| name | dev-tech-debt-review |
| description | Detect AI/agentic-specific anti-patterns that traditional linters miss. Analyzes tool/agent boundary violations, prompt debt, context window issues, testing patterns, and more. Returns scored findings with remediation guidance. |
dev.tech-debt-review
Detect AI/agentic-specific tech debt that traditional linters miss.
User Input
$ARGUMENTS
You MUST consider the user input before proceeding (if not empty).
Usage Modes
| Mode | Command | Scope | When to Use |
|---|
| Changed Files | /dev.tech-debt-review | Files changed vs main | Before PR (default) |
| Full Audit | /dev.tech-debt-review --all | Full src/ directory | Release gate, quarterly review |
| Category Focus | /dev.tech-debt-review --category=boundary | Specific category | Targeted cleanup |
Categories: boundary, prompt, context, testing, error, state, observability, subagent
Goal
Detect AI/agentic-specific anti-patterns that traditional linters miss:
| Category | What It Finds | Why It Matters |
|---|
| Tool/Agent Boundary | Tools making decisions, encoding thresholds | Violates Constitution VII |
| Prompt Debt | Hardcoded prompts, no versioning, injection risks | Maintenance nightmare |
| Context Window | Unbounded data, no truncation, raw dumps | Performance/reliability |
| Testing | Live LLM in unit tests, missing VCR, no evals | CI flakiness, cost |
| Error Handling | No retry/backoff, missing timeouts | Reliability |
| State Management | Unbounded conversation history | Memory leaks |
| Observability | No tracing, missing metrics | Debugging blind spots |
| Subagent | Depth > 1, unclear boundaries | Architecture violation |
Traditional linters catch syntax and style. This skill catches semantic anti-patterns specific to agentic systems.
Operating Constraints
STRICTLY READ-ONLY: Do not modify any files. Output analysis and recommendations.
SEMANTIC ANALYSIS: Read and understand code patterns, don't just grep.
EVIDENCE-BASED: Every finding must cite file:line and code snippet.
SCORING: Aggregate findings into 0-100 score for PR gate decisions.
Constitution Alignment
This skill validates adherence to project principles:
| Principle | What We Check |
|---|
| I. Local-First | No data transmission, security of prompts |
| III. Causal-First | Recommendations trace to evidence |
| VII. Intelligent Tooling | Tools provide data, agent provides judgment |
| VIII. Compounding Value | Historical tracking of tech debt |
| IX. Agent-Aware | Context optimized for agent cognition |
Reference: Constitution Mapping
Execution Steps
Phase 0: Scope Determination
You handle this phase directly.
Parse user input to determine scope:
-
If --all flag present: Full codebase audit
find src -name "*.ts" -type f | grep -v ".d.ts" | grep -v "node_modules"
-
If --category=X flag present: Filter to that category's patterns only
-
If specific file path provided: Analyze that file
ls -la <provided-path>
-
Default (no args): Changed files only
git diff --name-only main...HEAD | grep -E '\.tsx?$'
Report mode to user:
--all mode: "Running FULL CODEBASE audit on N files"
--category=X: "Running targeted audit for category: X"
- Default: "Reviewing N files changed vs main"
Load context:
- Read
.specify/memory/constitution.md for principle references
- Check for previous review at
.claude/reviews/tech-debt-*.json for trend comparison
Phase 1: Parallel Category Analysis (Subagents)
Invoke 8 category-specific analyzers IN PARALLEL (single message, multiple Task calls).
Each subagent focuses on one category. Use Haiku for single-file analysis, Sonnet for cross-file patterns.
Task(boundary-analyzer, model=haiku, "Analyze files for Tool/Agent Boundary violations.
PATTERNS TO DETECT:
- JUDGMENT_IN_TOOL: Functions that return judgments ('good', 'bad', 'low', 'high')
- ORCHESTRATION_IN_TOOL: Tools that decide when/how to call other tools
- THRESHOLD_ENCODING: Hardcoded thresholds that encode decisions (if score < 0.3)
- DECISION_RETURN: Functions returning 'suggestedAction', 'shouldDo', etc.
FILES: [list]
For each finding, return:
- pattern: The pattern code (e.g., JUDGMENT_IN_TOOL)
- file: Absolute path
- line: Line number
- snippet: The offending code (max 5 lines)
- evidence: Why this matches the pattern
Return JSON array of findings. Return empty array if no issues.")
