Skip to content

v0.2.0

Choose a tag to compare

@luochenw luochenw released this 17 Aug 18:31
· 51 commits to main since this release

Voice dictation, a floating status bubble, real concurrent sessions, and an experimental cross-session messaging capability.

Highlights

  • Voice dictation with a wake word: on-device speech recognition with a configurable wake word, live transcript streaming into the composer, and voice commands to end/cancel a turn. First mic use auto-enables the wake word; wake-triggered tasks dispatch into their own background session instead of hijacking the one you're looking at, with a floating bubble and notification that jump straight to the dispatched session.
  • Floating always-on-top bubble: a small status window that stays visible across spaces — live scene while a session runs, an unseen-completion dot when something finished while you were away.
  • Real concurrent sessions: per-session run state and live sidebar dots, so multiple sessions genuinely running at once no longer bleed into each other's status.
  • Settings overhaul: official vs. custom provider configuration cleanly separated, API key values visible inline (password-style, revealable) instead of write-only, desktop-app-style toggle rows.
  • Cross-session relay (experimental, off by default): a Settings toggle mounts a Host-side plugin giving the model list_sessions / send_to_session / notify_session_on_done tools, so one session can discover and message another directly. Off by default; enabling it lets your sessions coordinate with each other, not just with you. See the design note for the full threat-model discussion (trust-boundary disclaimers in every relayed message, rate limiting, human approval still required on the receiving end for anything sensitive).
  • Composer, transcript, and sidebar polish: full skill roster in the slash menu, native /export and /model, imported-folder chips, follow-scroll fixes, moon-transit running indicator, and a good number of smaller conversation/animation fixes.

Download (Apple Silicon)

DeepSeek-Harness-0.2.0-arm64.zip — requires macOS 13+ on Apple Silicon. Intel Macs: build from source (see below).

Gatekeeper note. This build is ad-hoc signed and not notarized (no Apple Developer account behind it). macOS will warn on first launch:

  1. Unzip, drag DeepSeek Harness.app to /Applications, double-click — macOS will block it.
  2. Open System Settings → Privacy & Security, scroll down, click Open Anyway.
  3. Or from a terminal: xattr -rd com.apple.quarantine "/Applications/DeepSeek Harness.app"

If you'd rather not trust an un-notarized binary (reasonable!), building from source takes one command:

./scripts/build-macos-app.sh

(needs Node + npm install -g @deepseek-ai/dsh on the build machine; the built app is self-contained.)

Upgrading from 0.1.0

Drop the new .app in over the old one — no config migration needed. The cross-session relay toggle defaults to off even on an existing install.

Verification

  • ./scripts/test-macos-native.sh — typecheck + native contract checks
  • ./scripts/build-macos-app.sh — full build, including a manifest integrity check that now also covers in-house plugin packages
  • ./scripts/test-macos-native.sh --smoke — boots the bundled Host and verifies read-only Host APIs
  • The cross-session relay feature was additionally verified end-to-end against a real model in an isolated Host config copy at every milestone, plus a dedicated multi-angle code review pass before merging — see the design note for details and what's deliberately deferred.