dsh-deepcanary

When you run multi-agent workflows inside DeepSeek Harness, events from Sessions, Tools, Subagents, and the Host stream in constantly. Most of them are noise, but a missed signal—a Host becoming unreachable, a tool stuck in a failure loop, context running dry—can be costly. dsh-deepcanary answers exactly one question: does this moment deserve your attention? It does not re-order agents or perform high-impact operations on DSH's behalf; it simply gets attention allocation right.

The plugin uses a four-level attention model. C0 stays quiet, C1 enters the Inbox, C2 interrupts and writes to the Inbox, and C3 fires a strong alert. Adjacent events sharing the same root cause are merged into a Decision Bundle, and a rolling hourly budget, deduplication window, and quiet-hours policy keep reminder noise low. C3 escalation requires authoritative Host or Runtime evidence and does not depend on model judgment, so critical alerts are never silently swallowed.

In the DSH Web page the plugin defaults to a single sidebar entry and never obscures your workspace. Clicking it opens a resizable floating Inbox panel where you can acknowledge, snooze, mute, give validity feedback, or jump to the relevant thread. All panel copy follows DSH's Chinese or English locale in real time. The plugin also registers nine model tools and ten same-origin Web APIs for automated policy dry-runs, de-identified outcome logging, and diagnostics. State files store metadata and redacted summaries only; prompts, model outputs, credentials, and full session contents remain on the DSH side.

If you orchestrate several parallel agents in DSH and want an evidence-driven, replayable, well-bounded alerting layer without the noise, dsh-deepcanary is a lightweight but rigorous attention supervisor built for that job.