llm-review

name	llm-review
description	Multi-LLM consultation wrapper for external AI reviews. Supports Gemini and GPT with hardcoded model names to prevent agent modifications.
user-invocable	true

LLM Review - Multi-AI Consultation Skill

Provides a reliable interface for calling external LLMs (Gemini, GPT) with fixed model configurations.

Purpose

1. Prevent model name changes by agents (hardcoded in Python)
2. Standardize API protocols (different APIs handled internally)
3. Provide consistent JSON output for easy parsing

---

Usage

python3 ~/.claude/skills/llm-review/llm_client.py \
  --provider [gemini|openai] \
  --phase [consultation|review|regulation] \
  --prompt "Your prompt" \
  --output /tmp/result.json

Parameter	Required	Description
--provider	Yes	gemini or openai
--phase	Yes	consultation (fast), review (thorough), regulation (thinking + web search)
--prompt	Yes*	Prompt text
--prompt-file	Yes*	Read prompt from file
--include-files	No	Append source files to prompt (space-separated, supports file:start-end)
--max-file-lines	No	Max lines per included file (0 = unlimited, default: 0)
--output	No	Output file (default: stdout)

--include-files Examples

# Attach full files
--include-files output.py utils.py complex.py

# Attach specific line ranges
--include-files "complex.py:100-300" "utils.py:1-200"

# Mix full files and ranges
--include-files engine.py "complex.py:100-500" utils.py

IMPORTANT: When consulting about code, ALWAYS use --include-files to attach the relevant source files. Without it, the external LLM only sees file names, not the actual code, and can only give generic advice.

---

Models (HARDCODED - DO NOT CHANGE)

Phase	Gemini	OpenAI	Reasoning
consultation	gemini-3-flash-preview	gpt-5.4 (Chat Completions)	-
review	gemini-3.1-pro-preview	gpt-5.4 (Responses API)	medium
regulation	gemini-3.1-pro-preview	gpt-5.4 (Responses API, web_search)	high

---

Output Format

{
  "success": true,
  "provider": "gemini",
  "model": "gemini-3-flash-preview",
  "response": "LLM response text...",
  "tokens": {"prompt": 1234, "completion": 567, "total": 1801},
  "elapsed_seconds": 3.2
}

Error: "success": false, "error": "...", "error_type": "..."

---

Environment

Variable	Description
GEMINI_API_KEY	Google AI Studio API key
OPENAI_API_KEY	OpenAI API key

Auto-loads from .env in current working directory, then walks up to workspace root.

Check keys: python3 ~/.claude/skills/llm-review/llm_client.py --check-keys

---

Error Handling

Error	Behavior
API key missing	Returns error JSON
Timeout / 429 / 502 / 503	Retries 3x with backoff (2s, 5s, 10s)
400 / 401 / 404	Immediate failure, no retry
review/regulation without --include-files	Prints [WARN] to stderr