observability-analyzer（可观测性分析器）

Observability Analyzer

Query Claude Code telemetry and generate insights from metrics, logs, and traces. Works with both default OTEL telemetry and enhanced hook-based telemetry.

Data Sources

Operations

query-metrics <promql>

Execute PromQL query against Prometheus.

query-metrics 'sum(claude_code_token_usage)[7d]'

query-logs <logql>

Execute LogQL query against Loki.

query-logs '{job="claude_code_enhanced", event_type="tool_call"} | json' --since 24h

analyze-errors

Detect and group error patterns from enhanced telemetry.

{job="claude_code_enhanced", event_type="tool_result", status="error"} | json

Output: Error types, frequencies, affected tools, recommendations.

analyze-performance

Identify slow operations and response sizes.

{job="claude_code_enhanced", event_type="tool_result"} | json | response_length > 50000

Output: Large responses, estimated token costs, slow patterns.

analyze-costs

Calculate token usage from content size estimates.

sum by (repo) (sum_over_time({job="claude_code_enhanced", event_type="context_utilization"} | json | unwrap estimated_session_tokens [24h]))

Output: Token estimates by repo, session costs, projections.

analyze-tools

Tool usage statistics and sequences.

sum by (tool) (count_over_time({job="claude_code_enhanced", event_type="tool_call"} | json [24h]))

Output: Call frequency, success rates, tool sequences, common patterns.

analyze-sessions

Session lifecycle and duration analytics.

{job="claude_code_enhanced", event_type="session_end"} | json

Output: Session durations, turn counts, tools per session, termination reasons.

analyze-conversations

Conversation and prompt analytics.

sum by (pattern) (count_over_time({job="claude_code_enhanced", event_type="user_prompt"} | json [24h]))

Output: Prompt patterns (question/debugging/creation/ultrathink), turn distribution.

analyze-subagents

Subagent/Task tool usage.

{job="claude_code_enhanced", event_type="tool_call", tool="Task"} | json

Output: Subagent types used, completion rates, parallel execution patterns.

analyze-skills

Skill invocation analytics.

sum by (skill_name) (count_over_time({job="claude_code_enhanced", event_type="skill_usage"} | json [24h]))

Output: Most used skills, skill usage by repo, trends.

analyze-context

Context window utilization.

{job="claude_code_enhanced", event_type="context_utilization"} | json | context_percentage > 50

Output: High utilization sessions, compaction events, token efficiency.

analyze-repos

Repository/project activity.

sum by (repo, tool) (count_over_time({job="claude_code_enhanced", event_type="tool_call"} | json [24h]))

Output: Activity per repo, tool usage by project, branch patterns.

generate-report

Comprehensive analysis report (all dimensions). Output: Markdown report with errors, performance, costs, sessions, conversations, tools.

Key Queries

Enhanced Telemetry (Loki)

# All events (last hour)
{job="claude_code_enhanced"} | json

# Session analytics
{job="claude_code_enhanced", event_type="session_end"} | json | duration_seconds > 300

# Tool errors
{job="claude_code_enhanced", event_type="tool_result", status="error"} | json

# High context usage
{job="claude_code_enhanced", event_type="context_utilization"} | json | context_percentage > 75

# Subagent spawns
{job="claude_code_enhanced", event_type="tool_call", tool="Task"} | json

# Skill invocations
{job="claude_code_enhanced", event_type="skill_usage"} | json

# Prompt patterns
{job="claude_code_enhanced", event_type="user_prompt"} | json | pattern="ultrathink"

# Tool sequences
{job="claude_code_enhanced", event_type="tool_call"} | json | line_format "{{.tool_name}} → {{.previous_tool}}"

# Context compaction
{job="claude_code_enhanced", event_type="context_compact"} | json

# Permission requests
{job="claude_code_enhanced", event_type="permission_request"} | json

Default OTEL (Prometheus)

# Total token usage (7 days)
sum(increase(claude_code_token_usage[7d]))

# Error rate by tool
sum by (tool_name) (rate(claude_code_tool_result{status="failure"}[1h]))

# P95 tool latency
histogram_quantile(0.95, claude_code_tool_duration_bucket)

# Daily costs
sum(increase(claude_code_cost_usage[24h]))

Event Types Reference

Grafana Dashboards

• Claude Code Overview - High-level metrics
• Tool Performance - Tool latencies and success rates
• Cost Analysis - Token usage and costs
• Error Tracking - Error patterns and trends
• Session Analytics - Session-level insights
• Enhanced Analytics - Model/skill/context/repo tracking
• Deep Analytics - Comprehensive conversation and tool analysis

Access: http://localhost:3000 (admin/admin)

Scripts

• scripts/query-prometheus.sh - PromQL query helper
• scripts/query-loki.sh - LogQL query helper
• scripts/analyze-errors.sh - Error analysis automation
• scripts/analyze-sessions.sh - Session analytics
• scripts/generate-report.sh - Full analysis report