tech-debt-analyzer（科技债务分析器）

Tech Debt Analyzer

Systematic technical debt inventory, prioritization, and remediation planning. Multi-pass analysis with confidence scoring and evidence-based findings.

Scope: Debt inventory and tracking only. NOT for code review (honest-review), refactoring execution, or dependency updates.

Canonical Vocabulary

Dispatch

Mode: Scan

Full codebase debt inventory. Run all 4 analysis scripts, aggregate results, assign categories and severities.

Scan Step 1: Project Profile

Run uv run python skills/tech-debt-analyzer/scripts/complexity-scanner.py <path> to get complexity metrics. Parse JSON output. Flag functions with cyclomatic_complexity > 10 as HIGH, > 5 as MEDIUM.

Scan Step 2: Dead Code Detection

Run uv run python skills/tech-debt-analyzer/scripts/dead-code-detector.py <path> to find unused code. Parse JSON output. Each unused item becomes a debt item (category: design, severity by confidence).

Scan Step 3: Dependency Staleness

Run uv run python skills/tech-debt-analyzer/scripts/dependency-staleness-checker.py <path> to check outdated packages. Parse JSON output. Deprecated packages are CRITICAL. Staleness > 365 days is HIGH.

Scan Step 4: Pattern Consistency

Run uv run python skills/tech-debt-analyzer/scripts/pattern-consistency-checker.py <path> to detect inconsistencies. Parse JSON output. Each inconsistency becomes a debt item (category: design).

Scan Step 5: AI-Augmented Analysis

After script-based detection, perform additional analysis:

1. Documentation gaps — scan for undocumented public APIs, missing README sections, stale comments
2. Test coverage gaps — Grep for untested modules, missing edge cases, test-to-code ratio
3. Infrastructure debt — outdated CI configs, missing linting, inconsistent tooling
4. Design smells — God classes, feature envy, shotgun surgery patterns

Assign confidence scores (0.0-1.0) per finding. Research-validate HIGH/CRITICAL items using Grep and codebase evidence.

Scan Step 6: Aggregate and Classify

Merge all findings into a unified inventory:

• Deduplicate across script outputs
• Assign categories from debt taxonomy (references/debt-taxonomy.md)
• Calculate debt score: sum of (severity_weight x confidence)
• Store baseline at ~/.{gemini|copilot|codex|claude}/tech-debt/<project-slug>-<date>.json

Present findings grouped by category, sorted by severity within each group.

Mode: Analyze

Targeted deep analysis of a specific file or directory. Run all 4 scripts scoped to the target. Apply the same 6-step scan process but with deeper per-function analysis. Include: function-level complexity breakdown, inline dead code, local pattern violations.

Mode: Prioritize

Rank debt items using risk x effort matrix. Load references/prioritization-framework.md.

For each debt item, estimate:

• Risk: blast radius x severity x confidence
• Effort: LOC affected x complexity x dependency count

Output a ranked list with priority labels (P0-P4).

Mode: Roadmap

Generate a phased remediation plan. Requires a prior scan (reads baseline from ~/.{gemini|copilot|codex|claude}/tech-debt/).

Phase structure:

1. Quick Wins (P0 + low-effort P1): immediate fixes, minimal risk
2. Structural (remaining P1 + high-risk P2): design improvements, refactoring
3. Maintenance (P2 + P3): documentation, test coverage, dependency updates
4. Strategic (P3 + P4): architecture changes, long-term improvements

Each phase includes: items, estimated effort, dependencies, success criteria.

Mode: Report

Render dashboard visualization. Requires a prior scan.

1. Read the most recent baseline from ~/.{gemini|copilot|codex|claude}/tech-debt/
2. Copy templates/dashboard.html to a temporary file
3. Inject findings JSON into the <script id="data"> tag
4. Open in browser or report the path

Dashboard sections: category pie chart, complexity heatmap, trend chart (if multiple baselines), prioritized backlog table.

Mode: Track

Compare current scan against previous baseline for longitudinal tracking.

1. Run a fresh scan (Mode: Scan steps 1-6)
2. Load previous baseline from ~/.{gemini|copilot|codex|claude}/tech-debt/<project-slug>-*.json (most recent)
3. Compute delta: new items, resolved items, changed severities, score trend
4. Present comparison report with trend indicators

State Management

• State directory: ~/.{gemini|copilot|codex|claude}/tech-debt/
• Create directory on first use with mkdir -p
• Filename: <project-slug>-<YYYY-MM-DD>.json
• Project slug: sanitized basename of the project root directory
• Store after every scan; track mode reads historical baselines
• Schema: {"project": str, "date": str, "score": float, "items": [...], "summary": {...}}

Reference Files

Load ONE reference at a time. Do not preload all references into context.

Debt Item Structure

Every debt item follows this format:

1. Location: [file:line] or [file:start-end] — exact source location
2. Category: design | test | documentation | dependency | infrastructure
3. Severity: CRITICAL | HIGH | MEDIUM | LOW
4. Confidence: 0.0-1.0 score with evidence basis
5. Description: What the debt is (1-2 sentences)
6. Impact: Why it matters (blast radius, risk)
7. Remediation: Recommended fix approach with effort estimate

Critical Rules

1. Run all 4 analysis scripts before presenting findings — partial scans are labeled as such
2. Every finding must have a confidence score backed by evidence (script output or codebase grep)
3. Confidence < 0.3 = discard; 0.3-0.7 = flag as uncertain; >= 0.7 = report
4. Never execute remediation — this skill inventories and plans, not fixes
5. Store baseline after every full scan — longitudinal tracking depends on it
6. Do not report style preferences as debt — only structural, behavioral, or maintainability issues
7. Deduplicate across script outputs — same file:line should not appear twice
8. Prioritize mode requires a prior scan — prompt user to run scan first if no baseline exists
9. Always present the debt score (aggregate metric) in scan output
10. Track mode must show delta (new/resolved/changed) — raw numbers without comparison are useless
11. Dead code detection requires high confidence (>= 0.8) — false positives erode trust
12. Load ONE reference file at a time — do not preload all references