* fix(desktop): show Hindsight memory provider
* feat(desktop): configure Hindsight memory provider
* fix(desktop): limit Hindsight modes to supported setup
* refactor(desktop): generic memory-provider config surface
Replace the bespoke Hindsight settings surface with a declarative,
schema-driven path so adding a memory provider is pure declaration —
no per-provider page, conditional, or endpoint.
- memory_providers.py: declarative registry. Each provider lists its
fields {key, label, kind, default, options, secret-vs-plain}. Hindsight's
mode is a select(cloud, local_external), so rejecting local_embedded
falls out of generic enum validation instead of a hand-written check.
- One generic endpoint pair GET/PUT /api/memory/providers/{name}/config.
GET returns declared fields + current values (secrets only as is_set,
never read back); PUT validates selects against their options, writes
plain fields to the provider config file, secrets to the env store,
and flips memory.provider.
- ProviderConfigPanel renders straight from the schema, replacing
hindsight-settings.tsx and the memory.provider === 'hindsight'
conditional in config-settings.tsx — same pattern as
toolset-config-panel.tsx off env_vars.
Scoped to memory providers; storage layout is unchanged so the runtime
Hindsight plugin reads the same config.json / HINDSIGHT_API_KEY / provider
keys as before. Tests cover the registry, endpoint behavior (defaults,
write+secret, select rejection, unknown provider, secret-never-returned),
and the generic panel.
The "💾 Self-improvement review" summary (skill/memory updated) was invisible
on two surfaces:
- Desktop Electron app had no review.summary event handler — skill/memory
writes happened silently. Now appends a persistent system message to the
transcript (matching the Ink TUI's persistent-line semantics, not a
transient toast that can be missed).
- tui_gateway (backs both 'hermes --tui' and the desktop) never read
display.memory_notifications, so it always behaved as 'on' and ignored a
user who set 'off'/'verbose'. Added _load_memory_notifications() (mirrors
the messaging gateway's bool->str normalization, defaults to 'on') and
wired it to agent.memory_notifications, matching gateway/run.py and the CLI.
Delivery chain now reaches all surfaces:
background_review.py -> background_review_callback -> review.summary event ->
desktop transcript / Ink TUI line / gateway message / CLI print.
* fix(desktop): recover stranded session windows when resume fails
Opening a session in a new window (or any routed resume) could latch the
thread loader on "session" forever — the reported "stays stuck loading,
even after a nap" bug. Two compounding causes:
1. use-session-actions.resumeSession's catch ran the REST transcript
fallback OUTSIDE its own try. When session.resume rejected AND the
fallback also threw (the common case on a wedged/unreachable backend),
the throw skipped setMessages and left activeSessionId null with an
empty transcript — exactly the state the loader gates on
(messagesEmpty && !activeSessionId), with no terminal/error state.
2. use-route-resume's self-heal could never re-fire: resumeSession sets
selectedStoredSessionIdRef synchronously at entry (before failing), so
stuckOnRoutedSession stays false, and on an already-open idle window
neither pathnameChanged nor gatewayBecameOpen fire again. The window
never retried — naps, focus, nothing recovered it.
Fix:
- Wrap the REST fallback in its own try so a fallback failure can't strand
the loader.
- Add $resumeFailedSessionId: armed on terminal resume failure, cleared at
the next resume's entry (and left clear on success).
- use-route-resume gains a bounded backoff auto-retry (4 attempts, 1s→8s)
that re-resumes while the routed session matches the failure flag, with a
fire-time liveness recheck so a recovered session isn't double-resumed.
Regression tests cover: fallback-wrap arming the flag without throwing,
flag cleared on success, retry fires on backoff, no retry for a
non-routed/recovered session, and the retry cap.
* feat(desktop): show error + manual Retry when resume retries exhaust
When a stranded session window's bounded auto-retry gives up (gateway
resume RPC + REST fallback fail through all MAX_RESUME_RETRIES attempts),
the loader latched forever. Add a $resumeExhaustedSessionId atom armed at
the give-up point so the chat view swaps the perpetual spinner for an
explicit error state + manual Retry button. Retry / reconnect / reselect
clears the latch and resets the auto-retry counter for a fresh cycle; a
route-change away from the stranded session also clears it.
