ia-simplifying-code

Simplifying Code

Principles

Changing an interface, exported name, persisted format, or path reaches past the import graph. Enumerate the producers, consumers, schemas, fixtures, generators, manifests, scripts and CI recipes, config references, and documents that carry the old identifier, and migrate them in the same pass. Close out by searching for the old identifier: zero hits, or one line accounting for each intentional remainder. Renames rot in the fixture holding the old key and the .env.example entry, neither of which any import graph contains. When the identifier is a public or exported API, Stop Conditions applies first -- confirm with the user, then enumerate; the sweep runs unprompted only for internal identifiers.

Process

1. Read first -- understand the full file and its dependents before changing anything. Apply Chesterton's Fence: if you see code that looks unnecessary but don't understand why it's there, check git blame before removing it. First understand the reason, then decide if the reason still applies.
2. Identify invariants -- what must stay the same? Public API, return types, side effects, error behavior
3. Identify targets -- find the highest-impact simplification opportunities. Impact = readability and maintainability; prioritize: control flow -> naming -> duplication -> types (see Smell -> Fix table)
4. Apply in order -- control flow → naming → duplication → data shaping → types. Structural changes first, cosmetic last
5. Verify -- confirm no behavior change: tests pass, types check, imports resolve
6. Pre-submit scope audit -- walk every changed line and ask "does the requested task explicitly require this line?" If no, revert it and list it as a follow-up under Residual Risks. Drive-by edits belong in a separate change, not the current patch. For the pre-edit complement on ambiguous-scope requests ("simplify my project"), see ia-verification-before-completion's Scope Confirmation gate.

Smell → Fix

AI Slop Removal

When simplifying AI-generated code, specifically target:

• Redundant comments that restate the code (// increment counter above counter++) -- delete them
• Unnecessary defensive checks for conditions that cannot occur in context -- remove the guard. Where the guard, retry, workaround, or flag counters an external hazard (a harness default, an upstream bug, a race, a platform quirk), "cannot occur" needs evidence: demonstrate the hazard's precondition is present and handled. A green suite is not that evidence when the run may never have triggered the hazard at all -- absence of failure and absence of the hazard look identical from the outside. If the precondition cannot be reproduced, keep the code and record the gap. Guards against conditions the type system already excludes need no such proof, provided the type is enforced at that boundary rather than merely declared -- deserialized payloads, unchecked API responses, and anything reached through a cast or assertion do not qualify
• Gratuitous type casts (as any, as unknown as T) -- fix the actual type or use a proper generic
• Over-abstraction (factory for 2 objects, wrapper around a single call, util file with 1 function) -- inline the code
• Inconsistent style that drifts from the file's existing conventions -- match the file
• Placeholder stubs (// ..., // rest of code, // similar to above, // continue pattern, // add more as needed) -- leave unsimplified code as-is rather than replacing it with stubs
• Redundant error wrapping (catch(e) { throw e; }, catch(e) { throw new Error(e.message); }) that strips the original stack for no reason -- remove the try/catch entirely and let errors propagate
• Verbose stdlib reimplementations (hand-rolled loops that replicate array_filter, Array.from, Collection::pluck(), itertools) -- replace with the stdlib/framework one-liner, but verify edge-case parity first: empty input, null/None guard, no-match default, zero-value path. The one-liner can silently differ from the loop (an empty-input crash, a missing no-match default, lost ordering) -- a structurally cleaner version that changes behavior on an edge case is not a simplification
• Hand-maintained guarantees the platform, framework, or a downstream layer already enforces (a manual retry wrapping a client that already retries, a hand-rolled TTL cache the ORM/query layer already provides, manual null-coalescing on a value the contract guarantees non-null) -- name the layer that owns the guarantee and what the code collapses to without it. Remove only when it preserves every output, error, side-effect, and ordering; cite the test or a direct comparison proving equivalence, since "it's already guaranteed" over-fires easily
• Copy-paste with variation -- before proposing a shared abstraction, check whether the duplicated construct can be eliminated by deriving it from an existing source of truth (a constant, an existing map, a generated value). Consolidate into a helper only when elimination isn't behavior-preserving and the duplication has already cleared the 3-occurrence gate (Smell → Fix); below that, leave it inline per Constraints

Stop Conditions

Stop and ask before proceeding when:

• Simplification requires changing a public API (function signatures, return types, exports)
• Behavior parity cannot be verified (no tests exist and behavior is non-obvious)
• Code is intentionally complex for domain reasons (performance-critical, protocol compliance)
• Scope implies a redesign rather than a simplification

Constraints

• Only simplify what was requested -- do not add features, expand scope, introduce new dependencies, or add speculative configurability or flexibility the request did not ask for
• Leave unchanged code untouched -- do not add comments, docstrings, or type annotations to lines that were not simplified
• Do not bundle unrelated cleanups into one patch -- each simplification should be a coherent, reviewable unit
• Do not introduce framework-wide patterns while simplifying a small local change
• Do not replace understandable duplication with opaque utility layers -- three similar lines are better than a premature abstraction
• Keep comments that explain intent, invariants, or non-obvious constraints. Remove comments that restate obvious code behavior.
• If a simplification would make the code harder to understand, skip it
• Watch for over-simplification: inlining too aggressively removes names that gave concepts meaning; combining unrelated logic into one function hides distinct responsibilities; removing abstractions that exist for testability breaks the test suite
• When unsure whether a block is dead code, ask instead of deleting

Verify

• Tests pass and types check after changes
• No behavior change (same inputs produce same outputs)
• Scope limited to requested files -- no drive-by cleanups
• Match test scope to the importer count surfaced in step 1 (Surface assumptions). Zero external importers: scoped tests on the changed paths. One or more external importers, or shared/utility code edited: run tests covering each importer. Run the full suite when the test runner has no path-scoping mechanism.

Orchestrator Mode (When Chained With Other Skills)

When this skill is invoked by an orchestrator that also runs ia-code-review, ia-writing-tests, or ia-verification-before-completion on the same scope, each sub-skill re-resolving scope independently wastes tokens and risks drift. Avoid this by resolving scope exactly once and passing a canonical block to every sub-skill.

Resolved scope format — the orchestrator builds this once, before dispatching any sub-skill:

## Resolved scope
Files:
- path/to/file-a.ts
- path/to/file-b.ts

Commit range: HEAD~3..HEAD (or "uncommitted")

Intent: [one-sentence description pulled from the user request or PR description]

Constraints:
- Preserve public API
- No behavior change
- [other constraints specific to this run]

Every chained sub-skill receives this block verbatim in its prompt and uses it as the source of truth — no re-running git diff --name-only, no re-parsing the user request, no independent scope resolution. Sub-skills accept --no-verify --no-report flags when chained so verification and reporting happen once at the end of the chain, not per-skill. The last sub-skill in the chain runs verification; the orchestrator trusts that result rather than re-verifying.

This prevents two failure modes: scope drift (sub-skill A simplifies one set of files, sub-skill B reviews a different set) and double work (every sub-skill rediscovers the same facts).

Integration

• ia-code-simplicity-reviewer agent -- analysis-only pass producing a simplification report (no code changes). Use before refactoring to identify targets.

Output

After simplifying, report:

• Scope touched: files and functions modified
• Key simplifications: what changed and why (one line each)
• Verification: tests pass, types check, no behavior change
• Residual risks: assumptions made, areas not touched that may need attention