메뉴
Use when asked to audit code for bugs without fixing them. Scans the repo with 3 independent expert judges (weighted scoring + expert veto), proves root causes, and generates a report. NEVER modifies code — read-only audit.
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
Use when delegating coding, review, analysis, testing, or repo-maintenance work to OpenAI Codex CLI from Claude Code or another agent, especially when the user asks to run codex, get a Codex review, use codex exec, save Codex output, stream JSON events, or run autonomous codebase work in a local repository.
Use when the task is to FIND ideas, solutions, root causes, workarounds, prior art, or benchmarks ONLINE — researching a problem/idea/solution on the web — NOT a simple single-doc API lookup. Searches across MANY languages in parallel: spawns one research subagent per language (English, Chinese Simplified, Chinese Traditional, French, Russian, Hebrew, Spanish), each searching native technical terms and native forums, then synthesizes. Triggers: "research this", "search online for", "find a solution/idea/approach", "how do people solve", "is there prior art", "find workarounds", "deep multilingual web research".
Use when the user says "bg-it" or "spawn this in background" or "run this as bg agent" — spawns a background claude session for ANY prompt type or archetype, including interactive ones. Claude Code only.
Micromanager (`/mm`) - Living PM skill that maintains continuous HANDOFF.md task state, repo-level PM dashboards, scheduling fields for target dates and risk, implementation prompts, file inbox updates from agents and CEO recommendations, durable insights, and task close-out. Auto-detects fresh vs continuing vs update vs dashboard mode; never writes production code itself; preserves the Section 0 continuation contract.
Send text messages and files to a configured Telegram chat (group, channel, or DM with bot). Supports forum/topic groups. Reads bot token + chat ID from ~/.claude/.env so the skill itself contains no secrets.
Hybrid LLM-engineering research — synthesizes a podcast-grounded answer from the Langtalks corpus with current web search results. Triggers on /langtalk, "langtalks", "what does langtalks say about", and LLM-system architecture/trade-off questions where named-practitioner provenance matters.
| name | bh-audit |
| description | Use when asked to audit code for bugs without fixing them. Scans the repo with 3 independent expert judges (weighted scoring + expert veto), proves root causes, and generates a report. NEVER modifies code — read-only audit. |
Scan any repo for bugs, verify with 3 independent parallel expert judges (weighted scoring + expert veto), deduplicate, prove root cause, and generate a report.
This skill is STRICTLY READ-ONLY. It NEVER fixes, edits, or commits anything.
/bh/bh (bug-hunter) | /bh-audit (this skill) | |
|---|---|---|
| Finds bugs | Yes | Yes |
| Fixes bugs | Yes (TDD + auto-fix) | NEVER |
| Creates commits | Yes | NEVER |
| Code review gate | Yes | No (nothing to review) |
| DoD gate | Yes | No (nothing to verify) |
| Output | Fixed code + report | Report only |
Use /bh when you want bugs found AND fixed.
Use /bh-audit when you want bugs found and reported — a human decides what to do.
This skill MUST NOT:
This skill MUST ONLY:
/babysitter:call Run the generic/bh-audit process.
Process file: processes/bh-audit.js (or ~/.a5c/processes/bh-audit.js)
Inputs: projectDir=<CWD>, scanTarget=both
| Input | Default | Description |
|---|---|---|
projectDir | CWD | Path to the repo root |
testCmd | auto-detect | Override test command |
categories | all 6 | Bug categories to scan |
scanTarget | both | current / dev / both |
severityFilter | low | Minimum severity to include |
maxFindings | 30 | Max findings per scan |
logic, security, memory-lifecycle, error-handling, performance, thread-safety
sql-logic, data-integrity, resource-config, pipeline-logic, test-gaps, conventions, contract-drift
DETECT project
-> [SETUP DEV WORKTREE if scanning dev]
-> RUN TESTS (detect failures)
-> SCAN (N categories IN PARALLEL)
-> DEDUP
-> VERIFY (3 independent expert judges IN PARALLEL)
|-- Judge 1: Software Engineer (separate agent)
|-- Judge 2: Data/Infra Engineer (separate agent)
|-- Judge 3: Security Specialist (separate agent)
-> MERGE SCORES (weighted average + expert veto)
-> PROVE ROOT CAUSE
-> REPORT (markdown + Hebrew summary)
-> [CLEANUP WORKTREE]
Each judge runs as a separate parallel agent with its own context — true independence.
| Judge | Expert Categories |
|---|---|
| Software Engineer | logic, error-handling, test-gaps, conventions, contract-drift |
| Data/Infrastructure Engineer | sql-logic, data-integrity, resource-config, pipeline-logic |
| Security & Systems Specialist | security, memory-lifecycle, performance, thread-safety |
| Condition | Result |
|---|---|
| Domain expert >=80 | Verified (expert veto) |
| Weighted average >=50 | Verified |
| Weighted average 30-49 | Needs Attention |
| Weighted average <30 | Dismissed |
Saved to <projectDir>/BH-AUDIT-REPORT.md:
Each finding tagged [current] or [dev] by source branch.