🔭 Infra Observability — DSH-Plugin

結構化觀測層:真實的工具/技能使用記錄(tools/result)、技能目錄審計、看門狗。零模型自報——平台記什麼,報告就寫什麼。

A DeepSeek Harness host plugin that replaces model self-reporting with structural recording:

• tools/result listener — every real tool execution (name + outcome) is recorded by the platform into a ring buffer and per-day aggregates.
• agent/error / agent/status — error totals and live-agent tracking.
• Watchdog — timer-driven alert when agent errors spike (≥ N in 5 min).
• Three tools for the model and humans:

Tool	What it reports
usage_report	Per-day tool/skill aggregates (calls / ok / err) + recent executions
audit_skills	Catalog totals, provider spread, warnings (missing description, non-kebab names), today's skill usage
infra_health	Uptime, records kept, error counts, running agents, catalog size, watchdog state

Dashboard (browser)

The client half registers an 「觀測」view tab (order 20, right of chat / trajectory) rendering a live per-session dashboard from the infraView session projection: turn counts, failure count, tool/skill TOP lists with success-rate bars, and the recent execution trail. Data updates as events commit — no refresh needed.

Install

Path A — Host composition (recommended, always-on)