Skip to content

v0.2.168

Choose a tag to compare

@github-actions github-actions released this 02 Sep 00:28
· 42 commits to main since this release
7079c56

v0.2.168 hardens voice reliability, makes A2A delegation durable across sessions, and adds observability for the new delegation path.

Voice & Delegation

  • Verbal cancel phrases now cancel live delegated tasks, clearing pending input-required questions and marking local handles canceled.
  • When an A2A task parks on input-required, ORBIS routes your next utterance as the answer to that task, preserving task ID and context.
  • Reconnecting to a session requeries all live outbound tasks, delivering results missed while away exactly-once and re-arming input-required questions.
  • The /a2a/callback endpoint now correlates push-backs by task ID to mark durable rows terminal, preventing duplicate delivery on subsequent requeries.
  • A new durable outbound-task registry persists task handles in memory schema v5, surviving barge-ins, disconnects, and restarts.
  • Loopback A2A delegates now register push notification URLs, allowing results to be pushed to /a2a/callback even after dispatch timeouts.
  • The presence loop now yields after two spoken check-ins instead of narrating indefinitely, speaking one explicit yield line before going quiet.
  • A dead-turn recovery line now plays when an LLM completion emits no text or function calls, replacing silence on truncated or empty responses.
  • Reasoning tags are stripped from streamed LLM text before TTS, preventing leaked reasoning blocks from being spoken aloud.

Performance & Observability

  • Every delegation dispatch now generates a Langfuse span and success/error counters, with A2A trace headers stamped on outbound requests for end-to-end trace stitching.
  • Session tool logs are now persisted to sessions.tool_calls, recording up to 500 function calls per session.
  • The micro-LLM filler tier now has a 300-token budget and 4.0s timeout, fixing silent failures when reasoning models consume the token allowance.
  • Canned fallback lines now play when micro-LLM progress generation fails or times out, ensuring the presence loop never falls into dead air.

Distribution

  • Docker runtime is verified before publish, and exact Docker image imports are gated.
  • Docker publication is now bound to verified refs, isolating PR Docker verification permissions.