v1.8.1

Fixed

• Session-header drawer toggle: 0.2.2's masked titleCluster made the
  absolutely positioned control untappable (present, sized, never hit). The
  phone branch now turns that mask off and the toggle keeps its composer-gutter
  place at x=16.
• Session-delete route: it registered an exact /api path on webServer, and
  the launch-token fence is a /api prefix route an exact route shadows, so the
  handler ran unauthenticated. It now calls connection.requestRejection() and
  fails closed with 503 when the host publishes no such service.
• Settings sheet: one content column instead of five left edges; the tab strip
  fades only a cut edge; the panel is a column instead of the host's grid (the
  drag handle had been auto-placed into the 36px header cell); a row's switcher
  no longer overflows the column; the panel honours its own height cap.
• Response compression now installs only when the host does not compress: 0.2.2
  ships compression@1.8.2 and the deployed profile enables it, so the
  process-wide patch was dead weight.
• The contract scan hit-tests interactive controls and drives a real session
  first (hit=15 skip=33 -> hit=26 skip=25 on the same page).

Changed

• Peer ranges accept the running DSH prerelease (>=0.2.2-0 <0.3.0-0,
  cordis ~4.0.6-alpha.1); devDependencies pin the same release so pnpm verify
  type-checks against the host the plugin runs on.
• Removed the inert dsh.client.inject entry (dsh-client-ui-primitives is a
  platform seed, not a boot row).

What's Changed

• fix: make the plugin work on DSH 0.2.2-alpha.2 (toggle, delete-route auth, settings sheet) by @ddtcorex in #89

Full Changelog: v1.8.0...v1.8.1

v1.8.0

Added

• Composer clearance for the landscape keyboard band — while the frame
  carries data-mobile-nav-composer-tight, the sticky seat outranks the
  opaque session header (z-index: 31) and the stats row yields its height
  to the editing row, so rotating to landscape with the keyboard up no
  longer hides the input row. Pure geometry plus hysteresis, fail-open,
  complements the iOS-only keyboard lift. Ported from dsh-web-mobile
  v3.2.2 (#246).
• Manual ask-card fullscreen with answer scroll-back — an expand button
  in the card header grows the card to the measured header-to-composer
  band (only when that band beats the host min(60vh, 520px) cap), and the
  focused answer field is scrolled back into the body intersect visual
  viewport, so typing stays visible with the keyboard up. Manual toggle
  only, fail-open to the host default. Ported from dsh-web-mobile
  v3.2.2 (#246).

Fixed

• Drawer menu taps stay inside the drawer — a panelRow button that
  opens a menu (aria-haspopup) no longer counts as navigation, so the
  More menu no longer closes the drawer and floats over the session.
  Ported from dsh-web-mobile v3.2.2 (#246).
• Ask-card overflow convergence and 44px touch targets — the card body
  pins overflow-x: hidden, the question title caps at
  min(23vh, 155px) with internal scroll, and the header actions widen to
  a 20px gap with 44x44 hit boxes. Ported from dsh-web-mobile
  v3.2.2 (#246).

What's Changed

• feat(mobile): port upstream v3.2.2 phone batch by @ddtcorex in #87
• chore(release): 1.8.0 by @ddtcorex in #88

Full Changelog: v1.7.4...v1.8.0

v1.7.4

Fixed

• Large JSON responses no longer block the host event loop — response
  compression runs off the event loop (brotliCompress / gzip via the
  libuv thread pool); end() returns before headers are written and the
  reply completes from the zlib callback. Callers that flush headers during
  the window win (body replays verbatim), compression errors degrade to
  identity, and the callback never throws (process-wide patch). Fixes the
  first-run notice stuck at saving behind a multi-megabyte response.
  Ported from dsh-web-mobile v3.2.1 (#233).
• The image lightbox keeps a full-viewport hit area on tablet — the
  tablet dialog cap (min(100vw - 32px, 720px) centered) matched the host
  ImageLightbox through its :not(...) branch (stage last child is the
  close button > svg), leaving a dead strip on the right side. A positive
  div[role="dialog"][aria-modal="true"]:has(> :first-child > img) override
  restores full-bleed after the cap. Ported from dsh-web-mobile v3.2.1
  (#234).