Distinct from $resumeFailedSessionId (armed during the backoff window) so
the error UI only appears once auto-recovery has actually given up, not
mid-retry. Adds i18n strings across en/ja/zh/zh-hant and 3 tests covering
latch-arms-on-exhaustion, stays-clear-while-retries-remain, and
clears-on-route-change.
* fix(desktop): address review on stranded-resume recovery layer
Follow-up to review on #47655 (PR head 253bfc0e3). Four issues on the
recovery layer:
1. (blocking) Arm $resumeFailedSessionId only when the transcript is still
empty after the REST fallback ($messages.get().length === 0), matching the
atom's documented contract and the loader's messagesEmpty gate. Previously
armed on any resume-RPC reject regardless of fallback outcome, so a window
that recovered its history via REST still auto-retried and, on exhaustion,
blanked the visible transcript behind the error overlay.
2. Reset the bounded-retry attempt counter on the $resumeExhaustedSessionId
armed->cleared edge so a manual Retry / reconnect / reselect on the SAME
stranded session gets a fresh backoff cycle, not a single one-shot attempt
that immediately re-arms the error. (Keyed on the exhausted latch rather
than the resumeFailedSessionId null->value transition the review suggested:
the auto-retry loop itself toggles resumeFailedSessionId every cycle, so
keying the reset there would defeat the MAX_RESUME_RETRIES cap. Only
resumeSession clears the exhausted latch, making its clear edge the
unambiguous manual-retry signal.)
3. Advance retryAttemptRef only when the timer actually dispatches a resume,
not at schedule time. Prevents unrelated dep changes during the 1s-8s
backoff window (transient gatewayState flip, non-stable resumeSession) from
burning attempts and hitting MAX with fewer than 4 real resume attempts.
4. Drop unrelated blank-line-only insertions in store/session.ts and
use-session-actions.ts to keep the diff tight.
Tests: +3 (RPC-fails-REST-succeeds-no-arm; manual-retry-fresh-cycle;
no-attempts-burned-on-dep-churn). All 19 resume tests + full session-hook
suite (65) pass; tsc --noEmit clean.
---------
Co-authored-by: Teknium <127238744+teknium1@users.noreply.github.com>
A failed turn leaves a red error banner inline in the transcript. These
errors are renderer-local state (never persisted) and stay pinned to the
message until the session is reloaded, so a stale, no-longer-relevant
error (e.g. a transient provider/inference error) lingers with no way to
clear it.
Add an 'x' dismiss button inside the existing MessagePrimitive.Error
block. Clicking it clears the error from BOTH the live $messages view
and the per-runtime session cache — the view first, because
preserveLocalAssistantErrors re-grafts any still-errored message it finds
in the view onto the next session.info flush, so clearing only the cache
would let the heartbeat resurrect the banner. A bare error placeholder
(no streamed content) is dropped entirely; a turn that streamed partial
output before failing keeps its text and just sheds the error.
The control only renders when an onDismissError handler is wired, so
secondary/embedded Thread usages are unaffected. Adds the dismissError
string to all four locales (en/ja/zh/zh-hant) and two behavior tests.
Co-authored-by: Teknium <127238744+teknium1@users.noreply.github.com>
A turn that ends in an error (e.g. an out-of-funds state) was being
re-rendered in unrelated threads. On a warm thread switch the on-screen
`$messages` still belongs to the previously viewed thread, and
`flushPendingViewState` fed it into `preserveLocalAssistantErrors`, which
grafted the prior thread's failed turn onto the newly opened one. Because
the polluted view then became the next switch's baseline, the error
cascaded into every thread the user visited.
Only carry local errors across a view flush when the on-screen baseline is
the same session being flushed; the cached state we publish already retains
that session's own errors. Also surface the turn error as a global toast
even when the failing turn ran in a background thread, since the error
blocks all subsequent interactions until the user acts.
