@linxiraos/pi-catalog 1.2.0 → 1.3.0

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

@@ -8,12 +8,27 @@ provider "cursor" {
 	}
 
 	thinking-mode "effort"
-	// Cursor's default route selects a differently priced backing model per request.
+	// Cursor relays its upstream's image count (Claude 600, GPT 1,500) and took
+	// a 64 MB body; 60 images at the 500 KiB resize target stay near 31 MB.
+	image-count-budget 60
+	// Cursor's default route selects a differently priced backing model per
+	// request, Claude and GPT included, so it keeps Claude's image count and
+	// their 1024px edge; Gemini and Composer read 1024px pages as well as
+	// larger ones.
 	models "default" {
 		pricing-status "variable"
+		image-count-budget 20
+		image-max-edge 1024
 	}
 
+	// Cursor re-encodes every image for its Claude and GPT upstreams to fit
+	// 1024×1024 before the model reads it; Gemini and Composer read larger
+	// images as sent.
 	class "anthropic" {
+		image-max-edge 1024
+		// Past 20 images Claude rejects any image edge over 3,000 px, checked
+		// before Cursor downsizes.
+		image-count-budget 20
 		family "fable" {
 			requires-cursor-tool-schema-projection #true
 			// Cursor advertised Context window 300k (Max 1M recovered elsewhere).
@@ -56,6 +71,8 @@ provider "cursor" {
 		}
 	}
 	class "openai" {
+		// GPT images are fitted to 1024×1024 like Claude's (see above).
+		image-max-edge 1024
 		revision ">=5.6 <5.7" {
 			// Cursor advertised GPT-5.6 Sol/Terra/Luna Context window 272k;
 			// raises unlabeled -fast variants from the 200k default.
@@ -433,6 +450,26 @@ provider "cursor" {
 			cache-write 0
 		}
 	}
+	// Gemini 4 Argon: $2/$10, cache read $0.10. Other Models bill at the
+	// model's API price, so the card follows Google's list-price increase to
+	// $4/$20, cache read $0.20, from 2026-12-01 (midnight US Pacific).
+	models "gemini-4-argon" {
+		cost-patch {
+			input 2.0
+			output 10.0
+			cache-read 0.1
+			cache-write 0
+			effective-rates {
+				list-price {
+					effective-from "2026-12-01T08:00:00Z"
+					input 4.0
+					output 20.0
+					cache-read 0.2
+					cache-write 0
+				}
+			}
+		}
+	}
 	// GLM 5.2: $1.40/$4.40, cache read $0.26.
 	models "glm-5.2" {
 		cost-patch {
@@ -711,3 +748,15 @@ provider "cursor" {
 		}
 	}
 }
+
+// The same 1024px edge for targets that name only Cursor's wire (`{ api, id }`,
+// no host). priority=-1: a fallback, so the equally specific provider rule
+// above wins when both match.
+on-api "cursor-agent" {
+	class "anthropic" priority=-1 {
+		image-max-edge 1024
+	}
+	class "openai" priority=-1 {
+		image-max-edge 1024
+	}
+}

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

@@ -1,4 +1,4 @@
-// Catalog entry for "deepinfra"; no deployment-contract compat rules are needed.
+// Catalog entry for "deepinfra".
 
 provider "deepinfra" {
 	default-model "deepseek-ai/DeepSeek-V4-Flash-0731"
@@ -9,6 +9,8 @@ provider "deepinfra" {
 		image "openai-images"
 		tts "openai-speech"
 	}
+	// Same-provider image-generation swap target for chat models.
+	image-model "black-forest-labs/FLUX-2-pro"
 
 	seed base-url="https://api.deepinfra.com/v1/openai" bundle="always" {
 		model "black-forest-labs/FLUX-2-pro" name="FLUX.2 Pro" api="openai-images" {

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

@@ -1,4 +1,4 @@
-// Catalog entry for "devin"; no deployment-contract compat rules are needed.
+// Catalog entry and deployment contract for "devin".
 
 provider "devin" {
 	default-model "swe-1-6"
@@ -45,7 +45,7 @@ provider "devin" {
 	// (PRO reports $0 for SWE-2, SWE-1.7, and GLM-5.2 High). `cost-fallback`
 	// seeds the enterprise-tier list price only when upstream reports no
 	// token price, so accounts whose discovery does carry dimensions keep
-	// their real rates.
+	// their real rates. None of these models publishes a cache-write rate.
 
