DSH Live Voice preview — consent-bound manual Qwen audio turns

[Jstn-1g]

DSH Live Voice

I have published an experimental live-voice add-on for DeepSeek Harness:

https://github.com/Jstn-1g/dsh-live-voice

The preview deliberately starts with one bounded manual Qwen audio turn. It is
designed around an exact-Session consent boundary: microphone audio is sent only
after a visible destination/authority disclosure and a separate recording
gesture. The protocol is designed to export no Harness history, files, Workspace
instructions, tool schemas, memory, or arbitrary text. Assistant output remains preview-only;
a final user transcript can become an editable composer draft only through
another explicit action, and the plugin never submits or invokes tools.

Install the current preview

dsh plugin --profile web add github:Jstn-1g/dsh-live-voice#v0.3.0-preview.3

The immutable prerelease
points to exact commit 10bd8bf504cf17fb24523f2c18e9f8a586a41167;
its tarball SHA-256 is
016ce06f37e635413734cdb2aa814d1714651b92ff8ac92306bd56dfd64af33d.

Verified so far

• 21 test files / 221 tests, both Host/browser typechecks, script checks,
  builds, package lint, and the 27-file package audit passed on the release
  commit; protected CI is green.
• The exact packed release passed the official-CLI composition on a clean
  dsh-v0.1.1-rc.2 source build: Workspace/Session binding, disclosure, bounded
  fake-provider PCM input/output, both final transcripts, completed turn, and
  provider disposal.
• The exact release also passed the pinned dsh-v0.1.2-alpha.1 source-build
  authentication smoke: unauthenticated voice upgrade rejected with 401,
  private launch-token exchange, authenticated Web/RPC paths, and one bounded
  fake-provider turn. The published peer range admits this exact alpha only and does not infer later alpha builds.
• The official served-Web BFCache receipt covers real save/restore rather than
  reload, DSH stream reconnection, active teardown, draft preservation, and
  fresh Session/consent binding with synthetic audio and a fake provider.
• Independent security/diff and release-claim reviews found no blocker.

Contributions and exact test receipts wanted

This is a preview, not ChatGPT Live parity or a marketplace-ready release. The
remaining claimable paths are tracked in the
release gate:

• physical microphone/speaker, permission, and OS indicator teardown;
• independent reproduction of the served-Web BFCache receipt;
• one credential-backed Qwen roundtrip and live export-boundary receipt;
• packaged Desktop restart/uninstall cleanup in a throwaway VM; and
• a portable carrier for Harness's documented file:// + IPC Desktop path.

Please use a disposable DSH profile and, only for the live-provider gate, a
non-production credential with a strict quota. Never post credentials, captured
speech, Workspace content, transcripts, or unsanitized logs.

Upstream integration path

Because the Harness repository currently disables external pull requests, I
prepared a focused upstream-ready authenticated-upgrade documentation and
real-network regression patch on a fork. It adds no API and does not conflate
browser authentication with plugin consent or Session/Workspace authority:

master...Jstn-1g:deepseek-harness:docs/authenticated-upgrade-route

The exact commit is e4b7142a819e226159eb06205000855e85eb62c3; focused
tests, documentation checks, full lint/build/typecheck, and independent review
passed. Maintainers can cherry-pick or adapt it, and the existing owner comment
records the exact test matrix and asks whether this is the extension contract
the project wants documented upstream.

DSH Live Voice is MIT-licensed and is an independent community project, not an
official DeepSeek AI, Alibaba Cloud, or OpenAI product.

[denial123789]

Added the dsh-live-voice preview to the handbook ecosystem guide with explicit boundaries for recording gesture, destination disclosure, provider/session isolation, transcript promotion, and teardown. It remains a preview/discovery lead, not a marketplace or production recommendation. https://github.com/sandbaseai/deepseek-harness-handbook/releases/tag/v0.5.396

[Jstn-1g]

A concrete compatibility step is now merged: DSH Live Voice PR #18 makes the custom voice WebSocket participate in Harness browser-session authentication before any plugin carrier/consent checks run.

The published smoke hard-pins the exact clean dsh-v0.1.2-alpha.1 source tag (cd5ef814…), verifies its official Web artifact receipt, installs the packed plugin through the official CLI, requires an unauthenticated voice upgrade to return 401, privately exchanges the launch token for the Harness cookie, and completes authenticated root/client/workspace/session/voice paths plus one bounded fake-provider turn. CI, the existing rc.2 packed composition, independent exact-diff review, and focused auth/secret review all passed. The sanitized receipt is recorded in issue #14.

I then followed the extension seam through current master / dsh-v0.1.2-alpha.1 (cd5ef814…). The existing public contract is intentional: HostConnectionHandle.requestRejection() accepts HTTP or upgrade requests, and the official API Gateway authenticates through it before WebServer.registerUpgrade() hands the socket to a protocol.

Because external pull requests are currently disabled, I prepared a focused, fully verified upstream patch on my fork instead of opening one:

• reviewable comparison
• exact commit e4b7142a8

It adds bilingual dsh-client-connection guidance for authenticated custom upgrade routes and a real-network sibling-route regression proving 401 without a browser session, 403 for an untrusted Host or mismatched Origin, 101 with the authority-bound cookie, no pre-auth protocol acceptance, and route/protocol cleanup on plugin disposal.

Verification: focused Vitest 13/13; documentation typecheck 79 blocks; documentation sync 32/32; full lint/build/typecheck; independent exact-diff review clean.

The patch deliberately adds no new API and does not treat browser authentication as plugin consent, Session authority, or Workspace membership. The published Live Voice peer range also remains on rc.2; none of this is presented as broad alpha, live-provider, physical-device, BFCache, or packaged Desktop validation.

Maintainers can cherry-pick e4b7142a8 or adapt the comparison. Is this the supported extension contract you want plugin authors to rely on and the project to document upstream?

[Jstn-1g]

Compatibility update: v0.3.0-preview.3
is now published and supersedes preview.2. It fixes a real Session-isolation
defect: DSH client-side navigation could unmount both voice controls while the
old Session's shared controller continued holding its socket, browser capture
path, provider turn, playback, and transient state.

The immutable hotfix uses exact-Session UI leases and passed a clean
dsh-v0.1.2-alpha.1 official-Web browser regression: real BFCache restoration,
DSH stream reconnection, an active Session A-to-B SPA switch without
pagehide, old capture/socket/provider teardown, fresh consent, and zero
post-cleanup audio. The run used synthetic Web Audio and a fake loopback
provider. Node 22.22.2 and 24.19.0 full checks passed (21 files / 221 tests), and
the exact rc.2 packed composition was rerun successfully.

This updates two boundaries in my earlier comment: the package now admits
0.1.2-alpha.1 exactly, and the alpha BFCache/SPA path is directly tested;
later alpha builds are still not inferred. Credential-backed Qwen, physical
microphone/speaker behavior, packaged Desktop, and independent reproduction
remain open. Exact install command, commit, checksum, and evidence are in the
release notes. Focused tester reports for those remaining gates are especially
welcome; please do not post credentials, recordings, transcripts, or private
Workspace content.