* fix(desktop): keep the pre-session model pick selected in the picker
The composer picker derived its "current" row from `model.options ?? store`,
so model.options always won. Pre-session that query returns the PROFILE
DEFAULT, not the sticky composer pick — so selecting a model before a session
exists left the checkmark (and the picker's "current" line) on the default,
making the pick look ignored even though the pill updated.
Add `currentPickerSelection()`: with a live session the gateway's model.options
is authoritative; pre-session the sticky `$currentModel`/`$currentProvider`
wins, falling back to options. Wire it into ModelMenuPanel and ModelPickerDialog.
* feat(desktop): global reasoning/speed defaults in Settings → Model
The composer picker is now sticky-UI/per-session only and never writes the
profile default (#46959), but Settings → Model had no reasoning/speed control
and `agent.reasoning_effort` wasn't in the curated config surface at all
(`service_tier` was buried in Advanced) — so there was nowhere to set the
profile default that crons/subagents/messaging resolve from.
Add capability-gated Reasoning (effort) + Fast controls beside the main model,
gated by the applied model's reported capabilities (reasoning defaults on, fast
off when unreported — same as the composer). They read/write `agent.reasoning_effort`
and `agent.service_tier` by round-tripping the config record, matching the
gateway's value semantics (service_tier "fast"/"priority"/"on" ⇒ fast).
* refactor(desktop): don't open the reasoning select from its row label
A <label> wrapping the Select forwarded text clicks to the trigger, opening
the dropdown unexpectedly. Plain row for reasoning; Fast stays a <label> so
clicking its text toggles the switch (expected for a checkbox-like control).
* feat(desktop): stream subagent replies into watch windows
A desktop watch window resumes a child session lazily (no full agent) and
mirrors the parent-relayed `subagent.*` events into native child-session
stream events. The child's streamed reply text was never relayed, so the
window sat blank while the subagent "talked".
- delegate_tool: forward the child's `run_conversation` stream tokens up the
progress relay as `subagent.text` (inert under CLI/TUI — their progress
handlers ignore non-tool event types; only a gateway watch window mirrors it).
- server: mirror `subagent.text` -> `message.delta` on the child sid only, and
skip the parent emit (per-token frames are meaningless on the parent session,
which shows the child via the spawn tree). Demote `subagent.start` to a
one-time goal header and drop the noisy `subagent.progress` mirror — tools
already mirror natively.
- server: guard `_start_agent_build` so a lazy watch session spectating an
in-flight child stays lazy; incidental RPCs were upgrading it to a full
agent mid-stream and silently killing the mirror.
* fix(desktop): keep watch-window chat clear of titlebar chrome
Secondary windows (new-session scratch, subagent watch, cmd-click pop-out)
hide the titlebar tool cluster + session header, so the transcript ran to the
window's top edge and streamed text slid up under the OS traffic lights.
- Gate the hidden chrome on `isSecondaryWindow()` everywhere (app-shell,
chat header, thread list) instead of the narrower new-session flag.
- Add a fixed opaque drag-strip at the top of the secondary-window transcript:
content padding alone scrolls away with the text, so the strip masks
anything behind it and keeps the window draggable like the main header.
* fix: WSL subagent window
* fix: subagent window top padding
---------
Co-authored-by: Austin Pickett <pickett.austin@gmail.com>
Co-authored-by: Teknium <127238744+teknium1@users.noreply.github.com>
Clicking a model row in the composer dropdown now commits and closes the menu
(via a close context); the hover-revealed reasoning/fast submenu stays open to
tweak. The pill shows a quiet braille loader instead of literal "No model"
until one resolves, and steer takes over the mic slot while typing into a
running agent.
The picker no longer touches the profile default. Model/effort/fast live as
plain UI state persisted in localStorage, so a pick follows across Cmd+N and
restarts instead of snapping back. New chats ship that state through
session.create as per-session overrides; live chats still scope switches to the
current session. Settings -> Model remains the only surface that writes the
profile default.
The gateway now accepts those session.create overrides, builds the agent with
them directly, reflects them in the immediate session.info payload, and writes
the chat's own model_config into the lazy DB row so reconnect/resume restores
that chat instead of the global default.
