@oh-my-pi/pi-catalog 18.14.2 → 18.16.0

package/src/compat/rules/README.md

@@ -258,6 +258,8 @@ Opted in: Muse Code `muse-spark-1.3`, `muse-spark-1.3-contributor`, and the four
 
 The routed-subscription registry axes (`upstream-rotation`, `region-upstreams-global|us|eu`, `region-limits-eu`, `credit-rates`, `list-price-from`, `routing-family`, `policy-aliases`, `entitlement`, `default-reasoning-off`) describe a gateway whose proxy fans one model out to several upstreams. They are read through `src/compat/factory-droid.ts` by Factory Droid discovery and its request provider, not materialized by `buildModel`. A provider-wide `region-upstreams-*` rule is the upstream serving table and a model rule replaces it for that region. `list-price-from "<provider>" ["<id>"]` shows a bundled row's list price beside the subscription's own billing; unlike seed values it is resolved at runtime and degrades to the seed's zero cost when the row is gone. See [`providers/factory-droid.kdl`](providers/factory-droid.kdl), whose wire and billing pool per model are `api-routes` and `quota-tiers` rules in `runtime/behavior.kdl`.
 
+`credit-rates` accepts the same optional `effective-rates` block of named dated cards as `cost-patch`, parsed by the same code: each card has a distinct ISO UTC `effective-from` and any of `input`/`output`/`cache-read`, and the latest card at or before the moment the model browser renders overwrites the legs it names. A vendor credit promotion is two cards, the discounted rate from its start and the base rate again from its end, so a card left behind after the end changes nothing. See [`providers/factory-droid.kdl`](providers/factory-droid.kdl).
+
 ### Time-based pricing
 
 The catalog object axis `time-based-cost` materializes into `ModelCost.timeBased`. It uses **named child objects**, not KDL arrays or repeated anonymous windows. For example, inside a matching provider/model scope:

package/src/compat/rules/auth/_order.kdl

@@ -10,4 +10,4 @@ login-order "openai-codex" "anthropic" "zai" "zai-coding-plan" "kimi-code" "open
 	"qianfan" "venice" "siliconflow" "siliconflow-cn" "synthetic" "nanogpt" "wafer-serverless" "coreweave" \
 	"vercel-ai-gateway" "cloudflare-ai-gateway" "cloudflare-workers-ai" "snowflake" "litellm" "kilo" "zenmux" "opencode-zen" "opencode-go" \
 	"yolo-auto" "tavily" "kagi" "exa" "parallel" "typesafe" "ollama" "ollama-cloud" "lm-studio" "llama.cpp" "vllm" \
-	"gmi-cloud" "stencil" "singularityapi-dev" "singularityapi-tech" "coralbricks"
+	"gmi-cloud" "stencil" "singularityapi-dev" "singularityapi-tech" "coralbricks" "modular" "mistral" "scaleway"

package/src/compat/rules/auth/mistral.kdl

@@ -1,3 +1,7 @@
+// Browser sign-in mirrors the Mistral Vibe CLI and mints a regular API key billed
+// against the plan's Vibe Code quota; the hook returns it as a plain API key.
 auth "mistral" {
 	name "Mistral"
+	env "MISTRAL_API_KEY"
+	login "custom" hook="mistral-vibe-sign-in"
 }

package/src/compat/rules/auth/modular.kdl

@@ -0,0 +1,9 @@
+auth "modular" {
+	name "Modular"
+	login "api-key" {
+		auth-url "https://console.modular.com/api_keys"
+		instructions "Create an API key in the Modular Console"
+		prompt "Paste your Modular API key" placeholder="..."
+		validate "models-endpoint" label="Modular" url="https://api.modular.com/v1/models"
+	}
+}

package/src/compat/rules/auth/scaleway.kdl

@@ -0,0 +1,17 @@
+auth "scaleway" {
+	name "Scaleway"
+	login "api-key" {
+		auth-url "https://console.scaleway.com/iam/api-keys"
+		instructions "Create an API key on the Scaleway console and paste its Secret Key (SCW_SECRET_KEY)"
+		// Scaleway secret keys are UUID-shaped opaque strings with no `sk-`
+		// prefix, so the placeholder must not teach a credential shape the
+		// console never issues.
+		prompt "Paste your Scaleway secret key" placeholder="..."
+		normalize "strip-bearer"
+		// `/v1/models` is key-protected: it answers 401 without a bearer key
+		// and 403 to a rejected one (verified 2026-10-21), so login validates
+		// the pasted key there. The probe bills nothing, unlike a
+		// chat-completions round-trip against a prepaid balance.
+		validate "models-endpoint" label="Scaleway" url="https://api.scaleway.ai/v1/models"
+	}
+}

