v0.15.0

v0.15.0

v0.15.0 is a quota-legibility, free-roam, and platform-reliability release. The
subscription quota ring becomes readable at a glance — identity-based colors, vendor
glyphs, and a readout that follows whichever window is actually raising the alarm.
Free roam gains two independent ways to bound where the pet wanders. The OpenCode
family, Codex, macOS, and Windows runtime paths each closed a set of real-world
defects reported against v0.14.0.

This release also prepares the project to take over WinGet manifest publishing, after
discovering that x64 users had been receiving the ARM64 installer since v0.6.2, and
welcomes two first-time contributors.

Subscription Quota And Usage

• Identity-based ring colors and vendor glyphs (#863) — healthy rings are colored
  by source identity rather than headroom, so the rolling and weekly windows stay
  distinguishable instead of collapsing into one thick green band. The Claude coin now
  carries the vendor mark, taken from the same MIT icon upstream that supplies the
  other Lobe Icons agent glyphs; every other runtime icon is byte-for-byte unchanged,
  with provenance and hashes updated in source-manifest.json and the MIT attribution
  list in NOTICE.md.
• The readout follows the alert (#864) — the coin used to always print the rolling
  window while the ring colored itself from the tightest window, so a coin could read
  "1% 5h" while the inner ring sat at 61% amber. The title now hands over to the
  tightest window once it crosses the warning threshold, so in the common case the
  digits report the alert instead of leaving color to carry it alone. A flashback
  briefly replays the rolling number at the moments it is most likely to be asked
  about; color remains the only alert channel during that ~1.6s window, in the 60-85%
  band that does not pulse, and under prefers-reduced-motion. Each coin's glyph is
  scaled to its own artwork rather than one shared zoom.
• Used / Remaining display modes (#789) — a persisted preference selects whether
  the ring reports consumption or headroom, without changing quota ingestion or the
  warning thresholds. The same change consolidates reusable Settings buttons, warning
  dialogs, segmented choices, and dropdown behavior, and stabilizes Settings scrolling
  and dropdown lifecycles — including the theme-accessory case that previously lost its
  real scroll range after a selection. Thanks to @YOIMIYA66.
• Correct context windows for custom Claude models (#809, issue #797) — when local
  Claude quota collection is enabled, context usage is read from the Claude statusline
  rather than inferred, so a custom model no longer displays a context length that does
  not match the model actually in use. Collection stays opt-in (Settings → General)
  because it installs Clawd's own statusline; an occupied third-party statusline is
  preserved rather than taken over, and Clawd falls back to transcript-derived context
  when registration does not succeed.

Free Roam

• Constrain to axis (#795, issue #686) — an opt-in setting restricts idle wandering
  to horizontal or vertical movement only, for users who prefer grid-aligned motion
  over diagonal drift. Default off, so existing behavior is unchanged. Thanks to
  first-time contributor @weed33834.
• Optional roam fence (#810) — free roam can be bounded to a rectangle expressed as
  fractions of the work area, so the pet can be kept to, say, the bottom-right quadrant.
  The fence is re-read on every target pick, so edits apply live without a restart, and
  the minimum hop distance scales down so small fences still produce reachable targets.
  Thanks to @anthonyonazure.
• Hardened axis targets (#819) — axis-constrained targets are clamped against the
  same bounds as unconstrained ones, and the movement-style wording is clarified.

OpenCode Family

• Real session titles in the HUD (#841, issue #829) — OpenCode sessions show their
  actual title instead of the project folder, and the title updates after OpenCode
  renames a session. Titles are bounded, kept out of logs, stripped of Unicode
  bidirectional formatting marks before display, and carry no telemetry stamp. Thanks
  to @xiaoshidefeng.
• No premature idle during long active work (#853, issue #850) — an active agent
  could fall back to idle mid-task and release the session; the stale floor is now
  scoped to interactive sessions and no longer fires while tools are still running.
  Thanks to @PeterShanxin.
• Ordered state delivery and directory-scoped disposal (#855, #858) — /state
  requests for one session are serialized so an older lifecycle request can no longer
  arrive after a newer rename and restore the stale title, and
  server.instance.disposed clears only the disposing directory's sessions instead of
  every cached session process-wide. Sustained state updates queued behind an
  unresponsive local endpoint are coalesced to the latest snapshot and the retained
  backlog is hard-capped. Lifecycle and metadata barriers retain their order within
  that bound; SessionEnd replaces stale queued snapshots without overtaking the
  active request. Regression coverage locks these ordering and stale-floor boundaries.
• Session cwd bound to its owning session (#798, issue #796) — a session's working
  directory can no longer be attributed to a different session.