 	// SWE-2: enterprise promo is 75% off list through 2026-12-31; list
 	// pricing ($3/$15, $0.30 cache read) applies from 2027-01-01.
@@ -54,14 +54,14 @@ provider "devin" {
 			input 0.75
 			output 3.75
 			cache-read 0.075
-			cache-write 0.75
+			cache-write 0
 			effective-rates {
 				list-price {
 					effective-from "2027-01-01T00:00:00Z"
 					input 3
 					output 15
 					cache-read 0.3
-					cache-write 3
+					cache-write 0
 				}
 			}
 		}
@@ -74,7 +74,7 @@ provider "devin" {
 			input 0.5
 			output 2.5
 			cache-read 0.2
-			cache-write 0.5
+			cache-write 0
 		}
 	}
 	models "glm-5-2" {
@@ -82,7 +82,118 @@ provider "devin" {
 			input 1.4
 			output 4.4
 			cache-read 0.26
-			cache-write 1.4
+			cache-write 0
 		}
 	}
+
+	// Cache-write rates come from Devin's published price card
+	// (https://docs.devin.ai/desktop/models, `modelCostData`);
+	// `GetCliModelConfigs` and the v2alpha models endpoint carry none. Chat
+	// usage reports cache-write tokens for Claude and GPT-5.6+ (Gemini
+	// reports none), and the card bills them at 1.25x input. Earlier GPT,
+	// Grok, Kimi, GLM, DeepSeek, and SWE rows publish no cache-write rate.
+	// The GPT selector stops at the next major generation, like the tiers
+	// below, so an unpublished generation records no cache-write price.
+	class "anthropic" {
+		cache-write-rate 1.25
+	}
+	class "openai" {
+		family "gpt" {
+			revision ">=5.6 <7" {
+				cache-write-rate 1.25
+			}
+		}
+	}
+	// residue: Adaptive is a router uid with no model identity; its card
+	// bills cache writes at the input rate.
+	models "adaptive" {
+		cache-write-rate 1.0
+	}
+
+	// Long-context tiers come from the same page's `longContextFamilies`,
+	// which no Devin endpoint carries. A tier bills the whole request once
+	// total prompt tokens (uncached input, cache reads, and cache writes)
+	// exceed the threshold. GPT and Gemini Pro tiers double input and cache
+	// rates and raise output 1.5x; Grok doubles every leg. Cache writes
+	// follow `cache-write-rate`. Each selector stops at the next major
+	// generation, and Gemini Pro skips 3.0: the page lists GPT 5.4–6.1,
+	// Gemini 2.5 and 3.1 Pro, and Grok 4.5–4.7, so an unlisted generation
+	// carries no tier.
+	class "openai" {
+		family "gpt" {
+			// Every GPT-6 lane, Fast (`-priority`) lanes included.
+			revision ">=6 <7" {
+				long-context-cost {
+					input-threshold 272000
+					multiplier 2.0
+					output-multiplier 1.5
+				}
+				// mini and nano share the class rank but have no published tier.
+				models token="mini" {
+					long-context-cost {}
+				}
+				models token="nano" {
+					long-context-cost {}
+				}
+			}
+		}
+	}
+	// residue: GPT-5.4 through 5.6 tiers cover only the standard lanes; the
+	// Fast (`-priority`) lanes and GPT-5.4 mini share their class rank.
+	models "gpt-5-4" "gpt-5-4-none" "gpt-5-4-low" "gpt-5-4-medium" "gpt-5-4-high" "gpt-5-4-xhigh" \
+		"gpt-5-5" "gpt-5-5-none" "gpt-5-5-low" "gpt-5-5-medium" "gpt-5-5-high" "gpt-5-5-xhigh" \
+		"gpt-5-6-sol" "gpt-5-6-sol-none" "gpt-5-6-sol-low" "gpt-5-6-sol-medium" "gpt-5-6-sol-high" \
+		"gpt-5-6-sol-xhigh" "gpt-5-6-sol-max" \
+		"gpt-5-6-terra" "gpt-5-6-terra-none" "gpt-5-6-terra-low" "gpt-5-6-terra-medium" "gpt-5-6-terra-high" \
+		"gpt-5-6-terra-xhigh" "gpt-5-6-terra-max" \
+		"gpt-5-6-luna" "gpt-5-6-luna-none" "gpt-5-6-luna-low" "gpt-5-6-luna-medium" "gpt-5-6-luna-high" \
+		"gpt-5-6-luna-xhigh" "gpt-5-6-luna-max" {
+		long-context-cost {
+			input-threshold 272000
+			multiplier 2.0
+			output-multiplier 1.5
+		}
+	}
+	class "gemini" {
+		family "pro" {
+			revision ">=2.5 <3" {
+				long-context-cost {
+					input-threshold 200000
+					multiplier 2.0
+					output-multiplier 1.5
+				}
+			}
+			revision ">=3.1 <4" {
+				long-context-cost {
+					input-threshold 200000
+					multiplier 2.0
+					output-multiplier 1.5
+				}
+			}
+		}
+	}
+	// residue: Gemini 2.5 Pro's legacy uid carries no model identity.
+	models "MODEL_GOOGLE_GEMINI_2_5_PRO" {
+		long-context-cost {
+			input-threshold 200000
+			multiplier 2.0
+			output-multiplier 1.5
+		}
+	}
+	class "xai" {
+		family "grok" {
+			revision ">=4.5 <5" {
+				long-context-cost {
+					input-threshold 200000
+					multiplier 2.0
+				}
+			}
+		}
+	}
+}
+
+// Devin's Cascade catalog reports each model's thinking surface; a row
+// without one carries no reasoning ladder.
+on-api "devin-agent" {
+	trust-explicit-thinking-only #true
 }

