Runpod Flash Flash

Flash Framework Explorer Skill

When to Use

Use this skill when:

• Exploring the runpod-flash framework codebase
• Understanding class hierarchies and relationships
• Finding where methods or classes are defined
• Checking what decorators are used in the codebase
• Getting a quick overview of file structure

Workflow

Step 1: Query the Code Intelligence Database FIRST

Before reading any files, use the MCP code intelligence tools:

Finding a symbol:

• Use find_symbol tool with the symbol name

Understanding a class:

• Use get_class_interface tool with the class name

Exploring a file:

• Use list_file_symbols tool with the file path

Finding decorated functions:

• Use find_by_decorator tool (e.g., decorator="remote")

Listing all classes:

• Use list_classes tool

Step 1.5: NEVER Use Bash Commands for MCP Tool Tasks

PROHIBITED patterns - use MCP tools instead:

❌ DO NOT use tail, grep, or cat for test output analysis ✅ DO use parse_test_output MCP tool after running tests

❌ DO NOT use grep or find to search for symbols/classes/methods ✅ DO use find_symbol, list_classes, or find_by_decorator MCP tools

❌ DO NOT use Read tool to scan files for class definitions or interfaces ✅ DO use get_class_interface or list_file_symbols MCP tools

Step 2: Only Read Full Files When Necessary

After querying the code intelligence database, only use the Read tool if:

• You need to understand the implementation details
• The database doesn't have the specific information you need
• You need to see the full context around a symbol

Analyzing Test Results

After running tests (make test-unit, make test, pytest), ALWAYS use parse_test_output MCP tool to analyze results:

Good - Use MCP tool (~200 tokens):

1. Run: make test-unit
2. Pass output to parse_test_output MCP tool
3. Get structured failures, coverage, passed tests, summary
4. No manual parsing needed

Bad - Using bash commands (~20,000+ tokens):

1. Run: make test-unit > output.txt
2. Use tail, cat, or grep on output file
3. Manually parse failures and coverage
4. Excessive token usage, poor context extraction

Examples

Good - Query first:

1. Use find_symbol with "ServerlessEndpoint"
2. Review the signatures and locations
3. Only read the full file if you need implementation details

Bad - Reading files directly:

1. Read entire src/runpod_flash/core/resources/serverless.py (500+ tokens)
2. Search manually for ServerlessEndpoint

Good - Parse test output:

1. User runs: make test-unit
2. Get test output, use parse_test_output tool
3. Get structured result with failed tests, summary, coverage

Bad - Manual test output parsing:

1. User runs: make test-unit
2. Use tail -50 output.txt to see failures
3. Use grep "FAILED" output.txt to find test names
4. Manually parse coverage numbers

Benefits

• 85% token reduction for symbol exploration tasks (vs reading full files)
• 99% token reduction for test result analysis (200 tokens vs 20,000+ with bash)
• Faster responses - no need to parse large files or output
• Better context - see symbols across multiple files
• Focused reading - only read what you actually need
• Structured data - test failures, coverage, passed tests parsed automatically

Available MCP Tools

Code Intelligence Tools:

• find_symbol - Search for any symbol by name
• list_classes - List all classes with signatures
• get_class_interface - Get methods and properties of a class
• list_file_symbols - List all symbols in a file
• find_by_decorator - Find symbols with specific decorators
• parse_test_output - Parse test results and extract failures, coverage, summary

Important Notes

• The code intelligence database is updated by running make index
• If you get unexpected results, the index might be stale
• Database contains: classes, functions, methods, decorators, type hints, docstrings
• Database does NOT contain: implementations, comments, full file content
• Always use parse_test_output for test results - never use tail/grep/cat on test output
• Always use symbol tools - never use Read to search for class definitions

Source: runpod/flash — distributed by TomeVault.