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.
- 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.
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.
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.
The v0.19.0 regression, its root cause in 223b254, the four decisions
(native landing, narrow-viewport park, jump-back always expands, bottom
workbench keeps its own flows), the change list, and the evidence — including
the environment trap that the PATH `dsh` on this machine is 0.1.2-rc.1 and
cannot register the `file:` tarball into a scratch profile.
Baseline moves to DSH 0.1.5-rc.1 — the 0.1.5 line's first release
candidate, and the npm `latest` dist-tag now: peer floor ^0.1.5-rc.1,
devDependency pins, the CI mount-lane pin and dsh.plugin.json engines
all follow. 0.1.5-alpha.2 users stay on v0.19.0-alpha.1.
The upstream delta from alpha.2 is small: 373 changed files, mostly
version bumps, sidebar preview polish and test snapshots. Exactly two
changes touch this plugin, and only one needs work:
- `SidebarRightGuideEntry.description` is back (optional). The native
guide renders it only while it lists at most 4 entries; a longer list
drops every description. The plugin restores `TabDescriptor.description`
and the six built-in `guideDesc*` lines (20 dictionaries) verbatim from
0b6ac5d^, but deliberately does NOT restore the old generic
`nativeGuideDesc` fallback: the host has no fallback of its own, so a
descriptor that declares none now ships no `description` field at all
and renders as a clean title-only capsule.
Consequence worth knowing: the plugin contributes six guide entries, so
the default composition is over the 4-entry limit and shows titles only.
The mount lane proves the restored field end to end by shrinking the
enabled set to three types through the plugin's own settings route
(scratch profile only) and requiring the Files capsule to grow its line.
- The built-in `files` type declares a guide description again and uses a
coloured folder glyph; the plugin owns that kind, so nothing to do.
Verified unchanged from alpha.2 (so nobody re-checks them): the global
main-panel model, the file address grammar (util/workspace-path changed
only its package.json), the native tab-body host contract (.paneBody is
still a block scroller with a definite height), the core/agent/session/
subagent host APIs and the ui-primitives export surface.
Restore caveat fixed in the lane: the settings route merges patches
key-wise, so undoing a temporary `tabsEnabled` change by posting the
original (empty) map leaves the disabled keys behind — the e2e restores
every key it touched explicitly.
Baseline moves to DSH 0.1.5-alpha.2 only (peer floor ^0.1.5-alpha.2,
CI pin, manifest engines); alpha.1 users stay on v0.19.0-alpha.0.
Host-contract changes handled:
- the native guide entry lost its `description` field (entries are
icon+title capsules now): `TabDescriptor.description` and the six
`guideDesc*` keys plus the `nativeGuideDesc` fallback are removed.
- the file address grammar changed: `fileAddressFor` always returns a
session-scoped address (absolute paths keep their leading `/`),
`parseFileAddress` is prefix-based and strips `?`/`#`. The plugin's
purity-gated mirror follows.
- global main panels replaced the `conversation` root slot with the
keyed `main` slot (`main.conversation`): the bottom workbench's
center-column locator now accepts both keys and skips
`display: contents` slot hosts.
Fixes a pre-existing fill bug the alpha.2 probe made obvious: the native
tab body host is a block scroller with a definite height, so tab roots
that only declared `flex: 1` collapsed to content height and the side
chat composer drifted off the pane bottom. The native adapter now wraps
every tab body in a `height: 100%` column flex host.
Also promotes the dependencies `dsh-client-ui-primitives@0.1.5-alpha.2`
dropped from its manifest (anser / shiki / mdast-util-* / micromark-* /
katex / @shikijs/langs) and the newly unmet `dsh-session-persistence`
peer into devDependencies.
Escape the .pdf extension in the op-target regex (the unescaped /.pdf$/
matched any path ending in pdf) and fold the copy-pasted toggle button
into PaneToggle; record the sixth deliverable in the design doc.
Records the five features shipped on this PR (md reading mode incl.
mermaid, highlight expansion, secret redaction, html render mode), their
deliberate boundaries (redaction heuristics, saved-file render
semantics), and the 2026-09-06 standards pass.
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.
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.
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.
DSH 0.1.2-alpha.1 removed the client `ctx.connection.api` face (Remote
gateway migration), so every Side Chat transcript poll (`sessions.history`)
threw TypeError into the view's catch and was swallowed — the tab rendered
an empty transcript forever, silently (no console error, so the mount lane
never caught it). The successor APIs are not a drop-in either:
session/follow|page reject plain session addresses for origin:'subagent'
logs (agent-busy fence) and paging cannot read past the current cursor.
Move the transcript to the plugin's own wire — one transport for every
DSH generation:
- new `sidechat.events` route: live threads read the in-memory
`agent.session.events`; cold threads read `sessionPersistence.inspect`
(both seams exist since 0.1.0-rc.8; alpha.1's inspect returns the
expanded logical log, never the packed chunk-row encoding). The
inherited seed is cut host-side (`threadOwnLogEvents`), polls ride an
`afterSeq` delta, and the response is tail-capped at 8000 events (the
old client walk ceiling).
- SideChatView.fetchThread is single-path; collectOwnEvents (the client
walk) is gone, and so are the ctx.connection mirror and inject entry
(no consumer left).
- transcriptRows keeps re-cutting at session/end-seed as defense in
depth.
Verification:
- unit: sidechat.events route (seed cut, afterSeq delta + never
resurrecting the seed, cap, cold/live sources, 404/503/bad-request),
full suite 1104 passed; typecheck and build (purity gate) green.
- mount e2e vs a real dsh-v0.1.2-alpha.1 source build: 14/14, including
the Side Chat tab sweep polling sidechat.events (performance resource
assertion) and live + post-dispose (persisted) route reads.
- mount e2e vs PATH dsh 0.1.0-rc.8: 14/14 (no regression).
- New persisted built-in pref workspaceFence (default true) under the
files card's secondary settings modal; host reads it per request via
the settings seam (fenceEnabledOf — absent/malformed stays ON).
- ensureWorkspacePath/ensureWorkspaceWritePath gain a fence param:
disarmed skips the containment assert but keeps realpath resolution;
wired through fs.tree/fs.read/fs.write, the media + html routes, and
the upload path (writeWorkspaceUpload fence input).
- Editor load failures and file-tree level errors show the localized
reason plus a one-click global off (FenceErrorNotice) that flips the
pref, adopts it into the store, and auto-retries the failed operation
(reloadSeq / retryDir) — the raw wire text is never shown.
- GitView's two client-side pre-checks (hidden open action + no-op
defense) respect the pref; store threaded through TreePanel/FileTree.
- i18n: 4 new keys across all 21 dictionaries; tests: fence-off route
coverage, prefs parsing, schema default, editor toggles list;
docs: guide §8/§8.1 + plans/2026-08-29 design record.
Menghuan1918