dsh-infra-observability

In multi-agent runtimes, the real success or failure of tool calls and the compliance state of the skill catalog have long depended on the model's own verbal self-reporting, which is imprecise and hard to audit. dsh-infra-observability replaces this pattern with platform-level structural recording: every tool execution's name and outcome is written by the host directly into a ring buffer with per-day aggregates, agent error and live status are tracked in parallel, and a watchdog checks error counts on a fixed interval, raising an alert the moment the five-minute threshold is exceeded.

The plugin exposes three query tools for both the model and operators: usage_report returns per-day tool and skill call, success, and error totals plus a recent execution trail; audit_skills summarises catalog size, provider spread, and flags entries missing descriptions or using non-kebab-case names; infra_health reports process uptime, record capacity, error counts, running agent count, and watchdog state. On the browser side it registers an Observability tab that renders a live per-session dashboard with turn counts, failure totals, tool and skill TOP lists with success-rate bars, all updating as events commit without a manual refresh.

It is aimed at DeepSeek Harness platform admins and SREs who need auditable tool-call records, skill-catalog governance, and automated anomaly alerting in production multi-agent deployments.