What's Changed

• fix(mobile): compress large JSON off the event loop by @ddtcorex in #84
• fix(mobile): keep the image lightbox full-viewport on tablet by @ddtcorex in #85
• chore(release): 1.7.4 by @ddtcorex in #86

Full Changelog: v1.7.3...v1.7.4

v1.7.3

Fixed

• Chrome for Android no longer pans the whole page when the keyboard opens
  — the viewport meta gains interactive-widget=resizes-content, opting
  into the WebView keyboard model (layout viewport shrinks instead of the
  browser panning the document and scrolling the top bar off-screen).
  Engines that do not recognise the field (iOS Safari, Firefox Android)
  ignore it. Consequence folded in: a shrinking layout viewport collapses
  the classic innerHeight - visualViewport.height keyboard inset to ~0,
  so both keyboard readers now compare against the shared resting baseline
  (restingViewportHeight, one home in core/composer-dom.ts) — otherwise
  the tap-time blur fires while the keyboard is up and the composer jumps.
  Ported from dsh-web-mobile v3.2.0 (#226).
• The stats strip is marked again on hosts that split StatsPills — the
  slow path hunts in two tiers (_root first, then [data-slot] inside the
  composer stack) sharing one isStatsStripCandidate predicate, so a future
  host reshape fixes one place. Tier 2 is forward-looking for the 0.2.1
  split shape (unverifiable live: the installed host is 0.2.0-rc.2, where
  tier 1 short-circuits) — confirm on a real 0.2.1 host before relying on
  it. Ported from dsh-web-mobile v3.2.0 (#216).
• The drawer keeps an opaque fill under wallpaper plugins — gated on
  body[data-dsh-bg], light #ffffff (this file's own bg-base fallback)
  and dark #151517 (upstream's bg-base fallback for the same token); the
  no-wallpaper path keeps resolving the theme token. Appearance under a real
  wallpaper install still needs on-device re-check. Ported from
  dsh-web-mobile v3.2.0 (#229).
• Edge-zone touches no longer lose their native scroll on iOS — the
  swipe touchmove prevent is gated on the horizontal axis lock
  (shouldPreventTouchMove); pre-lock touches scroll natively
  (touch-action: pan-y already covers the horizontal race). Ported from
  dsh-web-mobile v3.2.0 (#223).

Audited, no change

• Header chips need no 0.2.1-split fix: the lineage trigger is already a
  shape anchor (_trigger[aria-haspopup][aria-expanded], not _root) and
  the hover scope already a marker + role (data-lineage-root,
  role="tree"); the jobs chip is already role/slot-identified. The
  _titleRow 36px reserve sits on a non-scroll flex row for an out-of-flow
  toggle, so padding-right is correct there (the #217 margin lesson applies
  to scroll containers only).

What's Changed

• fix(mobile): reference dsh-web-mobile v3.2.0 batch by @ddtcorex in #82
• chore(release): 1.7.3 by @ddtcorex in #83

Full Changelog: v1.7.2...v1.7.3

v1.7.2

Fixed

• Settings sheet probe follows the Account menu — since DSH 0.2.1
  (account login) Settings no longer lives in the drawer: the drawer holds
  Plugins / Automation tasks / the Account menu, and Settings is a menuitem
  inside it. The probe clicked a drawer Settings entry that no longer exists
  and aborted at sheet.open; it now drives Account menu → Settings, with
  all geometry assertions unchanged. Shipped code needed no change (sheet
  bottom-anchoring, horizontal nav pills, close hit-test, drawer footer slot
  and toolbar reparent all verified live on 0.2.1-rc.2).

Fixed

• In-place TPS text rewrites wake the stats task again — the shared
  full-tree observer never opened characterData, so frames where React
  rewrites the TPS text node without rebuilding it woke no task and the
  readout sat stale. The observer now delivers those records through an
  explicit characterData dirty key (mutationDirtyKey in the DOM-free
  core, fail-open to the tree sentinel), and the stats task declares that
  key next to '*'. Ported from dsh-web-mobile v3.1.0 (C08).
• A throwing host selectPanel no longer wedges the panel exit — the
  call runs after the exit state flips but before the fallback timer arms,
  so a host error left the exit stuck with no safety net. The fallback now
  installs in a finally while the error still propagates. Ported from
  dsh-web-mobile v3.1.0.