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

@@ -691,6 +691,12 @@ provider "factory-droid" {
 		thinking-requires-effort #true
 		thinking-default-level "medium"
 	}
+	// The native 922K window is authoritative; the class-wide gpt-5.4 1M patch does not apply.
+	class "openai" {
+		models "gpt-5.4" {
+			limits-patch {}
+		}
+	}
 	models "gpt-5.4-fast" {
 		routing-family "openai"
 		upstream-rotation "openai" "azure_openai"

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

@@ -38,4 +38,19 @@ provider "firepass" {
 		thinking-mode "effort"
 		thinking-efforts "low" "high" "max"
 	}
+	// The Fire Pass router shares Fireworks' 32,768 Kimi K2 output cap (#1849).
+	class "kimi" {
+		family "k2" {
+			max-tokens-ceiling 32768
+		}
+		family "k2.5" {
+			max-tokens-ceiling 32768
+		}
+		family "k2.6" {
+			max-tokens-ceiling 32768
+		}
+		family "k2-thinking" {
+			max-tokens-ceiling 32768
+		}
+	}
 }

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

@@ -24,6 +24,32 @@ provider "fireworks" priority=-1 {
 	class "unknown" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 	}
+	// Replaces the Fireworks Kimi output caps in provider-models/openai-compat.ts
+	// (#1849). The K2 family ships long reasoning traces Fireworks bounds at
+	// 32,768 output tokens; the control plane reports no output budget, so the
+	// cap is also the fallback. K2.7-Code serves its full 65,536 (verified with
+	// a 58,971-token completion), which seeds rows that report none.
+	class "kimi" {
+		family "k2" {
+			max-tokens-fallback 32768
+			max-tokens-ceiling 32768
+		}
+		family "k2.5" {
+			max-tokens-fallback 32768
+			max-tokens-ceiling 32768
+		}
+		family "k2.6" {
+			max-tokens-fallback 32768
+			max-tokens-ceiling 32768
+		}
+		family "k2-thinking" {
+			max-tokens-fallback 32768
+			max-tokens-ceiling 32768
+		}
+		family "k2.7-code" {
+			max-tokens-fallback 65536
+		}
+	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
 	models "deepseek-v4-flash" {
 		requires-reasoning-content-for-all-assistant-turns #true