package/src/compat/rules/providers/factory-droid.kdl

@@ -141,24 +141,54 @@ provider "factory-droid" {
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=922000 max-tokens=128000
 		}
+		model "gpt-6-astra-ultrafast" name="GPT-6 Astra Ultra Fast" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=922000 max-tokens=128000
+		}
 		model "gpt-6.1-sol" name="GPT-6.1 Sol" {
 			reasoning #true
 			input "text" "image"
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=922000 max-tokens=128000
 		}
+		model "gpt-6.1-sol-ultrafast" name="GPT-6.1 Sol Ultra Fast" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=922000 max-tokens=128000
+		}
+		model "gpt-6.1-terra" name="GPT-6.1 Terra" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=922000 max-tokens=128000
+		}
 		model "gpt-6-sol" name="GPT-6 Sol" {
 			reasoning #true
 			input "text" "image"
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=922000 max-tokens=128000
 		}
+		model "gpt-6-sol-fast" name="GPT-6 Sol Fast Mode" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=922000 max-tokens=128000
+		}
 		model "gpt-6-luna" name="GPT-6 Luna" {
 			reasoning #true
 			input "text" "image"
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=922000 max-tokens=128000
 		}
+		model "gpt-6-luna-fast" name="GPT-6 Luna Fast Mode" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=922000 max-tokens=128000
+		}
 		model "gpt-5.6-sol" name="GPT-5.6 Sol" {
 			reasoning #true
 			input "text" "image"
@@ -339,12 +369,36 @@ provider "factory-droid" {
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=131072 max-tokens=131072
 		}
+		model "nemotron-3.5-super" name="Nemotron 3.5 Super" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=934464 max-tokens=65536
+		}
 		model "nemotron-3-ultra" name="Nemotron 3 Ultra" {
 			reasoning #true
 			input "text"
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=136464 max-tokens=65536
 		}
+		model "muse-spark-open" name="Muse Spark Open" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=908928 max-tokens=131072
+		}
+		model "muse-spark-1.2" name="Muse Spark 1.2" {
+			reasoning #true
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=908928 max-tokens=131072
+		}
+		model "beam-1.0" name="Beam 1.0" {
+			reasoning #true
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=1015808 max-tokens=32768
+		}
 		model "deepseek-v4.1-flash-1017" name="DeepSeek V4.1 Flash 1017" {
 			reasoning #true
 			input "text" "image"
@@ -446,7 +500,10 @@ provider "factory-droid" {
 	// Per-model registry entries: live-routing family, upstream rotation (first
 	// entry is the default `x-api-provider`), region overrides, Standard Credits
 	// rates, list-price source, org-policy aliases, the effort ladder (an off/none
-	// rung means `thinking-requires-effort #false`), and account gates.
+	// rung means `thinking-requires-effort #false`), and account gates. A credit
+	// promotion is two dated `effective-rates` cards copied from the droid
+	// release's model registry: the discounted `input` from its start and the
+	// base again from its end.
 	models "claude-fable-5.1" {
 		routing-family "anthropic"
 		upstream-rotation "anthropic" "vertex_anthropic" "bedrock_anthropic"
@@ -484,7 +541,7 @@ provider "factory-droid" {
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
 		thinking-requires-effort #true
 		thinking-default-level "medium"
-		entitlement { feature-flag "claude_opus_5_5_fast"; base-variant "claude-opus-5-5" }
+		entitlement { base-variant "claude-opus-5-5" }
 	}
 	models "claude-opus-5" {
 		routing-family "anthropic"
@@ -621,6 +678,20 @@ provider "factory-droid" {
 		thinking-requires-effort #true
 		thinking-default-level "medium"
 	}
+	models "gpt-6-astra-ultrafast" {
+		routing-family "openai"
+		upstream-rotation "openai"
+		region-upstreams-eu
+		credit-rates { input 24; output 5 }
+		thinking-efforts "low" "medium" "high" "xhigh" "max"
+		thinking-requires-effort #true
+		thinking-default-level "medium"
+		entitlement {
+			feature-flag "gpt_6_astra_ultrafast"
+			base-variant "gpt-6-astra"
+			requires-explicit-opt-in #true
+		}
+	}
 	models "gpt-6.1-sol" {
 		routing-family "openai"
 		upstream-rotation "openai" "bedrock_openai" "databricks"
