workspace-era metadata that does not match their
published code —
- dsh-subagent@0.1.2-alpha.3 imports `admitPromptContent` from
@deepseek-ai/dsh-attachment without declaring that dependency, so
resolution floats to a stale 0.1.1-rc.1 copy that lacks the export
(broke 5 test files at load time);
- six transitive peers (dsh-scope, dsh-session-projection,
dsh-system-prompt, dsh-user-approval, dsh-util-time, dsh-code-runtime)
drifted to pre-alpha.3 versions against the ^0.1.2-alpha.3 ranges the
alpha.3 packages declare;
- the alpha.3 line peers on @deepseek-ai/cordis ^4.0.2 (was ^4.0.1).
Fix: pin the seven stragglers via pnpm-workspace.yaml overrides (with
a note to drop them once upstream fixes its metadata), narrow the
cordis peer to ^4.0.2, and bump the baseline pins — CI mount gate,
dsh.plugin.json engines, peer/devDependencies, README badges, plus
SIDEBAR_SERVICE_VERSION in lockstep.
Verified on the npm-published @deepseek-ai/dsh@0.1.2-alpha.3:
typecheck clean, 107 test files / 1137 tests green, `pnpm peers check`
reports zero unmet peers, and the real-host mount smoke passes 14/14.
The main conversation's tool row head is a 16px leading slot with the
tool-kind glyph at 14 (GenericToolCard's variant table), swapped for an
error StateDot on failed rows. Mirror it: bash/pwsh -> IconApiOutline14,
read/web_fetch -> IconBrowseOutline16, edit/write -> IconEditOutline16,
grep/glob/web_search -> IconSearchOutline16, anything else ->
IconSparkle16, all from the ui-primitives module seed; CollapsibleRow
gains the matching icon slot. Thinking/injection rows keep their bare
chrome (the host's equivalents are a different row family).
Rendering-layer upgrade for the Side Chat transcript (data already rides
the raw events; the mapping just stopped discarding it):
- Tool rows with a derivable structure now render the host's
ui-primitives atoms instead of a generic double <pre>: bash calls map
onto TerminalBlock (command from call args, output + exit pill
recovered from the result's trailing exit marker), edit/write map onto
DiffBlock (call-time literal replacement, refined by the result meta's
contextual hunks), read maps onto ReadBlock (line-window meta, same
semantic narrowing the host applies). Every narrowing declines to the
generic row on malformed input.
- turn/end now emits a turnSummary row: token usage aggregated from
assistant/message.usage (output summed across steps, input from the
last request) plus the envelope-time wall duration, printed with the
main conversation's formatTokens/formatDurationMs cadence.
- The client entry injects 'connection' (DSH 0.1.2-alpha.2+) and the
thread view shows the host's ConnectionIndicator while the Remote
transport is down; the reconnect control calls reconnect() and a
disconnected -> connected transition pulls the transcript immediately
instead of waiting for the next poll tick.
19 new locale keys propagated across zh/en and all 19 third-language
dictionaries; unit coverage in sidechat-transcript.spec and a new
sidechat-view render spec (jsdom). Real-host mount smoke: 14/14.
Guochao