Codex Runtime

• Terminal turns fenced across event sources (#831, issue #821) — after a Codex turn
  ends, late PreToolUse / PostToolUse events can no longer pull the pet back into a
  typing state, and token_count telemetry is separated from session liveness so a
  metadata refresh cannot keep a stuck working session from timing out.
• Bounded rollout log reads (#820, issue #817) — rollout log consumption is bounded
  and recovery retries are hardened, closing a main-process error path.
• No flicker during Codex Pet drag transitions (#804, issue #620) — drag direction,
  release, and state changes switch animation rows inside the already-loaded SVG
  document rather than rebuilding the spritesheet, removing the intermittent transparent
  frame. (#803 reverts an earlier direct push of the same fix so it could land through
  review.)

Platform: macOS

• Menu bar and Dock settings restored (#851) — the menu-bar and Dock visibility
  preferences apply and persist correctly again.
• Pinning against physical display bounds (#826, issue #241) — edge pinning uses
  physical display bounds, so the pet can reach the bottom of the screen when the Dock
  is present and bottom pinning stays on-screen.

Platform: Windows

• Server-side process-chain resolution (#837, issue #694) — native Windows process
  ancestry queries (NtQueryInformationProcess primary, Toolhelp comparison) move
  eligible /state and Codex /permission process-metadata resolution into the server,
  with per-agent legacy | shadow | b1a-authoritative capability routing and bounded
  shadow parity diagnostics. In authoritative mode the legacy hook-side PowerShell
  snapshot is structurally skipped.
• Hardened FFI initialization and per-registry Koffi caching (#839, #840, issue
  #838) — process-query FFI initialization no longer fails open, and Koffi bindings are
  cached per registry rather than re-resolved.
• Target-native Koffi packaging and foreign-native audit (#824, issue #763) — each
  release target keeps exactly one architecture-matching koffi.node, and a native
  inventory audit rejects every foreign-architecture binary except the
  electron-builder-managed Windows ia32 elevate.exe helper.

Agents And Diagnostics

• QwenWork (千问办公) (#843) — a hook-only, state-only integration modeled on
  QoderWork. The pet reflects QwenWork lifecycle state while permission decisions stay
  entirely in QwenWork's native flow: PermissionRequest / PermissionDenied are
  observation-only and the hook always returns {}. Detection, install, and uninstall
  go through ~/.QwenWorkCN/settings.json. Windows and macOS only — QwenWork currently
  ships no official Linux client. Thanks to @xiaoshidefeng.
• Hermes WSL Pair and Unpair (#842, issue #540) — pairing and unpairing work against
  WSL targets.
• Unreviewed Codex hooks explained in Doctor (#854) — Doctor distinguishes "hooks
  are missing" from "hooks are installed but still need review in Codex /hooks", which
  was the actual cause behind several "no permission prompt appears" reports.

Dashboard, Remote Access, And Notifications

• Persisted Sessions/Dashboard window bounds (#807, #814, issue #801) — the window
  remembers its position and size, with hardened persistence against invalid or
  off-screen bounds. Thanks to @KaiC5504.
• Serialized Codespaces SSH transport (#845, issue #546) — concurrent Codespaces
  gh cs ssh --stdio activity no longer breaks Remote SSH deploy/probe on Windows.
• Safe Remote SSH reconnects (#808, issue #800) — reconnect after a dropped
  connection recovers instead of failing permanently.
• Safe HTML in Telegram messages (#802, issue #766) — message rendering escapes
  agent-controlled content instead of emitting it into Telegram's HTML parse mode.
• Distinct default completion sound (#833) — built-in themes now use a dedicated,
  softened completion cue instead of sharing the ordinary confirmation sound.

