rule-engine-extract

name	rule-engine-extract
description	Extract reusable migration rules from code changes. Works on interactive fixes (before/after comparison), git history, or any pair of SQL files. Triggers: extract rule, deduce rule, learn from fix, create rule from fix, rule extraction.

Extract Rules

Analyze changes to converted SQL files and extract reusable migration rules. Works in two modes:

• Interactive fix mode — the user fixed a file; compare before/after to extract a rule
• Git history mode — analyze committed changes retroactively

Prerequisites

• Rule engine set up (../SKILL.md)
• Snowflake connection active

Mode Selection

If the user says something like "I fixed X, extract a rule" or "deduce a rule from my fix" → use Interactive Fix Mode.

If the user says "analyze recent changes" or "extract rules from git history" → use Git History Mode.

---

Interactive Fix Mode

Step 1: Get the "After" SQL

Read the current converted file (the fixed version):

cat <project_dir>/snowflake/<type>/<schema>/<name>.sql

Or ask the user which file they fixed.

Step 2: Get the "Before" SQL

Find the changes in the file — use git diff, git log, or any other method to obtain the previous version. If no history is available, ask the user.

Step 3: Analyze the Diff

Compare before and after to identify distinct fix patterns. For each pattern:

1. What changed — describe the transformation (before → after)
2. Why — what Snowflake compatibility issue does this fix?
3. Is it mechanical? — can a regex handle it, or does it need AI context?

If mechanical (direct find/replace):

• Propose replacement_mode = 'regex'
• Write match_pattern (detection regex)
• Write replacement_find and replacement_replace (substitution regex)

If context-dependent:

• Propose replacement_mode = 'ai'
• Write match_pattern (detection regex)
• Write ai_context describing: what to look for, how to fix it, edge cases to watch for

Always generate examples from the actual before/after.

Step 4: Check for Duplicates

Use the find_similar_rules tool with query set to a description of the pattern.

If a similar rule exists:

This pattern matches existing rule "". No new rule needed.

Done.

Step 5: Confirm with User

Present the proposed rule:

New rule from your fix:

• Name: Variable binding colon prefix
• Mode: ai
• Match pattern: LANGUAGE\s+SQL
• Context: In Snowflake LANGUAGE SQL procedures, variables in SQL statements need : prefix...
• Example: WHERE col >= p_start_date → WHERE col >= :p_start_date

Save this rule? [Yes / Edit / Skip]

Step 6: Create the Rule

Use the create_rule tool with the confirmed parameters.

Step 7: Propagate (Optional)

After creating the rule, offer to find other affected code units:

Rule created. Check which other code units it applies to?

If yes → Load ../propagate/SKILL.md with the new rule ID.

---

Git History Mode

Step 1: Identify Target Objects

Accept object names from the user, or discover from recent commits to find modified SQL files under the project's snowflake/ directory.

Present the list:

Found N modified SQL files. Analyze all, or select specific files?

Step 2: Extract Diffs

For each target file, retrieve the diff showing what changed.

Step 3: Analyze Changes

For each diff, identify distinct fix patterns. Classify each:

Category	Example	Likely Mode
Direct substitution	ISNULL( → COALESCE(	regex
Syntax rewrite	TOP N → LIMIT N	regex
Structural change	EXEC proc @p → CALL proc(:p)	ai
Logic adjustment	Date arithmetic rewrite	ai

Step 4: Deduplicate Across Objects

Merge identical patterns across files. Two patterns are the same if they describe the same source→target transformation.

Step 5: Check for Existing Rules

For each unique pattern, use find_similar_rules to check for duplicates. Remove already-covered patterns.

Step 6: Confirm with User

For each new pattern, ask:

New pattern: EXEC schema.Proc @p1 → CALL schema.Proc(:p1) Found in: dbo.CalcDiscount (2 occurrences)

Save as a reusable migration rule? [Yes / No]

Step 7: Insert New Rules

Use the create_rule tool for each confirmed pattern with created_from = "deduced_from_history".

Step 8: Find Other Affected Code Units

For each new rule, use reverse_search_rules to find other matching objects and report the count.

Summary

Extraction complete:

• Analyzed N files
• Found P distinct patterns
• K already covered by existing rules
• Created R new rules
• New rules match Q additional code units