@linxiraos/pi-catalog 1.1.26 → 1.2.0

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

@@ -111,6 +111,12 @@ provider "factory-droid" {
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=180000 max-tokens=32000
 		}
+		model "claude-haiku-5-5" name="Haiku 5.5" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=872000 max-tokens=128000
+		}
 		model "claude-haiku-4-5-20251001" name="Haiku 4.5" {
 			reasoning #true
 			input "text" "image"
@@ -273,6 +279,12 @@ provider "factory-droid" {
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=1007232 max-tokens=32768
 		}
+		model "mistral-large-4" name="Mistral Large 4" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=524288 max-tokens=64000
+		}
 		model "mistral-medium-3.5" name="Mistral Medium 3.5" {
 			reasoning #true
 			input "text" "image"
@@ -400,7 +412,7 @@ provider "factory-droid" {
 	models "claude-fable-5.1" {
 		routing-family "anthropic"
 		upstream-rotation "anthropic" "vertex_anthropic" "bedrock_anthropic"
-		region-upstreams-eu
+		region-upstreams-eu "vertex_anthropic"
 		credit-rates { input 4; output 5; cache-read 0.025 }
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
 		thinking-requires-effort #false
@@ -410,7 +422,7 @@ provider "factory-droid" {
 	models "claude-fable-5" {
 		routing-family "anthropic"
 		upstream-rotation "anthropic" "vertex_anthropic" "bedrock_anthropic"
-		region-upstreams-eu
+		region-upstreams-eu "vertex_anthropic"
 		credit-rates { input 4; output 5 }
 		list-price-from "anthropic"
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
@@ -439,7 +451,7 @@ provider "factory-droid" {
 	models "claude-opus-5" {
 		routing-family "anthropic"
 		upstream-rotation "anthropic" "vertex_anthropic" "bedrock_anthropic" "azure_anthropic" "snowflake"
-		region-upstreams-eu "bedrock_anthropic"
+		region-upstreams-eu "vertex_anthropic" "bedrock_anthropic"
 		credit-rates { input 2 }
 		list-price-from "anthropic"
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
@@ -458,7 +470,7 @@ provider "factory-droid" {
 	models "claude-opus-4-8" {
 		routing-family "anthropic"
 		upstream-rotation "anthropic" "vertex_anthropic" "bedrock_anthropic" "azure_anthropic" "snowflake"
-		region-upstreams-eu "bedrock_anthropic"
+		region-upstreams-eu "vertex_anthropic" "bedrock_anthropic"
 		credit-rates { input 2 }
 		list-price-from "anthropic"
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
@@ -477,7 +489,7 @@ provider "factory-droid" {
 	models "claude-opus-4-7" {
 		routing-family "anthropic"
 		upstream-rotation "anthropic" "vertex_anthropic" "bedrock_anthropic"
-		region-upstreams-eu "bedrock_anthropic"
+		region-upstreams-eu "vertex_anthropic" "bedrock_anthropic"
 		credit-rates { input 2 }
 		list-price-from "anthropic"
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
@@ -506,9 +518,10 @@ provider "factory-droid" {
 	models "claude-sonnet-5-5" {
 		routing-family "anthropic"
 		upstream-rotation "anthropic" "vertex_anthropic" "bedrock_anthropic" "azure_anthropic"
-		credit-rates { input 0.8; output 5 }
+		credit-rates { input 0.8; output 5; cache-read 0.05 }
+		list-price-from "anthropic"
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
-		thinking-requires-effort #true
+		thinking-requires-effort #false
 		thinking-default-level "high"