Changed

• DSH peer floor raised to 0.2.0-rc.2 (>=0.2.0-rc.2 <0.3.0-0 on every
  @deepseek-ai/dsh-* peer, owner decision 2026-10-07). Hosts older than
  0.2.0-rc.2 are no longer supported. Measured: all 13 peers publish
  0.2.0-rc.2, which pnpm resolves under default prerelease semantics and the
  DSH gate accepts with includePrerelease.

Removed

• Pre-0.2.0-rc.2 host compat in the session-delete row resolution: the
  displayTitle legacy label match and the tolerance for a missing
  session.untitled dictionary key. untitledLabel is now a required input.

• Pre-0.2.0-rc.2 host compat in session deletion: the refusal of a live
  session on a host without an agent disposal face (every supported host's
  Agent has cancel and whenIdle) and the flat persistence.list() entry
  shape (supported hosts return snapshots carrying .header). A live session
  that does not converge to idle is still refused with 409 session-busy.

• Pre-0.2.0-rc.2 host compat in the client: the rc-era and Regular host
  icon candidates (the host exports the Medium weight), the dropped
  SessionListState.current and sessions.clear members, and the manager
  snapshot items read in session row resolution.

• Legacy ?mobile-nav-debug=1 diagnostics alias, only
  ?dsh-maestro-mobile-debug=1 shows the floating debug bar now.

What's Changed

• chore(mobile): raise the DSH peer floor and remove old host shims by @ddtcorex in #76
• fix(mobile): prefix locale namespace and pin streaming fast path by @ddtcorex in #77
• fix(mobile): wake stats-line on in-place TPS text rewrites by @ddtcorex in #78
• fix(mobile): arm panel-exit fallback even when selectPanel throws by @ddtcorex in #79
• fix(mobile): open Settings via the Account menu in the sheet probe by @ddtcorex in #80
• chore(release): 1.7.2 by @ddtcorex in #81

Full Changelog: v1.7.1...v1.7.2

v1.7.1

Fixed

• Deleting a session no longer reports failure when it succeeded — the
  host moves the session to trash then aborts the reply, so the browser
  fetch rejects with Failed to fetch for a delete that landed. The
  session list is now the judge: on a rejected request the list is
  re-read bounded (4 attempts x 350ms) and only an id that survives every
  attempt shows the error row.
• Opening the model / reasoning-level menu no longer raises the soft
  keyboard — the host drills into the model pane and focuses its search
  field from a passive effect. The field's focus method is shadowed, armed
  on capture-phase pointerdown strictly before the effect can run; a real
  tap on the field still focuses natively. No hashed-class fallback, per
  the hash-prefix ban.
• The iOS composer stays above the form-assistant bar while typing —
  the host selection-scroll helper mixes layout coordinates with the visual
  caret rect once the iOS keyboard forces a nonzero window scroll, dropping
  the seat behind the assistant bar on every keystroke. The seat is lifted
  from screen-space quantities only (seat bottom vs visualViewport
  height), fail-open on pinch or scroll-channel drift. The active path
  needs a real iPhone; headless cannot verify it.
• Dragging a selection handle across the header no longer jumps the
  viewport to the top of the session — while a conversation selection
  exists the header seat exits hit-testing (pointer-events only, no
  layout change), so the drag stays in message text. Complements the
  data-mobile-nav-dragging drawer-yield mark on a different axis.

