observability-pattern-detector（可观测性模式检测器）

Observability Pattern Detector

Automated pattern recognition and anomaly detection in Claude Code telemetry data from enhanced hooks.

Data Source

Primary: {job="claude_code_enhanced"} in Loki

Operations

detect-failures

Group similar failures and identify patterns.

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

Algorithm: Group by error_type → Calculate frequency → Rank by impact. Output: Failure patterns with occurrences, affected tools, first/last seen, trend.

detect-slowness

Identify large response patterns (proxy for slowness).

{job="claude_code_enhanced", event_type="tool_result"} | json | response_length > 100000

Algorithm: Flag responses >100k chars → Group by tool → Identify patterns. Output: Slow operations with response sizes, affected tools.

detect-anomalies

Statistical anomaly detection in sessions.

{job="claude_code_enhanced", event_type="session_end"} | json | turn_count > 50

Methods: High turn count, long duration, many errors per session. Output: Anomalous sessions with metrics, likely cause.

detect-trends

Long-term trend analysis.

sum(count_over_time({job="claude_code_enhanced", event_type="tool_call"} [1d]))

Metrics: Tool usage trend, error rate trend, session frequency trend. Output: Trends with direction (increasing/decreasing/stable), rate.

detect-waste

Identify inefficiencies (redundant operations).

{job="claude_code_enhanced", event_type="tool_call"} | json | line_format "{{.tool_name}}:{{.previous_tool}}"

Patterns:

• Multiple reads of same file (Read→Read)
• Repeated failed operations
• Excessive Glob before Read
• Many small edits vs one large edit Output: Waste patterns with occurrences, recommendations.

detect-conversation-patterns

Analyze user prompt patterns.

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

Patterns:

• Question frequency (pattern="question")
• Debugging sessions (pattern="debugging")
• Creation tasks (pattern="creation")
• Ultrathink usage (pattern="ultrathink") Output: Conversation style distribution, trends.

detect-tool-sequences

Identify common tool call sequences.

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

Common Patterns:

• Glob → Read (file discovery)
• Read → Edit (modify after read)
• Grep → Read (search then open)
• Task → Task (parallel agents) Output: Sequence frequencies, unusual patterns.

detect-subagent-patterns

Analyze Task tool usage patterns.

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

Patterns:

• Subagent types distribution
• Parallel spawning patterns
• Subagent success rates Output: Subagent usage analytics, recommendations.

detect-context-issues

Identify context window problems.

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

Patterns:

• Frequent auto-compaction
• High context usage sessions
• Large response accumulation Output: Context management issues, optimization suggestions.

detect-permission-patterns

Analyze permission request patterns.

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

Patterns:

• Frequent permission requests
• Permission types distribution
• Permission denials Output: Permission friction points, automation opportunities.

detect-repo-patterns

Repository activity patterns.

sum by (repo) (count_over_time({job="claude_code_enhanced", event_type="tool_call"} | json [7d]))

Patterns:

• Most active repos
• Tool usage by repo
• Error rates by repo Output: Project-level insights, cross-repo comparisons.

Example Output

{
  "failure_patterns": [
    {
      "pattern_id": "file_not_found",
      "signature": "File does not exist",
      "occurrences": 23,
      "affected_tools": ["Read", "Edit"],
      "trend": "stable",
      "recommendation": "Add file existence check before operations"
    }
  ],
  "tool_sequence_patterns": [
    {
      "sequence": "Glob → Read → Edit",
      "occurrences": 156,
      "context": "Standard file modification flow"
    }
  ],
  "conversation_patterns": [
    {
      "pattern": "debugging",
      "percentage": 35,
      "avg_turns": 12,
      "common_tools": ["Bash", "Read", "Grep"]
    }
  ],
  "context_issues": [
    {
      "issue": "auto_compaction_frequent",
      "sessions_affected": 5,
      "recommendation": "Use more focused queries, split large tasks"
    }
  ]
}

Pattern Detection Queries

Failure Patterns

# Group errors by type
sum by (error_type, tool) (count_over_time({job="claude_code_enhanced", event_type="tool_result", status="error"} | json [24h]))

# Error timeline
{job="claude_code_enhanced", event_type="tool_result", status="error"} | json | line_format "{{.timestamp}} {{.tool_name}}: {{.error_type}}"

Tool Sequence Patterns

# Most common transitions
{job="claude_code_enhanced", event_type="tool_call"} | json | previous_tool != "" | line_format "{{.previous_tool}} → {{.tool_name}}"

Session Anomalies

# Long sessions
{job="claude_code_enhanced", event_type="session_end"} | json | duration_seconds > 3600

# High error sessions
{job="claude_code_enhanced", event_type="session_end"} | json | error_count > 5

# High turn sessions
{job="claude_code_enhanced", event_type="session_end"} | json | turn_count > 30

Context Patterns

# Auto compactions
{job="claude_code_enhanced", event_type="context_compact", trigger="auto"} | json

# High utilization
{job="claude_code_enhanced", event_type="context_utilization"} | json | context_percentage > 80

Scripts

• scripts/detect-failures.sh - Failure pattern detection
• scripts/detect-anomalies.sh - Statistical anomaly detection
• scripts/detect-trends.sh - Trend analysis
• scripts/detect-sequences.sh - Tool sequence analysis
• scripts/generate-pattern-report.sh - Full pattern report