Clawd on Desk: baixar versions anteriors

v1.5.0
Clawd v1.5.0 gives each phone its own pairing, moves Linux .deb updates to an
official APT source, and animates the pet while Codex delegated subagents work.
It also tidies the AI Coding Tools and Permissions & approval settings, adds
support for community themes with author credits, shows Codex plan progress, and
keeps Clawd's local server running when its default ports are unavailable.
Linux updates
- Official APT source (#1255, #1288; related #1221) — .deb installations
keep update notifications, but the button now reads "Update with APT" and opens
the setup guide instead of running a privileged installer, which could fail
when no desktop authorization window was available. The opt-in source at
https://apt.clawdondesk.com/ supports Debian 12/13 and Ubuntu 22.04/24.04 on
amd64. Its signing key is limited to this source through Signed-By, and a pin
limits the source to Clawd's two packages. A release notification can arrive a
few minutes before the source has the new version. AppImage keeps its in-app
updater, including when the folder it unpacks into is reached through a
symbolic link. Other packaged Linux installs, such as an unpacked directory,
show "Open Downloads" instead. Thanks to @Kris20030907 for reporting the failed
update.
> Linux .deb updates now use APT. Existing users can add the official source once using the migration guide, even if the old in-app update failed. Quit Clawd before updating, then reopen it. AppImage keeps its in-app updater.
> Linux .deb 改用 APT 更新。 旧用户可按迁移指南接入一次官方软件源，无需先修好旧版应用内更新。更新前请退出 Clawd，完成后重新打开。AppImage 保留应用内更新。
Phone pairing
- One token per phone (#1279; issues #487, #1281) — each phone now has its
own token, so locking the phone or restarting Clawd no longer leaves it holding
a token that was replaced while it was away and forces a new pairing. To pair,
open the pairing link on the phone and type the 8-character code from
Settings → Permissions & approval → LAN approval. The code is single-use and
valid for 10 minutes; "Refresh pairing code" makes a new one. A Safari tab and
an "Add to Home Screen" app pair separately and count as two devices, up to 20
including phones that have not connected yet. Not tested on a real iPhone.
Thanks to @Bynlk, whose real-iPhone testing in #663 showed why the old shared
token could not be rotated safely.
- Paired devices list (#1280) — the same card lists paired phones with their
last-seen time. "Revoke" removes one phone, and "Revoke old token" cuts off
every phone still using a link from before the update. "Regenerate Token" is
now part of "Reset phone access".
- Safer phone connections (#1289) — a phone connection that Clawd rejects,
for example because of a wrong token, could raise an unhandled error in Clawd's
main process if it sent a malformed or oversized message while closing. Such
connections are now handled safely, and a message from a phone may be at most
4,096 bytes (the limit was about 100 MiB). Phone preview is off by default, and
this needed a misbehaving device on your network.
Codex
- Delegated subagents animate the pet (#1270, #1283, #1284) — while a local
Codex session has delegated subagents working, the pet shows the subagent
animation (headphones for one, juggling for two or more in the default theme)
and the HUD row shows the session as running. If the parent turn ends first, its
completion cue (sound, taskbar flash, happy animation) is held back and plays
once, when the last subagent really finishes or the parent exits. It is dropped
if the parent starts a new turn or is interrupted first, and a subagent that
stays silent for 20 minutes is released. Clawd also reads recently active
session logs first, which cut the delay before a subagent shows up from about
6–10 seconds to 0.09–1.38 seconds in a small macOS measurement. This covers
local sessions only; WSL and Remote SSH Codex sessions are unchanged. Known
difference: only the pet and the HUD wait for subagents. The Dashboard,
Telegram and Slack completion messages, phone 

v1.4.0
Clawd v1.4.0 reorganizes Settings into five focused tabs, reduces idle CPU use
with sleep-aware rendering, and adds KiroCrew and Qoder CN integrations. It also
adds opt-in external permission review, read-only phone approval summaries,
local OpenCode/MiMo question answers, and herdr pane focus.
Settings
- Five-tab Settings redesign (#1218, #1220, #1179; related #969) — splits
General into General, Pet, Notifications, Sessions, and Permissions & approval;
moves Remote approval into the Permissions & approval tab with On this machine,
IM channels, and LAN approval sub-pages; preserves existing values and routes
old Remote approval navigation to IM channels. Quota reminders now link into
Sessions. Thanks to @YOIMIYA66.
Power and animation
- Sleep when unattended (#1206; issue #1202) — lets the pet roam briefly
after you leave and then sleep, while the falling and sleeping poses stop
consuming animation work. On Windows, the whole Electron tree averaged about
0.6% of one core in the one-hour post-fix measurement; the same-configuration
v1.3.0 baseline averaged about 23% over its 15-minute control run. Thanks to
@trvny for the report.
- Deep sleep pauses by default (#1212) — collapsing and sleeping pause even
when the optional low-power switch is off. The paused segment measured about
0.3% on macOS and 0.13% on Windows at 175 Hz, compared with about 5–6% and
22% before. These are whole-tree, single-core CPU percentages from the Clawd
theme; the macOS measurements used a 50 Hz display.
- Object-channel rendering and 24-step sampling (#1210, #1213, #1214;
test coverage #1216) — moves Clawd's thinking, long-lived, typing, and building
visuals through the Electron-verified <object> path and a 24-step sampling
grid without changing their appearance. On macOS, CPU fell about 57–73% for
active visuals and about 30% for idle visuals; on Windows at 175 Hz, idle fell
about 55% and active visuals were flat to about 24% lower. The same work
keeps long Windows theme configuration out of the renderer command line.
- Theme-specific sleep behavior (#1210, #1212, #1213, #1214) — Cloudling
switches to a static image on sleep; Calico and the official Hash Sage and
Whale-chan themes keep playing; third-party object-channel sleep visuals freeze
after roughly five seconds. These optimizations are enabled by default for
Clawd; external themes can opt in through rendering.objectChannelFiles /
rendering.svgChannel and rendering.discreteSampling in their rendering
declarations.
Permissions and approval
- Optional external review command (#1215; issue #995; follow-up #1223) —
runs a user-configured argv command before an automatic allow. Exit 0 keeps the
allow; every other outcome, including timeout, missing command, concurrency or
payload limits, stops automatic allow: it normally leaves the request to the
human card, while Codex Auto's claimed WSL/remote Bash/apply_patch requests
return to Codex's native approval. It cannot approve an ask, deny a request,
rewrite input, or grant session trust. The feature is off by default and
command-only: edit permissionReviewCommand and permissionReviewTimeoutMs
in prefs after quitting Clawd. DND, disabling the agent, disabling that
agent's approval takeover, or quitting Clawd returns an in-flight request to
native flow; turning off only the global local bubble does not when a
Telegram/Feishu remote-only channel remains available. Thanks to @hanzhe-one.
- Codex Auto for WSL and Remote Bash/apply_patch (#1188; issue #1186) —
Clawd answers a request immediately when its global automation would already
allow it; requests that still need a person remain in Codex, and Auto does not
show a blocking Clawd bubble for these non-local requests. Because Clawd cannot
read a remote transcript, this exception overrides Codex's automatic reviewer.
Thanks to @V1staz for the report.
- Local OpenCode and MiMo v1 questions (#1163, #1177; issue #757) — for
OpenCode 1.18.31+ in the 1.x line and MiMo 0.1.15+ in the 0.1.x line, local
q