What's Changed

• docs(mobile): correct row id, slots, paths and test commands by @ddtcorex in #70
• ci: call the reusable workflows at their new repository name by @ddtcorex in #71
• fix(mobile): prose floor at the host secondary tier, and opt code out of iOS text autosizing by @ddtcorex in #72
• fix(mobile): port upstream v3.0.5 phone fixes by @ddtcorex in #74
• chore: release 1.7.1 by @ddtcorex in #75

Full Changelog: v1.7.0...v1.7.1

v1.7.0

Added

• Opening or switching to a session no longer raises the soft keyboard:
  the host's InputBar focuses the Lexical editor from a React effect keyed on
  sessionId (InputBar.tsx:178-182), so every switch handed the WebView a DOM
  focus and both iOS and Android put the IME over half the screen. A new
  session-focus-guard effect arms an 800 ms window on a real current-session
  change and suppresses that focus in two layers, because one is not enough: an
  own-property shadow over the editor's focus(), plus a synchronous
  focusin-capture blur. The blur must be synchronous because a remount's host
  focus runs in the commit's synchronous layout-effect phase, before any
  MutationObserver microtask, so the freshly mounted editor is focused while
  it still carries no shadow; a deferred blur does not help either, since the
  IME has already begun by the next macrotask. The window arms only on a real
  switch (retainedBy.mainView > 0, the same rule session-menu.ts:73 uses,
  because 0.1.6-alpha.2 dropped SessionListState.current), so list churn
  cannot flash the guard, and a pointerdown on the editor closes the window
  early so a user's own tap always opens the keyboard. The guard owns its own
  data-mobile-nav-session-guard marker and deliberately does not touch
  data-mobile-nav-focus-shadow, which composer-plus-toggle.ts owns on
  documentElement and composer-focus-release.ts:269 reads. Added
  pnpm probe:session-focus, an INFO probe that reports the DOM half of the
  guard against a live host and exits 0 unconditionally. Gated on a device: the
  soft-keyboard behaviour cannot be settled headless.

Fixed

• createFocusShadow is now reference counted per element. The composer's
  + menu shadow and the new session guard can hold the same editor at once, and
  they were independent instances: whichever restored first ran
  delete element.focus and unblocked native focus() for the other, so the
  surviving guard still reported armed while the keyboard rose inside the
  window it believed it owned. Holds now live in a WeakMap<FocusableLike, number>
  and only the last holder deletes the override, which protects every existing
  caller; a repeated restore() stays a no-op, so a stranded hold can never
  block focus forever.

What's Changed

• fix(mobile): keep the phone font floor on prose only by @ddtcorex in #64
• fix(mobile): drive the panel probe with a real touch on a clear screen by @ddtcorex in #65
• fix(mobile): re-anchor dsh peer ceiling to 0.3.0-0 by @ddtcorex in #66
• fix(mobile): resolve a title-less session row after the 0.2.0-rc.2 label move by @ddtcorex in #67
• feat(mobile): stop a session switch from raising the phone keyboard by @ddtcorex in #68
• chore: release v1.7.0 by @ddtcorex in #69

Full Changelog: v1.6.2...v1.7.0

v1.6.2

Fixed

• The session-row Delete item is back on 0.1.7 hosts — the host composes
  the row menu from a slot list and 0.1.7 prepended "Pin session", so the menu
  holds four items and the old labels.length === 3 signature stopped matching:
  Delete disappeared with no error anywhere. The gate is now an inclusive label
  signature (rename + fork + archive present, the unarchive twin rejected, extra
  items allowed), read from the live host locale so wording changes stay
  harmless. Verified live on 0.1.7-rc.2: five items with Delete injected;
  pnpm probe:session-delete 7/1/0 (the skip is the destructive step, which
  needs an explicit session id).
• The Settings sheet is a bottom sheet again — the overlay guard was written
  as :has() nested inside :has(), which is an invalid selector: Chrome
  throws and drops the whole rule from the CSSOM with no error, so the four
  overlay rules never applied a single declaration and the sheet rendered as a
  floating centred card (equal 56.5px gaps on an 844px viewport) since
  2026-08-30, while every source-text assertion kept passing. Flattened to
  overlay:has(navList); measured after: 4 overlay rules in the CSSOM,
  align-items: flex-end landed, panel bottom flush with the viewport. Guards:
  tests/css-selector-validity.test.ts fails the build on the nested shape and
  pnpm probe:settings-sheet asserts the CSSOM rule count plus the
  bottom-anchored geometry (8/8) — a text regex cannot see a rule the browser
  rejected.
• An oversized delete body is answered, not cut off — the route destroyed
  the socket at the overflow point, racing the error response so the client saw
  an empty reply instead of the reason. The reader drains to the natural end
  and the route answers 413 payload-too-large; the cap stays at 64 KiB.
  Verified live after the host restart: a 70 KB body returns the 413 JSON.
• end(callback) no longer writes the callback's source into the response
  — the deferred-response replay forwarded the end() tail blindly: a function
  argument was buffered as body text (and the callback never fired), an
  encoding token was written as a second body chunk, and a buffered write()'s
  callback was dropped. Arguments are now classified (function = callback,
  encoding = consumed by the buffering, buffered write callbacks replay once,
  in order, right after the real end()).
  tests/compress-response.test.ts drives a real HTTP server through the patch
  and fails all three behaviours against the pre-fix code.

