FAQ · Argus

Short answers to the things people ask first. For depth, follow the links into the rest of the KB.

What models does Argus use?

A registry of 15 frontier reviewers — GLM‑5.2, MiniMax M3, DeepSeek V4 Pro, Kimi K2.6, MiMo‑V2‑Pro, Qwen 3.6‑Plus, Grok 4.20, plus the Codex, Claude, and OpenCode CLIs. See the full reviewer roster.

Do I need all the API keys?

No. The public default route is OpenRouter, so a single OPENROUTER_API_KEY covers most reviewers. Direct provider keys (ZAI_API_KEY, MINIMAX_API_KEY, DEEPSEEK_API_KEY) are optional — add them only if you want direct-API routing. See the environment variables table.

OpenRouter vs direct API — what’s the difference?

OpenRouter is one key in front of many models — simplest to set up. Direct APIs use each provider’s own endpoint, which can be cheaper and lets you use subscriptions you already pay for. Three reviewers are dual-route and can use either; the route_preference knob picks which they try first. See Routing preference.

How do I switch to direct API?

Any of three ways (precedence: CLI flag > env var > config):

python scripts/verify.py --all --prefer-direct   # per-run flag
export ARGUS_ROUTE_PREF=direct                    # per-shell env var
# or set  defaults.route_preference: direct  in config.yaml

Pair it with the direct profile when your OpenRouter balance is depleted.

Is my code sent anywhere? What about privacy?

Your diff is sent to whichever reviewers you select, via their provider APIs or CLIs — that is inherent to LLM review. What Argus does not do: it never writes your API keys to disk. Keys live in your environment and are forwarded to subprocesses only at dispatch time. Choose a privacy-appropriate roster for sensitive code.

Does Argus work on Windows?

Yes — it is developed on Windows 11 (with Git Bash). Use py -3.12 to invoke Python and set PYTHONIOENCODING=utf-8 for non-ASCII output. One known caveat: the Gemini CLI reviewer is currently disabled pending a Windows .cmd tree-kill re-test — use gemini-or (the OpenRouter route) meanwhile.

How much does a review cost?

Usually cents. A typical review stays well under the $0.50 warning gate; the reference benchmark of 12 calls per reviewer across the suite totaled about $0.42. Cost gates warn at $0.50 and hard-block at $2.00 for reviews ($10 / $30 for benchmarks). Paid-CLI reviewers count as $0 in the estimate. See cost gates.

How do I add a reviewer?

Add an entry under reviewers in config.yaml with its route(s) and model ID(s); for a new provider, add a client in install_aichat.py and forward its $<PROVIDER>_API_KEY. It’s config-only in the common case — see Contributing.