@@ -629,7 +700,28 @@ provider "factory-droid" {
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
 		thinking-requires-effort #true
 		thinking-default-level "medium"
-		entitlement { feature-flag "gpt_6_1_sol" }
+	}
+	models "gpt-6.1-sol-ultrafast" {
+		routing-family "openai"
+		upstream-rotation "openai"
+		credit-rates { input 4.8; output 5; cache-read 0.05 }
+		thinking-efforts "low" "medium" "high" "xhigh" "max"
+		thinking-requires-effort #true
+		thinking-default-level "medium"
+		entitlement {
+			base-variant "gpt-6.1-sol"
+			requires-explicit-opt-in #true
+		}
+	}
+	models "gpt-6.1-terra" {
+		routing-family "openai"
+		upstream-rotation "openai"
+		credit-rates { input 0.4; output 5; cache-read 0.05 }
+		list-price-from "openai"
+		thinking-efforts "low" "medium" "high" "xhigh" "max"
+		thinking-requires-effort #false
+		thinking-default-level "medium"
+		entitlement { feature-flag "gpt_6_1_terra" }
 	}
 	models "gpt-6-sol" {
 		routing-family "openai"
@@ -641,6 +733,15 @@ provider "factory-droid" {
 		thinking-requires-effort #false
 		thinking-default-level "medium"
 	}
+	models "gpt-6-sol-fast" {
+		routing-family "openai"
+		upstream-rotation "openai"
+		credit-rates { input 1.6; output 5 }
+		thinking-efforts "low" "medium" "high" "xhigh" "max"
+		thinking-requires-effort #false
+		thinking-default-level "medium"
+		entitlement { base-variant "gpt-6-sol" }
+	}
 	models "gpt-6-luna" {
 		routing-family "openai"
 		upstream-rotation "openai" "bedrock_openai" "azure_openai"
@@ -651,11 +752,33 @@ provider "factory-droid" {
 		thinking-requires-effort #false
 		thinking-default-level "medium"
 	}
+	models "gpt-6-luna-fast" {
+		routing-family "openai"
+		upstream-rotation "openai"
+		credit-rates { input 0.08; output 5 }
+		thinking-efforts "low" "medium" "high" "xhigh" "max"
+		thinking-requires-effort #false
+		thinking-default-level "medium"
+		entitlement { base-variant "gpt-6-luna" }
+	}
 	models "gpt-5.6-sol" {
 		routing-family "openai"
 		upstream-rotation "openai" "bedrock_openai" "azure_openai" "snowflake" "databricks"
 		region-upstreams-eu "openai"
-		credit-rates { input 2; output 5 }
+		credit-rates {
+			input 2
+			output 5
+			effective-rates {
+				promotion {
+					effective-from "2026-08-22T00:00:00Z"
+					input 1.6
+				}
+				standard {
+					effective-from "2026-11-22T00:00:00Z"
+					input 2
+				}
+			}
+		}
 		list-price-from "openai"
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
 		thinking-requires-effort #false
@@ -664,7 +787,20 @@ provider "factory-droid" {
 	models "gpt-5.6-sol-fast" {
 		routing-family "openai"
 		upstream-rotation "openai" "azure_openai"
-		credit-rates { input 4; output 5 }
+		credit-rates {
+			input 4
+			output 5
+			effective-rates {
+				promotion {
+					effective-from "2026-08-22T00:00:00Z"
+					input 3.2
+				}
+				standard {
+					effective-from "2026-11-22T00:00:00Z"
+					input 4
+				}
+			}
+		}
 		policy-aliases "gpt-5.6-fast"
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
 		thinking-requires-effort #false
@@ -693,7 +829,7 @@ provider "factory-droid" {
 	}
 	models "gpt-5.5" {
 		routing-family "openai"
-		upstream-rotation "openai" "bedrock_openai" "azure_openai" "snowflake" "databricks"
+		upstream-rotation "openai" "bedrock_openai" "azure_openai" "databricks"
 		region-upstreams-eu "openai"
 		credit-rates { input 2; output 6 }
 		list-price-from "openai"
@@ -816,7 +952,20 @@ provider "factory-droid" {
 		routing-family "google"
 		upstream-rotation "google"
 		region-upstreams-eu "google"
-		credit-rates { input 0.6; output 5 }
+		credit-rates {
+			input 0.6
+			output 5
+			effective-rates {
+				promotion {
+					effective-from "2026-08-17T21:10:19Z"
+					input 0.3
+				}
+				standard {
+					effective-from "2027-01-01T08:00:00Z"
+					input 0.6
+				}
+			}
+		}
 		list-price-from "google"
 		thinking-efforts "low" "medium" "high"
 		thinking-requires-effort #true
@@ -826,7 +975,20 @@ provider "factory-droid" {
 		routing-family "google"
 		upstream-rotation "google"
 		region-upstreams-eu "google"
-		credit-rates { input 0.6; output 5 }
+		credit-rates {
+			input 0.6
+			output 5
+			effective-rates {
+				promotion {
+					effective-from "2026-08-17T21:10:19Z"
+					input 0.3