v1.3.0
Clawd v1.3.0 improves how the pet follows Claude Code tool batches, Codex
compaction and approval routing, and WorkBuddy conversations. DeepSeek Harness
gains desktop support alongside its web integration. This release also adds
optional quota reminders and improves session history, official theme downloads,
pet sizing, Footprints, and Linux AppImage shutdown.
Claude Code
- Thinking between tool batches (#1123, #1150, #1153, #1154; issue #1018) —
returns the pet to thinking after a correlated tool batch, instead of typing
until the turn ends. Also fixes misordered queued prompts, tool results,
subagent callbacks, and delayed tool accounting. Requires Claude Code
2.1.280 or later. Thanks to @200780381.
- Session titles (#1127; issue #1125) — shows AI-generated Claude Code CLI
and VS Code extension titles in the HUD and Dashboard. Manual renames take
priority; a prompt's first line remains a temporary fallback. Thanks to
@draintovmasyan783-creator for the report.
- Design reactions (#1084) — the Clawd theme holds a palette and brush
during a /design turn and shows heart eyes when it completes. Requires
Claude Code 2.1.265 or later. Thanks to @sanzanazaman.
- Windows recovery and hook repair (#1108, #1070; issue #874) — keeps
restart recovery available when a long-idle Claude session starts working
again through its live PID cache, and automatically migrates owned env-style
hooks when a usable local Node executable can be found. Thanks to
@200780381 and @hanzhe-one.
Codex
- Auto approval routing (#1156, #1167, #1169, #1068, #1101; issue #1155) —
makes Auto the default and moves existing Intercept/Native settings to Auto
once. Clawd handles only local human-reviewed shell/file requests with exact
session/turn evidence; automatic reviewers, MCP/app requests, remote/WSL
requests, and uncertain evidence stay in Codex, without Telegram or Feishu
approval cards. Long turns and unchanged settings repeated after compaction
can retain that evidence within the bounded scan. Closing an unanswered
Codex bubble or cleaning it up does not manufacture a denial. Intercept
remains available in Settings → Agents → Codex. Thanks to @PeterShanxin
and @200780381.
- Compaction (#1110, #1129, #1160; issue #1109) — starts sweeping at
compaction start and stops sweeping on success or interruption.
Compaction yields to other conversations' thinking and work,
while their completion cues still play. Unreviewed hooks retain the JSONL
completion-only sweep. Thanks to @200780381 and @hanzhe-one.
- Session cards and titles (#1158, #1111; issue #1103) — hides recognized
Desktop ambient-suggestion threads while keeping ordinary side chats, and
retires ended sessions on SessionEnd unless a replyable completion mapping
still needs the row. Local native titles and renames refresh through the
normal monitor poll, usually within about 1.5 seconds. Thanks to
@200780381 for tracing the extra Windows row.
- Live progress and roaming (#1162) — keeps a current local Codex turn
alive while genuine model/tool progress continues during a long wait. Busy
sessions now pause automatic roaming for every agent. Thanks to @200780381.
- New-hook review notices (#1169) — when only the new state hooks await
review, the Windows tray notice explains that an already-trusted approval
hook still works and points to /hooks; Doctor reports pending review.
DeepSeek Harness And Buddy Sessions
- DeepSeek Harness web and desktop (#1128, #1130, #1131, #1132, #1133,
#1169; issue #1119) — supports the verified 0.1 and 0.2 minor families,
with 0.2 starting at 0.2.0-rc.2, and manages web plus macOS/Windows desktop
profiles from one Settings/Doctor row. Open the desktop app once to initialize
its profile. Clicking a desktop session opens the app window, without
selecting a particular conversation; untouched starts remain in Dashboard
until activity reveals them in the HUD. Persistent notices explain restarts
and, after a host-family change, Uninstall followed by Install. Only
Clawd-owned bridge

## v1.2.0

Clawd v1.2.0 adds experimental MiniMax Code support, OpenCode 2.x permissions,
DeepSeek Harness session details, safer optional approval automation, and new
theme behavior. Idle bubbles in the built-in Clawd theme now mirror on the right
half of the screen, where the pet appears by default. This release also improves
session recovery, hook health, notifications, and remote Codex monitoring.

### Agents And Sessions

• **MiniMax Code** (#1038, #1049) — adds an experimental, state-only integration
through Clawd's managed local plugin at `~/.minimax/plugins/clawd-state`.
MiniMax's hook timeout does not support blocking approvals, so Clawd does not
show permission bubbles for it. Enable the plugin in MiniMax with
`mcode plugin enable clawd-state@local` or its plugin panel. Thanks to
@xiaoshidefeng.
• **OpenCode 2.x** (#1045, #1053, #1067, #1079) — detects a v1, v2, or unknown
host before changing its `plugin` and `plugins` registrations. The v2 plugin
uses the new event API and a blocking permission hook. If host detection
fails, `CLAWD_OPENCODE_HOST` can pin the host version. Windows detection now
handles installation paths with non-ASCII characters. Thanks to
@xiaoshidefeng and @gzx19990101.
• **OpenCode approvals** (#1075, #1079; issue #1065) — reaches wildcard-bound
`opencode web` and `serve --hostname 0.0.0.0` over loopback, withdraws pending
bubbles when a session is interrupted, and scans compound shell commands such
as `a && b` for destructive-operation reminders and warning badges.
• **DeepSeek Harness** (#1046) — supports 0.1.5-rc.1 and rc.3 and forwards
session titles and context usage. Thanks to @ypjn.
• **Claude recovery and completion** (#1061, #1063; issue #1060) — idle sessions
no longer resume as working after a Clawd restart. Normally ended sessions,
including turns that ended while background work continued, stay ended after
a reboot. A trailing `SubagentStop` no longer cancels completion animation
or notification. Thanks to @KaiC5504.
• **Codex session cards** (#1032, #1074, #1079; issue #1073) — groups remote
Desktop rollouts by thread, omits the `memories` consolidation worker, and
prevents remote monitor restarts from inventing idle sessions or reviving
finished turns.
• **Session identity and process fixes** (#1031, #1051, #1052, #1055, #1056;
issue #908) — probes Claude transcripts by raw session ID, preserves Cursor
delivery time through slow metadata lookup, recognizes Linux Node processes
named `MainThread` and retitled Kimi Code processes, and avoids probing WSL
session PIDs on the Windows host. Thanks to @hanzhe-one.

### Dashboard, Themes, And Desktop

• **Claude hook health** (#1059, #1064; issue #898) — the Agents card reflects
hooks displaced by tools such as CC Switch, paused automatic repair, repeated
repair failures, and missing scripts, with an explanation when attention is
needed. Windows shows one tray notice when automatic repair pauses. Thanks
to @hanzhe-one.
• **Desktop and Settings fixes** (#724, #1071) — the first click reaches
Settings or Dashboard when Clawd is in the macOS background. Custom app
launchability uses one rule, and saving a permission URL waits for sync to
finish. Thanks to @200780381 and @52mzd.
• **Theme animations** (#1050, #1076) — opted-in idle animations mirror on the
right half of the screen; the built-in Clawd idle bubble opts in. Mini mode
gains selectable peek hold and sleep peek visuals, plus idle visuals that
appear only when selected. Thanks to @KaiC5504.
• **Whale-chan** (#1077) — adds an optional official theme downloadable from
Settings; its media is not in the installer. Rights in the original character
design, setting and upstream materials remain with their creators; the theme
credits ZipZipPipe and 上善无形 as named by
[Neko3000/deepseek-whalechan](https://github.com/Neko3000/deepseek-whalechan).
New animation and effects by 鹿鹿 ([@rullerzhou-afk](https://github.com/rullerzhou-afk))
are licensed under CC BY-NC-SA 4.0; see the theme package for f