 		entitlement { feature-flag "claude_sonnet_5_5" }
 	}
@@ -540,6 +553,16 @@ provider "factory-droid" {
 		thinking-requires-effort #false
 		default-reasoning-off #true
 	}
+	models "claude-haiku-5-5" {
+		routing-family "anthropic"
+		upstream-rotation "anthropic" "vertex_anthropic" "bedrock_anthropic" "azure_anthropic"
+		credit-rates { input 0.04; output 5 }
+		list-price-from "anthropic"
+		thinking-efforts "low" "medium" "high" "xhigh" "max"
+		thinking-requires-effort #false
+		thinking-default-level "medium"
+		entitlement { feature-flag "claude_haiku_5_5" }
+	}
 	models "claude-haiku-4-5-20251001" {
 		routing-family "anthropic"
 		upstream-rotation "anthropic" "vertex_anthropic" "bedrock_anthropic"
@@ -562,7 +585,7 @@ provider "factory-droid" {
 	}
 	models "gpt-6.1-sol" {
 		routing-family "openai"
-		upstream-rotation "openai"
+		upstream-rotation "openai" "bedrock_openai" "databricks"
 		region-upstreams-eu "openai"
 		credit-rates { input 0.8; output 5; cache-read 0.05 }
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
@@ -589,7 +612,6 @@ provider "factory-droid" {
 		thinking-efforts "low" "medium" "high" "xhigh" "max"
 		thinking-requires-effort #false
 		thinking-default-level "medium"
-		entitlement { feature-flag "gpt_6_luna" }
 	}
 	models "gpt-5.6-sol" {
 		routing-family "openai"
@@ -749,6 +771,7 @@ provider "factory-droid" {
 	models "gemini-3.8-flash" {
 		routing-family "google"
 		upstream-rotation "google"
+		region-upstreams-eu "google"
 		credit-rates { input 0.6; output 5 }
 		list-price-from "google"
 		thinking-efforts "low" "medium" "high"
@@ -758,6 +781,7 @@ provider "factory-droid" {
 	models "gemini-3.7-flash" {
 		routing-family "google"
 		upstream-rotation "google"
+		region-upstreams-eu "google"
 		credit-rates { input 0.6; output 5 }
 		list-price-from "google"
 		thinking-efforts "low" "medium" "high"
@@ -767,6 +791,7 @@ provider "factory-droid" {
 	models "gemini-3.6-flash" {
 		routing-family "google"
 		upstream-rotation "google"
+		region-upstreams-eu "google"
 		credit-rates { input 0.6; output 5 }
 		list-price-from "google"
 		thinking-efforts "low" "medium" "high"
@@ -776,6 +801,7 @@ provider "factory-droid" {
 	models "gemini-3.5-flash" {
 		routing-family "google"
 		upstream-rotation "google"
+		region-upstreams-eu "google"
 		credit-rates { input 0.6 }
 		list-price-from "google"
 		thinking-efforts "minimal" "low" "medium" "high"
@@ -801,6 +827,19 @@ provider "factory-droid" {
 		thinking-requires-effort #false
 		thinking-default-level "high"
 	}
+	// Factory bills the Standard list-price ceiling, above the bundled Mistral
+	// row, so no list price is borrowed.
+	models "mistral-large-4" {
+		routing-family "factory"
+		upstream-rotation "mistral"
+		region-upstreams-us
+		region-upstreams-eu "mistral"
+		credit-rates { input 0.544; output 3.08; cache-read 0.103 }
+		thinking-efforts "high"
+		thinking-requires-effort #false
+		thinking-default-level "high"
+		entitlement { feature-flag "mistral_large_4" }
+	}
 	models "mistral-medium-3.5" {
 		routing-family "factory"
 		upstream-rotation "mistral"
@@ -889,7 +928,6 @@ provider "factory-droid" {
 		thinking-efforts "low" "high" "max"
 		thinking-requires-effort #false
 		thinking-default-level "high"
-		entitlement { feature-flag "deepseek_v4_1_flash" }
 	}