+				}
+				standard {
+					effective-from "2027-01-01T08:00:00Z"
+					input 0.6
+				}
+			}
+		}
 		list-price-from "google"
 		thinking-efforts "low" "medium" "high"
 		thinking-requires-effort #true
@@ -972,6 +1134,15 @@ provider "factory-droid" {
 		thinking-requires-effort #true
 		thinking-default-level "xhigh"
 	}
+	models "nemotron-3.5-super" {
+		routing-family "factory"
+		upstream-rotation "baseten"
+		credit-rates { input 0.08; output 2.5; cache-read 0.2 }
+		thinking-efforts "high"
+		thinking-requires-effort #false
+		thinking-default-level "high"
+		entitlement { feature-flag "nemotron_3_5_super" }
+	}
 	models "nemotron-3-ultra" {
 		routing-family "factory"
 		upstream-rotation "baseten" "fireworks"
@@ -981,6 +1152,34 @@ provider "factory-droid" {
 		thinking-requires-effort #false
 		thinking-default-level "high"
 	}
+	models "muse-spark-open" {
+		routing-family "factory"
+		upstream-rotation "baseten"
+		credit-rates { input 0.5; output 3.4; cache-read 0.12 }
+		thinking-efforts "low" "medium" "high" "xhigh" "max"
+		thinking-requires-effort #true
+		thinking-default-level "high"
+		entitlement { feature-flag "muse_spark_open" }
+	}
+	// Factory serves no `minimal` rung, so this ladder outranks the Meta class ladder.
+	models "muse-spark-1.2" priority=1 {
+		routing-family "factory"
+		upstream-rotation "fireworks"
+		credit-rates { input 0.5; output 3.4; cache-read 0.12 }
+		list-price-from "meta"
+		thinking-efforts "low" "medium" "high" "xhigh"
+		thinking-requires-effort #true
+		thinking-default-level "high"
+	}
+	models "beam-1.0" {
+		routing-family "factory"
+		upstream-rotation "baseten"
+		credit-rates { input 0.52; output 3.3077 }
+		thinking-efforts "low" "medium" "high" "xhigh" "max"
+		thinking-requires-effort #true
+		thinking-default-level "high"
+		entitlement { feature-flag "beam_1_0" }
+	}
 	models "deepseek-v4.1-flash-1017" {
 		routing-family "factory"
 		upstream-rotation "fireworks" "baseten" "databricks"
@@ -1029,7 +1228,6 @@ provider "factory-droid" {
 		thinking-efforts "high"
 		thinking-requires-effort #true
 		thinking-default-level "high"
-		entitlement { feature-flag "minimax_m3" }
 	}
 	models "grok-4.7" {
 		routing-family "xai"
@@ -1109,8 +1307,8 @@ provider "factory-droid" {
 		// Native declares every Responses tool `strict: false`; the field must reach the wire.
 		supports-strict-mode #true
 	}
-	models "glm-*" "kimi-*" "deepseek-*" "nemotron-*" "inkling" "mistral-medium-3.5" "mistral-large-4" "qwen3.8-max" \
-		"minimax-m3" {
+	models "glm-*" "kimi-*" "deepseek-*" "nemotron-*" "muse-spark-open" "muse-spark-1.2" "beam-1.0" "inkling" \
+		"mistral-medium-3.5" "mistral-large-4" "qwen3.8-max" "minimax-m3" {
 		clamp-output-to-model-max #true
 		completions-reasoning-mode "none"
 		completions-reasoning-history "omit"
@@ -1136,6 +1334,10 @@ provider "factory-droid" {
 			completions-reasoning-mode "opt-in"
 		}
 	}
+	// Beam tolerates tool-call turns that arrive without reasoning.
+	models "beam-1.0" {
+		completions-allow-missing-thinking #true
+	}
 	models "deepseek-*" {
 		on-upstream "baseten" priority=10 {
 			completions-reasoning-mode "forced-on"
@@ -1241,7 +1443,7 @@ provider "factory-droid" {
 		fast-mode #true
 	}
 	// Refusals retry server-side only on the first-party Anthropic upstream.
-	models "claude-fable-5" "claude-fable-5.1" {
+	models "claude-fable-5" "claude-fable-5.1" "claude-opus-5-5" {
 		on-upstream "anthropic" {
 			server-side-fallback-models "claude-opus-5"
 		}
@@ -1283,7 +1485,8 @@ provider "factory-droid" {
 	}
 	// GPT models whose ladder has an Off/None rung send the literal `none`
 	// tier on every upstream.
-	models "gpt-5.2" "gpt-5.6-sol" "gpt-5.6-sol-fast" "gpt-5.6-terra" "gpt-5.6-luna" "gpt-6-sol" "gpt-6-luna" {
+	models "gpt-5.2" "gpt-5.6-sol" "gpt-5.6-sol-fast" "gpt-5.6-terra" "gpt-5.6-luna" "gpt-6.1-terra" "gpt-6-sol" \
+		"gpt-6-sol-fast" "gpt-6-luna" "gpt-6-luna-fast" {
 		reasoning-disable-mode "none-effort"
 	}
 	models "gpt-*" {
@@ -1306,6 +1509,9 @@ provider "factory-droid" {
 	models "gpt-*-fast" {
 		responses-service-tier "priority"
 	}
+	models "gpt-*-ultrafast" {
+		responses-service-tier "ultrafast"
+	}
 	models "grok-*" {
 		supports-reasoning-summary #false
 	}