* fix(skills): guard recursive skill delete against tree-escape
Port from Kilo-Org/kilocode#11240. Their issue #11227 lost a user's entire
working directory: a built-in-skill sentinel location resolved to the server
cwd and the skill-removal endpoint ran a recursive delete on it.
Hermes' /skills uninstall path (skills_hub.py) is already hardened, but the
agent-facing skill_manage(action='delete') path did a bare
shutil.rmtree(skill_dir) with no last-line validation. Add _validate_delete_target():
refuse to rmtree a path that (1) isn't strictly inside a known skills root,
(2) is a skills root itself, or (3) is reached via a symlink/junction.
Tests: 4 cases (normal delete works; symlinked dir, skills-root, out-of-tree
all refused). E2E verified with real symlink + file I/O.
* feat(desktop): allow /browser connect on a local gateway
/browser was hardcoded as terminal-only in the desktop slash palette, so
the chat GUI rejected it with "only available in the terminal interface."
The TUI already drives the live CDP connection via the browser.manage RPC.
Wire the same RPC into the desktop dispatcher as a /browser action handler,
gated to local-gateway connections ($connection.mode !== 'remote'). connect
mutates BROWSER_CDP_URL (and may launch Chrome) in the gateway process, so
it's only meaningful when that process runs on this machine; a remote
gateway gets a clear "local gateway only" message instead.
External providers (Claude Code) store creds outside Hermes, so the
disconnect API refuses them. The backend now hands the GUI a per-OS
`disconnect_command` that clears the credential the same way the CLI's
logout does (macOS Keychain entry + ~/.claude/.credentials.json), and
the misleading "use claude setup-token" hint is corrected.
Settings → Providers offers a Disconnect button for these: it confirms,
leaves Settings, and runs the removal command in the embedded terminal
via a new runInTerminal() (queues onto $terminalInjection; the terminal
pane flushes and clears it once its session is live). The expanded list
also gets its own "Other providers" header so it no longer reads as
grouped under "Connected". API-managed providers keep the one-click
(trash) disconnect.
Each model remembers its own reasoning effort / fast mode (localStorage,
like model-visibility): editing a model's effort/fast in the submenu
writes its preset, and selecting a model restores its preset onto the
session (capability-gated, Hermes defaults when unset). Every row shows
its own remembered settings (grayed), and the row label and edit submenu
read the same effective value so they can't disagree.
Presets are desktop-client state only — applyModelPreset() no-ops without
a live session id, so selecting a model can't fall through to the
gateway's persistent agent.reasoning_effort / agent.service_tier writes.
Inactive variant `-fast` edits stay preset-only: toggleFast() records
{ fast } on the base model and only swaps models when the row is active,
and selectFamily() honors a saved variant-fast preset by selecting the
`-fast` sibling id.
Relocate the model pill to the composer, left of the mic. A new
ModelPill reuses the live ModelMenuPanel dropdown verbatim (single
click target) and the formatModelStatusLabel "Model · Fast Med" label,
anchored to its right edge so the menu doesn't drift with model-name
length. modelMenuContent now flows to ChatView instead of
useStatusbarItems, and the status-bar model-summary item is removed;
the pill subscribes to the model atoms directly and falls back to the
full picker when the gateway is closed.
On a remote gateway connection, agent-written files live on the gateway
host, not the desktop's disk, so the Artifacts view's file:// hrefs failed
("Invalid external URL") and image thumbnails broke.
Make mediaExternalUrl() remote-aware in one place: in remote mode it
rewrites gateway-local paths to GET /api/files/download (a new endpoint
that streams the file as a Content-Disposition: attachment). The artifacts
view now resolves through it, and so do the existing chat-media and
generated-image callers, for free.
The download endpoint stays auth-gated; auth_middleware additionally
accepts the session token as a ?token= query param for this one path so a
shell/browser-opened download (which can't set the session header) still
authenticates — the same query-token tradeoff as the /api/pty WebSocket.
It is NOT added to PUBLIC_API_PATHS.
Salvages #46663 (which carried ~19k lines of CRLF noise and made the
endpoint public). Reimplemented on a clean LF base with the security hole
closed and tests added.
