The Telemetry view answers the third question: is it working well? It uses the same history as the Me view, read for speed, errors and health instead of output.
The headline numbers
The top row gives six signals at once: the median turn duration with its 95th percentile, the API error rate, the tool success rate, compactions with the context before and after, abandonment (sessions you interrupted) and output speed in tokens per turn.
Turn duration
A turn is one prompt from start to finish, tools included. Turn duration sorts every turn into buckets from under two seconds to over fifteen minutes. p50 and p95 show the trend: by day for short ranges, by week past 45 days, so one quiet day does not look like an incident. A rising p95 with a flat p50 usually means a few long agentic turns, not a slower model.
Tool success and permission decisions
Tool success rate lists every tool with its calls, a 30-day sparkline, its success rate and how many calls were rejected. A green check means 97% or better, amber means watch it, red means low.
Permission decisions shows, for built-in tools and connectors, how many calls went ahead and how many you rejected or a rule denied. Under it, Skill activations splits skills by who started them.
Errors, connectors and compactions
API error rate breaks errors down by class, such as rate limits and overloaded servers, with a count per day. Connector health gives each MCP server a dot and its success rate. Compactions shows how often the context filled up and how much of it each compaction kept.
Where the time goes
Where the time goes splits the time of every session into waiting on the model, running tools, and retries. A large retry share points at rate limits or network trouble rather than at Claude.
At the bottom, the OpenTelemetry reference lists the metric names Claude Code exports when CLAUDE_CODE_ENABLE_TELEMETRY=1 is set, for building the same charts in Grafana or Datadog.