 	models "deepseek-v4-flash-0731" {
 		routing-family "factory"
@@ -999,7 +1037,8 @@ 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" "qwen3.8-max" "minimax-m3" {
+	models "glm-*" "kimi-*" "deepseek-*" "nemotron-*" "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"
@@ -1039,7 +1078,7 @@ provider "factory-droid" {
 	}
 	// The replay codec follows the selected route, including EU GLM routes.
 	on-upstream "mistral" {
-		models "glm-*" "mistral-medium-3.5" {
+		models "glm-*" "mistral-medium-3.5" "mistral-large-4" {
 			completions-reasoning-mode "effort"
 			completions-reasoning-history "omit"
 			mistral-reasoning-content-parts #true
@@ -1058,6 +1097,7 @@ provider "factory-droid" {
 	// Messages contract on every Factory Anthropic route: non-thinking-led
 	// history drops thinking, disabled thinking omits the field, and the effort
 	// beta rides only where the route requires it alongside output effort.
+	// Where disabled thinking is sent, replayed history drops its thinking.
 	on-api "anthropic-messages" {
 		strip-thinking-history #true
 		disabled-thinking "omit"
@@ -1073,6 +1113,12 @@ provider "factory-droid" {
 			disabled-thinking "disabled"
 		}
 	}
+	// Adaptive models the registry marks `supportsDisabledThinking` send the
+	// explicit disabled form at Off on every upstream.
+	models "claude-sonnet-4-6" "claude-sonnet-5" "claude-opus-4-6" "claude-opus-4-7" "claude-opus-4-8" \
+		"claude-opus-4-8-fast" "claude-opus-5" "claude-opus-5-fast" "claude-haiku-5-5" {
+		disabled-thinking "disabled"
+	}
 	models "claude-*" "atlas-07-21" "aster-07-15" {
 		anthropic-thinking "adaptive-summarized"
 	}
@@ -1091,9 +1137,15 @@ provider "factory-droid" {
 		anthropic-thinking "budget-effort"
 	}
 	// Always-thinking models keep adaptive thinking on disabled turns, on every upstream.
-	models "claude-fable-5.1" "claude-opus-5-5" "claude-opus-5-5-fast" "claude-sonnet-5-5" {
+	models "claude-fable-5.1" "claude-opus-5-5" "claude-opus-5-5-fast" {
 		disabled-thinking "adaptive"
 	}
+	// Sonnet 5.5 rejects `disabled`: Off is `between_tools`, pinned to high
+	// effort because `between_tools` rejects xhigh and max.
+	models "claude-sonnet-5-5" {
+		disabled-thinking "between-tools"
+		between-tools-effort "high"
+	}
 	// Fast tiers are distinct SKUs served with `speed: "fast"`.
 	models "claude-opus-5-5-fast" "claude-opus-5-fast" "claude-opus-4-8-fast" {
 		fast-mode #true
@@ -1113,6 +1165,13 @@ provider "factory-droid" {
 	on-upstream "bedrock_anthropic" {
 		anthropic-tool-streaming-beta #true
 	}
+	// Azure Foundry enforces the preserved-thinking prefix check on Factory's
+	// account, so prefix-bound models (`thinking-prefix-binding`) ask it to drop
+	// stale signed blocks instead of failing. The other Anthropic upstreams
+	// predate default enforcement; the field would opt them into it.
+	on-upstream "azure_anthropic" {
+		supports-thinking-binding-controls #true
+	}
 	// xAI and Google sessions lock their upstream.
 	on-upstream "xai" "google" {
 		routing-session-lock #true
@@ -1132,6 +1191,11 @@ provider "factory-droid" {
 			reasoning-disable-mode "none-effort"
 		}
 	}
+	// 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" {
+		reasoning-disable-mode "none-effort"
+	}
 	models "gpt-*" {
 		omit-max-output-tokens #true
 		responses-verbosity "low"

