Observability

Observability

obsidian-tc is observable from day one. The Observability layer is the last stage of the dispatch pipeline and always fires — on success and on every error path. Emission is non-blocking: a slow sink drops the span/event with an [obsidian-tc:obs-drop] log line rather than stalling the call. Full design: docs/G2.4-observability.md.

Five emitters per tool call

Emitter	What it produces
OpenTelemetry	A span obsidian_tc.tool_call tagged {tool, vault, caller, status, args_hash} with end-to-end duration, exported over OTLP
Prometheus	Increments obsidian_tc_tool_calls_total{tool, vault, status} and records obsidian_tc_tool_duration_seconds{tool, vault}
CloudEvents / MORGIANA	A structured event per call, async fire-and-forget
JSONL trace	Appends to the current session's per-vault trace file
event_log row	A SQLite insert for local debug replay

Configuration

"observability": {
  "otel": { "enabled": true, "endpoint": "http://localhost:4318" },
  "prometheus": { "enabled": true },
  "morgiana": { "endpoint": "", "token": "" },
  "retention": { "tracesDays": 90, "eventLogDays": 30 }
}

Prometheus endpoint

When HTTP transport is enabled, metrics are scrapeable at GET /metrics. The get_metrics admin tool returns the same counters as structured JSON for clients without a Prometheus scraper.

JSONL traces

Per-vault, per-session: <cacheDir>/traces/<YYYY-MM-DD>/<session_id>.jsonl, rolled daily and retained 90 days. Sessions are opened with start_session and closed with end_session; get_session_traces replays events from a session or date window. Because each row carries args_hash (not raw args), traces are replayable without leaking note contents.

event_log

A per-vault SQLite table for local debug replay. Naturally isolated per vault (one DB each); no cross-vault analytics in v1. get_metrics can scope to one vault or aggregate across all.

Health

server_health round-trips the full transport → auth → ACL → audit path and returns liveness + build info — use it as a readiness probe.