@linxiraos/pi-catalog 1.1.26 → 1.1.27

package/src/compat/resolve.ts

@@ -179,7 +179,7 @@ function overlayEffortMapAxis(
 }
 
 function effortList(value: unknown): readonly Effort[] | undefined {
-	if (!Array.isArray(value) || value.length === 0) return undefined;
+	if (!Array.isArray(value)) return undefined;
 	const out: Effort[] = [];
 	for (const entry of value) {
 		const effort = THINKING_EFFORTS.find(candidate => candidate === entry);
@@ -505,7 +505,7 @@ function detectOpenAICompat(
 		// API-conditional: this completions-only Copilot exclusion cannot be a
 		// provider rule without changing Copilot Responses rows.
 		supportsReasoningParams: provider !== "github-copilot",
-		supportsSamplingParams: !(facts.is("openai") && (facts.family("o-series") || facts.revGte("5"))),
+		supportsSamplingParams: true,
 		supportsPenaltyAndStopParams: !(isGrok && reasoningCapable),
 		reasoningEffortMap: {},
 		supportsUsageInStreaming: !isCerebrasHost,
@@ -598,26 +598,12 @@ function detectOpenAICompat(
 	};
 }
 
-const DSML_HEALING_PROVIDERS: Record<string, true> = {
-	ollama: true,
-	"ollama-cloud": true,
-	nvidia: true,
-	deepseek: true,
-	fireworks: true,
-	nanogpt: true,
-	"opencode-go": true,
-	openrouter: true,
-	// Transparent gateways / user-configured hosts forward the upstream model's
-	// native chat template unchanged, so a deepseek-classed model behind them
-	// still emits DSML tool-call envelopes and needs the DSML healer.
-	litellm: true,
-	nous: true,
-};
-
 /**
  * Default leaked-markup healer. Kimi/DeepSeek dedicated grammars are keyed on
  * identity class; official OpenAI heals nothing; everything else defaults to
- * the generic `thinking` healer.
+ * the generic `thinking` healer. DSML is DeepSeek's own tool-call grammar, so
+ * every host serving a DeepSeek model gets it — gateways, local backends, and
+ * custom providers alike.
  */
 function detectStreamMarkupHealing(
 	provider: string,
@@ -628,7 +614,7 @@ function detectStreamMarkupHealing(
 	// Kimi ids keep the generic healer, matching the census.
 	const isKimiK2 = facts.is("kimi") && facts.identity.family?.startsWith("k2") === true;
 	if (provider === "kimi-code" || provider === "moonshot" || isKimiK2) return "kimi";
-	if (facts.is("deepseek") && DSML_HEALING_PROVIDERS[provider] === true) return "dsml";
+	if (facts.is("deepseek")) return "dsml";
 	if (isOfficialOpenAIEndpoint(provider, baseUrl)) return undefined;
 	return "thinking";
 }
@@ -749,8 +735,15 @@ function resolveOpenAIResponsesPolicy(
 		supportsPromptCacheBreakpoints,
 		promptCacheBreakpointTtl: supportsPromptCacheBreakpoints ? "30m" : undefined,
 		strictResponsesPairing: isAzure || provider === "github-copilot",
-		supportsImageDetailOriginal: !isXaiHost && !modelMatchesHost(hostModel, "githubCopilot"),
+		// Azure's provider id alone only implies support while its endpoint is
+		// resolved at runtime; an explicit non-Azure baseUrl is a proxy, like Codex.
+		supportsImageDetailOriginal:
+			isOpenAIUrl ||
+			hostMatchesUrl(baseUrl, "azureOpenAI") ||
+			(isAzure && !baseUrl) ||
+			hostMatchesUrl(baseUrl, "openaiCodex"),
 		supportsReasoningSummary: !isXaiHost,
+		statefulResponses: undefined,
 		supportsAllTurnsReasoningContext: false,
 		supportsConfigurationUpdate: false,
 		supportsSteering: false,
@@ -759,7 +752,7 @@ function resolveOpenAIResponsesPolicy(
 		thinkingLoopGuard: undefined,
 		reasoningEffortMap: {},
 		supportsReasoningParams: true,
-		supportsSamplingParams: !(facts.is("openai") && (facts.family("o-series") || facts.revGte("5"))),
+		supportsSamplingParams: true,
 		supportsPenaltyAndStopParams: !isXaiHost,
 		thinkingFormat,
 		reasoningDisableMode: resolveReasoningDisableMode(thinkingFormat),
@@ -798,6 +791,7 @@ function resolveOpenAIResponsesPolicy(
 			PROXY_OPENAI_COMPAT_PROVIDERS[backendProvider] !== true &&
 			(LOCAL_OPENAI_COMPAT_PROVIDERS[backendProvider] === true || hasLocalLoopbackBaseUrl(baseUrl)),
 		supportsObfuscationOptOut: isOpenAIUrl || provider === "openai",
+		storeResponses: false,
 		officialEndpoint: isOfficialOpenAIEndpoint(provider, baseUrl),
 		harmonyLeakMitigation: false,
 		rejectRootObjectUnion: false,
@@ -859,6 +853,7 @@ function pickResponsesOnly(compat: ResolvedOpenAIResponsesCompat): ResponsesOnly
 		strictResponsesPairing: compat.strictResponsesPairing,
 		supportsImageDetailOriginal: compat.supportsImageDetailOriginal,
 		supportsObfuscationOptOut: compat.supportsObfuscationOptOut,
+		storeResponses: compat.storeResponses,
 		supportsAllTurnsReasoningContext: compat.supportsAllTurnsReasoningContext,
 		supportsConfigurationUpdate: compat.supportsConfigurationUpdate,
 		supportsSteering: compat.supportsSteering,
@@ -867,6 +862,7 @@ function pickResponsesOnly(compat: ResolvedOpenAIResponsesCompat): ResponsesOnly
 		cacheControlFormat: compat.cacheControlFormat,
 		requiresReasoningOffJuiceInstruction: compat.requiresReasoningOffJuiceInstruction,
 		supportsReasoningSummary: compat.supportsReasoningSummary,
+		statefulResponses: compat.statefulResponses,
 		isVercelGatewayHost: compat.isVercelGatewayHost,
 	} satisfies ResponsesOnlyCompat;
 }
@@ -906,7 +902,7 @@ function resolveAnthropicPolicy(
 		supportsThinkingBindingControls: false,
 		supportsBetweenToolsThinking: false,
 		supportsForcedToolChoice: !requiresThinkingEnabled && !facts.family("fable", "mythos"),
-		supportsSamplingParams: !facts.anthropicAdaptiveGenAtLeast("4.7"),
+		supportsSamplingParams: true,
 		requiresToolResultId: false,
 		requiresThinkingEnabled,
 		replayUnsignedThinking:
@@ -1162,6 +1158,7 @@ function resolveThinkingPolicy<TApi extends Api>(
 	if (compat !== undefined && "trustExplicitThinkingOnly" in compat && compat.trustExplicitThinkingOnly === true) {
 		return undefined;
 	}
+	if (rule.efforts?.length === 0) return undefined;
 	const config: ThinkingConfig = {
 		mode: rule.mode ?? defaultThinkingMode(spec, facts),
 		efforts: rule.efforts ?? fallbackEfforts(spec, compat),

