Fix broken dashboard panels (Active Sessions, Tool Success Rate, API Error Rate) + add usage panels

[404pilo]

Summary

Three panels in the bundled Grafana dashboard never resolve against a real stack. This PR fixes their root causes and adds six high-value panels. Every query was verified against a live OTel Collector + Prometheus + Loki stack receiving real Claude Code telemetry.

Bugs fixed

1. "Active Sessions" always shows 0

The panel uses increase(claude_code_session_count_total[1h]), but session.count is a one-shot counter — it ticks once at session start and never updates. The Collector's Prometheus exporter drops idle series after the default metric_expiration (~5m), so the series disappears and increase() has nothing to compute.

Fix: derive active sessions from the continuously-updated active_time metric:

count(count by (session_id) (increase(claude_code_active_time_seconds_total{job="otel-collector"}[1h]) > 0))

…and raise the exporter's metric_expiration to 2h so one-shot counters (session, lines_of_code, commit, pull_request, code_edit_decision) survive between updates. Without this, "Lines of Code", "Code Changes", and "Development Activity" also flatline once idle.

2. "Tool Success Rate" and "API Error Rate" show "No data"

Both pipe Loki events through | json. But Claude Code's OTLP log events put the event name in the log body (e.g. claude_code.tool_result) and every field (success, tool_name, status_code, …) in structured-metadata labels. | json tries to parse the body, throws JSONParserErr on every line, and silently zeroes the panels.

Fix: filter the labels directly — | success="true", sum by (status_code) (…) — no | json stage.

New panels

• Cache Hit Ratio — cacheRead / (cacheRead + input); cost efficiency at a glance
• Cost by query_source — main vs subagent vs auxiliary (surfaces multi-agent cost runaway)
• Token Spend by Model
• Active Time: User vs CLI
• Tool Decisions: Accept vs Reject — uses the decision label on claude_code.tool_decision
• API Latency P95 by Model

Test plan

• claude-code-dashboard.json is valid JSON (29 panels)
• Every PromQL/LogQL query returns data against a live Prometheus (:9090) and Loki (:3100)
• Collector restarts cleanly with metric_expiration: 2h
• Grafana provisions the dashboard with no errors

🤖 Generated with Claude Code