Move the plugin onto the DSH 0.1.5-alpha.1 baseline (peer floor
^0.1.5-alpha.1, CI pinned to the alpha dist-tag) and fix the two host
contract changes that alpha line introduced.
Baseline:
- all @deepseek-ai/dsh-* peer/dev deps -> 0.1.5-alpha.1; optional peer
@deepseek-ai/dsh-client-ui-sidebar-right (declared, unused until the
native-surface phase); devDep @deepseek-ai/dsh-agent-loop (inbox fold)
- dsh.plugin.json engines >=0.1.5-alpha.1; version 0.19.0-alpha.0;
SIDEBAR_SERVICE_VERSION follows
Live assistant stream (assistant/chunk is gone):
- 0.1.5 publishes in-flight model deltas as process-local
agent/assistant-stream frames instead of durable assistant/chunk events
and embeds the settled stream in assistant/message / assistant/attempt.
src/assistant-live.ts folds the frames into a bounded per-session buffer;
sidechat.events serves the current attempt as `live` rows (replaced every
poll, never a delta) and the transcript renders them; the inherited
in-progress snapshot reads message content / attempt stream / live frames.
Cold session reads (persistence.inspect is gone):
- 0.1.5 reads a stored session through an explicit read handle
(open(id,'read') -> handle.read() -> close()). src/session-store.ts owns
that protocol; every cold read (sidechat transcript, cold-resume preset,
session.cwd, changes replay) goes through it.
Dead code and docs:
- drop the remote.session.openWorkspacePath interception: 0.1.5's chat opens
files through ctx.sidebarRight.openResource instead, so the wrapper had no
caller and would hijack the host's "open in app" gesture; the
interceptOpenPath pref, its settings row and its locale keys go with it
- AGENTS.md §3 + external-plugin-guide.md describe the new baseline
- vitest setup bridges Node's localStorage accessor to jsdom's store (the
pre-existing 26 failures on Node >= 22)
- desktop-layout e2e seeds through host RPC and types into the Lexical
composer (0.1.5 replaced the textarea) instead of driving the directory
picker
Verified: pnpm typecheck / lint / build / test (124 files, 1327 tests) /
check:consumer-types, plus pnpm test:mount against a real
@deepseek-ai/dsh@0.1.5-alpha.1 host (18/18).
Path layer: patterns now match with a right boundary (must not continue
into [a-z0-9]) and list only store-plural forms (credentials/secrets/
passwords, not tokens), so tokenizer.ts / dev.environment.ts /
style.keys.ts / api-tokens.md are no longer masked whole while
.env.local / credentials.json / mytoken.yaml still are.
Content layer: drop the bare key/auth/token/pass/bearer field names —
they label far more ordinary content (keyboard keys, CSS custom
properties, lexical tokens) than secrets; qualified names stay and
Bearer schemes remain covered by BEARER_RE.
Adds the over-masking regression cases; fixes the spec's source import
extension to .ts (repo convention).
Every .md preview renders through MarkdownDocument (htmlInfo is ungated),
but its markdown segments passed the raw source into MarkdownText, whose
chat-security stance drops non-http(s) image destinations — so local
markdown-syntax images (, reference-style definitions)
always fell back to alt text. Only HTML <img src> tags were rewritten
(postProcessSanitized).
Apply rewriteLocalImageUrls to each markdown run (segment + appended
reference definitions) inside MarkdownDocument's prepared memo, through
the same media trust fence the sanitized HTML leaves already use.
Idempotent for already-absolute media URLs; remote destinations
untouched. Also correct the stale "non-null only for HTML documents"
comment on TextEditor's htmlInfo.
.html/.htm/.xhtml op previews (read / write / edit, non-error) gain a
Render/Raw toggle in the preview pane head: the render view loads the
SAVED file through the same /sidebar/html route the editor's html viewer
uses, so relative assets (./style.css, img/x.png) resolve inside the
route and a segmented read still renders the whole document (the route
serves the file, not the op snapshot — the file-trace-style srcDoc
approach would render a cut-off document whose scripts never run).
The frame is ALWAYS sandboxed (HTML_IFRAME_SANDBOX, moved to the shared
html-preview.ts module — opaque origin, no allow-same-origin, no
top-navigation) plus the route's CSP sandbox header as
defense-in-depth; unlike the editor viewer this surface has no
no-sandbox escape hatch. The contract is pinned by new sandbox-spec
tests on the extracted HtmlRenderPreview component and the pane toggle
(off by default; absent for non-html and error targets).
Locales: zh/en + all 19 shipped third-language overrides (key-set-equal
gate green). Typecheck + eslint clean; sandbox/redact/diff-rows/frontmatter
suites 36/36 green. The full-suite host-side failures (fs-operations /
pty smoke / git / install-powershell) are pre-existing on the base
branch (verified via stash) and unrelated.
sidechat.start passed `seed` without `meta.isSeeded: true` +
`inheritedEventCount`, so dsh-session treated the whole seed as the
child's OWN events and the child's Inbox constructor replayed the
parent's `agent/inbox/spliced` history — inheriting whatever input sat
UNCLAIMED in the parent at the click moment:
- next-turn: a queued follow-up behind a still-running turn;
- next-step (the reported long-context case, parent queue empty): a
tool-result context / steering spliced between step boundaries.
The first side prompt then claimed [phantom, boundary, question] — the
stale parent message was sent to the model BEFORE the boundary and
logged as a user/message. The fix passes the exact marker pair the
host's own session.fork uses (api-session-controller); ownEvents() now
stops at the end-seed boundary and the inherited inbox replays empty.