Task(prompt-analyzer, model=haiku, "Analyze files for Prompt Debt.
PATTERNS TO DETECT:
- HARDCODED_PROMPT: Prompt strings embedded directly in code (not in prompt files)
- NO_VERSIONING: Prompts without version tracking or timestamps
- INJECTION_RISK: User input concatenated into prompts without sanitization
- PROMPT_SPRAWL: Same prompt logic duplicated in multiple places
FILES: [list]
Return JSON array of findings with pattern, file, line, snippet, evidence.")
Task(context-analyzer, model=haiku, "Analyze files for Context Window issues.
PATTERNS TO DETECT:
- UNBOUNDED_DATA: Arrays/objects passed to agent without size limits
- NO_SUMMARIZATION: Large data returned without truncation/summarization
- RAW_DUMP: Tool returning raw data structures instead of formatted summaries
- CONTEXT_BLOAT: Unnecessary verbose output in tool responses
FILES: [list]
Return JSON array of findings with pattern, file, line, snippet, evidence.")
Task(testing-analyzer, model=haiku, "Analyze files for Testing Anti-patterns.
PATTERNS TO DETECT:
- LIVE_LLM_IN_UNIT: Unit tests that make real LLM API calls
- MISSING_VCR: Integration tests without VCR recordings
- MISSING_EVAL: Behavioral scenarios without TruLens evals
- FLAKY_ASSERTION: Tests asserting on non-deterministic LLM output
- NO_MOCK: Tests calling real external services
FILES: [list]
ADR-0011 Reference: Unit=mocked, Integration=VCR, E2E/Evals=live
Return JSON array of findings with pattern, file, line, snippet, evidence.")
Task(error-analyzer, model=haiku, "Analyze files for Error Handling issues.
PATTERNS TO DETECT:
- NO_RETRY_LOGIC: LLM API calls without retry/backoff
- MISSING_TIMEOUT: API calls without timeout configuration
- SILENT_FAILURE: Errors caught but not logged/propagated
- UNHANDLED_REJECTION: Async operations without error handling
- GENERIC_CATCH: catch(e) without specific error handling
FILES: [list]
Return JSON array of findings with pattern, file, line, snippet, evidence.")
Task(state-analyzer, model=haiku, "Analyze files for State Management issues.
PATTERNS TO DETECT:
- UNBOUNDED_HISTORY: Conversation history without max length
- MEMORY_LEAK: Growing state structures without cleanup
- NO_CHECKPOINT: Long-running operations without checkpointing
- STALE_STATE: Cached state without invalidation strategy
FILES: [list]
Return JSON array of findings with pattern, file, line, snippet, evidence.")
Task(observability-analyzer, model=haiku, "Analyze files for Observability gaps.
PATTERNS TO DETECT:
- NO_TRACING: LLM calls without trace/span context
- MISSING_METRICS: No token counting, latency tracking
- NO_DEBUG_LOG: Complex operations without debug logging
- OPAQUE_ERROR: Errors without context for debugging
FILES: [list]
Return JSON array of findings with pattern, file, line, snippet, evidence.")
Task(subagent-analyzer, model=sonnet, "Analyze files for Subagent Architecture violations.
PATTERNS TO DETECT:
- DEPTH_VIOLATION: Subagents spawning sub-subagents (depth > 1)
- UNCLEAR_BOUNDARY: Subagent with multiple unrelated responsibilities
- MISSING_ISOLATION: Subagent sharing mutable state with parent
- RECURSIVE_SPAWN: Agents that can spawn themselves
FILES: [list]
Constitution C8 Reference: Subagent depth limited to 1
Return JSON array of findings with pattern, file, line, snippet, evidence.")
Wait for all subagents to return.
Phase 2: Constitution Cross-Reference
You handle this phase directly.
For each finding from Phase 1, map to Constitution principles:
Reference: Constitution Mapping
| Pattern | Principle | Severity Modifier |
|---|
| JUDGMENT_IN_TOOL | VII | +2 (direct violation) |
| ORCHESTRATION_IN_TOOL | VII | +3 (critical) |
| INJECTION_RISK | I | +5 (security) |
| DEPTH_VIOLATION | C8 | +4 (architecture) |
| LIVE_LLM_IN_UNIT | ADR-0011 | +3 (testing strategy) |
Phase 3: Severity Scoring
You handle this phase directly.