package/src/compat/rules/providers/github-copilot.kdl

@@ -10,6 +10,10 @@ provider "github-copilot" {
 	// tool schemas and does not support the context-management beta or field.
 	disable-strict-tools #true
 	supports-context-management #false
+	// Responses items from an earlier connection answer `401 input item does
+	// not belong to this connection`: a resumed session rebuilds prior turns
+	// until its first response.
+	connection-bound-native-history #true
 	class "anthropic" {
 		revision ">=4 <4.6" {
 			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
@@ -76,4 +80,29 @@ provider "github-copilot" {
 		}
 		max-context-window 1050000
 	}
+	// Claude Haiku 5.5: `billing.token_prices` reports its input/output rates
+	// but no cache legs, so discovered rows priced cache reads and writes at
+	// $0. Copilot's rate card bills $0.01 read / $0.125 write up to 100K and
+	// $0.05 / $0.625 on the long tier (the `-1m` sibling).
+	// https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing
+	// Copilot serves that tier as the `-1m` row (promotion target of the base
+	// row), so both rows opt out of the class-wide 100K 5x band: the base row
+	// would otherwise bill the band on top of the promotion, and the `-1m`
+	// row would charge 5x twice.
+	models "claude-haiku-5.5" {
+		cost-patch {
+			cache-read 0.01
+			cache-write 0.125
+		}
+		long-context-cost {
+		}
+	}
+	models "claude-haiku-5.5-1m" {
+		cost-patch {
+			cache-read 0.05
+			cache-write 0.625
+		}
+		long-context-cost {
+		}
+	}
 }

package/src/compat/rules/providers/google-antigravity.kdl

@@ -2,6 +2,9 @@
 
 provider "google-antigravity" {
 	default-model "gemini-3.1-pro"
+	// `fetchAvailableModels` is the account's served roster; bundled rows it
+	// omits (e.g. Claude 5.5 on plans without it) 404 on every request (#14328).
+	dynamic-models-authoritative #true
 	kind-apis {
 		image "google-gemini-cli"
 	}
@@ -64,7 +67,11 @@ provider "google-antigravity" {
 	models "gemini-3-pro" {
 		thinking-mode "google-level"
 	}
-	models "gemini-3-pro-image" {
+	// Antigravity discovery lists image SKUs beside chat models with no output
+	// modality, so they would build as chat. Accounts advertise either SKU in
+	// `imageGenerationModelIds` (gemini-3.1-flash-image replaced
+	// gemini-3-pro-image on some accounts, #11106).
+	models "gemini-3-pro-image" "gemini-3.1-flash-image" {
 		kind "image"
 	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.

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

@@ -4,8 +4,29 @@ provider "groq" {
 	default-model "openai/gpt-oss-120b"
 	env "GROQ_API_KEY"
 
+	// Groq validates assistant message properties and refuses `reasoning_content`
+	// (400 "property 'reasoning_content' is unsupported"), so the Kimi lineage's
+	// tool-call replay field, real or the "." placeholder, is never sent here.
+	requires-reasoning-content-for-tool-calls #false
 	thinking-mode "effort"
 	class "unknown" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 	}
+	// Groq accepts OpenAI `reasoning_effort` for Qwen, with `none` as its off
+	// value, not Qwen-native `enable_thinking` (400 "property 'enable_thinking'
+	// is unsupported"). Its documented Qwen ladder (Qwen 3.8) is low/medium/high;
+	// `minimal` answers 400 "invalid Qwen3.8 reasoning_effort".
+	class "qwen" {
+		thinking-format "openai"
+		reasoning-disable-mode "none-effort"
+		thinking-efforts "low" "medium" "high"
+	}
+	// Qwen 3.6 only takes `default` (thinking) or `none` (off), per its Groq
+	// model card.
+	models "qwen/qwen3.6-27b" {
+		thinking-efforts "high"
+		thinking-effort-map {
+			high "default"
+		}
+	}
 }

