multi-model-review

name	multi-model-review
description	Query AI models in parallel via OpenRouter and synthesize findings. Use when the user says "ask [model]", "get [model]'s opinion", "what does [model] think", "multi-model review", "get a second opinion", "council review", "validate this with [model]", or names any supported model alongside content to review. Requires OPENROUTER_API_KEY set in the environment.
when_to_use	Also invoke proactively for high-stakes work — security-sensitive code, architecture decisions before implementation, production-bound logic, or any case where a single model's blind spots could cause real harm.
allowed-tools	Bash Glob

multi-model-review

Prerequisites

• OPENROUTER_API_KEY set in environment
• httpx installed: pip install httpx

---

Available Models

Benchmark sources: DeepSWE (Datacurve, May 2026), SWE-bench Pro (Scale AI), GPQA Diamond, LMArena ELO, Artificial Analysis Intelligence Index.

Alias(es)	OpenRouter ID	Tier	Strengths
gpt, gpt5, codex	openai/gpt-5.5	Premium	DeepSWE #1 (70%), GPQA 92.8% — best overall coder
opus, claude	anthropic/claude-opus-4.8	Premium	LMArena #1, SWE-bench 88.6%, deep reasoning
sonnet	~anthropic/claude-sonnet-latest	Standard	Fast Anthropic, balanced quality/cost
gemini	google/gemini-3.1-pro-preview	Standard	GPQA #1 (94.3%), 1M ctx — best for reasoning/knowledge
flash	google/gemini-3.5-flash	Fast	Newest Google, fast + cheap
grok	x-ai/grok-4.3	Standard	Cost-efficient coding, real-time X/web data
deepseek	deepseek/deepseek-v4-pro	Budget	SWE-bench 80.6%, strong reasoning, low cost
free	deepseek/deepseek-v4-flash:free	Free	Zero cost
kimi	moonshotai/kimi-k2.6	Budget	Best open-weight (GPQA 87.6%)

Note on Gemini for coding: Gemini 3.1 Pro scores ~1% on DeepSWE (agentic coding benchmark) due to harness generalization issues. Prefer gpt or opus for code review tasks.

---

Step 1 — Parse the request

Given the user's message:

Extract model names (case-insensitive, any order):

• Scan for aliases from the table above
• Map each alias → its OpenRouter ID
• If no model named → ask: "Which models should I consult? e.g. gpt and claude, deepseek and kimi"
• Single model is valid

Extract the task (becomes --system):

• Identify the action phrase: everything after "to [verb]", "for [noun]", "on [topic]"
• Examples:
  • "ask gpt and claude to check for race conditions" → task: "check for race conditions"
  • "get gemini's opinion on this architecture" → task: "review this architecture"
  • "have deepseek and kimi audit for SQL injection" → task: "audit for SQL injection"
  • "ask opus and gpt" (no task) → no --system
• Wrap the extracted task: "You are an expert reviewer. [task]. Be specific, reference exact constructs, rate each finding: Critical / Warning / Suggestion."
• If no task phrase, omit --system entirely

Locate content:

• File path mentioned → --file <path>
• Content in conversation → --prompt "<content>"
• Neither → ask: "Please provide the content to review, or a file path."

---

Step 2 — Find the script

Locate via glob: **/multi-model-review/scripts/or_review.py

---

Step 3 — Run the script

python "<script_path>" \
  --models "<openrouter-id-1>,<openrouter-id-2>" \
  [--system "You are an expert reviewer. <task>. Rate findings: Critical / Warning / Suggestion."] \
  [--file "<path>" | --prompt "<content>"]

Optional flags:

• --max-tokens N — cap response length (default 2000)
• --timeout SEC — request timeout (default 120)

---

Step 4 — Synthesize responses

The script outputs JSON:

{
  "results": {
    "<model-id>": {
      "content": "...",
      "elapsed_s": 2.14,
      "prompt_tokens": 523,
      "completion_tokens": 412,
      "cost_usd": 0.013
    }
  },
  "meta": {
    "models": ["...", "..."],
    "total_elapsed_s": 2.14,
    "total_prompt_tokens": 1046,
    "total_completion_tokens": 824,
    "total_cost_usd": 0.027
  }
}

Apply this protocol:

Situation	Action
Both models agree	High-confidence finding — present as settled
One model unique	"Worth considering (flagged by [model] only)"
Models conflict	Surface both positions, explain tradeoff
Security finding from any model	Always escalate regardless of disagreement
Model returns ERROR:...	Note unavailable, weight remaining models accordingly

Bias awareness:

• Longer response ≠ more reliable. Weight specificity and concrete examples over length.
• Solo findings are unconfirmed, not dismissed.

Output format

## Multi-Model Review — [Task]
Models: [model-a] · [model-b]  |  Cost: $X.XX  |  Time: X.Xs

### Agreed findings  ✓ high confidence
- [finding]

### Solo findings  ? unconfirmed
- [model-a]: [finding]
- [model-b]: [finding]

### Conflicts  ↔
- **[topic]:** [model-a] says X / [model-b] says Y — [tradeoff note]

### Verdict
[1–2 sentence synthesis of the most actionable takeaways]

Omit any section with no entries. Omit cost if total_cost_usd is null.

---

Standalone usage

# Two models, task-driven system prompt:
python scripts/or_review.py \
  --models "openai/gpt-5.5,anthropic/claude-opus-4.8" \
  --file myfile.py \
  --system "Check for race conditions and resource leaks."

# Piped input, no system prompt (general review):
cat myfile.py | python scripts/or_review.py \
  --models "deepseek/deepseek-v4-pro,moonshotai/kimi-k2.6"

# Free review:
python scripts/or_review.py \
  --models "deepseek/deepseek-v4-flash:free" \
  --file myfile.py

# Custom tokens cap:
python scripts/or_review.py \
  --models "openai/gpt-5.5,x-ai/grok-4.3" \
  --file myfile.py --max-tokens 1500