Co-authored-by: qingshan89 <qs2816661685@gmail.com>
When a desktop/dashboard session had no agent built yet and the user explicitly
picked a provider in the model picker, config.set('model', ...) would first try
to initialize the agent from the (possibly broken) config default provider —
failing before the user's explicit switch could take effect, trapping them on a
misconfigured default.
config.set now pre-parses the model flags: if an explicit --provider is present
and no agent exists yet, it skips the default-provider agent build and routes
straight through _apply_model_switch with the explicit provider. _apply_model_switch
gained a parsed_flags passthrough (avoids double-parsing) and only falls back to
resolve_runtime_provider(requested=None) when no explicit provider was given.
The desktop hook now sends config.set instead of slash.exec for active-session
model changes, so errors from the selected provider surface to the user instead
of being swallowed.
Co-authored-by: rodboev <rod.boev@gmail.com>
Show a shimmering "Summarizing thread" label during auto-compaction, skip
the post-turn hydrate when compaction fired so the live transcript does not
collapse to the stored summary-only session.
Auto-compression rewrites history mid-turn, which made long threads look
like they reset. Re-tag the gateway lifecycle status as compacting and
surface it in the desktop thread loading indicators.
The turn-end sound is a notification concern, not an appearance one — relocate
the variant picker + preview from the Appearance tab to the Notifications tab
(its i18n keys move from settings.appearance to settings.notifications with it).
Adds a native OS notification system (Electron Notification, routed cross-OS)
distinct from the in-app toast feed. Before this, one hardcoded cue existed
(message.complete while document.hidden) with no settings or event coverage.
- Engine (store/native-notifications.ts): localStorage-backed prefs (master
switch + per-kind toggles) and a gated dispatcher over five kinds — approval,
input, turnDone, turnError, backgroundDone — with a 1s per-(kind,session)
self-evicting throttle.
- Gating: "backgrounded" = document.hidden OR !document.hasFocus(), so an
alt-tabbed window still counts as away. Completion kinds fire only when
backgrounded and for the active session (no spam from a busy gateway);
attention kinds (approval/input) also break through for off-screen sessions.
- Wired into real event sites (use-message-stream.ts): message.complete, error,
approval/clarify/sudo/secret.request; backgroundDone from composer-status at
the running -> exited transition.
- Click focuses the window and jumps to the originating session; approval
notifications carry Approve/Reject buttons that resolve in place over
approval.respond, mirroring the in-app Run/Reject bar.
- Settings: new Notifications panel (master + per-kind switches, test button
with real OS-result feedback). Full i18n (en/ja/zh/zh-hant).
* feat(desktop): add curated completion sound bank for turn completion
Replace the prior haptic-only completion cue with a curated Web Audio completion sound flow, defaulting to the minimal two-note comfort preset while keeping alternate presets available for quick iteration. Play the cue on every message completion event (including background sessions) so turn-end feedback is consistent across active and non-active chats.
* refactor(desktop): drop done1 byte sample from completion bank
Keep the curated Web Audio presets only; the embedded sample added bulk without shipping as the default cue.
* feat(desktop): expand completion sounds and add Appearance picker
Add fourteen synthesized turn-end presets with preview in settings, persisted variant selection, and softer default mixing for late-night use.
Co-authored-by: Cursor <cursoragent@cursor.com>
* refactor(desktop): dedupe completion-sound resolver, trim audio comments
Make the store the single source of truth for the variant default + range
validation and have the sound lib import it (one-way lib→store edge, no
cycle), instead of two divergent copies. Extract the shared white-noise
buffer used by the air/whoosh voices and cut the synth comments down to
why-only notes.
---------
Co-authored-by: Austin Pickett <pickett.austin@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(desktop): jump-to-approval pill for off-screen approvals
A blocked approval's only response surface is the inline Run/Reject bar on
the pending tool row. When that row is scrolled out of view the session looks
stalled with no visible action. Surface a composer-anchored "Approval needed"
pill only when an approval is pending AND its inline bar is scrolled away;
clicking scrolls the bar back into view. Preserves the deliberate inline (not
modal) approval design — the pill never duplicates the approve/reject controls.
