docs: omni-modal always-on agent direction (benchmark: OmniPro)

docs/ROADMAP_OMNI_STREAMING.md

@@ -0,0 +1,24 @@
+# Roadmap — omni-modal, always-on agent
+
+CogniFold is extending toward an **omni-modal, always-on agent**: a system that
+continuously watches an audio + video stream and **proactively decides when to act
+and what to say**, instead of only responding when it is explicitly asked.
+
+## What we're building
+- **Omni perception** — jointly attend to vision, speech, and **non-speech sound**
+  (environmental audio), since real-world cues are inherently multi-modal.
+- **Proactive & streaming** — choose the moment to respond with no polling or fixed
+  schedule, and hold up over **long-horizon** (minutes-scale) streams.
+- **In-model working memory** — keep and consolidate state online, extending
+  CogniFold's memory story from discrete events to a continuous audio-visual stream.
+
+## Primary benchmark
+**OmniPro — A Comprehensive Benchmark for Omni-Proactive Streaming Video
+Understanding** (Zhao et al., 2026; [arXiv:2605.18577](https://arxiv.org/abs/2605.18577)):
+2,700 human-verified samples, 9 sub-tasks across 3 cognitive levels, 84%
+audio-dependent, evaluated under a dual **Probe** / **Online** protocol. We also
+track transfer to OVO-Bench and StreamingBench.
+
+## Status
+Early — this note states the **direction**. Design details and results land
+incrementally.