ln-512-tech-debt-cleaner（ln 512 科技债务清理者）

Paths: File paths (shared/, references/, ../ln-*) are relative to skills repo root. If not found at CWD, locate this SKILL.md directory and go up one level for repo root. If shared/ is missing, fetch files via WebFetch from https://raw.githubusercontent.com/levnikolaevich/claude-code-skills/master/skills/{path}.

Tech Debt Cleaner (L3 Worker)

Type: L3 Worker

Automated cleanup of safe, low-risk tech debt findings from codebase audits.

Purpose & Scope

• Consume audit findings from docs/project/codebase_audit.md (ln-620 output) or ln-511 code quality output
• Filter to auto-fixable findings with confidence >=90%
• Apply safe fixes: remove unused imports, delete dead code, clean commented-out blocks, remove unsupported aliases
• Never touch business logic, complex refactoring, or architectural changes
• Create single commit with structured summary of all changes
• Invocable from ln-510 quality coordinator pipeline or standalone

Auto-Fixable Categories

NOT Auto-Fixable (skip always)

When to Use

• Use after code quality analysis when safe low-risk cleanup can be applied
• Standalone: After ln-620 codebase audit completes (user triggers manually)
• Scheduled: As periodic "garbage collection" for codebase hygiene

Inputs

• Pipeline mode (ln-510): findings from ln-511 code quality output (passed via coordinator context)
• Standalone mode: docs/project/codebase_audit.md (ln-620 output)

MANDATORY READ: Load shared/references/mcp_tool_preferences.md — ALWAYS use hex-line MCP for code files when available. No fallback to standard Read/Edit unless hex-line is down.

MANDATORY READ: Load shared/references/mcp_integration_patterns.md.

Use hex-line as the primary path for code files and hex-graph as the primary path for dead-code reference checks. Built-in Read/Edit/Grep are fallback only when the relevant MCP is unavailable.

Workflow

1. Load findings: Read docs/project/codebase_audit.md. Parse findings from Dead Code section (ln-626 results) and Code Quality section (ln-624 results).
2. Filter to auto-fixable:

- Category must be in Auto-Fixable table above - Severity must be LOW or MEDIUM (no HIGH/CRITICAL) - Effort must be S (small) - Skip files in: node_modules/, vendor/, dist/, build/, *.min.*, generated code, test fixtures

1. Verify each finding (confidence check): MANDATORY READ: Load shared/references/clean_code_checklist.md For each candidate fix: a) Read the target file at specified location b) Confirm the finding still exists (file may have changed since audit) c) Confirm removal is safe: Hex-line acceleration (if available): IF hex-line MCP server is available:

- For unused imports: grep codebase for usage (must have 0 references) - For unused functions: grep for function name (must have 0 call sites) - For commented-out code: verify block is code, not documentation - For unsupported aliases: verify no consumers remain d) Assign confidence score (0-100). Only proceed if confidence >=90 - Use outline(file_path) and discovery-first read_file() before manual cleanup edits. Re-read with edit_ready=true, verbosity="full" only when you need revision/checksums for edit_file. - Batch cleanup: When fixing >3 files with same pattern (e.g., unused import removal), use bulk_replace(dry_run=true) to preview, then bulk_replace() to apply. - Verified edits: After each fix, verify(file_path, checksums) to confirm no stale state. - Semantic dead-code check: Use find_references() before deleting exports, wrappers, aliases, or shims. - Fall back to per-file Edit or Grep only if the relevant MCP is unavailable.

1. Apply fixes with per-fix keep/discard (autoresearch pattern): MANDATORY READ: Load shared/references/ci_tool_detection.md for discovery hierarchy. Detect lint + typecheck commands once (reuse for all fixes). Group verified fixes by file. For each file (process files independently): If no lint/type commands detected: apply all fixes, skip per-file verification with warning, git add all modified files.

- Sort fixes within file by line number descending (bottom-up prevents line shift) - Apply ALL fixes for this file using Edit tool - Run lint/typecheck on the modified file - IF passes → git add {file} (status: keep) - IF fails → git checkout -- {file} (status: discard), log discarded fixes - Track per fix: file, lines removed, category, finding ID, status (keep/discard)

1. Create commit (kept fixes only):

- All kept files already staged via git add in step 4 - If zero files kept (all discarded): skip commit, report all failures - Commit message format: chore: automated tech debt cleanup Removed {N} auto-fixable findings from codebase audit: - {count} unused imports - {count} dead functions - {count} commented-out code blocks - {count} unsupported aliases Source: docs/project/codebase_audit.md Confidence threshold: >=90%

1. Update audit report:

- Add "Last Cleanup" section to docs/project/codebase_audit.md: ## Last Automated Cleanup **Date:** YYYY-MM-DD **Findings fixed:** N of M auto-fixable **Skipped:** K (confidence <90% or verification failed) **Build check:** PASSED / SKIPPED

Output Format

verdict: CLEANED | NOTHING_TO_CLEAN | ALL_DISCARDED
stats:
  total_findings: {from audit}
  auto_fixable: {filtered count}
  kept: {files that passed lint/typecheck}
  discarded: {files that failed lint/typecheck}
  skipped: {confidence <90 or stale}
fixes:
  - file: "src/utils/helpers.ts"
    line: 45
    category: "unused_function"
    removed: "formatDate()"
    finding_id: "MNT-DC-003"
    status: "keep"
  - file: "src/api/v1/auth.ts"
    line: 12
    category: "unsupported_alias"
    removed: "export { newAuth as oldAuth }"
    finding_id: "MNT-DC-007"
    status: "discard"
    discard_reason: "typecheck failed: Type error in auth.ts:15"
commit_sha: "abc1234" | null

Critical Rules

• Safety first: Never fix if confidence <90%. When in doubt, skip.
• Bottom-up editing: Always apply fixes from bottom to top of file to avoid line number shifts.
• Per-file keep/discard: If linter/type-checker fails for a file, revert only that file (git checkout -- {file}), keep other successful files.
• No business logic: Never modify function bodies, conditionals, or control flow.
• Explicit staging: Stage files by name, never git add. or git add -A.
• Idempotent: Running twice produces no changes if audit report unchanged.
• Git-aware: Only operate on tracked files. Skip untracked or ignored files.
• Exclusions: Skip generated code, vendor directories, minified files, test fixtures.

Runtime Summary Artifact

MANDATORY READ: Load shared/references/quality_summary_contract.md, shared/references/quality_worker_runtime_contract.md

Runtime profile:

• family: quality-worker
• worker: ln-512
• summary kind: quality-worker
• payload fields used by coordinators: worker, status, verdict, issues, warnings, artifact_path

Invocation rules:

• standalone: omit runId and summaryArtifactPath
• managed: pass both runId and exact summaryArtifactPath
• always write the validated summary before terminal outcome

Definition of Done

• Audit report loaded and parsed
• Findings filtered to auto-fixable categories
• Each finding verified with confidence >=90%
• Fixes applied bottom-up per file
• Build integrity verified (lint + type check) or skipped with warning
• Single commit created with structured message (or all reverted on build failure)
• Audit report updated with "Last Automated Cleanup" section
• Output YAML returned to caller

Reference Files

• Clean code checklist: shared/references/clean_code_checklist.md
• Audit output schema: shared/references/audit_output_schema.md
• Audit report template: shared/templates/codebase_audit_template.md

Version: 1.0.0 Last Updated: 2026-02-15