The inline bar mirrors its own viewport visibility via IntersectionObserver
(tracks scroll/resize/layout) and registers a scroll-into-view handler the pill
fires, mirroring the existing thread-scroll jump-button bridge.
Supersedes #45828.
* fix(desktop): morph jump-to-bottom into approval prompt; drop scroll bridge
Collapse the separate "jump to approval" pill into the existing
scroll-to-bottom control: when scrolled away from the bottom while an approval
is pending, it relabels to "Approval needed". A parked approval's inline
Run/Reject bar is always the bottom-most content, so the existing
scroll-to-bottom action lands the user right on it — one control, no collision.
This also fixes the layout corruption from the first cut: the pill called
native el.scrollIntoView(), which scrolls every scrollable ancestor including
the overflow:hidden chat shell containers. Those have no scrollbar to scroll
back and don't remount on session switch, so the composer stayed shoved and
the breakage persisted across sessions. Reusing requestScrollToBottom() (the
use-stick-to-bottom path) only touches the one designated scroll container.
Removes the now-unused approval-scroll store + IntersectionObserver wiring.
Large paste and Ctrl+A → Delete froze the composer for seconds — both routed
through Chromium's contenteditable editing pipeline (~O(n²) on multiline DOM).
- insertPlainTextAtCaret: Range + text/<br> fragment (paste path)
- deleteSelectionInEditor: range.deleteContents for non-collapsed Backspace/Delete
- Shared composerSelectionRange helper; both flush via flushEditorToDraft
Profiled live (47 KB / 122 paragraphs): paste 4474 ms → 13 ms; select-delete
1304 ms → 4 ms. Collapsed-caret deletes still native.
* fix(desktop): accept slash command on space at command stage
Pressing space on a no-arg slash command (e.g. /hermes-agent) fell
through to the arg-completion stage and dead-ended on "No matches"
instead of inserting the directive. Space now mirrors Tab/Enter while
the command name is still being typed: no-arg commands commit the chip,
arg-taking commands expand to their options step.
* fix(desktop): suppress arg popover for no-arg slash commands
Committing a no-arg command (`/hermes-agent `) re-detected the chip+space
as an arg query and re-opened the popover on "No matches". The arg-stage
menu now only opens when the command actually takes args.
* fix(desktop): polish slash arg completion (space/tab/click + typed args)
Unify Enter/Tab/Space accept of the highlighted item at both the command
and arg stages: no-arg commands commit a chip, arg commands expand to
options, and an arg option commits the full `/cmd arg` chip. A fully-typed
arg (which the backend completer drops from suggestions) now commits on
Space/Tab via the verbatim text instead of dead-ending, and the "No
matches" empty state is suppressed past a command's name. Space stays
slash-only so @ mentions keep a literal space.
* feat(cli): add --safe-mode troubleshooting flag
Inspired by Claude Code v2.1.169 (June 2026): run Hermes with all
customizations disabled to isolate setup problems from product bugs.
--safe-mode implies --ignore-user-config and --ignore-rules, and
additionally skips plugin discovery (hermes_cli/plugins.py) and MCP
server loading (tools/mcp_tool.py) via the internal HERMES_SAFE_MODE
env bridge.
* fix(desktop): keep composer usable during reconnect
Right-click → Color flashed open then closed: on dismiss the context menu
refocuses its trigger, which doubles as the popover anchor, so the picker
read it as a focus-outside event and closed itself. Suppress the menu's
close auto-focus so the picker survives. Long-press already worked since
it bypasses the menu lifecycle.
The image-generate tool showed a placeholder, then the model echoed a
(often different) image inline in its prose — a second, jarring copy in
the wrong place, dimmed as tool scaffolding, with a misplaced download
button.
Now the generated image lives only in the tool slot:
- Strip every embedded image/media link from the assistant prose of a
message that produced an image (the model frequently restates the
remote URL while the result holds the local path), preserving the
agent's words. Applied on hydration, live deltas, and completion.
