Logs and telemetry

Diagnostic log

~/.blitz/logs/blitz-YYYY-MM-DD.jsonl (owner-only, secrets masked, kept log.retain_days, 14 by default) records warnings, failed turns and errors, with trace IDs when telemetry is on. log.level (or BLITZ_LOG_LEVEL) is debug, info, warn, error or off. A background goroutine writes it, so logging never waits on the disk. The desktop app reads it in Settings › Logs, where a past day’s log can be deleted (today’s is still being written).

The audit log is separate: a record of what the agent did, not of Blitz’s own health.

OpenTelemetry

Telemetry is off by default, and nothing is sent unless you turn it on. With [telemetry] enabled = true (or BLITZ_TELEMETRY=1), traces, logs and metrics are exported over OTLP/HTTP to telemetry.endpoint, else OTEL_EXPORTER_OTLP_ENDPOINT, else http://localhost:4318. Other OTEL_EXPORTER_OTLP_* settings apply.

[telemetry]
enabled = true
endpoint = "http://localhost:4318"
# capture_content = false

One trace per prompt: a turn span (agent, model, token counts, cost) containing the ADK’s agent, model-call and tool spans, plus approval (time spent waiting on you), hook and compact spans.
Sessions are chains of turns, not one long trace, because sessions last days and resume in new processes. Every span carries the session as gen_ai.conversation.id, and each turn has a turn.index and a link to the previous turn, even after --resume.
No content by default. Prompts, replies, tool arguments and tool results aren’t exported; Blitz removes the ADK’s tool arguments and results from its spans before export. capture_content = true includes them, with secrets masked.
Metrics go to the same collector every 30 seconds: sessions, turns (by agent and outcome), tokens and estimated cost (by model), tool calls (by tool and outcome) and approval decisions (by action and how each was made: by you, a rule, the mode, a hook or the auto reviewer).
Export runs on background goroutines and gives up after 3 seconds at exit if the collector is unreachable.