package/src/compat/rules/providers/image-tokenization.kdl

@@ -0,0 +1,42 @@
+// Image billing for readers whose lineage has no `image-tokenization` rule in
+// classes/*.kdl, by the wire API carrying the request. priority=-1: these are
+// fallbacks, so a class-wide lineage rule (same specificity) always wins.
+
+// OpenAI-compatible wires bill like GPT-5.5: 32px patches × 1.2, `auto` sized
+// like `original`.
+on-api "openai-completions" "openai-responses" "openai-codex-responses" "azure-openai-responses" priority=-1 {
+	image-tokenization {
+		regime "openai-patch"
+		multiplier 1.2
+		auto "original"
+		low {
+			max-edge 512
+		}
+		high {
+			max-edge 2048
+			patch-budget 2500
+		}
+		original {
+			max-edge 6000
+			patch-budget 10000
+		}
+	}
+}
+
+// Google wires bill Gemini 3's fixed `media_resolution` budget per image.
+on-api "google-generative-ai" "google-gemini-cli" "google-vertex" priority=-1 {
+	image-tokenization {
+		regime "fixed"
+		tokens 1120
+	}
+}
+
+// Anthropic wires and every other adapter bill Claude's high-res tier, the
+// largest per-image bill, so an unpriced reader is never undercounted.
+on-api "anthropic-messages" "bedrock-converse-stream" "openrouter" "ollama-chat" "cursor-agent" "factory-droid-agent" "gitlab-duo-agent" "devin-agent" "apple-foundation-models" priority=-1 {
+	image-tokenization {
+		regime "anthropic-patch"
+		max-edge 2576
+		max-tokens 4784
+	}
+}

package/src/compat/rules/providers/judgment-questions.kdl

@@ -0,0 +1,53 @@
+// Most questions one System One judgment request may carry. A server that
+// validates the `questions` map rejects an oversized request (HTTP 400/422)
+// before judging anything, so the judge splits larger requests. Scoped by
+// adapter, not provider name: self-hosted servers are user-configured judge
+// rows under arbitrary provider ids.
+//
+// Unassigned, verified without a count cap (token budget only): TypeSafe jev
+// (docs.typesafe.ai/models: 64k tokens per request), the same jev behind
+// OpenRouter Decisions, Command Code and OpenCode Zen, Kev
+// (github.com/jaredpalmer/kev: "A request may carry any number of
+// questions"), and wev / OpenJev-4B (github.com/alanhuangyoo/OpenJev, which
+// answers under any requested name).
+//
+// Unassigned: OpenAI Decisions (`openai-decisions`); its guide
+// (developers.openai.com/api/docs/guides/decisions) documents no question cap.
+//
+// Not isolatable: openjev-sglang (github.com/ekzhang/openjev-sglang) caps at
+// 64 but answers as `jev-latest` or its upstream model name, the same ids as
+// uncapped hosts.
+on-api "typesafe" "openrouter-decisions" {
+	// residue: no taxonomy class covers self-hosted OpenJEV deployments. An
+	// `openjev-*` gateway answered HTTP 422 `at most 16 items`
+	// (https://github.com/can1357/oh-my-pi/issues/15081); which server that is
+	// remains unidentified, so unknown OpenJEV ids keep the smallest cap
+	// observed. Over-splitting costs requests; under-splitting fails them.
+	models "*openjev*" {
+		judgment-questions-max 16
+	}
+	// residue: razorback16/openjev and its Swift port (Algorythm-Canada/
+	// OpenJevSwift) reject more than OPENJEV_MAX_QUESTIONS, default 256, with
+	// `at most 256 questions per request`. These are the wire names their
+	// DiffusionGemma, Laya, Verdict and JevK5 backends answer under; exact ids
+	// outrank the glob above.
+	models "openjev-latest" "openjev-0.1" "laya-1.0" "verdict-1.4" "jevk5-0.2" {
+		judgment-questions-max 256
+	}
+}
+// Cloudflare Clef, Clef Flash and Clef Omni: the Workers AI input schema
+// bounds `questions` to 1-64 entries (`maxProperties: 64`,
+// developers.cloudflare.com/workers-ai/models/clef, /clef-flash, /clef-omni).
+on-api "workers-ai-systemone" {
+	models "@cf/cloudflare/clef*" {
+		judgment-questions-max 64
+	}
+}
+// OpenRouter Decisions serves its `cloudflare/clef*` rows from Workers AI (and
+// PrimeIntellect, whose cap is unpublished), so the Workers AI schema's 64
+// applies; scoped apart from the TypeSafe block above.
+on-api "openrouter-decisions" {
+	models "cloudflare/clef*" {
+		judgment-questions-max 64
+	}
+}