v1.1.0
Clawd v1.1.0 adds OMP and Grok Build integrations, downloadable official
themes, Dashboard Quick Select, durable session recovery, and richer remote
workflows. It also hardens local HTTP ingress, user-owned configuration,
OpenCode plugin delivery, Remote SSH metadata, and cross-platform release
checks.
Agents And Sessions
- OMP (oh-my-pi) (#1013) — adds a first-class, state-only OMP extension
with distinct session titles and a two-stage completion protocol. Completion
waits for OMP's final agent_end decision, and switching or branching a
session retires the old Dashboard row. Clawd skips installation when the
community bridge is already present. Thanks to @TalexDreamSoul.
- Grok Build (#1004) — adds marker-owned, state + Notification command
hooks while leaving every permission decision in Grok. Lifecycle mapping,
turn ordering, Windows command handling, and uninstall preserve unrelated
hooks. Thanks to @pu-1205.
- Safer OpenCode delivery (#1026) — materializes the managed OpenCode
plugin into a verified, content-addressed generation under the user's home
instead of registering files from the application bundle. Ownership,
locking, tuple options, interrupted cleanup, symlinked homes, and uninstall
all fail closed without taking over foreign entries, including when Windows
reports a missing child beneath an existing file ancestor.
- Durable session recovery (#1010) — stores bounded, profile-aware Claude
session history so eligible conversations can be resumed after a reboot.
Legacy rows without verifiable profile provenance remain visible but cannot
be resumed. Thanks to @TalexDreamSoul.
- Better session context (#945, #999, #1002, #1006) — Dashboard cards show the
active model, Qoder titles remain stable, OpenCode restores deferred context
usage, and Windows session focus/deep links are more reliable. Thanks to
@easyhak, @brantshin, @Tsdsj, and @Free-LZJ.
Dashboard, Themes, And Desktop
- Dashboard Quick Select (#972) — after you assign its macOS/Windows
shortcut in Settings, it opens the full Dashboard with frozen 1–9 session
slots. It preserves drafts and IME composition, handles scrolling and focus
handoff, and supports physical number-row keys across layouts without
treating NumLock-off navigation keys as session jumps.
- Downloadable official themes (#1020) — Settings can download, verify,
install, repair, and uninstall optional official themes from the separate
rullerzhou-afk/clawd-themes release catalog. Hash Sage remains outside the
Clawd source/package tree; install ownership and ZIP extraction are bounded,
and uninstall explains that customizations and managed sound overrides are
removed.
- Settings and desktop polish (#924, #965, #966, #998) — stabilizes theme
menus under zoom, unifies buttons and switches, restores failed quota
toggles, preserves accessible holiday labels, and cleans up abandoned drag
gestures.
- Windows fixes (#779, #805, #978, #1002, #1008, #1023) — improves
interactive Remote SSH console launch, Cursor hook paths and chat titles,
non-activating hit-window clicks, Codex deep links, editor focus, and
fullscreen detection around Windows 11 shell islands. Thanks to @VonSdite,
@mantertius, @Free-LZJ, and @KaiC5504.
Remote Workflows And Reliability
- Reply to Codex from Telegram (#970) — completion notifications can route
a bounded reply back to the authoritative Codex session. Queued replies are
bound to the originating session and cannot drift into another thread.
Thanks to @FuZoe.
- Hermes over Remote SSH (#975, #1003) — Remote SSH can deploy the Hermes
plugin to multiple targets, preserve secure transport ownership, and route
supported permission callbacks without weakening native fallback. Thanks to
@xfurqan0.
- Remote SSH visibility (#996) — Settings shows the next reconnect time and
clears an expired deadline as a retry begins. Remote hook metadata can no
longer inject a local Windows Terminal HWND. Thanks to @undefined-moe.
- Hook and session reliability (#902, #981, #993) — installers r

v1.0.0
Clawd v1.0.0 adds local activity Footprints, expandable permission cards and an
overflow queue, signed macOS in-app updates, and Windows fullscreen auto-hide.
It also improves Codex session recovery, Claude background completion, agent
installation detection, and Settings navigation.
Local Activity Footprints
- Private activity history (#963) — Settings → Footprints shows Today,
This week, This month, and This year. It counts accepted agent signals and
distinguishes recorded quiet periods from periods Clawd could not observe.
Unsupported metrics remain a dash, never an invented zero.
- Local storage only — recording is on by default, with 14 local days of
minimized event tickets and 400 local days of daily totals and coverage.
Footprints does not store prompts, responses, commands, paths, project names,
or raw session/turn/tool identifiers, and adds no network, export, or sharing.
- Recording controls — turn off Record footprints to pause, or Clear
footprint data to remove the managed history. Delayed completion counts cannot
cross a clear or off/on boundary. Normal completion feedback still works.
When preference recovery pauses recording, the page explains that state.
- Clock changes — historical local hours stay visible after timezone changes,
including coverage without activity; daylight-saving gaps and folds retain
their distinct meaning.
Permission Cards And Desktop
- Expandable details and overflow queue (#937, #944, #949, #950, #953) —
inspect longer requests, plans, and questions while preserving input drafts.
Answerable questions open in detail when space allows. The queue navigates
requests without making decisions. Global Allow/Deny shortcuts refuse an
overflow target that is clipped or obscured by the session HUD, including
queue loading and failure fallback.
- Independent bubble placement (#918, #934) — permission and update bubbles
can follow Clawd automatically or on a preferred side, or stay at one of four
primary-display corners, while avoiding other floating surfaces.
- Windows fullscreen auto-hide (#973) — optionally hides Clawd and floating
surfaces during fullscreen use, including newly arriving local permission
cards. Leaving fullscreen restores only requests still pending. Remote
approval and configured auto-close keep their existing behavior.
- Desktop reliability (#928, #955, #957) — improves permission cleanup,
macOS theme-specific Dock/menu-bar icons, and animation/geometry settlement.
- Settings polish (#923, #925, #926, #929, #930, #931, #932) — improves
language picking, collapsible sections, focus rings, idle visual selection,
and permission automation controls.
Agents And Sessions
- TraeCode / Trae CN (#886) — experimental state-only hooks, including
Windows Sandbox command handling. Enable hooks in Trae CN's own Settings;
approvals remain in Trae. Migration preserves unrelated nested hook entries.
- DeepSeek Harness (#938, #962) — supports exact rc.2 and rc.6 version
contracts, safer lifecycle cleanup, and POSIX GUI-launch discovery. Real
API-backed session approval was verified with rc.6 source runs on macOS and
Windows x64; rc.2 Windows packaged evidence covers install/web boot and direct
bridge-client round trips. These are separate validation levels. See the
DSH guide for the experimental scope.
- Codex (#954, #959, #960) — restores scoped-session focus, adds an opt-in
cold-start switch, and separates active-turn timeout from idle cleanup.
Upgrades retain a longer effective legacy timeout when the Codex-specific
preference is absent. Delayed question/output records cannot revive a closed
turn or extend another turn's lifetime.
- Windows Codex hook reliability (#986, #987) — avoids Windows Defender
false positives caused by the retired PowerShell launcher, while retaining
managed-hook validation and migration. Thanks to @eugenewang5425.
- Claude completion and installation detection (#958, #961) — improves
background subagent completion and avoids mistaking leftover directories for
i

