메뉴
Checks the codebase for deprecated imports, broken module imports, and common anti-patterns. Use after major refactoring, after adding new modules, or before committing large changes. Takes ~2 minutes. Non-destructive — read-only.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Audit and selectively re-run the deterministic analysis engine across all 26 corpus architectures. Default (no args) shows a staleness check — which reports are missing metadata, which arch types need engine fixes re-applied, and the current corpus score table. Pass --ai-only to re-run only AI/agentic architectures, --all to re-run everything, or a single arch name to re-run one. Use after engine changes (self_validation.py, ground_truth_generator.py, rapids_driven_controls.py) to confirm scores are consistent and no regressions landed.
Run TATB (TA Test Benchmark) on a single architecture and print the four-rubric scorecard — Threat-Relevant, TTP-Accurate, Risk-Defensible, Plan-Actionable. Uses the same scoring logic as the dashboard TATB tab. Pass an architecture name as argument, or omit for the most recent report. Read-only. ~3 seconds (includes MITRE API fetch).
Score all architectures in /report/ across the four TATB rubrics and print a visual corpus map — bar table, heatmap, and pipeline stage attribution. Shows which architectures are Excellent/Solid/Weak/Draft on each dimension and identifies which pipeline stages to fix for the largest corpus-wide gain. No UI needed. ~10-15 seconds (MITRE fetch batched once).
Run Expert Review (MoE critics + ScrumMaster) on any architecture that has an existing analysis. Three modes — sequential (3 core critics), full_moe (5 critics + SM, generates Plan-Actionable data), or single-critic (re-run one critic by name). Streams live critic results with scores, gaps, and final confidence. ~60–120s for full_moe, ~30–60s for sequential.
TATB feedback loop — observe corpus or single-arch TATB scores, diagnose weak signals, propose targeted pipeline stage fixes, gate on human approval, apply changes, re-run TATB to verify improvement, and log the before/after delta to docs/DECISIONS.md. The agent does NOT autonomously commit code; every code change requires explicit human approval.
Runs a fast ~15 second integration sanity check — verifies dependencies, rate limiter, MITRE data load (≥835 techniques), OpenRouter API key, API connection, and one real embedding call. Use at session start or after installing dependencies to confirm everything is working before coding.
| name | check-deprecation |
| description | Checks the codebase for deprecated imports, broken module imports, and common anti-patterns. Use after major refactoring, after adding new modules, or before committing large changes. Takes ~2 minutes. Non-destructive — read-only. |
| allowed-tools | Bash(python3:*) Bash(grep:*) Bash(pytest:*) Bash(source:*) |
Runs 5 checks in sequence. Exit code equals the number of issues found.
bash "$(git rev-parse --show-toplevel)/.claude/skills/check-deprecation/scripts/check-deprecation.sh
from agentic.llm import * (should be from agentic.llm_client import LLMClient)import <module> for all tracked modules; reports failuresDeprecationWarning as errortests/unit/ with pytestlitellm.completion() calls; raw os.getenv("OPENROUTER_API_KEY") outside helper.pyCore: chatbot.modules.ground_truth_generator, chatbot.modules.threat_analyst, chatbot.modules.threat_report, chatbot.modules.exhaustive_mitigation_mapper, chatbot.modules.self_validation, chatbot.modules.residual_risk, chatbot.modules.completeness_validator, chatbot.modules.mitre, chatbot.modules.mitre_embeddings, chatbot.modules.embeddings, chatbot.modules.rate_limiter
Patterns: chatbot.modules.patterns.ai_pattern, chatbot.modules.pattern_registry
SSP: chatbot.modules.ssp_mapper
API: chatbot.api.app, chatbot.api.routes.streaming, chatbot.api.routes.reports
LLM: agentic.llm_client, agentic.helper
# Deprecated
from agentic.llm import generate_response
# Correct
from agentic.llm_client import LLMClient
client = LLMClient()
response = client.generate(prompt="...")