dsh-maestro-observe

When you run a Maestro multi-plugin orchestration, every turn, tool call, and error flashes by in milliseconds. Without a persistent local record, debugging failures and tracking token spend becomes guesswork. dsh-maestro-observe closes that blind spot with a local-first design: it captures turn, tool, and error signals into a single SQLite database on your own machine, then exposes them through a query tool and a loopback RPC so you can pull records for any time window, session, or channel on demand. Five core operations come ready to use: trace filters by session, tool, kind, and timestamp; cost aggregates token usage per day or per session, groupable by tool or session; latency returns p50/p95/p99 percentiles; errors are grouped by tool and normalized signature; and health gives a one-call snapshot of every plugin and channel status. Budget operations let you set token ceilings and check against them, while config operations manage retention days, digest schedule, spike thresholds, and other runtime knobs.

Before anything hits disk, a redact pipeline replaces secret-shaped fields with a placeholder and truncates overly long detail strings, so credentials never persist in plain text. If you want proactive visibility, the optional maestroNotifier service hooks in cleanly: a daily Telegram digest fires at the schedule you set, and a spike alert triggers the moment the error rate crosses an N/M-minute threshold. Skip the notifier entirely and the plugin still works fine, both channels simply stay quiet and log a warning without throwing.

This plugin is aimed at developers and small teams running Maestro multi-plugin orchestrations in production who need local, zero-external-dependency observability. It ships no client UI (rendering is delegated to the dsh-maestro-dashboard Activity tab) and focuses on one job: collect signals, store them safely, and make them answerable.