v0.16.0
v0.16.0 is an agent-integration, remote-notification, accessory-geometry, and
macOS release-hardening update. It adds the first experimental DeepSeek Harness
bridge, manual ZCode permission approval, Slack notifications, Kimi quota
visibility, and a Spanish interface. It also hardens OpenCode/Codex integration,
preference recovery, Windows fullscreen detection, and Developer ID packaging.
Agent Integrations And Permissions
- DeepSeek Harness bridge (#876) — adds a Windows-first experimental,
plugin-only integration for the web profile, with managed immutable
generations, state events, and manual Allow Once / Deny handling for ordinary
approval requests. ask_user_question, automation, foreign packages, unknown
mutations, and unsupported DSH versions remain fail-closed. Thanks to
first-time contributor @RS-Nocsi.
- Manual ZCode approval (#880) — adds the blocking PermissionRequest hook
and desktop Allow/Deny bubble while deliberately keeping ZCode outside global
and per-session permission automation. Foreign hooks remain the sole owner,
and no-decision paths fall back to ZCode's native UI. Thanks to @liugou27.
- OpenCode JSONC and Desktop bridge support (#899, #900) — writes the plugin
into the effective JSON/JSONC configuration and supports OpenCode Desktop's
Node utility-process bridge. Permission forwarding now requires the live,
owner-only Clawd runtime identity instead of sending reverse-bridge credentials
to scanned ports, while preserving no-decision fallback to the native
permission flow when Clawd is unavailable or intentionally silent.
- Codex and Gemini detection fixes (#897) — installation discovery now
distinguishes real CLI evidence from directories created by other products or
by Clawd itself.
- Codex hook review stability and Claude hook repair (#870, #873) — keeps
official-hook trust stable across builds and safely handles environment-based
worktree hook paths.
Remote Approval, Notifications, And Quota
- Slack notification-only channel (#836, recovered and hardened in #909) —
sends bounded, ordered completion/error and permission-request announcements
by Incoming Webhook or bot token. Slack never makes approval decisions, link
unfurling is disabled, secrets remain outside prefs, and outbound summaries
avoid raw sensitive search content. Thanks to first-time contributors
@wang4433 and @shengmai-justin.
- Feishu/Lark approver lookup by email (#750) — resolves an approver through
the selected platform and binds the saved identity to that platform and App.
Thanks to first-time contributor @Cobb04.
- Safe Feishu/Lark upgrade guidance — configurations saved before platform
and approver provenance binding stay fail-closed. A one-time startup notice
and Doctor warning now explain the required repair instead of letting remote
approval disappear silently.
- Kimi subscription quota rings (#881) — adds bounded local quota refresh and
shared Dashboard/Orbit presentation for Kimi Code CLI.
- Safer session labels in remote output (#905, #909) — notifications use the
snapshot-owned display tag and suppress opaque internal workspace identifiers.
Desktop Runtime And Reliability
- Accessory-aware drag hitboxes (#866) — moving and animated accessories now
participate in the canonical geometry handshake, including mini mode and
holiday/accessory changes, so the interactive surface follows what is drawn.
Thanks to first-time contributor @CheeseAgent.
- Subagent activity tiers (#877) — juggling intensity follows live subagent
lifecycle identities instead of aggregate session count.
- Windows fullscreen detection (#889) — distinguishes maximized normal windows
from fullscreen applications so ordinary maximization no longer triggers the
fullscreen overlay policy. Thanks to @KaiC5504.
- Unreadable or damaged preferences are visible and still safe (#888, #891)
— Clawd no longer overwrites a prefs file it could not read. If readable
contents are malformed, it preserves the original as clawd-prefs.json.bak
and repairs the primary fi

v0.13.0
v0.13.0 is an agent-extensibility and runtime-reliability release. It adds
WorkBuddy and MiMo Code integrations, lets users register state-only custom HTTP
agents, adds a default idle-animation picker, brings remote approval to both
Feishu and Lark, surfaces Codex request_user_input questions on the desktop,
and hardens hook recovery and Windows/macOS window behavior.
This release welcomes seven first-time contributors.
New Integrations And Extensibility
- WorkBuddy support (#618) — optional Claude Code-compatible command hooks
support the current ~/.workbuddy-ai/settings.json location and the legacy
~/.workbuddy/settings.json location. WorkBuddy is state + Notification only:
approval remains in WorkBuddy's native sandbox and GUI, and Clawd never
registers a /permission hook. Thanks to first-time contributor @kkirito16.
- MiMo Code support (#607) — MiMo joins the opencode family through
@mimo-ai/plugin, with comment-preserving JSONC install/repair/uninstall,
deterministic handling of multiple config candidates, and the same
Allow/Always/Deny permission bridge and DND fallback as opencode. task child
sessions are headless and do not participate in the visible multi-session
animation fanout. Thanks to first-time contributor @jiaxuan1101.
- Custom HTTP agents (#652) — Settings can register local applications and
assign stable custom-... IDs for state-only /state reporting through the
runtime-discovered port. Registration does not install a hook or observe a
process automatically; disabled, removed, stale, and forged IDs fail closed,
while server-side session namespacing prevents two custom applications from
overwriting one another. Thanks to first-time contributor @200780381.
- Default idle animation picker (#679) — choose the idle visual used by a
theme while preserving eye-tracking capability rules. Thanks to returning
contributor @KaiC5504.
- Lark remote approval (#493) — the existing Feishu channel now supports
Lark tenants as a first-class platform, with real-tenant fixes and updated
event-subscription guidance.
Agent Runtime And Permission Reliability
- Codex desktop questions (#707) — supported request_user_input prompts
can appear as desktop question cards, including recovery after restart and
correct resolution cleanup. Recovery is bounded by age, file count, and byte
budget, keeps durable read positions, handles exact 1 MiB record boundaries,
and does not replay stale completion events. Thanks to first-time contributor
@JesmonX; rollout read-position and unavailable-session work also came from
first-time contributor @lurui1997 (#698, #700).
- Codex Desktop attribution and deduplication (#669) — Desktop originator
metadata is retained and Desktop sessions are counted separately where
needed. Thanks to first-time contributor @Dxy2326.
- Claude hook self-healing (#673) — all in-process Claude settings mutations
now run through one server-owned queue. Directory watching catches atomic
replacement, a periodic read-only health audit finds deleted scripts even
without an fs event, and three repeated failed repairs transition to a manual
fix state. Explicit Install/Fix only reports success after the entire declared
hook set verifies healthy.
- CodeBuddy permission ownership — local and explicitly configured custom
permission URLs remain supported, while unrelated HTTP hooks are preserved;
a bare hook named only clawd is never treated as owned.
- Kimi passive approval cues (#675, #680, #702, #703) — permission cards are
tool-aware, batched approval state is tracked without taking over Kimi's
native decision, dismissing a cue can focus the terminal, and the legacy
suspect heuristic is migrated safely. New legacy ~/.kimi installs now
default to persisted --permission-mode=suspect because current kimi-cli
versions do not emit explicit permission fields; an existing explicit choice
is preserved. Pre-approved gated commands lasting longer than about 0.8s may
briefly show a false-alarm cue. Set CLAWD_KIMI_PERMISSION_MODE=expli