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

@@ -1,7 +1,10 @@
-// Catalog entry for "litellm"; no deployment-contract compat rules are needed.
+// Provider-wire compat for "litellm".
 
 provider "litellm" {
 	default-model "claude-opus-5-5"
 	env "LITELLM_API_KEY"
 	discovery label="LiteLLM" allow-unauthenticated=#true
+
+	// The proxy groups request logs into sessions by this header.
+	session-header "x-litellm-session-id"
 }

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

@@ -42,8 +42,8 @@ provider "meta" {
 			supports-reasoning-effort #true
 			include-encrypted-reasoning #true
 		}
-		// Meta documents the `max` effort tier for Muse Spark 1.3 (standard) only;
-		// contributor tiers and other revisions stay on the 5-tier ladder.
+		// Meta documents the `max` effort tier for Muse Spark 1.3 standard and
+		// contributor tiers; other revisions stay on the 5-tier ladder.
 		model "muse-spark-1.3" name="Muse Spark 1.3" {
 			reasoning #true
 			input "text" "image"
@@ -60,7 +60,7 @@ provider "meta" {
 			cost input=0.1 output=0.2 cache-read=0.002 cache-write=0
 			limits context=1048576 max-tokens=131072
 			thinking-mode "effort"
-			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
+			thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
 			supports-reasoning-effort #true
 			include-encrypted-reasoning #true
 		}

package/src/compat/rules/providers/minimax-code-cn.kdl

@@ -6,11 +6,16 @@ provider "minimax-code-cn" priority=1 {
 
 	// Replaces the provider-id regex in both OpenAI baselines.
 	reasoning-deltas-may-be-cumulative #true
-	class "minimax" {
+	// residue: the legacy lightning SKUs survive only as previous-snapshot rows
+	// on the OpenAI-compatible route (models.dev and both live model lists omit
+	// them), so they keep that route's effort-mode thinking instead of the
+	// Anthropic adaptive rules the rest of the plan uses.
+	models "MiniMax-M2.1-lightning" "MiniMax-M2.5-lightning" {
 		thinking-mode "effort"
-	}
-	// residue: taxonomy ranks and exact globs do not isolate these models.
-	models "MiniMax-M3" {
-		thinking-efforts "minimal" "low" "medium" "high"
+		thinking-effort-map {
+			low "low"
+			medium "medium"
+			high "high"
+		}
 	}
 }

package/src/compat/rules/providers/minimax-code.kdl

@@ -6,11 +6,16 @@ provider "minimax-code" priority=1 {
 
 	// Replaces the provider-id regex in both OpenAI baselines.
 	reasoning-deltas-may-be-cumulative #true
-	class "minimax" {
+	// residue: the legacy lightning SKUs survive only as previous-snapshot rows
+	// on the OpenAI-compatible route (models.dev and both live model lists omit
+	// them), so they keep that route's effort-mode thinking instead of the
+	// Anthropic adaptive rules the rest of the plan uses.
+	models "MiniMax-M2.1-lightning" "MiniMax-M2.5-lightning" {
 		thinking-mode "effort"
-	}
-	// residue: taxonomy ranks and exact globs do not isolate these models.
-	models "MiniMax-M3" {
-		thinking-efforts "minimal" "low" "medium" "high"
+		thinking-effort-map {
+			low "low"
+			medium "medium"
+			high "high"
+		}
 	}
 }