package/src/compat/rules/providers/modular.kdl

@@ -0,0 +1,35 @@
+// Provider-wire compat for "modular". Modular Cloud serves open models on an
+// OpenAI-compatible API. The provider is catalog-only: models, prices, and
+// limits come from the models.dev `modular` entry. Reasoning controls differ
+// per model and were measured against the API.
+
+provider "modular" {
+	default-model "google/gemma-4-31b-it"
+	env "MODULAR_API_KEY"
+
+	thinking-mode "effort"
+	// The API takes OpenRouter's `reasoning` object for every model, so a model
+	// models.dev adds later works without its own block: on sends
+	// `reasoning.effort`, off sends `reasoning.enabled = false`.
+	thinking-format "openrouter"
+
+	// MiniMax M3.1 Flash takes a nested `reasoning.effort` and always thinks:
+	// `reasoning.enabled = false` still returns reasoning tokens.
+	models "MiniMaxAI/MiniMax-M3.1-Flash" {
+		thinking-requires-effort #true
+		thinking-efforts "low" "medium" "high" "xhigh" "max"
+	}
+	// MiniMax M3 and Gemma 4 31B switch reasoning with `reasoning.enabled` and
+	// ignore any effort, so the ladder is one on-rung beside off: off sends
+	// `enabled: false`, and on replaces the effort with `enabled: true`.
+	models "minimax/minimax-m3" "google/gemma-4-31b-it" {
+		thinking-efforts "high"
+		when-thinking {
+			extra-body {
+				reasoning {
+					enabled #true
+				}
+			}
+		}
+	}
+}

package/src/compat/rules/providers/sambanova.kdl

@@ -28,6 +28,8 @@ provider "sambanova" {
 	// (https://docs.sambanova.ai/docs/api-reference/chat-completions/create-chat-based-completion).
 	thinking-mode "effort"
 	thinking-format "openai"
+	// Replay the streamed `reasoning` field so turn-N prompts hit the gateway's prefix cache.
+	replay-reasoning-content #true
 	class "gpt-oss" {
 		thinking-efforts "low" "medium" "high"
 		thinking-upgrade-neutral #true