v0.11.0
v0.11.0 is a user-facing release focused on first-run setup, desktop behavior,
Codex reliability, and release hardening. It adds the first-run onboarding
wizard, free-roam idle movement, the dizzy spin reaction, Windows fullscreen
overlay handling, Codex official-hook health surfacing, Remote SSH
connect-on-launch, safer hook settings backups, and a broad set of fixes around
completion celebrations, low-power wake recovery, Reasonix, Telegram approvals,
and Ghostty focus.
This release also welcomes seven first-time contributors.
New Features
- First-run onboarding tutorial (#524) - New users, and existing users who have never seen the tutorial, get a one-time setup wizard covering language, agent install/cleanup, shortcuts, and key features. Any dismissal marks it seen so it does not reopen on later launches.
- Free roam mode (#467, #570) - When enabled, the pet can wander around the desk while idle, with cancellation on state changes, mini-mode guards, hit-window/HUD/bubble sync during movement, and target-selection hardening. Thanks to first-time contributor @zhaoxv210, with a follow-up flake fix from first-time contributor @Git-creat7.
- Dizzy spin reaction (#507) - Rapidly circling the cursor around Clawd can trigger the existing dizzy animation on themes that support it, with cooldowns and theme-capability gating so unsupported themes keep normal idle behavior. Thanks to first-time contributor @IatomicreactorI.
- Windows fullscreen overlay mode (#538, #562) - Clawd can float over foreground fullscreen apps while keeping the hit window non-activating, so borderless-fullscreen games are not kicked out by clicks or drags. The old settings toggle was removed because it could not reliably force the pet behind borderless games; the hidden pref remains as an escape hatch.
- Codex official hook health (#571) - Since Codex approval awareness now depends on official PermissionRequest hooks, the Agents tab and startup nudge now surface disabled, unreviewed, missing, or broken Codex hook states using the same detector as Doctor.
- Remote SSH connect-on-launch (#561) - Remote SSH profiles can now connect silently at startup after the local hook server is actually listening, so reverse tunnels target the real bound port even if Clawd had to drift away from 23333. Thanks to first-time contributor @undownding.
- Settings UI cleanup (#521, #539, #541) - Settings uses consistent inline SVG sidebar icons, reorganizes General by usage frequency, moves Claude hook management into Agents, spells out Chinese language labels, and folds Animation Map into Animation & Sound.
- Context and tray menu cleanup (#523) - The pet right-click and tray menus are grouped into clearer state/work/display/app sections, the dangerous auto-approval entry is renamed to Auto-approve all requests, and New Session is clarified as New Claude Session across supported languages.
Bug Fixes And Hardening
- Hook settings backups before injection (#531) - Clawd now backs up settings.json before injecting hooks and keeps the backup set bounded by a rolling window. Thanks to first-time contributor @chrono-meta.
- Codex JSONL approval heuristic removed (#571) - JSONL fallback no longer guesses approval prompts from shell calls. It remains for state and metadata fallback events such as web search, compaction, aborted turns, and stuck local-turn rescue. Official PermissionRequest is the only approval path.
- Codex Desktop approval inference fixed (#543) - Codex Desktop sessions no longer synthesize passive approval notifications from JSONL activity. Thanks to first-time contributor @quantai1314.
- CJK Stop payload byte-fit (#567) - Long Chinese/Japanese/Korean completion text no longer overflows /state byte limits and drops the happy completion animation; hook payloads byte-fit assistant_last_output while preserving structural completion fields.
- Completion happy fixes (#548, #555, #567) - Short completions, background-only Claude stops, and busy Stop deliveries now re

v0.8.1
New Features
- Qwen Code integration - Added first-class Qwen Code support via ~/.qwen/settings.json command hooks. Clawd now tracks Qwen Code sessions, working / completion-attention transitions, model and cwd metadata, and blocking PermissionRequest approvals through desktop permission bubbles. Windows hook commands use an encoded PowerShell wrapper so Node paths with spaces survive Qwen's cmd.exe /s execution path.
- Antigravity CLI (agy) integration - Added state-only Antigravity CLI support through ~/.gemini/config/hooks.json. Clawd reflects agy working / idle / attention state while intentionally leaving permissions to agy's native 5-option terminal menu; PreToolUse is not registered so users do not get double-confirmation loops.
- Copilot CLI hook auto-sync (#341) - Copilot CLI hooks now auto-register at Clawd startup like Claude Code / Codex / Cursor — no more manual ~/.copilot/hooks/hooks.json setup. Path resolution honors $COPILOT_HOME (trimmed; empty / whitespace falls back to ~/.copilot); 10 events are covered, including the previously missing preCompact so compaction shows the sweeping animation. Doctor now uses a dedicated copilot-hooks mode that scans bash / powershell / command fields, validates per-event coverage, and reports disableAllHooks: true as a warning without offering Fix so Clawd never overrides an explicit opt-out. Marker-based merge preserves other hook entries and other hooks/*.json files.
- Hardware Buddy groundwork (experimental) - Hardened Clawd's adapter for an external Clawstick runtime, including the Settings panel, permission-reply toggle, internal quick-command event path, test approval flow, Unicode hardening, missing-runtime status, and BLE link preservation when permission routing is toggled. Clawd v0.8.1 does not bundle, download, or advertise Hardware Buddy as an out-of-the-box user feature.
- Quick Commands plumbing - Added an internal validated event path that adapters can use to enqueue preset actions for the active session. The v0.8.1 Settings panel does not expose local preset buttons yet.
- Background update checks (#329, #345) - Clawd can now quietly check GitHub releases on a scheduler, remember a pending version, surface the hint in About/tray/update bubbles, reconcile yanked releases, and keep manual update checks explicit.
- Task-complete tray flash (#335) - The tray icon can flash when an AI task completes, making completion visible even when the pet is covered.
- Session HUD click-to-reveal (#327, #337, #339) - Replaced hover auto-hide with a more deliberate click-to-reveal interaction and added optional HUD state labels so live sessions are easier to scan.
- Claude Code API error events (#338) - Claude API errors are now surfaced as ApiError events so Dashboard/HUD history and visual state have a clearer failure signal.
- Clawd idle animation refresh - Refreshed Clawd idle / reading / bubble assets and added the new bubble idle animation and GIF preview.
Bug Fixes
- Pi agent restored to state-only mode (#322) - Pi extension reports lifecycle and tool activity only. Clawd no longer shows Pi permission bubbles, no longer calls Pi terminal confirmation, and preserves Pi's default YOLO execution behavior.
- Legacy Pi permission hook compatibility - Older in-memory Pi extension instances that still POST /permission now receive an allow response from Clawd, so they do not get blocked by stale bubble logic while the user is upgrading.
- Claude cold-start hook blocking (#349) - Fixed a startup path where Claude hook handling could block during cold start.
- Codex native prompt sound muted by default (#346) - Codex fallback/native prompt paths no longer play the desktop prompt sound by default.
- Remote SSH interactive password prompt unblocked (#351) - Reordered SSH options so ssh -o first-wins behavior does not suppress interactive password prompts.
- Remote SSH deploy shell safety (#332) - In-app one-click deploy now blocks Windows cmd.exe remote shell paths. The script

