Flatten Fake Layers

Flatten Fake Layers

Operationalizes the delete-fake-layers doctrine with a script instead of prose. The thesis: a model is bad at finding its own over-abstraction by reading code, but good at judging a concrete candidate the moment a tool points at it. So this skill does not ask you to scan for slop — it runs a detector that flags exact violations, and you apply one judgment per item.

What a "fake layer" is

A function/const that exists only to forward to another call, reshape arguments, or inject a dependency — adding no invariant, validation, lifecycle guarantee, or real ownership. The test (from delete-fake-layers): what becomes true after this layer runs? If the answer is "same data, different name" or "same call, different function", it is a fake layer.

Whole-program tools (knip, fallow) find code with zero callers. They cannot find the layer with one caller that should not exist. That is this script's gap to fill.

When NOT to flatten

A one-caller function earns its keep when it does work the caller would lose by inlining it — not when it merely reads nicer with a name. A nice name is a weak reason and usually an inline target anyway. Real work means it changes types or control flow, or owns a boundary:

• Tightens an optional — requireSession(req): Session narrows Session | undefined to Session (and throws if absent). Inlining loses the narrowing; the caller would re-handle undefined everywhere.
• Throws / asserts — centralizes a failure path and its error, so the caller stays on the happy path with a guaranteed type.
• Enforces an invariant, validates a boundary, or owns a lifecycle (e.g. middleware that sets a tightened context var for every downstream handler, try/finally cleanup).

This is the delete-fake-layers test: what becomes true after this layer runs? If a real guarantee becomes true (a narrowed type, a thrown error, a validated value), keep it — even at one caller, even when it is registered as middleware. If the honest answer is "same data, different name", inline it. The judgment per flagged item is therefore: does this do real work the caller would lose, or does it only forward/reshape/rename?

Step 1 — install and run the detector

cd scripts && npm install            # one-time: installs ts-morph + tsx
node_modules/.bin/tsx find-fake-layers.mts <path/to/tsconfig.json> [pathFilter] [--json]

pathFilter is an optional substring (e.g. src/server) to scope a large repo. --json emits the candidate list as structured evidence for a loop/goal model to consume.

The flagged list is ranked. It reports three shapes:

• di-factory — a factory whose parameter is captured inside the closure it returns (hand-rolled dependency injection). Almost always real slop: import the dependency directly and make the product a module-level value.
• alias — forwards its arguments unchanged to one project function ("same call, different name"). Inline it; replace the reference with the target.
• reshape — builds an object literal and forwards it, annotated with +N const fields (how many constant values it injects). Few const fields = clean inline-upstream target (push the constant to the caller); many = it is assembling a real config block, judge with care.

Suppressed (never flagged): predicates (is/has/should), React hooks (use*), try/finally and validation-boundary bodies, cross-package exports (public API), and framework-registered route handlers. See the TUNING block at the top of the script to adjust these per stack (router method names, monorepo layout, React/Zod signals).

Step 2 — flatten, one candidate at a time

For each flagged item, apply the judgment, then if it is slop use only two operations: delete and inline. Never introduce new indirection; the touched file's line count must not increase.

• di-factory: import the injected dependency directly (it is a static import everywhere else), drop the parameter, and turn the returned product into a module-level const/export.
• alias: delete it; point its single caller at the real function.
• reshape: move the injected constant fields to the call site (or let the upstream function default them) and delete the wrapper.

Step 3 — verify, then re-run

After each change run the project's type-check and lint (e.g. tsc --noEmit, the repo linter). Both must pass. Then re-run the detector — a flattened candidate disappears from the list. Repeat until the flagged list contains only items you have judged "keep". A closed list means done.

Running it as a goal loop

To run this autonomously across a whole codebase, feed prompts/goal.md to a goal/loop runner (Codex --goal, Claude Code, etc.). It instructs the model to run the detector, triage each candidate against the judgment, inline the real slop, verify, and re-run until the list is clean — with the script as the ground-truth feedback signal on every iteration.