tech-debt
Audit the codebase for structural improvement opportunities — SOLID/composability, code duplication, and unnecessary layering — using a parallel team of review agents. Ranks the 10 most impactful findings, excludes anything already in the open backlog, and walks the owner through them one at a time with full context; the owner decides per item whether to file it as a GitHub issue. Not tied to any issue; owner-invoked, never automatic.
rustyrazorblade · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add rustyrazorblade/skills --skill tech-debt技能说明
name: tech-debt description: Audit the codebase for structural improvement opportunities — SOLID/composability, code duplication, and unnecessary layering — using a parallel team of review agents. Ranks the 10 most impactful findings, excludes anything already in the open backlog, and walks the owner through them one at a time with full context; the owner decides per item whether to file it as a GitHub issue. Not tied to any issue; owner-invoked, never automatic. argument-hint: [optional: scope to a path/module — default is the whole repo]
tech-debt — structural audit → ranked findings → owner files what's worth it
Foreground, standalone: no worktree, no issue coupling, no code edits. A team of review agents reads the whole codebase looking for structural debt, ranks what it finds, and the owner decides — one finding at a time — what's worth turning into real backlog work.
Steps
-
Establish scope. Default is the whole repo. If the owner (or
argument-hint) named a path/module, scope the review to it and say so up front. -
Snapshot the existing backlog, once, before spawning anything. The whole point is to never re-surface what's already known:
gh issue list --state open --json number,title,url --limit 200Keep this list in context — you'll hand it to every lens agent below so they self-filter, and you'll do a final cross-check yourself before presenting anything to the owner.
-
Spawn three lens agents in parallel (one Agent tool call each, sent together — not sequential), each
subagent_type: general-purpose,model: opus(the strongest available tier — this is a deep, whole-codebase read, not a quick lookup). Give every lens the scope from step 1 and the open-issue list from step 2 verbatim, with the instruction to silently drop any candidate that duplicates an existing open issue's title/body rather than flagging it. Each lens returns its own ranked candidate list — file:line evidence, the concrete problem, and a concrete direction to fix it, never a vague "consider refactoring X":- SOLID / composability — "Is the codebase built from small, composable components, or from large ones doing several unrelated jobs? Find the worst violations: single-responsibility breaks (a class/module/function doing 2+ unrelated things), dependencies on concretions where an abstraction would decouple two things that shouldn't know about each other, and interfaces wide enough that most callers only use a slice of them. For each, name the smaller pieces it should split into."
- Code duplication — "Find logic duplicated across files/modules that belongs in one shared place — not textually identical-only, but the same behavior re-implemented with minor variations (a strong signal the abstraction was never extracted, or drifted after copy-paste). For each, name where the shared module/function should live and what call sites would move to it."
- Structure / layering — "Find unnecessary indirection: layers that just forward calls to the next layer with no logic of their own, wrapper types/interfaces with exactly one implementation and no test-seam justification, or a module boundary that's crossed so often in both directions it isn't really a boundary. For each, name what should collapse or merge."
Each lens caps itself at its own strongest ~5-8 candidates rather than padding to hit a number — a lens with fewer genuine findings should return fewer.
-
Merge, dedupe, and rank globally. Collect all three lenses' candidates. Two different lenses sometimes point at the same underlying spot (a god-class is both a SOLID violation and the reason its duplicated logic never got extracted) — merge those into one finding citing both angles rather than presenting it twice. Do your own pass against the step-2 backlog list too (lens-level self-filtering can miss a near-match) — drop anything that's substantially already an open issue. Rank what's left by impact: how much it currently costs (bug surface, onboarding friction, how often it's touched) weighed against how contained the fix is. Keep the top 10 — fewer if fewer genuinely impactful findings survive; never pad the list to reach 10.
-
Present findings one at a time — full context, owner decides each one before you show the next. Don't dump the ranked list upfront. For each:
[3 of 10] <one-line title> Lens: <SOLID | duplication | structure> Where: <file:line, file:line, ...> Problem: <what's actually wrong, concretely> Direction: <the shape of the fix — not a full design, just where it points> Impact: <why this one ranked where it did>Ask: file as a GitHub issue, skip, or adjust (owner can narrow/reshape it before it's filed). Never file anything without that per-item confirmation — the owner decides what enters the backlog, this skill only surfaces candidates.
-
File exactly what the owner accepted, drafted with a distinct
## Directionsection (not folded into Notes) — keep it as its own heading, verbatim; some downstream consumer of the filed issue may key off that section being present:gh issue create --title "<concise title>" --label "<P0|P1|P2|P3>" --label "status:ready" --label "type:tech-debt" \ --body "## Scope <what this touches, in/out> ## Acceptance criteria - [ ] <behavior is unchanged — existing tests still pass unmodified, no public signature/error-contract/CLI/config/serialized-output change> - [ ] <the structural outcome itself — e.g. 'X and Y are split into single-responsibility units'> ## Direction <the finding's Direction field, verbatim — the shape of the fix, where it points> ## Notes / context Lens: <SOLID | duplication | structure>. Where: <file:line, file:line, ...>. Problem: <what's wrong>."The first acceptanc