- One stable frame sized from the aspect_ratio arg up front, so the
diffusion placeholder and the decoded image share the same box and
crossfade with no layout shift; the box derives its height from the
true ratio on load (no letterboxing).
- Exempt generated images from the tool-block dim-until-hover rule.
- Extract a shared useImageDownload hook + ImageLightbox so the tool
image and markdown images share one implementation.
"Session busy" (4009) is the gateway's concurrency guard, not a user-facing
error. The queue already covers the deliberate "type while busy" case, so
the only leak was a submit racing the settle edge. Generalize the rewind
path's busy-retry into a shared `withSessionBusyRetry` and wrap every
`prompt.submit` (fresh send, session-resume resubmit, and rewind) so a
transient busy is ridden out within a bounded deadline and the call lands
silently. The fromQueue swallow stays as a backstop for the pathological
>deadline case.
A queued drain firing on the settle edge can race a not-yet-wound-down
turn and get a transient 4009 "session busy". Previously that appended a
red "session busy" error bubble (and toast) per attempt. For fromQueue
submits, swallow the busy error: release busy, keep the entry queued, and
let the composer's bounded auto-drain retry on the next idle.
A prompt typed mid-turn ("ghost bubble") could stick forever and never
send when the backend restarted/reconnected during the turn. Two fragile
assumptions in the composer queue drain caused it:
1. Drain fired ONLY on an observed busy true→false edge. A remount/
reconnect resets `previousBusyRef` to the current busy value, so the
settle edge is swallowed and the queue never drains. Replace
`shouldAutoDrainOnSettle` with the edge-independent `shouldAutoDrain`
(idle + non-empty), driven on the settle edge, on mount/reconnect, and
after a re-key. The drain lock still serializes sends.
2. The queue is keyed by `queueSessionKey || sessionId`. When a backend
resume mints a new runtime session id for the same conversation, the
entry strands under the dead key. Pass the *stable* stored id as
`queueSessionKey` so the composer can tell runtime churn from a real
session switch, and `migrateQueuedPrompts` re-keys pending entries on a
runtime-id change only (never on a deliberate switch).
Also make the drain resilient to a thrown/rejected onSubmit (e.g. a stale-
session 404): the entry stays queued and is retried on the next idle, with
a per-entry attempt cap (MAX_AUTO_DRAIN_ATTEMPTS) to avoid spin-loops and a
quiet toast once it gives up. A manual send clears the backoff.
Tests: composer-queue covers edge-free drain + re-key migration;
use-prompt-actions covers rejected-drain-keeps-entry + idle retry sends.
- Resume: fire the REST transcript prefetch and the session.resume RPC in
parallel, and skip the redundant message conversion + reconciliation
when the prefetch already hydrated the transcript.
- Haptics: web-haptics builds its AudioContext lazily on first trigger,
paying the ~850ms CoreAudio spin-up on the first streamStart haptic as
the first token paints. Open/close a throwaway context at idle so the
real one connects to an already-warm audio service.
During a token stream $messages is replaced ~30x/s. Subscribing the whole
chat view to it re-rendered the composer, runtime boundary, and every
message on every delta.
- Derive coarse facts (empty thread? tail is user?) via nanostores
`computed` atoms so per-token flushes don't re-render their consumers.
- Move the $messages subscription + runtime wiring into a dedicated
ChatRuntimeBoundary; the composer reads $messages imperatively.
- Drive message rows off stable useAuiState selectors and a lazy
getMessageText getter instead of eagerly materialized text.
- Feed ResizeObserver entry sizes into measureClamp / FadeText and dedupe
the style writes, killing the read-write-read reflow cascade.
The terminal looked soft/heavy on every platform because the xterm
Terminal was built with allowTransparency: true, which drops the WebGL
renderer's opaque fast-path and bakes glyphs as grayscale-alpha coverage
for compositing over a see-through canvas. Our surface (--ui-bg-chrome)
is opaque and withSurface already paints it, so transparency was pure
blur for no benefit — VS Code keeps it off too. Also drop the Medium
(500) base weight for normal/bold (400/700) to match VS Code's metrics,
and remove the now-unused JetBrains Mono Medium face + woff2.