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

@@ -14,6 +14,30 @@ provider "github-copilot" {
 	// not belong to this connection`: a resumed session rebuilds prior turns
 	// until its first response.
 	connection-bound-native-history #true
+	// Per-plan API hosts the Copilot token resolves to; the bundled base URL
+	// is `api.githubcopilot.com`.
+	endpoint-hosts "api.individual.githubcopilot.com" "api.business.githubcopilot.com" \
+		"api.enterprise.githubcopilot.com"
+	// `/chat/completions`, `/responses` and `/v1/messages` decode zstd, br and
+	// gzip request bodies (deflate is rejected), probed live 2026-10. GHE.com
+	// tenants (`copilot-api.<tenant>.ghe.com`) are unverified and stay plain.
+	request-encodings "zstd" "gzip"
+	request-encoding-rejection "invalid_request_body|[Ff]ailed to (?:parse|read) request"
+	// Replaces the provider-keyed strict Responses tool-call/result pairing.
+	strict-responses-pairing #true
+	// Replaces the provider check that disabled reasoning params on chat completions.
+	on-api "openai-completions" {
+		supports-reasoning-params #false
+	}
+	// Copilot caps GPT-5.4 below the lineage's 1M window (model-limits in
+	// runtime/behavior.kdl), so the class-wide window patch and uncapped
+	// window do not apply here.
+	class "openai" {
+		models "gpt-5.4" {
+			limits-patch {}
+			long-context-window-uncapped #false
+		}
+	}
 	class "anthropic" {
 		revision ">=4 <4.6" {
 			thinking-efforts "minimal" "low" "medium" "high" "xhigh"

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

@@ -9,6 +9,9 @@ provider "google-antigravity" {
 		image "google-gemini-cli"
 	}
 	web-search "gemini"
+	// No size rejection up to a 64 MB body; 60 images at the 500 KiB resize
+	// target stay near 41 MB. Gemini counts images only against its context.
+	image-count-budget 60
 
 	seed api="google-gemini-cli" base-url="https://daily-cloudcode-pa.googleapis.com" bundle="always" {
 		model "gemini-3-pro-image" name="Gemini 3 Pro Image" {
@@ -26,6 +29,29 @@ provider "google-antigravity" {
 	preserves-max-output-tokens #false
 
 	class "anthropic" {
+		// Claude here is Vertex Anthropic: one image whose base64 passes
+		// 5,242,880 bytes is a 400 "image exceeds 5 MB maximum" (half the
+		// direct API's 10 MiB), and 8001 px is a 400 like on the direct API.
+		// Declaring the contract makes the Cloud Code Assist wire clamp images
+		// for Claude the way `anthropic-messages` does; Gemini on the same
+		// gateway has no such limits and keeps them unset.
+		max-image-dimension 8000
+		max-image-payload-bytes 5242880
+		// The Claude upstream rejects a message over 30 MB and, past 20 images,
+		// any image edge over 3,000 px. The clamp above caps each image; with no
+		// byte budget on this wire, the count bounds the body.
+		image-count-budget 20
+		// Cache-warmer lifetime. Claude here is Vertex Anthropic, whose prompt
+		// cache has a five-minute default TTL refreshed on every hit; this wire
+		// sends no `cache_control`, so the 1h tier is never requested.
+		// https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/prompt-caching
+		// Gemini declares none: its implicit cache keeps entries for a
+		// load-dependent time ("caches always deleted within 24 hours"), so no
+		// fixed lifetime exists to warm against.
+		// https://cloud.google.com/blog/products/ai-machine-learning/vertex-ai-context-caching
+		prompt-cache {
+			short 300
+		}
 		family "opus" {
 			thinking-mode "budget"
 		}
@@ -59,6 +85,17 @@ provider "google-antigravity" {
 			}
 		}
 	}
+	// Cloud Code Assist bills a Gemini 2.5 or 3.x image a fixed 1,089 tokens
+	// from 32px to 2048px (extreme strips bill less); priority=1 outranks the
+	// lineage's 1,120 at equal specificity. verified: billing matched this rule
+	// on Gemini 2.5 Flash Lite, 3.1 Flash Lite and 3.8 Flash.
+	class "gemini" priority=1 {
+		image-tokenization {
+			regime "fixed"
+			verified #true
+			tokens 1089
+		}
+	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
 	models "claude-opus-4-6" {
 		thinking-efforts "minimal" "low" "medium" "high"

package/src/compat/rules/providers/google-gemini-cli.kdl

@@ -2,6 +2,7 @@
 
 provider "google-gemini-cli" {
 	default-model "gemini-3.1-pro-preview"
+	image-count-budget 200
 
 	class "gemini" {
 		// Cloud Code Assist rejects an unsigned first function call; the bypass

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

@@ -3,6 +3,7 @@
 provider "google-vertex" {
 	default-model "gemini-3.1-pro-preview"
 	allow-unauthenticated #true
+	image-count-budget 200
 