v0.8.0
New Features
- Telegram remote approval - Approve Claude Code / CodeBuddy / Codex / Pi permission requests from your phone via a Telegram bot. New Settings → Remote Approval tab covers bot token entry, recipient pairing, sidecar status, and a manual test send. Local approval bubbles remain the primary surface; Telegram is an opt-in remote channel that falls back gracefully when offline, disabled, or under Do Not Disturb. Full guide at docs/guides/telegram-approval.md.
- Pinned sidecar binary distribution - The Telegram approval sidecar (cc-connect-clawd) is now fetched from a pinned GitHub release tag with sha256 checksum verification. Both source launches (npm start) and packaged builds use the same pinned per-platform binaries. Windows x64, Windows arm64, macOS x64, macOS arm64, and Linux x64 are covered.
- Refreshed Clawd thinking animation - New thinking bubble (v25) with softer edge-flow pixels and a regenerated GIF preview.
- Normalized agent icons - All twelve agent icons (claude-code, codex, copilot-cli, cursor-agent, gemini-cli, hermes, kimi-cli, codebuddy, opencode, openclaw, pi, kiro-cli) are now exported at a consistent 64×64 PNG runtime spec with sources tracked under assets/source/agent-icons/ and a stable manifest hash.
Bug Fixes
- Remote SSH Windows node resolution (#293) - Fixed node binary detection and SSH probe / hook timeouts when deploying from a Windows host. Added a dedicated remote-ssh-node module and tightened hook resolution edge cases.
- Windows Terminal focus startup race - Settings window no longer steals the foreground Windows Terminal HWND during the startup window; focus reports are now gated by hook event whitelist.
- Mini-mode multi-monitor seam clipping (#294, by @Rladmsrl) - Pet no longer bleeds across display seams during entry, reload, or hover hit-window. Subsequent follow-ups hardened the clip for startup and reload phases and aligned mini seam hover after clipping.
- Settings dashboard positioning (#287, by @YOIMIYA66) - Dashboard now anchors correctly within Settings bounds (top edge, bottom edge, parent offset, floating bounds, source forwarding all re-aligned).
- Agent icon source hash stability (#285, by @YOIMIYA66) - Icon exporter now writes the manifest atomically and uses stable source hashes so subsequent exports don't drift.
- Claude hook lifecycle hardening - Tightened managed Claude hook marker matching, avoided blocking on clawd-only hook repair, and preserved third-party hooks when settings.json suddenly shrinks (#286, by @taehwanis).
- Cold-start Claude watcher seeding (#286, by @taehwanis) - Addressed review feedback for snapshot seeding so the suspicious-shrink heuristic doesn't false-positive on a fresh boot.
- Windows uninstall Claude hook cleanup - NSIS uninstaller now runs a PowerShell script that removes managed Claude hook entries from ~/.claude/settings.json, with a write-once backup.
- Telegram approval Settings repaint race - Fixed status badge flicker on the Telegram tab; tokenInfo and status refreshes are now sequence-guarded.
- Non-focusable HUD sessions - HUD sessions that should never receive terminal focus are now tagged so focus heuristics skip them.
Release & Packaging
- Pinned sidecar fetch workflow - New npm run fetch:sidecars and npm run verify:sidecars flows; prebuild* scripts now run the verifier before every electron-builder invocation.
- Tag-gated draft releases - GitHub Actions release workflow only fires on v* tags, creates GitHub releases as drafts, and marks hyphenated tags (e.g. v0.8.0-rc1) as prereleases. Release notes are pulled from docs/releases/release-<tag>.md.
- Per-platform packaging resources - package.json extraResources ships the full bin/cc-connect-clawd/ tree into the app bundle; Windows NSIS x64/arm64 artifacts remain architecture-specific.
- Sidecar upstream update policy - New docs/project/release-process.md documents how to bump the pinned sidecar release tag and refresh checksums.
Docs & Contributors
- Telegram appro

v0.7.1
New Features
- Hermes Agent plugin integration - Clawd now supports Hermes Agent via a Python plugin with session boundary hooks, terminal focus metadata, WebUI session metadata, a dedicated agent icon, and default enablement for new installs.
- Pi Agent extension integration - Pi is supported through a global extension under ~/.pi/agent/extensions, with permission bubbles for bash/write/edit operations and a dedicated agent icon.
- OpenClaw state-only integration - OpenClaw is supported through ~/.openclaw/openclaw.json plugin path for state reporting, with a session icon and icon attribution. Phase 1 covers local openclaw tui --local.
- Settings Doctor panel - New diagnostics modal in Settings that checks agent integration health, hook/plugin installation status, and feature coverage, with collapsible detail sections and attention indicators.
- Chinese (Traditional) translation (#280) - Full zh-TW locale coverage for Settings, Codex Pet dialogs, question navigation labels, and i18n tests.
- Session HUD auto-hide and pin (#210) - Session HUD now defaults to auto-hide and exposes a pin toggle; settings are collapsed into a summary group.
- One-click remote SSH tunnel (Phase 2) (#252) - Enhanced remote SSH deployment with improved profile management and IPC.
- Copilot CLI remote SSH deployment (#284) - Remote deploy now copies and registers Copilot CLI hooks when ~/.copilot/ exists on the remote host, marks those hooks as remote, and carries host prefix metadata while skipping local-only PID collection.
- User theme zip package import - Settings can now import user theme zip packages directly, with import queue semantics and validation.
- Copilot CLI session name forwarding (#267) - Copilot CLI hook now forwards the session name to Clawd for better session display.
- Permission auto-close fallback (#273) - Stuck hook sockets now auto-close after a configurable timeout instead of blocking indefinitely.
- Custom scrollbar styles (#243) - Settings window now uses custom scrollbar styles for a more polished look.
- Go to Terminal workspace focus - "Go to Terminal" now focuses the originating workspace when using Superset deep-link.
- cmux workspace switching (#259) - Session focus now supports cmux (terminal multiplexer), resolving cmux binary path via Spotlight (mdfind), matching panel UUID, and searching all cmux windows for the target TTY.
- Settings custom language picker (#283) - Language selection in Settings now uses an accessible custom listbox with keyboard support and polished light/dark styling.
Bug Fixes
- wmic deprecation (#268) - Replaced all Windows wmic process lookups with PowerShell Get-CimInstance across hooks, pi-extension, opencode-plugin, and state modules, fixing deprecation warnings and failures on newer Windows builds.
- macOS hook node path detection - Fixed node binary path resolution for Claude hooks on macOS.
- Windows PowerShell terminal focus - Fixed terminal focus failing when PowerShell was the active shell.
- Fullscreen window pet recovery - Pet now recovers correctly after fullscreen window switches on Windows.
- Hermes CIM process lookup - Replaced Windows process lookup with CIM snapshot for Hermes agent.
- Sound settings toggles - Hardened sound settings summary controls sync and compatibility.
- Theme card keyboard events - Isolated theme card button keyboard events to prevent interference.
- Summary chips compatibility - Preserved summary chip clearing compatibility and fixed gap rendering.
- Animation override controls - Stabilized animation override pending sync, wide hitbox affordances, and stale state cleanup.
- Bubble policy summary inline - Kept bubble policy summary inline instead of wrapping.
- Low power idle CPU - Fixed excessive CPU usage during low-power idle mouse-follow.
- Mini mode crabwalk - Fixed reverse mini crabwalk during pre-entry animation.
- Doctor panel polish - Multiple rounds of review feedback, collapsed detail sections, timestamp display, and attention check 

v0.7.0
New Features
- Gemini CLI hook-only integration (#224, #225, #226) - Gemini now uses the current hook path instead of the legacy session-file polling path, with agent Settings controls, installer updates, event mapping, PID detection, diagnostics, and targeted tests.
- Codex Pet compatibility MVP (#235) - Clawd can import Codex Pet packages, adapt their sprite atlas/state metadata into Clawd themes, sync imported pets into the theme list, refresh/remove imported pets, and switch to a newly imported pet from Settings.
- Fade theme switching (#237) - theme changes now fade the pet window between themes instead of snapping directly to the new asset set.
- Theme and animation Settings polish (#234, #236) - Settings tab switches are smoother, animation override timing sliders are easier to use, loading states are clearer, and imported Codex Pets are grouped separately from built-in and user themes.
- Session controls (#227, #229) - Session HUD and Dashboard now show renamed Codex session titles, expose stronger HUD display toggles, support hiding sessions, and avoid letting session cleanup push the pet into sleep.
- Windows Terminal focus hardening (#242) - terminal focusing now avoids ambiguous title matches, requires unique Windows Terminal tab matches, preserves direct terminal focus when it is already available, records richer diagnostics, and documents Windows Terminal's matching limits.
Bug Fixes
- General Settings switch stability (#238) - General tab switches now stay mounted during settings updates, preventing transient unmount/remount behavior while Settings patches are applied.
- Gemini diagnostics and event behavior - Gemini hook events now map consistently into Clawd states, installer tests isolate their home directories, and hook diagnostics are easier to interpret when setup fails.
- Codex runtime stability - Codex PID reachability checks are more reliable, renamed session titles are reflected in the UI, exit probe logging improves debugging, and same-named Codex Pet assets reload correctly after import or refresh.
- Renderer crash logging (#220) - renderer-side crash logging now guards against EIO write failures instead of letting logging errors cascade.
- Agent submenu toggle flash (#233) - toggling an agent from Settings no longer briefly flashes the submenu into an inconsistent state.
- Auto-start launch config (#240) - auto-start now shares the same Electron launch configuration path as normal startup, reducing drift between manual and automatic launches.
- DND permission fallbacks - Do Not Disturb now preserves agent-native permission fallback behavior instead of accidentally swallowing or auto-answering permission requests.
Refactoring & Tests
- Codex Pet importer coverage - adds importer, adapter, protocol, theme sync, and fixture coverage for atlas-based Codex Pet packages.
- Gemini hook coverage - adds hook and installer tests around hook-only behavior, event mapping, and isolated home directories.
- Windows focus coverage (#242) - adds focused regression tests for Windows Terminal title matching, direct terminal focus, ambiguous-match handling, session focus helpers, and Doctor log metadata.
- Settings browser coverage (#238) - expands Settings renderer tests around General tab patch paths and switch persistence during store updates.
- State, HUD, and Settings coverage - expands tests for session hiding, detached cleanup, notification gating, server ring buffers, animation override UI behavior, theme transitions, and Settings browser-environment behavior.
- Doctor and integration checks - refreshes Doctor integration descriptors and tests so the new hook/plugin paths are represented in diagnostics.
Docs & Contributors
- Release positioning - v0.7.0 is a feature release because it adds Gemini hook-only support, Codex Pet compatibility, theme transition behavior, and deeper Windows Terminal focus behavior on top of normal fixes.
- Windows Terminal limitations (#242) - docs now explain why Clawd only s