package/src/compat/rules/README.md

@@ -209,7 +209,7 @@ The directive vocabulary is closed and lives in **`src/compat/axes.ts`** — one
 The three value shapes are:
 
 - **Scalar**: exactly one KDL boolean, integer, float, or string argument and no children. `#null` is rejected.
-- **Array**: one or more scalar arguments and no children; it resolves to a JSON array. Axes marked `emptyArray` in `axes.ts` also accept a bare directive, which assigns an explicit empty list (`region-upstreams-eu` with no arguments: the region never serves the model).
+- **Array**: one or more scalar arguments and no children; it resolves to a JSON array. Axes marked `emptyArray` in `axes.ts` also accept a bare directive, which assigns an explicit empty list (`region-upstreams-eu` with no arguments: the region never serves the model). Bare `thinking-efforts` keeps a reasoning-capable deployment without selectable effort tiers off the effort dial.
 - **Object**: no arguments and a child block, including an empty block. Child names are kebab-case: an axis-directive spelling compiles to its resolved axis key (`template-reasoning-effort` → `qwenTemplateReasoningEffort`), anything else converts mechanically (`input-threshold` → `inputThreshold`); camelCase names are a compile error. `extra-body` payloads (top-level or nested) are the exception — their child names are literal wire JSON keys copied verbatim (`enable_thinking`). Each child is either one scalar or another object; arrays are not representable inside an object payload.
 
 A rule cannot assign the same resolved axis twice in one block.

