Functional Core, Imperative Shell (FCIS)

Overview

Core principle: Separate pure business logic (Functional Core) from side effects (Imperative Shell). Pure functions go in one file, I/O operations in another.

Why this matters: Pure functions are trivial to test (no mocks needed). I/O code is isolated to thin shells. Bugs become structurally impossible when business logic has no side effects.

When to Use

Use FCIS when:

• Writing any new code file
• Refactoring existing code
• Reviewing code for architectural decisions
• Deciding where logic belongs

Trigger symptoms:

• "Where should this function go?"
• Creating a new file
• Adding database calls to logic
• Adding file I/O to calculations
• Writing tests that need complex mocking

MANDATORY: File Classification

YOU MUST add pattern comment to every file containing runtime behavior:

// pattern: Functional Core
// pattern: Imperative Shell
// pattern: Mixed (needs refactoring)

If file genuinely cannot be separated (rare), document why:

// pattern: Mixed (unavoidable)
// Reason: [specific technical justification]
// Example: Performance-critical path where separating I/O causes unacceptable overhead

No file with runtime behavior without classification. If you create a file that contains functions, classes with methods, or orchestration logic without this comment, you have violated the requirement.

Exempt: Files Without Runtime Behavior

DO NOT add pattern comments to:

• Type-only files - files exporting only types, interfaces, or type aliases (no runtime code)
• Constants/enum-like files - static data declarations, no functions
• Barrel/index files - re-exports only (export * from './foo')
• Test files - tests exercise core/shell code but aren't themselves core or shell
• Generated files - machine-generated code
• Bash/shell scripts (.sh, .bash) - inherently imperative
• Configuration files (eslint.config.js, tsconfig.json, .env, etc.)
• Markdown documentation (.md)
• HTML files (.html)
• Task runner files (justfile, Makefile, etc.)
• Package manifests (package.json, pyproject.toml, etc.)
• Data files (JSON, YAML, CSV, etc.)

Note: If an exempt file grows to include runtime logic (e.g., a "types" file gains helper functions, or a constants file gains factory functions), it crosses the threshold and MUST be classified.

Classification applies to application source files containing runtime behavior (functions with logic, classes with methods, I/O orchestration).

File Type Definitions

Functional Core Files

Contains ONLY:

• Pure functions (same input -> same output, always)
• Business logic, validations, calculations, transformations
• Data structure operations
• Logging (EXCEPTION: loggers are permitted in Functional Core)

NEVER contains:

• File I/O (reading, writing files)
• Database operations (queries, updates, connections)
• HTTP requests or responses
• Environment variable access
• Date.now(), Math.random(), or other non-deterministic functions
• State mutations outside function scope

Logging exception: Functions MAY accept and use loggers. For unit tests, pass no-op loggers. This is the ONLY permitted side effect in Functional Core.

Test signature: Simple assertions, no mocks except logger (if used).

Imperative Shell Files

Contains ONLY:

• I/O operations: file system, database, HTTP, environment
• Orchestration: gather data -> call Functional Core -> persist results
• Error handling for I/O failures
• Minimal business logic (coordination only)

NEVER contains:

• Complex calculations
• Business rule validations
• Data transformations beyond format conversion

Test signature: Integration tests with real dependencies or test doubles.

Code Flow Pattern

1. GATHER (Shell):  Collect data from external sources
2. PROCESS (Core):  Transform input to output (pure)
3. PERSIST (Shell): Save results externally

Every operation follows this sequence. No exceptions.

Decision Framework

Before writing a function, ask:

digraph fcis_decision {
    "Writing a function" [shape=ellipse];
    "Can run without external dependencies?" [shape=diamond];
    "Does it coordinate I/O?" [shape=diamond];
    "Functional Core" [shape=box, style=filled, fillcolor=lightblue];
    "Imperative Shell" [shape=box, style=filled, fillcolor=lightgreen];
    "STOP: Refactor or escalate" [shape=octagon, style=filled, fillcolor=red, fontcolor=white];

    "Writing a function" -> "Can run without external dependencies?";
    "Can run without external dependencies?" -> "Functional Core" [label="yes"];
    "Can run without external dependencies?" -> "Does it coordinate I/O?" [label="no"];
    "Does it coordinate I/O?" -> "Imperative Shell" [label="yes"];
    "Does it coordinate I/O?" -> "STOP: Refactor or escalate" [label="no"];
}

Questions to ask:

• Can this logic run without file system, database, network, or environment?
  • YES -> Functional Core
  • NO -> Does it coordinate I/O or contain business logic?
    • I/O coordination -> Imperative Shell
    • Business logic + I/O -> STOP. Refactor or escalate to user.

Common Mistakes and Rationalizations

Red Flags - STOP and Refactor

If you catch yourself doing ANY of these, STOP:

• File I/O in a "pure" function (open, read, write, exists checks)
• Database passed as parameter to Functional Core (queries, updates, connections)
• HTTP requests in business logic (fetch, axios, requests)
• Environment variables in calculations (process.env, os.getenv)
• Math.random() or Date.now() in Functional Core (non-deterministic)
• Creating a file with runtime behavior without pattern classification comment
• Thinking "just this once" about mixing concerns

All of these mean: Extract I/O to Shell. Pass data to Core. Classify file correctly.

Implementation Patterns

Functional Core Pattern

# pattern: Functional Core

def calculate_total_with_tax(items, tax_rate, logger=None):
    """Pure calculation: same inputs always produce same output."""
    if logger:
        logger.debug(f"Calculating total for {len(items)} items")

    subtotal = sum(item['price'] * item['quantity'] for item in items)
    tax = subtotal * tax_rate
    total = subtotal + tax

    return {
        'subtotal': subtotal,
        'tax': tax,
        'total': total
    }

No I/O. No database. No file system. Only computation.