Regression tests run the REAL Session.create + Inbox in both
directions (old shape inherits the phantom, marked shape stays empty);
routes tests pin the two create options. Verified: pnpm typecheck /
test (5 pre-existing unrelated UI-spec failures on main unchanged) /
build.
The changes-tab op previews replay read/write/edit payloads into the DOM
(diff rows, read rows, markdown reading mode, error text) — when the model
reads a credentials file, that plaintext must not land on screen or in
inspectable DOM. New display-only redaction layer, on by default with a
one-click pane toggle (persisted in localStorage):
- path layer: credential-shaped filenames (.env / *secret* / *credential* /
*token* / *api-key* / private keys...) mask whole files line-wise
- content layer: ordinary files mask only secret-shaped spans — secret
field assignments keep the field name (api_key: [REDACTED]), plus bare
token prefixes (sk- / AKIA / ghp_ / xox / ya29.), Bearer headers, and
PEM private-key block headers
- redacted payloads flow through the same viewOp shape into diff rows,
read rows, markdown reading mode, and error text
Known edge (documented in README): an unquoted mermaid node label
containing a redacted secret breaks the diagram and falls back to source;
workaround is quoting the label. Session data is untouched — this is
display-only and not a boundary against same-context malicious plugins.
Locales: zh/en dictionaries + all 19 shipped third-language overrides
(key-set-equal gate green). 9 new unit tests; typecheck clean; targeted
suites 40/40 green. The fs-operations failure is pre-existing on the base
branch (environment-dependent, verified via stash) and unrelated.
Catalog entry names (15 tab + 4 viewer) and the dsh-desktop shell preset
desc were raw hardcoded zh strings rendered straight into the add-plugin
modals and the title-bar scheme dropdown. Mirror the existing desc
pattern exactly: PluginEntry.name / ShellPreset.desc widen to the
i18n-friendly string-or-() =>-string union, the catalogs carry lazy
t('pluginXxxName') / t('presetDshDesktopDesc') lookups, and the
add-plugin modal (render, search filter, aria-labels) plus the scheme
SelectMenu resolve them like every other copy. The plain-string arm of
the union stays as the raw-name fallback, and t() itself never yields
empty text (missing keys render the key), so a dict gap can never blank
a name.
20 new dictionary keys added to zh/en (locales.ts), ja (locales-ja.ts),
and all 18 third-language dictionaries — tests/locales.spec.ts enforces
key-set equality against zh for every shipped dict (precedent: 9cc642b),
so the 'leave other locales-* untouched' shortcut would fail the hard
parity gate; real translations are provided instead.
Spec follow-ups (behavior change: names/descs now render through t()):
- plugin-list.spec.ts: name non-empty assertion resolves the union.
- add-plugin-modal.spec.tsx: sentinel/office name assertions resolve
from the catalog (en locale in tests) instead of pinning zh literals.
- shell-presets.spec.ts: desc non-empty assertion resolves the union.
Graduate from the alpha track to the stable channel:
- Baseline: CI mount pin, dsh.plugin.json engines floor, and every
@deepseek-ai/* peer floor / devDependencies pin move to 0.1.2-rc.1
(published on the `next` dist-tag). alpha.5 -> rc.1 is a pure version
bump in the harness (all package.json "version" lines, zero source
delta), so no code adaptation; every host contract was re-verified
unchanged and the real-host mount smoke passes.
- dsh-client-locale resumed publishing; its trailing-pin exception is
gone — peers and devDeps align at 0.1.2-rc.1 (pnpm peers check clean).
- Version: fold the unpublished v0.18.1-alpha.0 / v0.19.0-alpha.0 bumps
into the v0.18.0 stable release; SIDEBAR_SERVICE_VERSION follows
package.json (asserted by tests/service.spec.ts).
- Docs: README/README_EN badges + channel notes + a full "since v0.17.1"
changelog entry (the two folded alpha numbers are annotated), AGENTS.md
§2/§3 baseline wording, external-plugin-guide version line, and the
e2e-host-protocol header's re-verified list gains 0.1.2-rc.1.
First ci-windows run exposed four platform assumptions in the fixtures:
- smoke: the scratch repo inherited the runner's system-wide
core.autocrlf=true (Git for Windows default), so every index restore
smudged LF to CRLF and broke the byte-exact assertions — pin
core.autocrlf=false in the fixture (eol policy is not what these
destructive-op tests are about).
- git / git-worktree: Windows TEMP is the 8.3 short form
(C:\Users\RUNNER~1\...) while git reports long paths, and
realpathSync does not expand 8.3 aliases — switch the canonicalizer
to realpathSync.native (libuv resolves via GetFinalPathNameByHandle,
which does); same trick as the existing macOS /var workaround.
- bundle-route: the ETag memo revalidates on stat (mtime+size) and
Windows updates file timestamps lazily, so the equal-size rewrite
still looked unchanged — change the length too.
- market-manifest: GNU tar (Git Bash) reads the drive-letter colon of
absolute Windows paths as a remote-host spec — run it from the pack
dir with the bare tarball name instead.
The explorer @ button appended plain text (@folder/file.ts), which DSH's
composer only decorates as a folder ref (@folder/) — the file name stayed
outside the link. Insert FILE references through DSH's own structured
insert-reference event instead, minting one chip for the whole reference
(displayed as @basename, serialized as the full @path). Directories keep the
plain @dir/ mention so folder completion still works.
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).
Menghuan1918