Added

• Deleted session logs are recoverable for 24 hours — the directory moves
  into <root>/.sessions-trash/ with every canonical payload renamed to
  <name>.trash plus a best-effort manifest.json, instead of rm -rf. The
  rename precedes the move because the host scans every directory under the
  storage root as a project directory, trash included, and a canonical payload
  name there breaks the whole session list; entries older than 24h are purged
  inline on a later deletion, never on a timer. A stash failure after a live
  teardown answers cleanup-failed with deletedLiveSession: true and still
  settles the workspace accounting, so the caller is never told a session is
  intact when it is already gone — and the client refreshes the list on that
  code for the same reason.
• A reparent-leak probe — pnpm probe:reparent-leak drives a Settings
  open/close cycle and a session switch and asserts no moved host node is
  orphaned and no page exception or console.error is raised (5/5 on rc.2).
  That measurement is why the three reparent sites stay: same hazard class as
  upstream's #104/#105, no observable fault here.

What's Changed

• fix(mobile): rc.2 session-menu compat, sheet anchoring, and delete-route hardening by @ddtcorex in #62
• chore: release v1.6.2 by @ddtcorex in #63

Full Changelog: v1.6.1...v1.6.2

v1.6.1

Removed

• Dead modules, and the composer marker got one home — a reachability +
  reference scan over src/ (every module reachable from the two entries, every
  export referenced somewhere, tsc --noUnusedLocals --noUnusedParameters clean
  on both faces) left four things to delete:
  components/BottomSheet.tsx and its styles/sheet.css.ts section — the
  component had no importer and its [data-mobile-sheet] rules therefore styled
  nothing; the walker that builds the client bundle never inlined it, so only the
  stylesheet section was being shipped. effects/overlay-backdrop-fab.ts — the
  legacy manual frame.appendChild backdrop/FAB task: nothing installed it, and
  because its fadeHook was the only thing that ever set it, the swipe-close
  call fadeOverlayOut() had been a silent no-op. isGestureConsumed — a
  gesture-guard.ts export referenced nowhere, not even by its own tests. The
  [data-composer-input] literal now lives once, in core/composer-dom.ts,
  together with the editorElement() copy both composer effects carried and the
  visual-viewport reader; five modules used to spell the marker out.
  probe:composer-plus 17/17, probe:pointer-gating 6/6 and
  probe:multi-width 33/33 after the deletions, pnpm test 223/223.