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

@@ -8,6 +8,6 @@ login-order "openai-codex" "anthropic" "zai" "zai-coding-plan" "kimi-code" "open
 	"xiaomi-token-plan-ams" "xiaomi-token-plan-cn" "firepass" "cline-pass" "factory-droid" "commandcode" "charm-hyper" "deepseek" "stepfun" "muse-code" "meta" "moonshot" \
 	"cerebras" "baseten" "helmcode" "fireworks" "together" "nvidia" "novita" "deepinfra" "huggingface" "perplexity" \
 	"qianfan" "venice" "siliconflow" "siliconflow-cn" "synthetic" "nanogpt" "wafer-serverless" "coreweave" \
-	"vercel-ai-gateway" "cloudflare-ai-gateway" "litellm" "kilo" "zenmux" "opencode-zen" "opencode-go" \
+	"vercel-ai-gateway" "cloudflare-ai-gateway" "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"

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

@@ -0,0 +1,11 @@
+// Snowflake Cortex REST API. Login runs Snowflake OAuth for local applications
+// (built-in SNOWFLAKE$LOCAL_APPLICATION client, PKCE, loopback); the account
+// host is prompted first, so the flow is a custom hook.
+auth "snowflake" {
+	name "Snowflake Cortex"
+	login "custom" hook="snowflake"
+	refresh hook="snowflake"
+	api-key-format "structured"
+	callback-port 54551
+	paste-code #true
+}

package/src/compat/rules/classes/anthropic.kdl

@@ -1,5 +1,70 @@
 // Model-lineage compat for "anthropic"; regenerated from the frozen census using class taxonomy selectors.
 