Packaging, Security, And Localization

• WinGet manifest self-publishing groundwork (#861, issue #860) — a prepare-only
  workflow generates the manifest komac would submit and gates it on an architecture
  contract. rullerzhou-afk.clawd-on-desk has been live in microsoft/winget-pkgs
  since 2026-04-20 without ever being submitted by this project: a third-party release
  tracker picked it up and, from v0.6.2 onward, produced manifests declaring two
  Architecture: x64 entries that both pointed at the ARM64 installer. The NSIS
  stub runs on x64, so the install reported success and the app then failed to launch.
  Automatic submission stays deliberately disabled until the upstream manifest is
  repaired by hand and komac's output is validated.
• npm dependency security hardening (#846) — dependency and package validation are
  tightened.
• Brazilian Portuguese (pt-BR) (#822, #827) — a full pt-BR locale, with references
  and language entry points synced. Thanks to first-time contributor @arismarioneves.
• Documentation accuracy (#828, #849) — AGENTS references are corrected, contributor
  credits updated, and runtime/safety boundaries synced with the code.

Contributors

Two first-time contributors landed changes in this release:

• @weed33834 — axis-constrained roam mode (#795)
• @arismarioneves — Brazilian Portuguese locale (#822)

Returning contributors: @xiaoshidefeng (#841, #843), @YOIMIYA66 (#789),
@PeterShanxin (#853), @KaiC5504 (#807), and @anthonyonazure (#810), whose accessory and
tint work shipped in v0.14.0 through co-author credit and who now lands a directly
authored change.

Upgrade Notes

• Launch Clawd once after upgrading so installed and enabled integrations can reconcile
  their managed hooks/plugins against the packaged v0.15.0 files.
• QwenWork is a new state-only integration and is not enabled by default. It requires
  the QwenWork desktop application on Windows or macOS.
• Quota ring display mode (Used / Remaining) defaults to the existing behavior; no
  action is required to keep the previous readout.
• Correct custom-model context windows require explicitly enabling local Claude quota
  collection in Settings → General. Clawd preserves an occupied third-party statusline
  and keeps transcript-derived context as the fallback when its own statusline cannot
  be registered.
• Free roam fence and axis constraint are both opt-in. With no fence file and the axis
  setting off, roaming behaves exactly as in v0.14.0.
• Windows installs that used a WinGet-provided package before this release may be on the
  ARM64 build regardless of their architecture. Reinstalling from the GitHub release
  asset for the correct architecture is the reliable path until the corrected WinGet
  manifest is live.

Known Limitations

• v0.15.0 Linux packaged artifacts have not been validated on real Linux hardware.
  Wayland/XWayland transparency, positioning, cursor tracking and tray behavior, plus
  MiMo JSONC permission and comment-preserving writes on POSIX, remain unverified for
  this release.
• Packaged macOS and Linux builds do not auto-update. Download future versions manually
  from GitHub Releases.

Validation Status

Release validation is complete. The whole-version release review, three independent
post-fix review passes, and the final cross-review found no remaining code-level
release blocker. The final local suite reported 7,751 tests: 7,716 passed, 35
platform/tool-conditioned tests skipped, and none failed. The v0.15.0 release
contract passed. The repository asset audit reported no errors and one non-blocking
tracked-tree size-budget warning.

The manual multi-platform Build & Release run for main commit
5cbe7d7f66383ce05da8a11e51f4f9b6a4e0666f completed successfully after retrying one
transient Linux dependency download failure. Windows, macOS, and Linux each reran the
full test suite and completed packaging, target-native payload audit, packaged Koffi
smoke, retired Telegram sidecar assertion, and exact-version updater metadata checks.
Windows produced separate x64 and ARM64 NSIS installers; their downloaded sizes and
digests matched latest.yml.

The online-built Windows x64 installer was installed over the system-level v0.14.0
installation on Windows 11 x64. The installer returned exit code 0; the installed
executable and uninstall registration both reported v0.15.0; the pre-upgrade prefs
hash was unchanged immediately after installation; and the installed resources kept
the expected unpacked hooks, agents, extensions, themes, single x64 Koffi addon, and
no retired Telegram sidecar. A normal packaged launch with an isolated fresh profile
returned a healthy local /state response, rendered an uncloaked pet, hit window,
Settings window, and first-run tutorial without an error dialog, and showed v0.15.0
in Settings -> About. Closing the tutorial through the OS persisted
tutorialSeen=true; a second cold start kept the pet and Settings visible and did not
reopen the tutorial. The candidate also bound the next available server port while a
development instance occupied the default port.

Windows ARM64 was validated structurally in CI but was not installed on this x64
machine. The final macOS and Linux packages were built and structurally validated in
CI but were not rerun on dedicated macOS or Linux hardware. Feature-specific
real-machine evidence recorded on the merged implementation PRs was retained rather
than recreating every external agent, remote host, approval service, and desktop
environment during the final package smoke.