Fixed

• The drawer's dimming layer fades with the slide-out again — the swipe
  layer's close commit called a fade that had been a silent no-op since the
  shell.overlay slot took over the backdrop (its hook's only setter lived in
  the legacy overlay task nothing installed), so the dimming snapped away ~280ms
  after the drawer had already left: the host flips data-sidebar-collapsed when
  the animation lands, and React unmounts the layer at that moment. The new
  effects/backdrop-fade.ts resolves the layer itself and drives the same two
  inline properties the retired task did — opacity and pointer-events, both
  !important so React's style prop and the layer's entry animation cannot
  outrank them — with the rect flushed between the transition and the target so
  the fade provably starts at opacity 1. Resolving the element instead of taking
  a registered hook is deliberate: that seam is what went stale, and a function
  that finds its own element cannot. No restore is needed, because the layer is
  conditionally rendered per open. Gate: pnpm probe:backdrop-fade (4 rows: the
  layer opens opaque, goes mid-fade while the drawer slides — 12 partial frames
  and a final opacity of 0 — lands gone with the drawer closed, and a
  non-animated backdrop tap still closes cleanly). A/B: with the fade call
  disabled the mid-fade row FAILs with the layer at opacity 1 for the whole
  755ms window, which is the reported symptom.
• The composer "+" stops raising the keyboard on the host's second focus —
  reported after the release above landed: tapping + with the keyboard dismissed
  still brought it up. The focus shadow was lifting too early — the host focuses
  the editor again from the effect that runs when its menu opens, which is after
  this plugin's bubble handler has read the menu as "not open" (mounted, not yet
  laid out). FOCUS_SHADOW_MIN_MS (700ms, still bounded by the 1.5s cap) makes
  the window a floor rather than a click-scoped one, and a focusin capture-phase
  blur takes back any focus that reaches the editor during the window, because a
  blur in a macrotask is too late (the IME has started). Both come from the
  community plugin mexiaosqwq/dsh-web-mobile's composer-keyboard-guard.ts,
  which measured the timing on a real iPhone; their v3.0.2 does not touch this
  path. probe:composer-plus gates the window (shadow-window-outlives-the-click,
  shadow-lifts-on-its-own), and the take-back rule is unit-tested
  (shouldTakeBackArmedFocus) because this host leaves the editor unfocusable
  while its menu is open, so a focus-driving probe row would pass either way.
  Refs: mexiaosqwq/dsh-web-mobile v3.0.0/v3.0.1.
• The composer no longer drops the keyboard while it is being typed in —
  reported from the phone as "the keyboard hides by itself" after the fix above
  landed. The release trusts the keyboard reading, and that reading was the
  classic innerHeight - visualViewport.height inset: on a page that cannot
  scroll (this shell is a full-height flex layout), iOS shrinks the LAYOUT
  viewport with the keyboard too, so both heights move together and the inset
  stays near zero while the keyboard is up. The reading now also compares against
  the tallest visualViewport.height seen recently (resetting on rotation) and
  treats either signal saying "up" as up, and a keystroke in the editor
  (beforeinput / input / keydown / compositionupdate) keeps the release
  away for 1.5s whatever the viewport says. probe:composer-plus gained
  focus-release-holds-when-both-heights-shrink (emulating the iOS case: both
  heights shrink by the same amount), A/B-validated against the inset-only
  reading, where it FAILs with the editor losing focus and the release count
  rising while the emulated keyboard is up.
• The iOS keyboard stops coming back on the first tap of the composer "+" —
  reported from the phone after the focus shadow landed. The shadow neutralises a
  focus() the host calls, and the tap that raises the keyboard focuses nothing
  at all: iOS keeps the composer's contenteditable as the active element once
  the keyboard goes away, and WebKit shows the keyboard for that retained
  editable on the next tap. The plugin now releases the focus while the keyboard
  is hidden — at install, on a 500ms heartbeat, on an editor focusin, on a tap,
  on visualViewport resize/scroll, and as soon as the "+" shadow lifts — and
  never during the tap (the uncompensated blur that bounced the composer row),
  never while the keyboard is up, never inside the 700ms grace after a finger
  lands on the editor, and never while the visual viewport is unreadable (a
  pinch-zoomed viewport reads as "no keyboard"). The blur's own viewport nudge is
  undone by restoring the scroll on the next frame when it moved by
  NUDGE_MAX_PX or less. The invariant is a state, not an event: measured, a
  focus can land on the editor with no focusin dispatched at all, which is what
  the heartbeat is for. probe:composer-plus runs the battery under an iPhone
  user agent (the release arms on detectIosWebKit) and gained three rows for
  it — precondition, release armed, and the release resuming after an editor tap
  — A/B-validated against the release inverted to never install for iOS, where
  they report the retained focus, releases=null, and the retained focus
  surviving a "+" tap.
• The composer probe no longer inherits the typing scene's fake viewport —
  the scene replaces window.visualViewport to emulate a keyboard and only put
  back an own descriptor, which Chrome does not have (the property lives on the
  prototype): the stub outlived the scene, every later row read "keyboard up",
  and the focus rows were silently disarmed. It now deletes the stub when there
  is nothing to restore, and a composer.focus-release-precondition row reports
  the keyboard inset it is asserting against. A probe throw also lands as a
  probe.crashed FAIL row with a SUMMARY instead of a stack, and the
  touch taps use the camelCase touchEnd CDP event type (touchend is rejected
  with -32602, which killed the run mid-battery).

