Presence And Attention Alignment Plan

Status: complete

Reference:

• docs/architecture/presence-and-attention.md
• docs/architecture/activity.md
• AGENTS.md core principle: shared context, soft guidance, no fake extra modes

Review Summary

The text side was already broadly aligned with the shared ACTIVE / AMBIENT model. The remaining drift was concentrated in the voice spoke, plus legacy settings compatibility and stale attention-era tests.

Confirmed Gaps

1. Voice reply admission still had multiple competing notions of “addressed to me”.

   • The shared decider, music overlay, and cancel/standing path were not reading the same cue source.

2. Voice generation still treated room-level ACTIVE like speaker-level follow-up continuity.

   • attentionMode === "ACTIVE" was suppressing eager/ambient handling even when the current speaker was not part of the live thread.

3. The classifier prompt was still missing the canonical shared-attention summary.

   • It had recency fragments, but not an explicit room-level ACTIVE / AMBIENT state plus current-speaker continuity.

4. Music wake behavior was still framed like a pseudo-mode in the classifier prompt.

   • The canonical design says music playback and wake latch are overlays, not a separate command mind.

5. Voice reply generation still threaded unused soft-addressing prompt context.

   • The generation prompt path was carrying voiceAddressingState even though the live prompt should ignore soft guesses.

Execution Plan

• Introduce a canonical voice attention context that exposes room-level ACTIVE / AMBIENT and keeps speaker-specific continuity as a separate signal.
• Refactor voice reply admission, prompting, snapshots, and dashboard types/UI to consume canonical attention terms.
• Make the ambient voice thought loop depend on AMBIENT attention instead of only silence-plus-transport heuristics.
• Remove unused voice addressing admission plumbing that no longer changes behavior.
• Remove legacy settings aliases from normalization and test helpers, then rewrite affected tests around the canonical fields only.
• Cut or rewrite legacy-label tests so the suite validates business behavior instead of obsolete internal wording.
• Run bun run test and bun run typecheck, then update this document to reflect final status and any remaining follow-up.

Completed Work

1. Canonical text attention and ambient thought behavior now use shared ACTIVE / AMBIENT semantics across admission, prompting, docs, settings normalization, and tests.

2. Voice attention now separates:

   • room-level attentionMode
   • speaker-level currentSpeakerActive
   • overlays such as music wake latch, pending command follow-up, and interruption recovery

3. Voice direct-address handling now uses one shared resolver.

   • deterministic direct address still uses exact wake-word / alias matching
   • softer name-cue detection now uses the canonical bot name only, which avoids alias-driven false positives in music gating and cancel standing

4. Voice generation no longer treats room-level ACTIVE as if the current speaker were automatically in-thread.

   • isEagerTurn now keys off currentSpeakerActive

5. The classifier prompt now explicitly reflects the canonical model.

   • shared attention state
   • whether the current speaker is in the active thread
   • pending command follow-up signal
   • interruption-recovery state
   • music wake as an overlay, not a separate mode

6. Vestigial prompt plumbing was removed from live voice generation.

   • voiceAddressingState is no longer threaded into generation-only conversation context

Verification

• bun run typecheck
• bun run test src/voice/voiceReplyDecision.test.ts src/voice/voiceMusicPlayback.test.ts src/voice/turnProcessor.test.ts src/voice/voiceSessionManager.addressing.test.ts --bail
  • completed as part of the default unit/integration suite run
  • result: 1086 pass, 0 fail

Completion Bar

This work is now complete:

• voice runtime surfaces ACTIVE / AMBIENT instead of legacy pseudo-modes
• overlays such as music wake latch and command follow-up are represented as overlays, not replacement attention modes
• voice thought generation only runs while ambient
• legacy normalization aliases used only for old saved settings are removed from the canonical runtime path
• affected docs and tests match the shipped model