Agent-Skills.md

Agent Skills: derot — truth-drift auditor

Truth-drift auditor — find & fix comment/doc/CI/version/dependency rot that contradicts the current code, with root + transitive verification (correct beats delete). Use when the user wants to remove stale comments, fix doc drift (CLAUDE.md/README/CHANGELOG), modernize CI publish auth, sync package versions, or question why a sub-library is used instead of a parent/successor package. Triggers on \"stale comments\", \"derot\", \"doc rot\", \"comments are out of date\", \"docs don't match the code\", \"clean up the docs\", \"why this dependency\", \"überarbeitung\".

UncategorizedID: ANcpLua/ancplua-claude-plugins/derot

Repository

ANcpLua/ancplua-claude-plugins
ANcpLuaLicense: MIT
51

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ANcpLua/ancplua-claude-plugins/tree/HEAD/plugins/derot/skills/derot

Skill Files

Browse the full folder contents for derot.

Download Skill

Loading file tree…

plugins/derot/skills/derot/SKILL.md

Skill Metadata

Name
derot
Description
"Truth-drift auditor — find & fix comment/doc/CI/version/dependency rot that contradicts the current code, with root + transitive verification (correct beats delete). Use when the user wants to remove stale comments, fix doc drift (CLAUDE.md/README/CHANGELOG), modernize CI publish auth, sync package versions, or question why a sub-library is used instead of a parent/successor package. Triggers on \"stale comments\", \"derot\", \"doc rot\", \"comments are out of date\", \"docs don't match the code\", \"clean up the docs\", \"why this dependency\", \"überarbeitung\"."

derot — truth-drift auditor

Rot = any place a repo's stated intent has drifted from what the code/build actually is. derot finds it, verifies it against ground truth, and fixes it — correcting where the knowledge is still useful, deleting only what is purely obsolete.

The five rot dimensions

  1. Comment rot — comments that lie about the code beneath them: stale version numbers, references to renamed/removed symbols/methods/files/diagnostic-IDs, resolved TODO/FIXME/HACK/"temporary"/"workaround for X", descriptions of old behavior, copy-paste comments describing a different member, dead-code comments.
  2. Doc rot — CLAUDE.md / README / CHANGELOG / workflow tables whose dependency versions, file lists, command names, or workflow names don't match reality. (Shipped past CHANGELOG entries are history, not rot — never flag them.)
  3. Infra rot — workflow comments that lie; unpinned actions (@v1 → SHA-pin with a # vN comment); hardcoded publish accounts → dynamic secrets; expired/long-lived API keys → OIDC trusted publishing.
  4. Version sync — pins/claims out of sync across Version.props / Directory.Packages.props / global.json / docs; published-vs-pinned drift; a property referenced but defined only by an injected SDK (annotate, don't "fix").
  5. Dependency rationalization — a narrow sub-package used where a parent/meta suffices; a superseded package still used after its successor shipped; redundant direct references already present transitively. This dimension needs external knowledge, not just the repo — see references/dependency-rationalization.md.

Discipline — non-negotiable

  1. Correct beats delete. If a comment explains a genuinely non-obvious why (a workaround's reason, an invariant), fix the wrong details and KEEP the explanation. Delete only when purely obsolete.
  2. Verify root + every transitive use. Before changing anything a comment/doc describes, read the code it refers to and everything that depends on it. Never "fix" a comment to match a wrong assumption — fix whichever of {comment, code} is actually wrong, and chase the change through every caller.
  3. Never invent. Can't confirm it's stale? Leave it; report it as unverified. Do not guess.
  4. Terse + behavioral corrections. Say what/why; don't restate the code.
  5. No drive-by behavior changes. Comments/docs + the named CI modernizations only — unless fixing rot exposes a real bug, in which case say so and fix the root cause.

Verification recipes (how a candidate becomes a confirmed finding)

  • Version claim → read Version.props / Directory.Packages.props / global.json, and the injected SDK props if a property is undefined locally; confirm the actual resolved value before calling a number stale.
  • Names a symbol / file / diagnostic ID / method / workflow → grep / ls to confirm it still exists and is spelled right.
  • Describes behavior → read the code directly under/around the comment and confirm they agree.
  • Dependency choice → see the references file: the NuGet graph for transitive-already-present + true subset; cited vendor docs / nuget-opensrc / microsoft-docs for succession.

Classify each candidate: confirmed-rot (wrong → fix/delete) · correct-keep-explanation (details wrong, why valid → corrected text that keeps the explanation) · obsolete-delete (purely obsolete → delete) · unverified (cannot confirm → leave + report).

Running a pass

Small scope: inline — scan → verify → fix → build → report.

Large repo: fan out rot-scout agents (read-only, one per area — they parallelise cleanly) and a dep-analyst for dimension 5; synthesize their verified findings into one apply-list; then YOU apply the edits, build/test, and commit. A Workflow is ideal for many areas. The orchestrator — not the scouts — performs the mutations, so comment removal stays under one reviewer's control.

Always finish with: build/test evidence (exact commands + results), a per-change root-cause summary, and a flagged, not changed list (secrets to delete, unverified candidates, dependency calls needing a human decision). derot proposes; the human disposes — especially for dependency swaps, which change the build.