What's Changed

• fix(mobile): use an installer-resolvable dsh peer floor by @ddtcorex in #56
• fix(mobile): stop the jobs chip from taking the lineage trigger's place by @ddtcorex in #57
• fix(mobile): scope the lineage chip's hover swallow to the chip, not to a stale hash by @ddtcorex in #58
• fix(mobile): hold the composer keyboard on iOS and drop dead modules by @ddtcorex in #59
• fix(mobile): fade the drawer backdrop with the slide-out again by @ddtcorex in #60
• chore: release v1.6.1 by @ddtcorex in #61

Full Changelog: v1.6.0...v1.6.1

v1.6.0

Added

• Sidebar panels have a way back to the conversation — a global panel
  (Plugins, Skills, …) replaces the main column and the host ships no exit:
  PanelRow's onClick is a bare selectPanel(id), so re-tapping the selected
  row re-selects it, and a panel page renders no session header — which is
  where the drawer toggle lives. One shared exit() now serves three routes:
  the system back key (one history entry armed while a panel owns the main
  column, given back when the panel leaves by another route so it cannot
  swallow the user's next back press), a re-tap of the already-selected panel
  row ([class*="panelRow"][aria-current="page"]), and the shell FAB, which
  reads as "back to conversation" there and moves to the top-left corner
  (data-mobile-nav-fab-mode="exit-panel") instead of the hero seat that used
  to float over the panel's own head. The host's panel-selection face is
  capability-probed, so a host generation without selectPanel leaves every
  route inert instead of throwing. Refs: mexiaosqwq/dsh-web-mobile v3.0.0/v3.0.1.
• pnpm probe:composer-plus, pnpm probe:panel-exit and
  pnpm probe:multi-width — three live gates for the phone interactions
  above: the composer menu across four taps (open/close/open/close), the panel
  FAB back face plus back-key and re-tap exits and their history bookkeeping,
  and the layout tiers (phone 320/360/390/430, tablet 768/1023, desktop 1280)
  including the header icon actually painting an svg path. All three were
  A/B-validated against the pre-fix behaviour so they fail on a build that
  lacks the fix. pnpm test gains tests/phone-tier-gating.test.ts, which
  locks the tier structure (a min-width: 768px block must also cap 1023 and
  require a coarse pointer; the tablet overrides must come after the mobile
  gate).