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

@@ -4,7 +4,31 @@ provider "mistral" {
 	default-model "devstral-medium-latest"
 	env "MISTRAL_API_KEY"
 
-	thinking-mode "effort"
+	// Public-preview aliases and current USD / 1M token pricing:
+	// https://docs.mistral.ai/models/mistral-large-4-0 (2026-10-06).
+	// Public preview is 50% off: input $0.68, output $2.09, cached input $0.07.
+	// List prices: input $1.36, output $4.18, cached input $0.14; update when preview ends.
+	// No separate maximum output-token limit is published.
+	seed api="openai-completions" base-url="https://api.mistral.ai/v1" bundle="always" {
+		model "mistral-large-4" name="Mistral Large 4" {
+			reasoning #true
+			input "text" "image"
+			cost input=0.68 output=2.09 cache-read=0.07 cache-write=0
+			limits context=1000000
+		}
+		model "mistral-large-4-0" name="Mistral Large 4" {
+			reasoning #true
+			input "text" "image"
+			cost input=0.68 output=2.09 cache-read=0.07 cache-write=0
+			limits context=1000000
+		}
+	}
+
+	// Mistral's chat-completions endpoint derives cache affinity from the
+	// top-level `prompt_cache_key` (https://docs.mistral.ai/studio/conversations/advanced/prompt-caching);
+	// without it, requests only hit replica-local implicit cache. Same shape as
+	// the kimi-code rule (#6049).
+	supports-prompt-cache-key #true
 	class "unknown" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 	}

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

@@ -31,6 +31,17 @@ provider "moonshot" {
 			input-modalities "text" "image"
 		}
 	}
+	// stencil.so reports a $3 K3 cache write; Moonshot bills only cache hits
+	// ($0.30) and misses (the $3 input rate), with no cache-write charge.
+	// https://platform.moonshot.ai/docs/pricing/chat-k3
+	models "kimi-k3" {
+		cost-patch {
+			input 3.0
+			output 15.0
+			cache-read 0.3
+			cache-write 0.0
+		}
+	}
 	// The /models roster carries no modality metadata for the named vision/vl
 	// SKUs; the bounded id token is the census signal.
 	models token="vision" {

package/src/compat/rules/providers/muse-code.kdl

@@ -15,6 +15,13 @@ provider "muse-code" {
 	// api.meta.ai/v1 accepts only tool_choice "auto"; "none", "required" and named
 	// choices are rejected with 400 (verified 2026-09-10), so omit the field.
 	supports-tool-choice #false
+	// Results can be stored so a dropped stream is recovered: verified 2026-09-29
+	// against api.meta.ai/v1, a stored run keeps generating after the client
+	// socket dies and `GET /responses/{id}` returns it once finished (~60s for
+	// a high-effort turn), while `store: false` runs 404 forever. Storage is
+	// opt-in (privacy: stored runs retain prompts/outputs on Meta's side): omp
+	// setting providers.muse-code.storeResponses or PI_MUSE_STORE_RESPONSES=1.
+	store-responses #true
 	// Subscription tier bills per request; the compact edit prompt cuts the
 	// recurring tool-schema bytes. Verified 2026-09-05 against api.meta.ai/v1:
 	// `custom` tools are rejected with 400 "`custom` tools are not supported

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

@@ -13,6 +13,9 @@ provider "ollama" priority=-1 {
 	// Replaces both Ollama empty-length finish baselines.
 	empty-length-finish-is-context-error #true
 	thinking-efforts "low" "medium" "high" "max"
+	// Ollama's OpenAI-compatible endpoints accept effort "none" to turn thinking
+	// off; the lowest ladder effort ("low") still thinks on e.g. Gemma 4.
+	reasoning-disable-mode "none-effort"
 	// Ollama continues a trailing assistant turn verbatim on /api/chat and
 	// /v1/responses alike; callers use it to commit an output prefix.
 	supports-assistant-prefill #true