 	// Replaces google-shared.ts stripping functionCall/functionResponse ids for Vertex.
 	supports-function-part-id #false
@@ -27,6 +28,10 @@ provider "google-vertex" {
 		}
 	}
 	class "anthropic" {
+		// Vertex rawPredict honors `cache_control.ttl: "1h"` (writes bill to
+		// `ephemeral_1h_input_tokens`, entries outlive 5m) without any beta; the
+		// `extended-cache-ttl-2025-04-11` beta is rejected in `anthropic_beta`.
+		supports-long-cache-retention #true
 		revision ">=4.5 <4.6" {
 			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		}

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

@@ -7,6 +7,7 @@ provider "google" {
 		image "google-generative-ai"
 	}
 	web-search "gemini"
+	image-count-budget 200
 
 	class "unknown" {
 		thinking-efforts "minimal" "low" "medium" "high"

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

@@ -1,6 +1,7 @@
 // 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.
+// fallbacks, so a class-wide lineage rule (same specificity) always wins. None
+// is `verified`: nobody measured the billing of the readers they price.
 
 // OpenAI-compatible wires bill like GPT-5.5: 32px patches × 1.2, `auto` sized
 // like `original`.

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

@@ -19,6 +19,13 @@ provider "local" {
 			limits
 			supports-tools #false
 		}
+		model "nemotron-streaming-en-0.6b" name="Nemotron Speech 0.6B (Streaming English)" {
+			reasoning #false
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits
+			supports-tools #false
+		}
 		model "whisper-base" name="Whisper Base" {
 			reasoning #false
 			input "text"
@@ -101,7 +108,7 @@ provider "local" {
 	models "kokoro" {
 		kind "tts"
 	}
-	models "parakeet-tdt-0.6b-v3" "whisper-*" {
+	models "parakeet-tdt-0.6b-v3" "nemotron-streaming-en-0.6b" "whisper-*" {
 		kind "stt"
 	}
 	models "lfm*" "falcon-*" "qwen*" "llama3.2:3b" "gemma-3-1b" {

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

@@ -54,6 +54,19 @@ provider "meta" {
 			supports-reasoning-effort #true
 			include-encrypted-reasoning #true
 		}
+		// Muse Wonder: dev.meta.ai/docs/models + pricing-rate-limits (2026-10-12).
+		// `max` documented; no contributor tier on Model API. Meta publishes no
+		// output cap, so it takes the Muse Spark ceiling.
+		model "muse-wonder-1.0" name="Muse Wonder" {
+			reasoning #true
+			input "text" "image"
+			cost input=4 output=20 cache-read=0.2 cache-write=0
+			limits context=1048576 max-tokens=131072
+			thinking-mode "effort"
+			thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
+			supports-reasoning-effort #true
+			include-encrypted-reasoning #true
+		}
 		model "muse-spark-1.3-contributor" name="Muse Spark 1.3 (C)" {
 			reasoning #true
 			input "text" "image"
@@ -72,7 +85,10 @@ provider "meta" {
 	// 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
+	// `/v1/models` carries no capability metadata: an id no seed or taxonomy
+	// knows is offered this ladder for the user to opt into.
 	class "unknown" {
+		thinking-unverified #true
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		thinking-mode "effort"
 	}

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

@@ -6,6 +6,8 @@ provider "moonshot" {
 	// as a fallback so China users need not learn MOONSHOT_API_KEY. (#2883)
 	env "MOONSHOT_API_KEY" "KIMI_API_KEY"
 	discovery label="Moonshot"
+	// Moonshot's China platform, selected with MOONSHOT_BASE_URL (#2883).
+	endpoint-hosts "api.moonshot.cn"
 
 	class "kimi" {
 		// Replaces the Moonshot Kimi thinking-format branch.