Fixed

• A second tap on the composer "+" closes the command menu — the button's
  onClick focuses the editor before toggling the launcher, and that focus
  re-enters controller.track(), whose first act is clearLauncher(); the
  host's "already open ⇒ dismiss" branch then compares against a null launcher
  and is unreachable, so every tap re-opened the menu (the host's own
  aria-expanded stays false while it is visibly open). The plugin now
  remembers the menu's state in the click capture phase and, in the bubble
  phase after React's handler, closes it through the host's own Escape path
  when it is still open. The opening tap is never touched.
• Tapping "+" releases the editor focus while the menu is open — the
  command menu needs no soft keyboard, but that same focus call re-raises the
  IME on Android and slides the composer row up under the user's second tap
  (measured upstream: visual viewport 754 → 471 about 170ms after the tap, with
  no DOM event reaching the page). A tap on "+" now drops the focus before the
  click fires and repeats the release at 120/320/640ms while the menu is on
  screen, cancelled the moment the user touches the editor.
• The two probe reds are gone — probe:panel-font now opens a session
  through the drawer (the injected dsh.sessions.current is only a hint on this
  host) and waits for the conversation history to render before measuring the
  content font axis: it used to sample while the flow still showed
  Loading history..., so the axis row was red against a placeholder. The axis is
  live on 0.1.7-alpha.2 (15px -> 22px, floor held). smoke:cdp no longer seeds
  localStorage['dsh.sessions.current'] with the requested session: the restored
  selection and the probe's own navigation opened the same session twice, the app
  released the first reference while the right sidebar awaited it, and the host
  logged Sidebar Session opening failed: ... is released — a page error from the
  probe's boot, not the plugin (with an unrestorable id the same drive logs zero).
  The scenario opens a session through the drawer instead, which also exercises
  the real header toggle. Result: probe:panel-font 12/12 and smoke:cdp
  17 pass / 1 state-skip / 0 fail with page.errors count=0.
• Tapping "+" no longer raises the keyboard on iOS — the Android half of
  this (blur the editor after the tap) is not enough there: iOS follows DOM
  focus, so once the host's focusDraftEditor inside the button's onClick has
  run the keyboard is up and a later blur() does not take it back (reported on
  iOS after the first fix landed). The editor's programmatic focus is now
  neutralised for the length of the interaction (armed on pointerdown and again
  in the click capture phase, which is what covers a click with no pointerdown
  at all), and it is always given back — on the next tap anywhere, as soon as the
  menu leaves the DOM, on a hard 1.5s cap, and on dispose — because an override
  that outlives the interaction is worse than the bug it prevents: the editor
  could never be focused again (the exact bug the community plugin fixed in
  v3.0.1). The blur the Android fix introduced is now conditional — it runs only
  while the keyboard is already hidden (shouldDropEditorFocus, read from the
  visual viewport). Blurring an editor whose keyboard is UP starts the hide
  animation, and the keyboard is the composer's floor, so the row slid down under
  the finger on the FIRST tap only (reported from the iPhone after the first iOS
  fix); blurring one whose keyboard is already down is the state the IME re-rises
  from, and there the blur is what keeps the row still. probe:composer-plus now
  gates all of it, including a typing-state scene (emulated keyboard) where the
  focus must survive the tap. A second phone report pinned the blur itself as the
  remaining movement: the composer bounced up and back within 10-20ms, which no
  keyboard can do (iOS takes ~250ms to show or hide one) but a programmatic blur
  on iOS can, because it nudges the visual viewport. iOS therefore never blurs —
  the focus shadow is the whole defence there — while Android keeps the blur,
  which is what stops the IME re-rising into a hidden-keyboard editor. The shadow
  is also armed on touchstart now (idempotent), because iOS can fire the touch
  before the pointer event.
• Host icons are resolved by name at runtime — the host's icon exports are
  generation-specific (IconXxxOutline16 on the 0.1.0-rc line versus
  IconXxxOutlineRegular / …Medium on 0.1.7) and the two generations share no
  names, so a static import resolves to undefined on the other one and React
  reports "Element type is invalid" for the whole plugin tree. The first
  candidate name the installed host actually exports now wins, and a name
  nothing matches renders nothing rather than crashing.
• The plugin stylesheet replaces its previous copy — a plugin re-applied in
  the same JS environment (client hot reload, or a second apply whose dispose
  never ran) stacked a second <style> tag with the same rules, and the older
  tag could win on source order inside the cascade: the symptom was "I changed
  the CSS and nothing moved" while the served bundle was correct. The mount now
  removes any tag carrying the plugin's data-plugin marker first, keeps the
  fresh tag last in <head> for its !important overrides, and never
  resurrects a disposed tag from the deferred re-append.

What's Changed

• chore: target DSH 0.1.7-alpha.2 by @ddtcorex in #51
• feat(mobile): panel exit, composer "+" fixes and icon/stylesheet hardening by @ddtcorex in #52
• fix(mobile): keep the iOS keyboard down when the composer "+" is tapped by @ddtcorex in #53
• fix(mobile): clear the two probe reds at their source by @ddtcorex in #54
• chore: release v1.6.0 by @ddtcorex in #55

Full Changelog: v1.5.0...v1.6.0