SpeechKit v0.75.5

Highlights

• Registered-agent Voice sessions can select native AssemblyAI or Deepgram media while preserving the registered agent's conversation, tools and authorization. Provider-supplied instructions and tools cannot replace that authority.
• Native sessions recheck the owner's current cloud-processing consent before provider setup and throughout the conversation. AssemblyAI additionally requires explicit consent for provider recording; revocation or an unavailable authority stops the session.
• Interruptions and client cancellation stop pending registered-agent work, including before the first audio response. Retried callbacks retain consumed turn identity to prevent repeating agent effects.
• The Android SDK carries the selected media provider through direct and coinstalled sessions. Explicit native selection requires coinstall contract version 3; older clients may continue omitting this choice.
• The public source export includes the complete native session runtime and consent reader. Linux checks exercise cancellation, signed consent admission and the exported server, MCP and CLI builds. Provider audio and device acceptance remain separate release checks.

Bug Fixes

• release: include native registered Voice scope in 0.75.5 (#767) (5d150c5)
• server: verify a signed, replay-guarded edge envelope (#757) (6f3ace1)
• voiceagent: skip the tool bridge for remote-agent sessions (#759) (764ea81)

Full Changelog: v0.75.3...v0.75.5