The registry has 18 reviewers. Three default-roster ones are dual-route
(a direct provider API plus OpenRouter) and are reordered by the
routing preference (the custom-only hermes is
dual-route too); the rest are single-route API reviewers or paid-CLI reviewers. Model versions are the headline of this
release: GLM‑5.3, MiniMax M3, and the new default
DeepSeek V4 Pro 0813. OpenRouter-routed model IDs, context
lengths and prices were verified against the live catalog on 2026‑09‑22, then
re-pinned for cost/ctx drift on 2026‑09‑29 via scripts/refresh.py
(no model-id changes); direct-provider
and CLI-provider slugs are not covered, because that API cannot see them — which is why
opencode-glm is still on glm-5.2.
| Reviewer | Route(s) | Notes |
|---|---|---|
glm dual-route | z.ai direct + OpenRouter (z-ai/glm-5.3) | Strong security + logic. |
minimax dual-route | MiniMax direct + OpenRouter (minimax/minimax-m3) | High precision. |
kimi | OpenRouter (moonshotai/kimi-k3, fallback kimi-k2.7-code) | Long-context, agentic. |
mimo | OpenRouter (xiaomi/mimo-v2.6-pro) | 1M context window. |
qwen | OpenRouter (qwen/qwen3.8-max-0902) | 1M ctx, conservative / high precision. |
grok | OpenRouter (x-ai/grok-4.7) | Current xAI flagship; 500K ctx, reasoning. |
grok-longctx | OpenRouter (x-ai/grok-4.20) | The only 2M-ctx reviewer — kept because every newer Grok has a smaller window. Not in any shipped profile; name it explicitly. |
deepseek dual-route default DeepSeek | DeepSeek direct + OpenRouter (deepseek/deepseek-v4-pro-0813) | 1.6T MoE, 49B active, ~1M ctx; reasoning + security. |
deepseek-v3.2 | OpenRouter | custom-only superseded by v4-pro. |
hermes dual-route | Nous direct + OpenRouter fallback | custom-only |
opencode-minimax | OpenCode CLI → minimax-coding-plan/MiniMax-M3 | custom-only MiniMax M3 billed to the OpenCode sub instead of the metered API. |
opencode-glm | OpenCode CLI → ollama-cloud/glm-5.2 | custom-only Still on 5.2 — the Ollama Cloud catalog is not visible from the OpenRouter API, so glm-5.3 there is unverified and this route has no fallback. |
gemini | Gemini CLI (paid sub) | disabled Windows .cmd tree-kill re-test pending; use gemini-or meanwhile. |
gemini-or | OpenRouter (google/gemini-3.8-flash) | Flash tier. Not benchmarked — the “~2s/call, best value” figure measured 2.5 Flash. |
codex | Codex CLI (paid sub, GPT-5.x) | Thorough, slow; OpenRouter fallback. |
claude | Claude CLI (paid sub) | Auto-added to profile rosters when host ≠ claude. |
opencode | OpenCode CLI (paid sub) | Top benchmark performer, slow cold start. |
copilot | GitHub Copilot CLI | disabled / eliminated agent harness returns prose, not our JSON schema. |
Keys are glm, kimi, qwen… — the model version lives in the
slug and display name only, so bumping a model never invalidates a saved profile, a --custom roster,
or a history.db row. The old version-named keys still work via the aliases: map in
config.yaml: --custom "glm-5.2,kimi-k2.6" resolves to glm,kimi, and naming
both an alias and its canonical key dispatches the reviewer once, not twice.
One deliberate exception: grok-4.20 aliases to grok-longctx, not
grok. The old key named the 2M-context model specifically, and grok is now Grok 4.7
at 500K — mapping it across would have silently shrunk an existing roster's context window.
The dual-route reviewers respect route_preference. CLI reviewers
(Codex / Claude / OpenCode / Gemini) keep their subscription primary with OpenRouter as a true fallback,
no matter the routing preference. See Routing preference.
Ready-made groupings of these reviewers live in Profiles.