Coding Agent Quality Rules (Galahad Principle)

Based on Jonathan Lange's "The Galahad Principle": https://jml.io/galahad-principle/

Core idea: getting to 100% yields disproportionate value—especially simplicity and trust. When checks are truly "all green", any new failure is a strong, unambiguous signal; "absence of evidence becomes evidence of absence".

Assess Before Applying

Before enforcing these rules strictly, understand the context:

1. Read project conventions: Check tsconfig.json, pyproject.toml, .eslintrc, setup.cfg, mypy.ini for existing standards
2. Gauge existing tech debt: If the codebase already has 500 any types, don't block progress on fixing all of them
3. Match scope to task: A quick bug fix ≠ a new feature ≠ a refactor

When working in a codebase that doesn't meet these standards:

• Don't make things worse: No new type escapes, no new skipped tests
• Opportunistically improve: Clean up what you touch
• Don't block the user's goal: Pragmatic progress beats ideological purity
• Use code ratchets to improve over time: Use the "code-ratchets" skill to improve patterns over time

Discovering project standards

TypeScript: Check tsconfig.json for strict, noImplicitAny, strictNullChecks. Match existing settings.

Python: Check for mypy.ini, pyproject.toml [tool.mypy], pyrightconfig.json. Note the strictness level.

General: Look at existing test files for patterns, existing code for style. When in doubt, match what's there.

Non-negotiables: never evade feedback

Treat type errors, test failures, pre-commit hooks, lint errors, and coverage warnings as helpful feedback. Fix root causes.

Forbidden by default (unless the user explicitly orders it)

• Type escapes / silencing
  • TypeScript: any, sketchy unknown laundering, unchecked casts, as any, @ts-ignore, disabling strict mode, weakening compiler flags
  • Python: # type: ignore, # pyright: ignore, # mypy: ignore-errors, cast() without justification, Any in public APIs, disabling type checkers
  • General: noqa, pragma comments to silence legitimate warnings
• Coverage gaming
  • TypeScript: /* istanbul ignore */, /* c8 ignore */, artificial exclusions in config
  • Python: # pragma: no cover, # coverage: skip, excluding entire modules from coverage config
  • General: "generated" file tricks, decorator/macro suppression, lowering coverage thresholds
• Faking results
  • Skipping CI steps and claiming success; "snapshotting" coverage; lowering thresholds; marking tests flaky to ignore them

When user requests conflict with these principles

If the user explicitly asks for a type escape, to skip tests, or similar:

1. Comply, but note the tradeoff: "Adding any here—this will need cleanup before the type system can catch errors in this area."
2. Offer alternatives briefly: "If you prefer, I could extract this to a small typed helper instead."
3. Don't lecture: One sentence, then move on.

The user owns the codebase. Your job is to inform, not obstruct.

Priorities

Type safety is part of correctness and outranks tests.

When tradeoffs exist, prioritize in this order:

1. Type safety / soundness
2. Correctness + meaningful tests
3. Clarity / maintainability
4. Performance
5. Backwards compatibility

Breaking changes are acceptable when they improve verifiability and simplify the system, but:

• Flag breaking changes explicitly to the user
• Prefer non-breaking improvements when effort is similar
• Consider migration paths for public APIs

Default workflow (when anything fails)

1. Read the failure output carefully.
2. Understand the context: Why does this code exist? What was the original intent? Check git history or ask if unclear.
3. Restate the real invariant being violated in plain English.
4. Fix the root cause (not the symptom).
5. Improve tests so the behavior is pinned and regressions get caught.
6. Refactor production code if needed to make it easy to type-check and validate.

Run checks in this order

1. Typecheck
2. Unit tests
3. Integration tests
4. Doc and End-to-End tests
5. Lint / pre-commit
6. Coverage

Goal: a repo where "all green" is normal, and any new red is a loud, trustworthy signal.

What makes a test meaningful

✅ Meaningful tests:

• Test observable behavior from the caller's perspective
• Would catch real regressions
• Document intent and edge cases
• Fail when actual bugs are introduced

❌ Not meaningful:

• Test implementation details (private methods, internal state)
• Duplicate what the type checker already verifies
• Assert only on mock interactions, not outcomes
• Pass regardless of whether the code works

The test: "If this test failed, would I learn something useful about a real bug?"

Coverage: aim for meaningful, not mechanical

• Do: Cover all business logic paths, edge cases, error handling
• Don't: Chase 100% by testing trivial getters or truly unreachable defensive code
• Legitimate exclusions exist: Platform-specific branches, debug-only code, abstract method stubs
• The bar: Would a failure in this line indicate a real bug? If yes, cover it.

Coverage comes from exercising real behavior, not from exclusion comments.

Handling flaky tests

If a test is genuinely flaky:

1. Identify the source: Time-dependence? Race condition? External service? Order-dependence?
2. Fix the non-determinism: Inject clocks, add synchronization, mock external calls, isolate state
3. If unfixable now: Quarantine in a separate test suite (not skipped, but run separately and tracked)
4. Never: Mark as "expected flaky" and leave in the main CI path

"Hard to test" means refactor

If something is hard to test or hard to type, treat it as a design smell.

Refactor towards:

• Smaller pure functions
• Explicit data flow, minimal global state
• Clear boundaries between logic and side effects
• Typed domain models over stringly-typed data
  • TypeScript: strong interfaces/types instead of Record<string, any>
  • Python: dataclasses, Pydantic models, or TypedDicts instead of dict[str, Any]

Mocks: use sparingly and explicitly

Avoid injecting mocks via monkeypatching or replacing system utilities by default.

Preferred approach:

• Make the function under test able to operate in multiple environments by passing in the substitutable operations explicitly (as function parameters or small interfaces)
• Only do this for operations that genuinely need substitution in tests: time, randomness, network, filesystem, process execution
• This makes the injection point explicit, documents what varies, and keeps tests honest

Examples:

TypeScript:

// ❌ Bad: hard-coded dependency, requires monkeypatching to test
function processOrder(orderId: string) {
  const now = new Date();
  const order = database.getOrder(orderId);
  // ...
}

// ✅ Good: explicit dependencies
function processOrder(
  orderId: string,
  deps: { getTime: () => Date; getOrder: (id: string) => Order }
) {
  const now = deps.getTime();
  const order = deps.getOrder(orderId);
  // ...
}

Python:

# ❌ Bad: hard-coded dependency, requires monkeypatching to test
def process_order(order_id: str) -> OrderResult:
    now = datetime.now()
    order = database.get_order(order_id)
    # ...

# ✅ Good: explicit dependencies
def process_order(
    order_id: str,
    *,
    get_time: Callable[[], datetime] = datetime.now,
    get_order: Callable[[str], Order] = database.get_order,
) -> OrderResult:
    now = get_time()
    order = get_order(order_id)
    # ...

Summary: what "good" looks like

• Types encode invariants; no "trust me" casts
• Tests assert observable behavior (not implementation trivia)
• Coverage comes from exercising real behavior, not exclusions
• If a thing can't be verified cleanly, refactor until it can
• Progress beats perfection; don't make things worse, do make things better