Base deductions per finding type:
Reference: Severity Criteria
| Category | Pattern | Base Deduction |
|---|
| Boundary | JUDGMENT_IN_TOOL | -8 |
| Boundary | ORCHESTRATION_IN_TOOL | -10 |
| Boundary | THRESHOLD_ENCODING | -6 |
| Prompt | HARDCODED_PROMPT | -3 |
| Prompt | INJECTION_RISK | -15 |
| Prompt | PROMPT_SPRAWL | -4 |
| Context | UNBOUNDED_DATA | -7 |
| Context | RAW_DUMP | -5 |
| Testing | LIVE_LLM_IN_UNIT | -10 |
| Testing | MISSING_VCR | -6 |
| Testing | MISSING_EVAL | -8 |
| Error | NO_RETRY_LOGIC | -6 |
| Error | SILENT_FAILURE | -8 |
| State | UNBOUNDED_HISTORY | -10 |
| State | NO_CHECKPOINT | -5 |
| Observability | NO_TRACING | -5 |
| Observability | MISSING_METRICS | -4 |
| Subagent | DEPTH_VIOLATION | -12 |
Scoring formula:
score = max(0, 100 - sum(deductions))
Score interpretation:
| Score | Grade | Action |
|---|
| 90-100 | A | Ship it |
| 80-89 | B | Minor cleanup before merge |
| 70-79 | C | Plan remediation sprint |
| 60-69 | D | Urgent fixes needed |
| < 60 | F | Block release |
Phase 4: Generate Report
You handle this phase directly.
Output format:
## Tech Debt Review Report
**Score**: XX/100 (Grade: X)
**Trend**: [+/-N vs last review] or [First review - no trend data]
**Files Analyzed**: N
**Scope**: [changed files | full codebase | category: X]
---
### Executive Summary
| Category | Findings | Impact | Constitution |
|----------|----------|--------|--------------|
| Boundary | N | -XX | VII |
| Testing | N | -XX | ADR-0011 |
| ... | ... | ... | ... |
| **Total** | **N** | **-XX** | |
---
### P0 - Block Release
These issues MUST be fixed before merge:
#### [PATTERN_CODE] in file.ts:123
**Code:**
```typescript
// The offending code snippet
Problem: [Why this is an anti-pattern]
Constitution: Violates Principle [X]
Remediation: [Specific fix guidance]
P1 - This Sprint
[Summary findings - less detail than P0]
P2 - Backlog
| Pattern | File | Line | Category |
|---|
| ... | ... | ... | ... |
Historical Trend
| Date | Score | Grade | Change |
|---|
| [today] | XX | X | - |
| [prev] | XX | X | +/-N |
Anti-Pattern Reference
See: AI Anti-Patterns Guide
---
### Phase 5: Save Historical Data
**You handle this phase directly.**
Save review results to `.claude/reviews/tech-debt-YYYY-MM-DD.json`:
```json
{
"version": "1.0.0",
"timestamp": "2026-01-26T14:30:00Z",
"scope": "changed_files|all|category:X",
"filesAnalyzed": 145,
"score": 78,
"grade": "C",
"categories": {
"boundary": { "findings": 3, "impact": -23 },
"testing": { "findings": 5, "impact": -30 },
...
},
"findings": [
{
"pattern": "JUDGMENT_IN_TOOL",
"category": "boundary",
"file": "src/tools/analyzer.ts",
"line": 42,
"snippet": "...",
"severity": 8
}
],
"trend": {
"previousScore": 72,
"previousDate": "2026-01-19",
"change": 6
}
}
Integration with Other Skills
| Skill | Relationship |
|---|
/dev.test-review | Complementary - test-review checks test quality, tech-debt-review checks testing anti-patterns |
/dev.verify-wiring | Sequential - verify-wiring first, then tech-debt-review |
/arch-review | Parallel - different focus areas |
/dev.pr | Gate - score threshold can block PR |
Suggested workflow:
/dev.verify-wiring → /dev.tech-debt-review → /dev.test-review → /dev.pr
Red Flags (Auto-Fail)
These patterns automatically set grade to F:
| Pattern | Why |
|---|
| INJECTION_RISK | Security vulnerability |
| DEPTH_VIOLATION | Architecture violation |
| Multiple ORCHESTRATION_IN_TOOL | Fundamental design issue |
References
Handoff
After running this skill:
- Grade A/B: Proceed to
/dev.pr
- Grade C: Review findings, plan fixes, proceed if time-boxed
- Grade D/F: Fix critical issues, re-run before PR