package/src/compat/rules/providers/scaleway.kdl

@@ -0,0 +1,179 @@
+// Provider-wire compat for "scaleway": Scaleway Generative APIs, an
+// OpenAI-compatible gateway at https://api.scaleway.ai/v1 serving serverless
+// open-weight chat models. Verified live 2026-10-21: /v1/models returns bare
+// OpenAI rows with no capability metadata, reasoning traces stream back on
+// the message `reasoning` field, and reasoning_effort accepts `max`, `xhigh`,
+// and `none` (the request-schema enum is stale; `none` answers with
+// reasoning_tokens 0), so the ladders below follow the model cards, not the
+// schema.
+// Docs: https://www.scaleway.com/en/docs/generative-apis/reference-content/supported-models/
+// Pricing (EUR/Mtok, pre-tax): https://www.scaleway.com/en/pricing/model-as-a-service/
+
+provider "scaleway" {
+	default-model "glm-5.3"
+	env "SCW_SECRET_KEY" "SCW_API_KEY"
+	// The key-protected `/v1/models` roster is the serverless catalog: a
+	// successful fetch prunes bundled rows Scaleway has retired (glm-5.2 is
+	// routed to glm-5.3 after 2026-11-16). Same-id rows on other hosts carry
+	// other deployments' limits and modalities (zai serves glm-5.3 at
+	// 1M/131k), so the generator must never backfill from them.
+	dynamic-models-authoritative #true
+	skip-cross-provider-reference-fills #true
+	discovery label="Scaleway"
+
+	// Documented serverless catalog (supported-models + pricing pages,
+	// 2026-11-02). `/v1/models` carries ids only, so these rows are the
+	// bundle in every regeneration and outrank the bare discovered rows; at
+	// runtime they are the reference discovery merges names, modalities,
+	// prices, and limits from. Prices are EUR list rates converted at the
+	// 2026-10-21 ECB reference rate (1 EUR = 1.1234 USD); Scaleway publishes
+	// no cache-write rate and cache-read only where listed. Limits read the
+	// documented "k" as 1,000 tokens, never overstating a serverless cap.
+	seed api="openai-completions" base-url="https://api.scaleway.ai/v1" bundle="always" precedence="seed" {
+		model "glm-5.3" name="GLM 5.3" {
+			reasoning #true
+			input "text"
+			cost input=2.02 output=6.18 cache-read=0 cache-write=0
+			limits context=256000 max-tokens=16000
+		}
+		model "glm-5.2" name="GLM 5.2" {
+			reasoning #true
+			input "text"
+			cost input=2.02 output=6.18 cache-read=0 cache-write=0
+			limits context=256000 max-tokens=16000
+		}
+		model "deepseek-v4.1-flash" name="DeepSeek V4.1 Flash" {
+			reasoning #true
+			input "text" "image"
+			cost input=0.67 output=2.02 cache-read=0.13 cache-write=0
+			limits context=256000 max-tokens=32000
+		}
+		model "deepseek-v4-flash-0731" name="DeepSeek V4 Flash 0731" {
+			reasoning #true
+			input "text"
+			cost input=0.45 output=0.9 cache-read=0.09 cache-write=0
+			limits context=256000 max-tokens=32000
+		}
+		model "qwen3.8-27b" name="Qwen3.8 27B" {
+			reasoning #true
+			input "text" "image"
+			cost input=0.67 output=3.71 cache-read=0.13 cache-write=0
+			limits context=256000 max-tokens=32000
+		}
+		model "qwen3.6-35b-a3b" name="Qwen3.6 35B A3B" {
+			reasoning #true
+			input "text" "image"
+			cost input=0.28 output=1.69 cache-read=0 cache-write=0
+			limits context=256000 max-tokens=32000
+		}
+		model "qwen3.5-397b-a17b" name="Qwen3.5 397B A17B" {
+			reasoning #true
+			input "text" "image"
+			cost input=0.67 output=4.04 cache-read=0 cache-write=0
+			limits context=250000 max-tokens=16000
+		}
+		model "qwen3-235b-a22b-instruct-2507" name="Qwen3 235B A22B Instruct 2507" {
+			reasoning #false
+			input "text"
+			cost input=0.84 output=2.53 cache-read=0 cache-write=0
