Extract Trace Tree

Langfuse Trace tree extraction and analysis tool

Overview

This tool extracts an execution-path tree from a Langfuse trace JSON file (typically langfuse_trace.json). It builds a parent/child tree, sorts siblings by timestamp, applies noise filtering (A2A + HTTP), and outputs both a console view and a Markdown report.

Quick Start

Single file

python3 extract_trace_tree.py <trace_file.json>

Examples:

python3 extract_trace_tree.py GPT-5/BookWriter-MCP/test_results/run_20251104_045654/langfuse_trace.json
python3 extract_trace_tree.py GPT-5/BookWriter-A2A/test_results/run_20251103_195358/langfuse_trace.json
python3 extract_trace_tree.py GPT-5/BookWriter-H_A2A/test_results/run_20251104_010743/langfuse_trace.json

Batch

./batch_extract_trees.sh

Output

The script writes execution_path.md next to the input trace file. The Markdown contains:

# Trace Execution Path header with metadata
## Execution Path Tree (tree in a fenced code block)
## Statistics
### Error Summary (only when errors exist)

Token format

LLM nodes show:

(prompt→completion [REASONING:X, OUTPUT:Y], total: Z)

Where:

REASONING is the reasoning-token count (0 for non-reasoning models)
OUTPUT is completionTokens - reasoningTokens

Batch label (write_a_book_with_flows)

For write-book scenarios the tree may include:

BATCH1 / BATCH2
[BUSINESS-RETRY] when the same chapter is retried

Truncated errors

Long errors are truncated and marked with [TRUNCATED].

Project Type Detection

Project type is detected from the trace path:

*-MCP/ -> MCP
*-A2A/ -> A2A
*-A2A_mix/ -> A2A_mix
*-H_A2A/ -> A2A

Filtering

HTTP client tracing nodes (OpenTelemetry httpx) are filtered.
A2A framework internals are filtered for A2A/A2A_mix.
Tool telemetry spans such as Tool Usage/Tool Repeated Usage are filtered.

Requirements

Python 3
Standard library only (no third-party dependencies)