+// Images bill ⌈w/28⌉·⌈h/28⌉ visual tokens on every host. Claude 4.7+ and
+// Fable/Mythos read up to 2576px and 4,784 tokens; older lines fit 1568px and
+// cap at 1,568 tokens, downscaling past it.
+// https://platform.claude.com/docs/en/build-with-claude/vision#evaluate-image-size
+// The class default is the high-res tier, so ids without a revision
+// (`claude-opus-latest`, `~anthropic/claude-sonnet-latest`) and future lines
+// are never undercounted; only Opus/Sonnet/Haiku revisions known to predate
+// 4.7 take the standard tier. Separator-collapsed revisions parse as two
+// digits (`claude-opus-45` is Opus 4.5 → 45), so 10–46 is standard too.
+class "anthropic" {
+	image-tokenization {
+		regime "anthropic-patch"
+		max-edge 2576
+		max-tokens 4784
+	}
+	family "opus" {
+		revision "<4.7" {
+			image-tokenization {
+				regime "anthropic-patch"
+				max-edge 1568
+				max-tokens 1568
+			}
+		}
+		revision ">=10 <47" {
+			image-tokenization {
+				regime "anthropic-patch"
+				max-edge 1568
+				max-tokens 1568
+			}
+		}
+	}
+	family "sonnet" {
+		revision "<4.7" {
+			image-tokenization {
+				regime "anthropic-patch"
+				max-edge 1568
+				max-tokens 1568
+			}
+		}
+		revision ">=10 <47" {
+			image-tokenization {
+				regime "anthropic-patch"
+				max-edge 1568
+				max-tokens 1568
+			}
+		}
+	}
+	family "haiku" {
+		revision "<4.7" {
+			image-tokenization {
+				regime "anthropic-patch"
+				max-edge 1568
+				max-tokens 1568
+			}
+		}
+		revision ">=10 <47" {
+			image-tokenization {
+				regime "anthropic-patch"
+				max-edge 1568
+				max-tokens 1568
+			}
+		}
+	}
+}
+
 class "anthropic" {
 	family "fable" {
 		revision ">=5.1" {
@@ -7,6 +72,30 @@ class "anthropic" {
 			thinking-prefix-binding #true
 		}
 	}
+	// Adaptive Claude (Opus 4.7+, Sonnet/Fable/Mythos 5+) rejects explicit
+	// sampling params (`temperature`, `top_p`, `top_k`) with a 400 on every
+	// serving host. The `<10` bound keeps whole-number revisions such as
+	// `claude-opus-45` (Opus 4.5) out of the range.
+	family "opus" {
+		revision ">=4.7 <10" {
+			supports-sampling-params #false
+		}
+	}
+	family "sonnet" {
+		revision ">=5 <10" {
+			supports-sampling-params #false
+		}
+	}
+	family "fable" {
+		revision ">=5 <10" {
+			supports-sampling-params #false
+		}
+	}
+	family "mythos" {
+		revision ">=5 <10" {
+			supports-sampling-params #false
+		}
+	}
 	// On-demand compaction (compact-2026-09-04) works on the Claude API, Vertex,
 	// and Amazon Bedrock's `/anthropic` routes, not Bedrock Converse.
 	// Foundry / Claude Platform on AWS are first-party Anthropic routes whose
@@ -145,13 +234,19 @@ class "anthropic" {
 			revision ">=5" {
 				supports-per-message-effort #true
 			}
+			// Opus 5.5 is prefix-bound like Sonnet 5.5; without `drop_block` a
+			// changed system prompt or tool set fails the replay with a 400.
+			// Bounded for the same reason as the forced-tool-choice rule above.
+			revision ">=5.5 <6" {
+				supports-thinking-binding-controls #true
+			}
 		}
 	}
 	on "aimlapi" "gitlab-duo" "kilo" "nanogpt" "venice" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 	}
 	on "amazon-bedrock" "anthropic" "cloudflare-ai-gateway" "github-copilot" "gitlab-duo" \
-		"google-antigravity" "google-vertex" "opencode-zen" "vercel-ai-gateway" "zenmux" {
+		"google-antigravity" "google-vertex" "opencode-zen" "snowflake" "vercel-ai-gateway" "zenmux" {
 		family "sonnet" {
 			revision ">=3.7 <4.6" {
 				thinking-mode "budget"
@@ -159,13 +254,13 @@ class "anthropic" {
 		}
 	}
 	on "amazon-bedrock" "anthropic" "cloudflare-ai-gateway" "github-copilot" "gitlab-duo" \
-		"google-vertex" "opencode-zen" "vercel-ai-gateway" {
+		"google-vertex" "opencode-zen" "snowflake" "vercel-ai-gateway" {
 		family "haiku" {
 			thinking-mode "budget"
 		}
 	}
 	on "amazon-bedrock" "anthropic" "cloudflare-ai-gateway" "github-copilot" "gitlab-duo" \
-		"google-vertex" "opencode-zen" "vercel-ai-gateway" "zenmux" {
+		"google-vertex" "opencode-zen" "snowflake" "vercel-ai-gateway" "zenmux" {
 		family "opus" {
 			revision ">=4.5 <4.6" {
 				thinking-mode "anthropic-budget-effort"
@@ -173,7 +268,7 @@ class "anthropic" {
 		}
 	}
 	on "amazon-bedrock" "anthropic" "cloudflare-ai-gateway" "github-copilot" "google-vertex" \
-		"opencode-zen" "openrouter" "vercel-ai-gateway" "zenmux" {
+		"opencode-zen" "openrouter" "snowflake" "vercel-ai-gateway" "zenmux" {
 		family "opus" {
 			revision ">=4.6 <4.7" {
 				thinking-efforts "low" "medium" "high" "max"
@@ -181,7 +276,7 @@ class "anthropic" {
 		}
 	}
 	on "amazon-bedrock" "anthropic" "cloudflare-ai-gateway" "github-copilot" "google-vertex" \
-		"opencode-zen" "vercel-ai-gateway" "zenmux" {
+		"opencode-zen" "snowflake" "vercel-ai-gateway" "zenmux" {
 		revision ">=4.7 <45.1" {
 			thinking-mode "anthropic-adaptive"
 		}
@@ -212,13 +307,13 @@ class "anthropic" {
 		}
 	}
 	on "anthropic" "cloudflare-ai-gateway" "github-copilot" "google-vertex" "opencode-zen" \
-		"openrouter" "vercel-ai-gateway" "zenmux" {
+		"openrouter" "snowflake" "vercel-ai-gateway" "zenmux" {
 		revision ">=4.7 <45.1" {
 			thinking-efforts "low" "medium" "high" "xhigh" "max"
 		}
 	}
 	on "anthropic" "cloudflare-ai-gateway" "github-copilot" "google-vertex" "opencode-zen" \
-		"vercel-ai-gateway" "zenmux" {
+		"snowflake" "vercel-ai-gateway" "zenmux" {
 		family "sonnet" {
 			revision ">=4.6 <5" {
 				thinking-efforts "low" "medium" "high"