v0.6.3
New Features
- Clawd Doctor - adds a diagnostics modal in Settings for local server health, agent integrations, permission bubble policy, theme health, and live hook connectivity checks. Doctor can copy a redacted diagnostic report, open the debug log, re-run checks, restart Clawd when the local server is unhealthy, and repair supported integration issues such as missing hooks or disabled Codex hook features.
- Cloudling built-in theme - ships a third built-in theme with full state coverage, mini-mode assets, direct sleep/DND transitions, pointer-aware idle and mini-eye tracking, theme-specific hitboxes, reaction assets, preview GIFs, state-mapping docs, and bundled source/spec notes for future polish.
- Japanese localization (#197) - adds Japanese UI strings across menus, Settings, Dashboard, permission bubbles, Doctor, and a new Japanese README.
- Session HUD completion bell (#217) - completed sessions now show an unread bell in the HUD until the row is opened/focused, then the bell reappears on the next completed run.
- iTerm2 tab-level focus (#216) - clicking a session can now switch to the matching iTerm2 tab on macOS instead of only raising the iTerm2 window.
- Codex permission mode control (#200) - Codex can stay in native mode so AutoReview/native terminal approval remains in charge, or switch to intercept mode for Clawd Allow/Deny bubbles. DND, disabled agents, and disabled bubbles now return no-decision so Codex falls back to its native prompt.
- Settings-native Claude hook confirmations (#180) - Claude hook disconnect/disable decisions now happen inside the Settings modal, keeping the hook management flow in one place.
- Animation override preview improvements - Settings preview posters now render through an isolated async preview surface, and animation override data merging is split out for safer future changes.
- Theme authoring upgrades (#207) - themes can define file-specific hitboxes, users can override wide-hitbox behavior per asset, a Clawd working-typing boss variant asset was added for future theme work, and the theme guide now documents Cloudling's built-in-only scripted SVG runtime boundary.
Bug Fixes
- AskUserQuestion and elicitation bubbles (#204, #206, #209, #223) - keyboard focus now lands in elicitation bubbles, ArrowUp/ArrowDown can move between preset options and the Other textarea, stale bubbles are dismissed when the user answers in the terminal, initial auto-selection was removed so Enter cannot submit an unread default choice, and Settings now shows the built-in bubble keyboard mappings.
- Long permission tool names (#198) - long MCP tool labels now stack below the title, ellipsize cleanly, and marquee on hover instead of wrapping the bubble title into an unreadable column.
- Codex desktop cleanup (#192) - stale Codex desktop sessions now emit SessionEnd, preventing long-lived desktop runs from accumulating idle zombie sessions until restart.
- Codex subagent handling - official hook role extraction, fallback monitor classification, remote callback handling, and conservative default-role behavior now suppress subagent completion noise instead of incorrectly completing the main session.
- Disabled agents during hook sync (#194) - startup integration sync, Claude hook watcher recovery, and hook ingress now respect the agent gate instead of re-syncing or accepting events for disabled agents.
- Async Claude hook management (#219) - Claude hook installation, version probing, settings actions, and related UI pending states are now async/serialized to avoid blocking Settings or racing broadcasts.
- Mini mode edge-pinned resizing (#213) - resizing now works while the pet is pinned in mini mode, including proportional sizes and cached work-area handling near screen edges.
- Session HUD and bubble placement polish - fixes HUD anchoring, keeps update/permission surfaces better aligned with the pet, and avoids stale elicitation or permission UI lingering after terminal-side progress.
- Cloudling pol

v0.6.2
New Features
- Session Dashboard & Session HUD — Clawd now tracks every AI coding session it sees and surfaces them through two complementary surfaces: a multi-row clickable HUD docked next to the pet (so you can glance at all live sessions while continuing to work) and a full Session Dashboard (opened from the pet shortcut, tray, or Settings) with second-level elapsed timers, session aliases for naming runs, and an always-on session-id card badge replacing the older showSessionId toggle. The legacy "Sessions" submenu has been retired in favor of the dashboard.
- Codex CLI official hook integration — Codex CLI now uses proper state hooks and permission approvals as the primary path on every supported platform — including Windows, which previously had to rely on JSONL polling. State hooks drive the pet's working/notification state directly, permission approvals route through the same bubble UI as Claude Code / CodeBuddy (with safe fallback guards for stale prompts), and the existing JSONL polling is kept as a fallback for older Codex builds that don't expose hooks yet.
- Windows ARM64 native installer (#179) — first-class native .exe installer for Windows on ARM, alongside the existing x64 installer.
- Windows-on-ARM x64 → ARM64 migration prompt (#183) — packaged x64 builds running under ARM64 translation now detect the situation and prompt the user to download the native ARM64 installer directly from the latest GitHub Release, never silently cross-grading through electron-updater.
- Low power idle mode (#175) — opt-in mode that further reduces CPU/GPU work while the pet is idle, on top of the platform-wide idle improvements in #167.
- Categorized bubble controls (#165) — permission, notification, and update bubbles are now configurable as separate categories in Settings, with hardened policy controls and clearer auto-close behavior.
- Per-agent Notification hook toggle (#155) — each agent can now decide independently whether the Notification hook raises a bubble. Master-off greying logic for sub-switches (#156, #157, #158) makes the resulting state legible at a glance.
- Theme-aware update visual flow (#159) — the update process visuals now adapt to the active theme, with hardened overlay rendering and a restored PermissionRequest resync path.
Bug Fixes
- Taskbar entries lingering after Explorer restart (#185 + follow-up 09b80fc) — floating pet/helper windows could re-appear in the Windows taskbar after explorer.exe was restarted, because shell state is recreated and creation-time skipTaskbar no longer holds. A shared src/taskbar.js helper now reasserts setSkipTaskbar(true) via the existing Windows topmost watchdog, with the same treatment extended to the context-menu owner window.
- Multi-question AskUserQuestion bubbles (#191) — elicitation prompts with several option groups now use a stepper and clamp overflowing form content inside the bubble, keeping labels, radio controls, and the Submit action reachable.
- Bubble auto-close timers sync with Settings (#173) — changing the auto-close duration in Settings now retimes already-visible bubbles instead of leaving them on the old schedule.
- Settings collapsible interaction stability (#171) — Settings collapsible sections no longer flicker / lose focus during quick toggles.
- Updater bubble anchors across themes (#164) — update bubbles anchor consistently regardless of the active theme.
- Session HUD shadow shell rebalanced (#172) — HUD shadow geometry retuned so the docked HUD sits cleanly against the pet across themes.
- Linux Electron sandbox explicitly disabled (#168) — sandbox state is now declared rather than implicit, with matching startup notes.
- hideBubbles drops the HTTP connection for Claude Code / CodeBuddy (#162) — hiding a permission bubble now also closes the underlying HTTP connection, so dismissed prompts no longer keep the agent waiting.
- Update bubble overlap with permission bubbles — the update bubble now yields layout space when a permission bubble

