session-analyzer

name	session-analyzer
description	Analyze Claude Code sessions for metrics and retrospection. Parses JSONL session files to extract message counts, tool usage, token consumption, and user feedback patterns. Used by daily-retrospection for budget tracking.
disable-model-invocation	false

Session Analyzer

Analyze Claude Code sessions to extract metrics for retrospection and budget tracking.

When to Use

• During daily/weekly retrospection to understand session costs
• To analyze specific sessions for debugging or review
• To track patterns in tool usage, token consumption, and user feedback
• For budget monitoring and cost optimization

Scripts

All scripts are located in this directory (.claude/skills/session-analyzer/):

Script	Purpose
query_pulses.sh	Query today's completed pulses from SQLite
get_todays_sessions.py	List sessions modified today
analyze_session.py	Full session analysis with metrics
todays_summary.py	Quick summary of today's sessions
full_session_analysis.py	Compact one-line analysis
bulk_analysis.py	Analyze all today's sessions
skill_invocation_analysis.py	Count skill invocations across sessions

Data Sources

1. Pulse Queue Database

Location: ~/.reeve/pulse_queue.db

Query completed pulses:

./query_pulses.sh

Note: The session_id column in pulses table may be empty for some pulses. The actual session files are created by Claude Code independently.

2. Sessions Index

Location: ~/.claude/projects/[YOUR_PROJECT_HASH]/sessions-index.json

Structure:

{
  "version": 1,
  "entries": [
    {
      "sessionId": "uuid-here",
      "fullPath": "~/.claude/projects/[YOUR_PROJECT_HASH]/uuid.jsonl",
      "fileMtime": 1769190637807,
      "firstPrompt": "The initial prompt...",
      "summary": "Auto-generated summary",
      "messageCount": 4,
      "created": "2026-01-23T06:46:56.260Z",
      "modified": "2026-01-23T06:47:06.575Z",
      "gitBranch": "master",
      "projectPath": "[YOUR_DESK_PATH]",
      "isSidechain": false
    }
  ]
}

3. Session JSONL Files

Location: ~/.claude/projects/[YOUR_PROJECT_HASH]/{sessionId}.jsonl

Each line is a JSON object. Key types:

• "type": "user" - User message
• "type": "assistant" - Assistant response (contains usage data)
• "type": "progress" - Progress events (hooks, etc.)
• "type": "summary" - Session summary

Workflow

Step 1: Get Today's Sessions

cat ~/.claude/projects/[YOUR_PROJECT_HASH]/sessions-index.json | python3 get_todays_sessions.py

Step 2: Analyze a Session File

cat ~/.claude/projects/[YOUR_PROJECT_HASH]/{SESSION_ID}.jsonl | python3 analyze_session.py

Step 3: Generate Summary Report

After analyzing sessions, format the output:

## Session Analysis: {session_id}
- Messages: X (user: Y, assistant: Z)
- Tool calls: N
- Input tokens: ~K (includes cache)
- Output tokens: M
- Duration: Xms
- Feedback signals: [list any detected]

Quick One-Liner: Today's Sessions Summary

cat ~/.claude/projects/[YOUR_PROJECT_HASH]/sessions-index.json | python3 todays_summary.py

Quick One-Liner: Full Session Analysis

cat ~/.claude/projects/[YOUR_PROJECT_HASH]/${SESSION_ID}.jsonl | python3 full_session_analysis.py

Bulk Analysis: All Today's Sessions

cat ~/.claude/projects/[YOUR_PROJECT_HASH]/sessions-index.json | python3 bulk_analysis.py

Skill Invocation Analysis

python3 skill_invocation_analysis.py

Token Cost Estimation

Rough pricing (check current rates):

• Claude Opus: ~$15/M input, ~$75/M output
• Claude Sonnet: ~$3/M input, ~$15/M output
• Cache reads are discounted ~90%

Formula for daily cost estimate:

# Approximate cost calculation
input_cost = (input_tokens / 1_000_000) * 15  # Opus rate
output_cost = (output_tokens / 1_000_000) * 75
total_cost = input_cost + output_cost
print(f"Estimated cost: ${total_cost:.2f}")

Feedback Pattern Detection

The analyzer looks for these signals that may indicate corrections or issues:

Pattern	Interpretation
"no," / "no "	Direct rejection
"wrong"	Error identified
"actually"	Correction incoming
"fix" / "error"	Problem found
"don't" / "stop"	Unwanted action
"should be" / "instead"	Correction
"undo" / "cancel"	Reversal needed

High feedback signal counts may indicate:

• Misunderstanding user intent
• Errors in execution
• Need for better prompting
• Learning opportunities

Integration with Daily Retrospection

Use this skill in /daily-retrospection to:

1. Get list of today's completed pulses
2. For each pulse with a session, analyze the JSONL
3. Aggregate totals (messages, tools, tokens, feedback)
4. Calculate estimated cost
5. Log insights to Diary/

Example output for retrospection:

## Daily Session Metrics - 2026-02-05

| Metric | Value |
|--------|-------|
| Sessions | 8 |
| Total messages | 142 |
| Tool calls | 87 |
| Input tokens | ~1.2M |
| Output tokens | ~45K |
| Est. cost | $21.38 |
| Feedback signals | 3 |

Notable sessions:
- c431664c (87 msgs): Complex skill creation, 11 tool calls
- 33f86f7c (26 msgs): Morning briefing, standard flow

Feedback detected:
- [fix] in session abc123: "fix the formatting..."