package/src/compat/rules/classes/deepseek.kdl

@@ -92,8 +92,10 @@ class "deepseek" {
 	models "deepseek-flash" priority=10 {
 		strip-image-input #false
 	}
-	on "ollama-cloud" "nvidia" "deepseek" "fireworks" "nanogpt" "opencode-go" "openrouter" "litellm" "nous" {
-		stream-markup-healing-pattern "dsml"
-	}
+	// DSML is the model's own tool-call grammar, not a host quirk: any server
+	// running DeepSeek's chat template without (or with a broken) tool parser
+	// leaks it into content — hosted gateways, local backends, and custom
+	// providers alike. On hosts that parse it the healer sees no markup.
+	stream-markup-healing-pattern "dsml"
 	thinking-loop-guard "deepseek"
 }

package/src/compat/rules/classes/gemini.kdl

@@ -118,3 +118,15 @@ class "gemini" {
 		}
 	}
 }
+
+// Gemini 3+ bills a fixed `media_resolution` budget per image regardless of
+// pixels; the default for images is HIGH, 1,120 tokens.
+// https://ai.google.dev/gemini-api/docs/media-resolution
+class "gemini" {
+	revision ">=3" {
+		image-tokenization {
+			regime "fixed"
+			tokens 1120
+		}
+	}
+}

package/src/compat/rules/classes/glm.kdl

@@ -107,7 +107,7 @@ class "glm" {
 	// the intentional overlap with census host-dialect residues such as
 	// baseten's `zai-org/GLM-5*`.
 	models "*glm-5.3" "*glm-5.3-air" "*glm-5.3-turbo" "*glm-5.3-flash" "*glm-5.3:*" \
-		"*glm-5.3-flash-uncensored" "*glm-5.3-free" "*glm-5.3-highspeed" "*glm-5.3-flash-lab" priority=10 {
+		"*glm-5.3-flash-uncensored" "*glm-5.3-free" "*glm-5.3-highspeed" "*glm-5.3-fast" "*glm-5.3-flash-lab" priority=10 {
 		thinking-default-level "max"
 		thinking-efforts "low" "high" "max"
 		thinking-requires-effort #true

package/src/compat/rules/classes/kimi.kdl

@@ -115,5 +115,12 @@ class "kimi" {
 			native-kimi-k3-reasoning #true
 			clamp-output-to-model-max #true
 		}
+		// Native K3 accepts forced tools on the Anthropic surface too;
+		// omitting thinking leaves its mandatory provider default intact.
+		on-api "anthropic-messages" {
+			family "k3" {
+				supports-forced-tool-choice #true
+			}
+		}
 	}
 }

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

@@ -3,13 +3,13 @@
 class "meta" {
 	// Exact ids on purpose: only an exact selector upgrades hosts whose
 	// discovery misreports a Muse Spark revision as non-reasoning.
-	// muse-spark-1.3 exposes the max effort tier; older/contributor revisions
-	// keep the 5-tier ladder.
-	models "muse-spark-1.3" {
+	// muse-spark-1.3 standard and contributor expose the max effort tier;
+	// older revisions keep the 5-tier ladder.
+	models "muse-spark-1.3" "muse-spark-1.3-contributor" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
 		thinking-mode "effort"
 	}
-	models "muse-spark-1.1" "muse-spark-1.2" "muse-spark-1.2-contributor" "muse-spark-1.3-contributor" {
+	models "muse-spark-1.1" "muse-spark-1.2" "muse-spark-1.2-contributor" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		thinking-mode "effort"
 	}

package/src/compat/rules/classes/openai.kdl

@@ -1,9 +1,15 @@
 // Model-lineage compat for "openai"; regenerated from the frozen census using class taxonomy selectors.
 