+			limits context=250000 max-tokens=16000
+		}
+		model "gemma-4-26b-a4b-it" name="Gemma 4 26B A4B IT" {
+			reasoning #true
+			input "text" "image"
+			cost input=0.28 output=0.56 cache-read=0 cache-write=0
+			limits context=256000 max-tokens=32000
+		}
+		model "mistral-medium-3.5-128b" name="Mistral Medium 3.5 128B" {
+			reasoning #true
+			input "text" "image"
+			cost input=1.69 output=8.43 cache-read=0 cache-write=0
+			limits context=180000 max-tokens=16000
+		}
+		model "mistral-small-3.2-24b-instruct-2506" name="Mistral Small 3.2 24B Instruct 2506" {
+			reasoning #false
+			input "text" "image"
+			cost input=0.17 output=0.39 cache-read=0 cache-write=0
+			limits context=128000 max-tokens=32000
+		}
+		model "llama-3.3-70b-instruct" name="Llama 3.3 70B Instruct" {
+			reasoning #false
+			input "text"
+			cost input=1.01 output=1.01 cache-read=0 cache-write=0
+			limits context=100000 max-tokens=16000
+		}
+	}
+
+	thinking-mode "effort"
+
+	// Scaleway takes OpenAI reasoning_effort on every lineage it serves —
+	// never Qwen-native enable_thinking — and documents `none` as the off
+	// value on every reasoning model it serves over Chat Completions, so
+	// thinking-off must send `none`: omitting the field would fall back to
+	// the server default (medium) and keep reasoning on.
+	class "qwen" {
+		thinking-format "openai"
+		reasoning-disable-mode "none-effort"
+	}
+	class "glm" {
+		reasoning-disable-mode "none-effort"
+	}
+	class "deepseek" {
+		reasoning-disable-mode "none-effort"
+	}
+	class "gemma" {
+		reasoning-disable-mode "none-effort"
+	}
+	class "mistral" {
+		reasoning-disable-mode "none-effort"
+	}
+
+	// The ladders below are cascade rules rather than seed thinking so they
+	// also reach bare discovered rows that have no bundled reference yet.
+	// GLM 5.3 diverges from the lineage glob (classes/glm.kdl priority=10:
+	// low/high/max, thinking cannot be disabled): Scaleway's card documents
+	// none/low/high/max and `none` is verified to disable reasoning, so this
+	// residue outranks the glob; the default effort stays the lineage's `max`.
+	models "glm-5.3" priority=20 {
+		thinking-upgrade-neutral #true
+		thinking-efforts "low" "high" "max"
+		thinking-requires-effort #false
+	}
+	// GLM 5.2 documents none/high/max, dropping the lineage's
+	// minimal/low/medium tiers.
+	models "glm-5.2" priority=20 {
+		thinking-upgrade-neutral #true
+		thinking-efforts "high" "max"
+		thinking-default-level "max"
+		thinking-requires-effort #false
+	}
+	// DeepSeek Flash ladders match the lineage's wire-exact low/high/max;
+	// `none` is documented on both cards. Restated here because the
+	// vendor-suffixed -v4.1-flash and dated -0731 ids have no same-id census
+	// row the family ladder could ride in on.
+	models "deepseek-v4.1-flash" "deepseek-v4-flash-0731" {
+		thinking-upgrade-neutral #true
+		thinking-efforts "low" "high" "max"
+	}
+	// Qwen cards: 3.8 keeps its native xhigh tier; 3.6 and 3.5 stop at high.
+	models "qwen3.8-27b" {
+		thinking-upgrade-neutral #true
+		thinking-efforts "low" "medium" "xhigh"
+	}
+	models "qwen3.6-35b-a3b" "qwen3.5-397b-a17b" {
+		thinking-upgrade-neutral #true
+		thinking-efforts "low" "medium" "high"
+	}
+	// Gemma 4 documents none/low/medium/high.
+	models "gemma-4-26b-a4b-it" {
+		thinking-upgrade-neutral #true
+		thinking-efforts "low" "medium" "high"
+	}
+	// Mistral Medium 3.5 documents none/high only; the server default
+	// (medium) is not a valid tier, so the omp default pins high.
+	models "mistral-medium-3.5-128b" {
+		thinking-upgrade-neutral #true
+		thinking-efforts "high"
+		thinking-default-level "high"
+	}
+}