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

Excuse/Thought Pattern Reality What To Do

"Just one file read in this calculation" File I/O = side effect. Not Functional Core. Extract to Shell. Pass data as parameter.

"Database is passed as parameter, so it's pure" Database operations are I/O. Not pure. Move to Shell. Core receives data, not DB connection.

"This validation needs to check if file exists" File system check = I/O. Not Functional Core. Shell checks file, passes boolean to Core validation.

"Small HTTP call, won't hurt" HTTP = side effect. Breaks purity guarantee. Shell makes request, Core processes response data.

"Need Date.now() for timestamp calculation" Non-deterministic. Not pure. Shell passes timestamp as parameter.

"Logging is a side effect, should remove" WRONG. Logging is explicitly permitted. Keep logger. This is the exception.

"This function does both logic and I/O, but it's simpler" Mixed concerns = untestable without mocks. Split into Core (logic) + Shell (I/O). Test Core simply.

"File classification is overhead" Prevents entire classes of bugs. Non-negotiable. Add classification comment. Takes 10 seconds.

"I'll refactor later" Later never comes. Do it now. Classify and separate now.

"Performance requires mixing" Prove it with benchmarks. Usually wrong. Separate first. Optimize with evidence. Mark Mixed (unavoidable) with justification.

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.

Imperative Shell Pattern

pattern: Imperative Shell

def process_order(order_id, db, logger): """Orchestrates: gather -> process -> persist."""

# GATHER: Collect data from external sources
items = db.get_order_items(order_id)
tax_rate = db.get_tax_rate_for_order(order_id)

# PROCESS: Call Functional Core (pure logic)
result = calculate_total_with_tax(items, tax_rate, logger)

# PERSIST: Save results externally
db.update_order_total(order_id, result['total'])

return result

Shell is thin. Core does heavy lifting. Testable separately.

Mixed (Needs Refactoring) - Bad Example

pattern: Mixed (needs refactoring)

def calculate_and_save_total(order_id, db): """BAD: Mixes calculation with I/O. Hard to test.""" items = db.get_order_items(order_id) # I/O subtotal = sum(item['price'] for item in items) # Logic tax_rate = db.get_tax_rate_for_order(order_id) # I/O tax = subtotal * tax_rate # Logic total = subtotal + tax # Logic db.update_order_total(order_id, total) # I/O return total

Testing this requires database mocks. Fragile. Refactor using patterns above.

Logger Exception Details

Loggers are EXPLICITLY PERMITTED in Functional Core.

pattern: Functional Core

def validate_order(order_data, logger=None): """Pure validation with logging.""" if logger: logger.info(f"Validating order {order_data.get('id')}")

errors = []

if not order_data.get('items'):
    errors.append("Order must have items")

if order_data.get('total', 0) < 0:
    errors.append("Total cannot be negative")

if logger and errors:
    logger.warning(f"Validation failed: {errors}")

return {'valid': len(errors) == 0, 'errors': errors}

For unit tests: Pass no-op logger or None. Function remains pure for testing.

Refactoring Patterns

Common patterns for separating concerns:

Extract Pure Core from Impure Functions

Symptom: Function mixes I/O with logic

BEFORE - hard to test

def process_order(order_id: str) -> None: order = db.fetch(order_id) # I/O discount = calculate_discount(order) # Pure logic total = apply_discount(order, discount) # Pure logic db.save(order_id, total) # I/O

AFTER - pure core extracted

def calculate_order_total(order: Order, rules: DiscountRules) -> Decimal: """Pure function - easy to test.""" discount = calculate_discount(order, rules) return apply_discount(order, discount)

def process_order(order_id: str) -> None: """Thin I/O wrapper.""" order = db.fetch(order_id) total = calculate_order_total(order, get_discount_rules()) db.save(order_id, total)

Return Values Instead of Mutating

Symptom: Methods mutate in place, making before/after comparison hard

BEFORE - mutation

def sort_tasks(tasks: list[Task]) -> None: tasks.sort(key=lambda t: t.priority)

AFTER - returns new value

def sorted_tasks(tasks: list[Task]) -> list[Task]: return sorted(tasks, key=lambda t: t.priority)

Add Missing Inverse Operations

Symptom: One-way operation exists but no inverse for testing roundtrips

BEFORE - only encode exists

def encode_message(msg: dict) -> bytes: return msgpack.packb(msg)

AFTER - add decode for roundtrip testing

def decode_message(data: bytes) -> dict: return msgpack.unpackb(data)

Replace Hardcoded Dependencies

Symptom: Functions use globals or hardcoded config, can't test edge cases

BEFORE - uses global

def validate_input(data: str) -> bool: return len(data) <= CONFIG.max_length

AFTER - dependency injected

def validate_input(data: str, max_length: int) -> bool: return len(data) <= max_length

Refactoring Priority

Pattern Impact Effort Priority

Extract pure core HIGH Medium Do first

Add missing inverse HIGH Low Quick win

Return instead of mutate MEDIUM Low Easy improvement

Inject dependencies MEDIUM Medium When testing blocked

Refactoring Checklist

When you find mixed concerns:

• Identify pure computations (logic, calculations, validations)

• Extract pure code to Functional Core file

• Identify I/O operations (file, database, HTTP, environment)

• Keep I/O in Imperative Shell file

• Shell gathers data, calls Core, persists results

• Add pattern classification comments to both files

• Test Core with simple assertions (no mocks except logger)

• Test Shell with integration tests

If you cannot separate: Escalate to user with specific technical justification. Don't assume mixed is necessary.

Summary

FCIS in three rules:

• Functional Core: Pure functions only. No I/O except logging. Easy to test.

• Imperative Shell: I/O coordination only. Minimal logic. Calls Core.

• Classify every file with runtime behavior. Type-only files, constants, barrels, tests, and generated files are exempt.

When in doubt: Can it run without external dependencies? -> Functional Core. Otherwise -> Imperative Shell.

Logging exception: Loggers permitted everywhere. Pass no-op logger for unit tests.

Mixed concerns = refactoring needed. Extract, separate, classify. Do it now, not later.