pi-log-analyzer

Pi Log Analyzer

Analyze Pi Agent's structured logs (.log/*.jsonl), providing multi-dimensional statistics and issue diagnosis.

Log Format

Logs are in JSONL format, one JSON object per line:

{
  "schema": "2.0.0",
  "timestamp": "2026-05-07T00:46:30.942Z",
  "level": "info|warn|error",
  "tag": "auto-extract|checkpoint|vector|...",
  "message": "日志消息",
  "context": {
    "role": "default|bw|psm|jly|zero",
    "sessionId": "uuid",
    "cwd": "/path",
    "pid": 12345
  },
  "meta": {
    "model": "provider/model-name",
    "duration_ms": 1234,
    ...
  },
  "traceId": "tr-xxx"
}

Quick Analysis

Basic Statistics

Run the built-in analysis script:

python3 ~/.pi/agent/skills/pi-log-analyzer/scripts/analyze.py [目录路径] [天数]

Parameters:

• 目录路径: log directory, default .log/
• 天数: analyze the last N days, default 3

Example:

# 分析最近 3 天
python3 ~/.pi/agent/skills/pi-log-analyzer/scripts/analyze.py

# 分析最近 7 天
python3 ~/.pi/agent/skills/pi-log-analyzer/scripts/analyze.py .log/ 7

# 分析指定目录
python3 ~/.pi/agent/skills/pi-log-analyzer/scripts/analyze.py /path/to/logs 5

Output Content

The script generates the following analyses:

1. Overview Statistics: number of log entries, number of sessions, time span
2. Level Distribution: count and proportion of info/warn/error
3. Module Activity: frequency of occurrence of each tag
4. Role Activity: log volume of each role
5. Model Usage: models called and their counts
6. Time Distribution: hourly activity heatmap
7. Warning Details: specific content and patterns of warn/error

Manual Analysis

For more detailed analysis, you can use Python to read the logs:

import json
from collections import Counter

with open('.log/2026-05-09.jsonl') as f:
    for line in f:
        d = json.loads(line)
        # 访问字段：d['level'], d['tag'], d['meta']['model'], etc.

Common Tag Meanings

Tag	Meaning
auto-extract	Automatic memory extraction
checkpoint	Scheduled save point
daily-memory	Daily memory write
vector	Vector index operation
pending	Pending item
repair	Memory repair
embedding	Embedding generation
knowledge	Knowledge base operation

Common Issue Diagnosis

auto-extract parse failed

Cause: The memory extraction result returned by the LLM does not match the expected format

Typical Patterns：

• Returning reasoning process instead of JSON
• JSON is truncated
• Contains <think> tags

Recommendations：

• Check whether the model supports structured output
• Prefer DeepSeek V4 Flash (lowest failure rate)

Model Call Failure

Check the meta.model field to confirm the model used, and compare with the models.json configuration.

Use Cases

1. Daily Inspection: quickly check whether recent activity is normal
2. Troubleshooting: locate the root cause of warnings/errors
3. Performance Analysis: identify high-load periods and modules
4. Model Evaluation: count usage and success rate of each model
5. Capacity Planning: estimate resources based on session count and log volume