multi-step-analysis

name	multi-step-analysis
description	Use when the user asks for analysis that requires decomposing a complex question into multiple queries (trend analysis, comparisons, aggregation across entities)

Multi-Step Analysis Skill

How to Run Analysis (CRITICAL)

Call the Python analysis modules:

# Decompose a complex question into sub-queries
python3 -m multi_step_analysis.lib.decompose --question "对比2025年和2026年Q1的采购金额变化"

# Cross-reference results from multiple rounds
python3 -m multi_step_analysis.lib.decompose cross-reference \
  --results-dir /tmp/mcp_results/

Execution Flow

Step 1: Decompose

python3 -m multi_step_analysis.lib.decompose --question "<QUESTION>"

Breaks the complex question into 2-5 sub-queries with round numbers.

Step 2: Execute Sub-queries

For each sub-query: use common.mcp_client to generate_query -> validate_and_execute.

Step 3: Cross-reference Results

python3 -m multi_step_analysis.lib.decompose cross-reference --results-dir /tmp/

Finds patterns, trends, or anomalies across sub-query results.

Step 4: Synthesize

Present findings with severity levels:

• INFO: Within normal range
• WARNING: Exceeds soft threshold
• ALERT: Exceeds hard threshold

CRITICAL: All numbers must come directly from query results. Do NOT calculate values not in the database.

Step 5: Audit

Write one audit log entry capturing all sub-queries and the final analysis.

Constraints

• Max 8 query rounds per analysis
• Always show evidence (which query produced which data)
• Mark anomalies with severity level