Deburr Edge Cases Skill: Make coding agents systematize their local code

A new structured code quality pass, 'deburr-edge-cases', helps coding agents systematically remove AI-generated cruft from branches before merge, targeting defensive checks, redundant validation, and speculative abstraction while preserving load-bearing seams.

| name | deburr-edge-cases | |---|---| | description | Strip AI-codegen cruft from a branch — defensive checks the type system should make unrepresentable, redundant validation, speculative abstraction, no-op wrappers — while preserving load-bearing seams. Use after a feature lands and before merge, when asked to "simplify", "de-ornament", "remove cruft", "make erroneous states unrepresentable", or "clean up" a diff. | A structured pass to remove accumulated cruft from a branch — particularly the failure mode where code solves, at a narrow local layer, a problem that should be addressed globally or structurally. The canonical example: instead of enforcing design intent in the type system making erroneous states unrepresentable , the code accumulates defensive runtime checks to detect and flag bad states that a better type would forbid at compile time. This is a quality pass, not a bug hunt. Every change must be behavior-preserving, with the existing test suite as the safety net. It complements a correctness-focused review which hunts bugs ; route bug findings there, not here. If you trip over a genuine correctness bug mid-pass, note and surface it separately — do not fix it as part of this pass. A recognizable shape — an unwrap or , a single- impl trait — is a prompt to ask whether the construct earns its keep here , not a finding in itself. Reason about intent and invariants, not form. - After a feature branch lands, before merge — especially work assembled by several contributors or agents, which tends to accrete per-phase scaffolding that reads as cruft once the whole is in view. - When the user asks to "simplify", "de-ornament", "remove cruft / ceremony / boilerplate", "make bad states unrepresentable", "tighten the types", or "clean up" a set of changes. Defensive coding the type system should forbid. Runtime checks, branches, or asserts guarding states a tighter type would make unrepresentable — for example a tri-state Option<bool , a sentinel "" / -1 / 0 standing in for an Option /enum, a field validated on every read rather than made unconstructable-when-invalid, or a "this can't happen" guard. Redundant / over-eager validation. Re-checking an invariant already guaranteed upstream e.g. re-validating what a parse step already proved . Speculative generality / over-abstraction. A trait, generic, indirection layer, config knob, or helper with a single call site and no second implementor in view. No-op wrapper methods that just forward to another method with no distinct behavior. Ornamental error handling. Error/diagnostic variants for conditions that cannot occur, or a fan-out of variants where one would do. Dead-but-dressed-up scaffolding that is not a deliberate, documented forward seam. Do not strip things that earn their keep. Call them out as intentional rather than flagging them: Genuine architectural seams : a trait that is the documented extension point for future implementors even with one impl today , mirroring an established pattern in the codebase. The sole enforcement point : input validation that is the only guard e.g. validating tool-call arguments when the schema is advisory and never enforced at the boundary . Concurrency primitives matching the actual ownership an AtomicBool guard through a shared &self can't be replaced by take-by-value Option if the caller holds &self . Test seams : a generalization e.g. over AsyncWrite that has a real second caller in tests. Required safety rails : a bound or check mandated by a spec/requirement is not "added cruft" even if the current structure makes it hard to hit — keep it.- A DI/context struct that is over-built today but has a confirmed near-term second consumer. When unsure whether something is cruft or load-bearing: it is load-bearing until proven otherwise. Default to keeping. Finding no ornamentation is a valid and common result — especially on small or already-tight diffs. Report "clean" rather than inventing a finding to justify the pass. Run the steps directly, or route the fan-out and edits through whatever review/delegation process the repo otherwise follows. Keep edits one-at-a-time and behavior-preserving, and review each before moving on. Match effort to the diff: for a few files, scan and fix directly; the clustering and per-cluster fan-out below are for large, multi-author branches. Compute the cumulative branch diff against the merge base and list the changed source files, excluding