dsh-realtime-voice v0.12.0

Guarded hands-free follow-up submission for realtime voice.

• Follow-up speech captured during Harness reasoning or TTS is merged in the native composer.
• Submission waits for playback completion, speech end, merge window, configurable dwell, and a final commit guard.
• Continued speech resets the lease; long utterances cannot be cut by a guessed timeout.
• Manual edits, paste, clear, submit, stale phase/revision, voiceprint failure, cancel phrases, and sensitive commands revoke automatic submission.
• Adds composer phase/draftRev CAS, settings, docs, and regression tests.

Validation: npm run check; 89/89 tests; DSH creator-mode independent review PASS.

dsh-realtime-voice v0.11.0

Truly dynamic floor speech

• Healthy floor speech is now generated ephemerally by a lightweight model from the selected voice provider; it is no longer selected from a fixed candidate library.
• Composition starts in parallel with Harness at t=0. The floor delay only decides whether it may speak; the first Harness answer delta aborts and discards any stale cue.
• Harness remains the only reasoning, tool and result channel. Floor cues never enter Harness history and cannot see reasoning or tool payloads.
• Host-side privacy controls send only a redacted topic (up to 18 characters), a closed stage enum and up to three validated prior cues.
• Qwen uses the Workspace OpenAI-compatible Chat endpoint; OpenAI uses Responses. Both have model allowlists, 1.5s aborts and fail-closed output validation.

Verification

• 82/82 tests
• typecheck, production build and client bundle purity
• DSH creator-mode task session-aa40568a-2449-4d70-b777-24a850304615: PASS

Install

dsh plugin --profile web add https://github.com/zfu691531-hash/dsh-realtime-voice/releases/download/v0.11.0/dsh-realtime-voice-0.11.0.tgz

SHA-256: 53763b0de9648fd00de0d906527685055f954dcd67245f896da331cdf6727477

dsh-realtime-voice v0.10.0

Highlights

• Dynamic, trace-aware floor speech: each bridge line is composed from the current topic, user intent, Harness tool/retry stage and prior lines; it is not a fixed canned response.
• Optional Tencent Cloud voiceprint soft gate for Qwen PCM sessions. Enrollment occurs only after an actionable utterance, and enrollment speech is never sent to Harness.
• Rejected, unavailable or short voiceprint samples are staged in the native composer for explicit review instead of starting another Harness turn.
• Stronger lifecycle isolation: voiceprint requests cannot block ASR events, stale checks cannot submit after restart, PCM is bounded to 30 seconds and remains in memory only.

Verification

• 72/72 tests
• TypeScript typecheck
• production build
• client bundle purity check
• DSH creator-mode reviews: PASS

Install

dsh plugin --profile web add https://github.com/zfu691531-hash/dsh-realtime-voice/releases/download/v0.10.0/dsh-realtime-voice-0.10.0.tgz

SHA-256: cf52a165d7fecf4eb9d5f1aa74985c499fa3c5e9aef22848c56389619a0c7d58

Voiceprint is an interference-reduction signal only; it is not identity verification or authorization.

v0.9.0 — Natural voice floor management

Highlights

• clean first-paragraph voice output with no visible machine markers
• adaptive 800 ms floor cue for slow Harness turns; fast answers remain direct
• Harness trajectory guards for reasoning, LLM retry, and tool-call preamble invalidation
• Qwen TTS generation cancellation and connection timeout hardening
• native DSH bundle remains Harness-first: ASR/TTS never replace Harness reasoning or plugins
• new SECURITY, architecture package, example configuration, uninstall, and doctor documentation

Validation

• 54/54 automated tests
• typecheck, production build, and client-bundle purity
• DSH creator-mode contract review approved
• installed and loaded as 0.9.0 in a restarted real Harness web profile

v0.8.0 — Hybrid realtime turns

Restores idle ASR auto-submit through the native Harness composer while staging speech captured during reasoning or TTS for the next user-controlled turn. Tightens summary streaming so split, spaced, malformed, or missing boundaries can never leak marker fragments, reasoning, or detailed output into TTS.\n\nValidation: 45 tests and production client-bundle purity check pass.\n\nSHA-256: 94a5906d1771dacc6f4b226ebab0b27b600edef1a2fcd9eea2da3d52887f87c0

dsh-realtime-voice 0.7.1

Warning-free one-command DSH plugin release. The repository contains prebuilt runtime files, so GitHub installation needs no prepare script, TypeScript toolchain, or pnpm build authorization. React is resolved from the Harness client runtime as an optional host-provided peer.

dsh-realtime-voice 0.7.0

Independent DSH bundle plugin release. Includes prebuilt runtime files for one-command GitHub installation without prepare scripts or build authorization. Qwen/OpenAI provide realtime speech while the active Harness session remains the sole reasoning and tool-orchestration authority.