v0.6.1
New Features
- Kimi CLI support (#138, #153) — Clawd now supports Kimi CLI through auto-synced hooks in ~/.kimi/config.toml, shared session/state mapping, and permission-request signaling that preserves terminal fallback behavior
- Sound controls in Settings (#148, #151) — per-theme sound effect overrides, a dedicated Sounds subtab alongside Animation Overrides, and a global sound-effect volume slider make theme polish much more configurable without editing files
- Keep size across displays (#136) — optional size persistence now keeps the pet's pixel size stable when moving between monitors with different DPI / work-area characteristics
- Permission bubble elicitation improvements — AskUserQuestion bubbles now support an "Other" custom input path and Enter-to-submit behavior for preset-radio flows, making the approval UI usable for more real agent prompts
Bug Fixes
- Windows Claude Code hook execution (#147) — Clawd now registers Claude Code hooks with shell: "powershell" on Windows, avoiding the bash/MSYS/WSL command-path failures that caused MODULE_NOT_FOUND hook errors for some users
- Kiro CLI on Windows (#145) — Kiro hook execution and template sync now work on Windows instead of assuming Unix-like shell behavior
- Startup position recovery — launch-time geometry recovery now checks the saved display snapshot before regularizing bounds, materializes fallback bounds earlier, and more reliably pulls an off-screen pet back into view after monitor changes or docking/undocking
- Codex session rescue hardening — stale attach state, historical replay suppression, active-session windowing, and long-lived session detection were tightened so Codex polling is less likely to replay old work or miss slow sessions
- Settings window reliability and polish (#130, #131, #133, #134, #135, #137, #144) — Windows taskbar icon/AppUserModelID handling is fixed, taskbar relaunch reopens Settings correctly, size-preview lag/alignment is cleaned up, and macOS shortcut labels render more reliably
- Permission cleanup edge cases — terminal-approved permissions are cleared more precisely, cleanup matching is less aggressive, and hidden/stale permission UI is less likely to linger after the real workflow moved on
- Calico position alignment — idle-follow, collapsing, sleeping, and waking offsets were retuned so the built-in Calico theme no longer visibly jumps during DND sleep transitions and wake-up
- Sound override resilience — preview fallback, stale override reset, and user-picked filename display are now consistent even when an override file goes missing on disk
Refactoring & Docs
- Settings renderer split into focused modules — the large settings renderer was split into bootstrap/i18n/core plus tab modules, making future panel work easier to review and safer to change without cross-tab regressions
- Docs refresh — AGENTS/runtime notes, supported-agent guidance, and Codex/WSL clarification were updated to match the current integration model
Contributors
Huge thanks to everyone who shipped code, reviewed behavior, or improved release quality for v0.6.1:
- @hhhzxyhhh — Kimi CLI support groundwork (#138)
- @Rladmsrl — per-theme sound effect overrides (#148)
- @zeus6768 — sound effect volume slider (#151)
- @YOIMIYA66 — Windows Settings window/taskbar polish and size bubble UI follow-ups (#130, #131, #133, #134, #144)
- @JaeHyeon-KAIST — keep-size-across-displays and related Settings/runtime polish (#135, #136, #137)
Known Limitations
- Kimi permission classification is partly heuristic — explicit approval flags are honored directly, but some Kimi tool runs still rely on a short delay-then-promote heuristic to decide whether the pet should switch from working to notification
- Codex on Windows still uses JSONL polling — Windows Codex CLI hooks remain unavailable, so Codex state updates there still carry the existing polling delay

v0.6.0
New Features
- Settings panel overhaul (#95, #97) — introduces a dedicated multi-tab settings window and moves growing controls out of crowded menus. v0.6.0 adds General preferences, Agent manager, per-agent bubble toggles, Theme management, Animation overrides, Reaction cards, Global shortcuts, Edge pinning, and an About tab
- Theme ecosystem — fully pluggable theme system replacing hardcoded constants. Themes now define state-to-asset mapping, timings, hitboxes, sounds (#84, thanks @Rladmsrl), fade behavior, and capability metadata in theme.json; external themes load from the user config directory (%APPDATA%/clawd-on-desk/themes/ on Windows, ~/Library/Application Support/clawd-on-desk/themes/ on macOS, ~/.config/clawd-on-desk/themes/ on Linux) with mixed-format rendering (SVG + GIF + APNG + WebP); contributor tooling now includes a skeleton template, validation CLI (scripts/validate-theme.js), scaffold CLI, and capability badges
- Built-in Calico theme — ships a new calico cat theme with APNG/SVG mixed assets, eye/head tracking, mini-mode variants, and extensive sizing/positioning polish across different display setups
- Theme controls in Settings and menus — right-click theme switching now pairs with a Settings-based theme picker/delete flow, "Open Theme Folder", capability-aware animation overrides with import/export, per-file wide-hitbox toggle, reaction cards, and aspect-ratio warnings
- Sessions and permission UX upgrades — Sessions menu now shows official agent logos (#78, thanks @TaoXieSZ), real session_title names, and status badges; permission bubbles now show session folder + short id and support answer elicitation flows
- Display and windowing improvements (#77, #79, thanks @TaoXieSZ, @stickycandy) — proportional size mode (#77), Send to Display, edge pinning, virtual bounds support, smoother size slider/switch animations, space-switching flicker fix (#79), and a more predictable cross-screen drag path make multi-monitor setups behave much better
- Mini mode and update UX polish — adds mini-working typing animation (#121, thanks @sophie-haynes), improves mini entry timing, and replaces traditional update dialogs with pet-following update bubbles (#88, thanks @YOIMIYA66)
- Korean localization — completes Korean locale support alongside the existing English and Chinese experience
Bug Fixes
- DND no longer denies Claude Code permissions on behalf of user — DND mode now uses res.destroy() to let Claude Code fall back to its built-in chat confirmation instead of actively rejecting. Verified on CC 2.1.92+; older CC versions or future upstream changes to destroyed-connection handling may behave differently — if Clawd is in DND and CC hangs on a tool, open the CC terminal and answer there
- Bubble focus steal fixes (#75, #80, #98, thanks @stickycandy, @Rladmsrl, @Kevin7Qi) — permission bubbles now open with show: false, use type: "panel" (NSPanel) on macOS, restore topmost state more carefully for floating bubbles, and preserve the previously focused app more reliably on macOS hotkey flows (#80)
- Cursor false error animation (#74, thanks @TaoXieSZ) — postToolUseFailure demoted from error to working for Cursor Agent, since tool failures (file not found, grep no match) are normal workflow, not real errors; task-level errors still trigger error state via stop with status === "error"
- High-DPI drag and display edge robustness (#103, #124, thanks @sefuzhou770801-hub, @PeterShanxin) — drag now anchors to the main-process DIP cursor, portrait displays keep the pet readable, work-area queries guard against empty-display enumeration (#93), and Windows taskbar-edge topmost behavior is reasserted more reliably. A few remaining screen.getAllDisplays() callsites in mini mode will be hardened in 0.6.1
- Theme runtime hardening — fixes theme sanitization, cross-platform basename handling, SVG url(#id) preservation, theme switching/pet geometry stability, theme selection persistence, fallback override cards, and asset picker sta