generated/state files lockfiles, task-tracker state unless relevant. This is the review surface. Partition the changed files into coherent clusters e.g. parse/types, registry/storage, config/state-modeling, wiring/transport . Take one cluster at a time — or, if the repo supports parallel read-only reviewers, one reviewer per cluster. For each, apply: - the precise lens above the five ornamentation categories + the preserve list , - the file cluster and the specific suspects to scrutinize, verify suspicious editor diagnostics against before reporting, cargo check --all-targets - a consistent finding shape: one-line title; confidence high/med/low that it's a real simplification, not a behavior change ; file:location ; the specific cruft; a crisp remedy prefer "replace runtime check X with type Y" ; behavior-preserving vs. behavior-changing. Separate clear wins from judgment calls , and name anything judged intentional / load-bearing so it is not re-litigated. The review stage reports findings only — it does not edit. Merge the findings. Drop: - anything behavior-changing that isn't clearly an improvement, - anything that removes a spec-required rail, - low-value cosmetic churn. Keep the type-system wins and genuine no-op removals. The headline finding is usually a single structural remodel e.g. tri-state Option<bool → Option<Resolved so the bad state is unrepresentable ; the rest are small isolated cleanups. Take one finding or one tightly-coupled pair at a time. Pin the exact contract before editing — especially for a structural refactor, write the target type design and the invariant truth-table the change must preserve, plus a test-migration map migrate tests to the new shape; never weaken assertions to pass . Execute in increasing invasiveness : small, isolated, behavior-preserving cleanups first; the big structural remodel last, on an already-cleaned base. After each change, verify cargo and review it before moving on. If no existing test exercises the changed path, the suite is not a safety net for it — add a characterization test first, or lower the finding's confidence and flag the gap. Before deleting a "no-op", confirm it is one. A classic trap: a let = &field, … discard or an allow that looks ornamental but actually suppresses a real dead code /lint warning e.g. derive Deserialize does not count as a read, and derive Debug is ignored by dead-code analysis — so fields only read via derives genuinely warn . If removing the "cruft" produces a warning or breaks a check, restore it and report rather than deleting into a broken build. Don't trade ornament for a warning. Run the full check suite on the whole de-ornamented stack: cargo fmt --check , cargo clippy --all-targets --all-features , cargo nextest run , ratchets check . Confirm no assertion was weakened or deleted to make a change pass — tests that only covered genuinely-removed code may go with it, but everything else must be migrated, not loosened. Tri-state collapse → unrepresentable bad state. enabled: Option<bool rewritten None → Some true at assembly and read via unwrap or false , then carried dead into a component that only exists when enabled. Fix: split the parse DTO from a resolved type; make the consumer field Option<ResolvedConfig with no , so "disabled" = enabled field None by construction. Deletes the rewrite, the accessor, the unwrap or , the dead field, and the explanatory comments that existed only to compensate for the confusing type. Sentinel → A helper returning an owned Option . String with "" meaning "absent", forcing a downstream is empty && … guard. Fix: return Option<&str ; the guard and an allocation vanish. Namespacing struct → free function. A zero-field struct FooThing; whose only purpose is to host one associated fn that is never called on an instance and implements no trait. Fix: make it a module-level free fn , delete the struct/impl. Distinguish from the real seam — the trait + concrete type it delegates to — which stays. No-op wrapper method. A trait default summarize that just returns self.list , with no override and one caller. Fix: delete it; the caller uses list . Reintroduce only if a second, genuinely-distinct behavior ever materializes. - Deleting something into a compiler/clippy warning verify first; restore if it warns . - Removing a spec-mandated safety bound because the current structure makes it hard to hit. - Stripping a documented forward seam or the sole enforcement point of an invariant. - Weakening or deleting test assertions to make a refactor pass — migrate them instead. - Doing the big structural remodel first, or batching everything into one giant unreviewable commit. - Treating the pass as a bug hunt — it is quality-only; route correctness findings to a dedicated bug-hunting review.