 class "openai" {
-	// Replaces the o-series mandatory-reasoning fallback.
+	// Replaces the o-series mandatory-reasoning fallback. o-series and GPT-5+
+	// reject explicit sampling params (`temperature`, `top_p`) with a 400 on
+	// every serving host (#5606), whichever API fronts them.
 	family "o-series" {
 		thinking-requires-effort #true
+		supports-sampling-params #false
+	}
+	revision ">=5" {
+		supports-sampling-params #false
 	}
 	// GPT wire ladders are model generations, not host contracts: <5.2 keeps
 	// the classic four-tier scale, 5.2+ drops minimal for xhigh, and 5.6+
@@ -81,3 +87,105 @@ class "openai" {
 		}
 	}
 }
+
+// GPT-5.2+ bills images in 32px patches × 1.2, sized per `detail` level:
+// https://developers.openai.com/api/docs/guides/images-vision#calculating-costs
+// Older lines carry no lineage rule and bill at their wire API's fallback
+// (providers/image-tokenization.kdl).
+class "openai" {
+	// gpt-5.6 sizing, shared by later lines. `<10` keeps separator-collapsed
+	// revisions (`openai-gpt-52-codex` parses as 52) off the 5.6+ sizing.
+	revision ">=5.6 <10" {
+		image-tokenization {
+			regime "openai-patch"
+			multiplier 1.2
+			auto "original"
+			low {
+				max-edge 512
+			}
+			high {
+				max-edge 2048
+				patch-budget 2500
+			}
+			original {
+				max-edge 65535
+			}
+		}
+	}
+	revision "=5.5" {
+		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
+			}
+		}
+	}
+	// gpt-5.4: `auto` sizes like `high`, and `low` keeps a 6,144-patch budget.
+	revision "=5.4" {
+		image-tokenization {
+			regime "openai-patch"
+			multiplier 1.2
+			auto "high"
+			low {
+				max-edge 2048
+				patch-budget 6144
+			}
+			high {
+				max-edge 2048
+				patch-budget 2500
+			}
+			original {
+				max-edge 6000
+				patch-budget 10000
+			}
+		}
+	}
+	// gpt-5.2 (and the 5.3 Codex SKU) size every level the same way.
+	revision ">=5.2 <5.4" {
+		image-tokenization {
+			regime "openai-patch"
+			multiplier 1.2
+			auto "high"
+			low {
+				max-edge 2048
+				patch-budget 6144
+			}
+			high {
+				max-edge 2048
+				patch-budget 6144
+			}
+			original {
+				max-edge 2048
+				patch-budget 6144
+			}
+		}
+	}
+	// residue: astra's limits differ from the rest of its revision.
+	models "*gpt-6-astra*" {
+		image-tokenization {
+			regime "openai-patch"
+			multiplier 1.2
+			auto "original"
+			low {
+				max-edge 512
+			}
+			high {
+				max-edge 65535
+				patch-budget 2500
+			}
+			original {
+				max-edge 65535
+			}
+		}
+	}
+}

package/src/compat/rules/providers/amazon-bedrock.kdl

@@ -36,8 +36,6 @@ provider "amazon-bedrock" {
 	}
 	class "openai" {
 		thinking-mode "effort"
-		// OpenAI models reject Converse image blocks nested inside toolResult content.
-		requires-tool-result-image-hoisting #true
 	}
 	class "deepseek" {
 		requires-reasoning-content-for-all-assistant-turns #true
@@ -51,11 +49,6 @@ provider "amazon-bedrock" {
 	class "unknown" {
 		thinking-efforts "minimal" "low" "medium" "high"
 		thinking-mode "budget"
-		// Opaque application-inference-profile ARNs carry no vendor signal, so an
-		// OpenAI-backed profile classifies here and would otherwise nest images in
-		// toolResult content and hit the same rejection. Hoisting is accepted by
-		// every Converse model measured, so it is the safe default for unknowns.
-		requires-tool-result-image-hoisting #true
 	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
 	models "moonshot.kimi-k2-thinking" {

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

@@ -28,12 +28,14 @@ provider "cursor" {
 		}
 	}
 	// GetUsableModels advertises no input modalities or context windows.
-	// K3 (including the bare `k3` alias Cursor serves) is natively
-	// multimodal and 1M-context.
+	// Reviewed full-context K3 aliases are multimodal and 1M-context.
+	// K3 family identity alone does not verify other Cursor SKUs such as k3-256k.
 	class "kimi" {
 		family "k3" {
-			input-modalities "text" "image"
-			context-window-floor 1000000
+			models "*kimi-k3*" "k3" "*/k3" {
+				input-modalities "text" "image"
+				context-window-floor 1000000
+			}
 		}
 		family "k2.7-code" {
 			// Cursor advertised Kimi K2.7 Code Context window 262k.
@@ -56,10 +58,6 @@ provider "cursor" {
 			context-window-floor 272000
 		}
 	}
-	models "k3" "*/k3" {
-		input-modalities "text" "image"
-		context-window-floor 1000000
-	}
 	models "default" {
 		context-window-floor 256000
 	}

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/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/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
 		}