Create Mre

MRE Extraction Workflow

This workflow guides the process of reducing a source file to its absolute minimum while preserving a specific bug (e.g., a compiler crash).

1. Setup and Verification

• Identify the target file to reduce. If this file is deep inside an installed library (e.g., in site-packages), copy it to tests/scratch/ (e.g. tests/scratch/MRE.py) to avoid accidentally corrupting your Python environment during the destructive reduction process and to keep temporary reduction artifacts out of the repository root.
• Identify the specific command that triggers the issue if possible.
• For macOS issues that may depend on Python distribution or CI packaging, establish the Python flavor matrix before reducing. Follow .agents/skills/reproduce-macos-python-flavors/SKILL.md.
• Augment the reproduction command so it always writes --report=tests/scratch/compilation-report.xml. If the command already uses --report=..., redirect it to this path. Keep the rest of the invocation as close as possible to the original failure as you can.
• Define the reproduction oracle before reducing. For crash bugs, this can be the same exception type, the same failing statement location, or the same relevant <exception> entry in tests/scratch/compilation-report.xml. For non-crash bugs, this should be a specific preserved signal such as an output line, warning text, optimization message, exit code, or diff against expected output. The oracle does not need to be byte-for-byte identical output; it needs to reliably indicate that the same bug is still present. For example, an extra-pass bug may use Module(s) '__main__' necessitate pass 3 as the oracle.
• Run the reproduction command to confirm the issue. If the crash occurs during a massive standalone build, it's highly recommended to pipe the output to a file (> tests/scratch/out.txt 2>&1) to easily search for the crash location while also inspecting tests/scratch/compilation-report.xml.

Handling Standalone Optimization Crashes

If the crash happens during optimization in standalone mode (often a Python traceback below nuitka/optimizations):

1. Check the output or tests/scratch/out.txt for lines starting with Problem with statement at ..., or inspect the <exception> tag in tests/scratch/compilation-report.xml.
2. Identify the filename involved.
3. Copy that file to tests/scratch/ (e.g. cp /path/to/ProblematicFile.py tests/scratch/MRE.py).
4. Run Nuitka in module mode on that copied file (e.g., bin/nuitka --mode=module --generate-c-only --report=tests/scratch/compilation-report.xml tests/scratch/MRE.py). Preserve the original failing invocation as closely as possible while reducing, while keeping the report output in tests/scratch/. Note: Use --generate-c-only to skip the C compilation phase, which is unnecessary if the crash occurs during Nuitka's optimization (Python) phase. This significantly speeds up the reduction cycle.
5. If the oracle still matches, specific reduction in module mode as described below is the best path forward.

2. Analysis

• Read the target file.
• Identify the core components suspected to be involved in the bug (e.g., specific imports like ctypes, specific Python constructs like classes or conditionals).

3. Iterative Reduction Loop

Repeat this process until the file is minimal:

1. Propose a reduction: Select a block of code to remove or simplify.
   • Strategy 1: Remove unused imports and helper functions.
   • Strategy 2: Mock external dependencies (e.g., replace import other_package with a dummy class/function) to allow removing the dependency.
   • Strategy 3: Simplify control flow (replace if complex_condition: with if True: or removing the branch entirely if possible, replace complex objects with simple ones).
   • Strategy 4: Rename long functions/variables to single letters (e.g., original_name -> f) once context is lost.
   • Strategy 5: For optimization issues, prioritize reducing the number of functions and methods. This is often crucial for isolating issues that occur during inter-procedural optimization or complex scope analysis.
   • Strategy 6: Remove default arguments from functions. This simplifies the function signature and can help isolate issues related to default value evaluation or parameter handling.
   • Strategy 7: Structural and Scoping Mutations. When dealing with compiler/optimization crashes, the bug is sometimes tied to the shape of the AST or variable scoping. Test subtle structural changes:
     • Replace attribute lookups in decorators (e.g., @module.decorator) with direct names (@decorator).
     • Remove or simplify locals() or globals() updates.
     • Simplify variable capturing in lambda functions, especially inside list/dict comprehensions.
     • Replace generator expressions with equivalent for loops to see if the implicit function scope is the trigger.
   • Consider Simplification: Don't assume the entire block must be removed.
     • Replace complex expressions with constants/literals.
     • Replace variable usage with direct values.
     • Replace function bodies with pass or a simple valid statement.
   • Avoid Confounding Variables: Ensure you are changing only one thing at a time. (e.g., don't remove an import and a global assignment in one step).
2. Apply the change.
3. Run the reproduction command and re-check the same reproduction oracle.
4. Evaluate result:

   • If the oracle still matches: Great! The removed code was irrelevant.

   • If the oracle no longer matches, disappears, or becomes ambiguous: The removed code was relevant. IMMEDIATELY REVERT the change to the scratch copy only. Never reset unrelated repository changes.

   • Validation:

     • Maximize Reduction: Even if you think a construct (like a function or loop) is needed, try to remove it or flatten it to verify.
     • Renaming: Always rename functions, variables, and classes to single short letters (e.g., f, x, C) to remove semantic meaning and ensure no special string-based handling is at play.

4. Final Polish

• Ensure the code is self-contained (no external requirements beyond standard library if possible).
• Format clean.
• Verify one last time.

5. Report

• Present the final MRE code block to the user.