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

@@ -6,6 +6,46 @@ provider "muse-code" {
 
 	// Muse Code shares Meta Model API's model capabilities and equivalent token pricing.
 	seed api="openai-responses" base-url="https://api.meta.ai/v1" bundle="always" {
+		// Subscription-only rows: Model API serves neither Muse Wonder Full nor a
+		// Wonder contributor tier, so no published SKU prices them; costs are
+		// API-equivalent spend. Standard rows carry Muse Wonder's Standard-mode
+		// list price ($4/$0.20/$20; an estimate for Wonder Full). Contributor rows
+		// carry the measured subscription-meter burn: Wonder contributor moved
+		// the meter 1% per ~61-64K input / ~12K output tokens vs ~23-25K / ~4.8K
+		// for standard Wonder, i.e. $1.45-1.63 input and $7.55-8.00 output at the
+		// $4/$20 list (measured 2026-10-13, #15522). Cached input is unmeasured
+		// and takes the same 0.4x ratio. All accept image input and the `max`
+		// effort over the subscription key (verified 2026-10-12).
+		model "muse-wonder-1.0-contributor" name="Muse Wonder (C)" {
+			reasoning #true
+			input "text" "image"
+			cost input=1.6 output=8 cache-read=0.08 cache-write=0
+			limits context=1048576 max-tokens=131072
+			thinking-mode "effort"
+			thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
+			supports-reasoning-effort #true
+			include-encrypted-reasoning #true
+		}
+		model "muse-wonder-full-1.0" name="Muse Wonder Full" {
+			reasoning #true
+			input "text" "image"
+			cost input=4 output=20 cache-read=0.2 cache-write=0
+			limits context=1048576 max-tokens=131072
+			thinking-mode "effort"
+			thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
+			supports-reasoning-effort #true
+			include-encrypted-reasoning #true
+		}
+		model "muse-wonder-full-1.0-contributor" name="Muse Wonder Full (C)" {
+			reasoning #true
+			input "text" "image"
+			cost input=1.6 output=8 cache-read=0.08 cache-write=0
+			limits context=1048576 max-tokens=131072
+			thinking-mode "effort"
+			thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
+			supports-reasoning-effort #true
+			include-encrypted-reasoning #true
+		}
 		models-from "meta"
 	}
 
@@ -27,7 +67,16 @@ provider "muse-code" {
 	// `custom` tools are rejected with 400 "`custom` tools are not supported
 	// on this endpoint", so apply-patch must stay a function tool here.
 	edit-prompt-variant "compact"
+	on-api "openai-responses" {
+		// Spark 1.3 Contributor warmed siblings hit 10/10 vs 0/10 with per-child keys (2026-10-12).
+		// Meta documents shared-prefix prompt_cache_key routing; keep the opt-in on these exact Muse Code IDs.
+		models "muse-spark-1.3" "muse-spark-1.3-contributor" "muse-wonder-1.0" "muse-wonder-1.0-contributor" "muse-wonder-full-1.0" "muse-wonder-full-1.0-contributor" {
+			task-sibling-prompt-cache #true
+		}
+	}
+	// Same capability-less `/v1/models` as Meta: unknown ids are opt-in.
 	class "unknown" {
+		thinking-unverified #true
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		thinking-mode "effort"
 	}

package/src/compat/rules/providers/native-tools.kdl

@@ -1,4 +1,7 @@
 // Cursor always exposes server-owned native operations, even with an empty tool catalog.
 on-api "cursor-agent" {
 	requires-native-tools #true
+	// MCP schemas live in Cursor's server-side catalog and enter model context
+	// only when dynamically selected, not as a fixed cost on every turn.
+	tool-schemas-in-context #false
 }

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

@@ -5,9 +5,14 @@ provider "ollama-cloud" {
 	env "OLLAMA_CLOUD_API_KEY"
 	discovery label="Ollama Cloud" oauth-provider="ollama-cloud"
 
+	// Legacy deployments can reject assistant thinking history (#1000).
+	strip-thinking-history #true
 	thinking-mode "effort"
 	// Same runtime as local Ollama: a trailing assistant turn is continued verbatim.
 	supports-assistant-prefill #true
+	// Replaces the provider checks in model-registry.ts and the generator:
+	// never send a caller output limit.
+	omit-max-output-tokens #true
 	class "unknown" {
 		thinking-efforts "minimal" "low" "medium" "high"
 	}
@@ -20,7 +25,12 @@ provider "ollama-cloud" {
 		thinking-efforts "minimal" "low" "medium" "high"
 	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
+	// #9693 verifies thinking history for these GLM-5.2 wire IDs only.
 	models "glm-5.2" {
+		strip-thinking-history #false
 		thinking-efforts "high" "max"
 	}
+	models "glm-5.2:cloud" {
+		strip-thinking-history #false
+	}
 }