@linxiraos/pi-catalog 1.1.27 → 1.2.0

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@linxiraos/pi-catalog",
-	"version": "1.1.27",
+	"version": "1.2.0",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"keywords": [
 		"ai",
@@ -90,12 +90,12 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@linxiraos/pi-omptype": "1.1.27",
-		"@linxiraos/pi-utils": "1.1.27"
+		"@linxiraos/pi-omptype": "1.2.0",
+		"@linxiraos/pi-utils": "1.2.0"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@linxiraos/pi-ai": "1.1.27",
+		"@linxiraos/pi-ai": "1.2.0",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/build.ts

@@ -101,9 +101,11 @@ function applyCatalogAssignments<TApi extends Api>(model: Model<TApi>, catalog:
 	if (serviceTierCost !== undefined) {
 		const flex = numberField(serviceTierCost, "flex");
 		const priorityTier = numberField(serviceTierCost, "priority");
+		const ultrafast = numberField(serviceTierCost, "ultrafast");
 		model.serviceTierCost = {
 			...(flex !== undefined && { flex }),
 			...(priorityTier !== undefined && { priority: priorityTier }),
+			...(ultrafast !== undefined && { ultrafast }),
 		};
 	}
 	const priority = catalog.priority;
@@ -174,40 +176,6 @@ export function applyCatalogCorrections(
 	model: Pick<ModelSpec<Api>, "cost" | "contextWindow" | "maxTokens" | "input">,
 	catalog: Record<string, unknown>,
 ): void {
-	const longContext = objectPayload(catalog.longContext);
-	if (longContext !== undefined) {
-		const inputThreshold = numberField(longContext, "inputThreshold");
-		const inclusive = Reflect.get(longContext, "inputThresholdInclusive") === true;
-		const multiplier = numberField(longContext, "multiplier");
-		const input = numberField(longContext, "input");
-		const output = numberField(longContext, "output");
-		const cacheRead = numberField(longContext, "cacheRead");
-		const cacheWrite = numberField(longContext, "cacheWrite");
-		const base = model.cost;
-		const hasTokenPrice = base.input !== 0 || base.output !== 0 || base.cacheRead !== 0 || base.cacheWrite !== 0;
-		if (inputThreshold !== undefined && multiplier !== undefined && hasTokenPrice) {
-			// Multiplier form: tier rates derive from the row's live list price.
-			model.cost = {
-				...base,
-				longContext: {
-					inputThreshold,
-					...(inclusive && { inputThresholdInclusive: true }),
-					input: base.input * multiplier,
-					output: base.output * multiplier,
-					cacheRead: base.cacheRead * multiplier,
-					cacheWrite: base.cacheWrite * multiplier,
-				},
-			};
-		} else if (
-			inputThreshold !== undefined &&
-			input !== undefined &&
-			output !== undefined &&
-			cacheRead !== undefined &&
-			cacheWrite !== undefined
-		) {
-			model.cost = { ...model.cost, longContext: { inputThreshold, input, output, cacheRead, cacheWrite } };
-		}
-	}
 	const patch = objectPayload(catalog.costPatch);
 	if (patch !== undefined) {
 		model.cost = { ...model.cost };
@@ -244,6 +212,41 @@ export function applyCatalogCorrections(
 			applyEffectiveFallbackRates(model.cost, Reflect.get(fallback, "effectiveRates"));
 		}
 	}
+	const longContext = objectPayload(catalog.longContext);
+	if (longContext !== undefined) {
+		const inputThreshold = numberField(longContext, "inputThreshold");
+		const inclusive = Reflect.get(longContext, "inputThresholdInclusive") === true;
+		const multiplier = numberField(longContext, "multiplier");
+		const input = numberField(longContext, "input");
+		const output = numberField(longContext, "output");
+		const cacheRead = numberField(longContext, "cacheRead");
+		const cacheWrite = numberField(longContext, "cacheWrite");
+		const base = model.cost;
+		const hasTokenPrice = base.input !== 0 || base.output !== 0 || base.cacheRead !== 0 || base.cacheWrite !== 0;
+		if (inputThreshold !== undefined && multiplier !== undefined && hasTokenPrice) {
+			// Multiplier form: tier rates derive from the row's final base card,
+			// after `cost-patch`/`cost-fallback` corrected the live list price.
+			model.cost = {
+				...base,
+				longContext: {
+					inputThreshold,
+					...(inclusive && { inputThresholdInclusive: true }),
+					input: base.input * multiplier,
+					output: base.output * multiplier,
+					cacheRead: base.cacheRead * multiplier,
+					cacheWrite: base.cacheWrite * multiplier,
+				},
+			};
+		} else if (
+			inputThreshold !== undefined &&
+			input !== undefined &&
+			output !== undefined &&
+			cacheRead !== undefined &&
+			cacheWrite !== undefined
+		) {
+			model.cost = { ...model.cost, longContext: { inputThreshold, input, output, cacheRead, cacheWrite } };
+		}
+	}
 	if (catalog.cacheReadAtInputRate === true) {
 		const { longContext } = model.cost;
 		model.cost = {
@@ -363,7 +366,8 @@ export function buildModel<TApi extends Api>(spec: ModelSpec<TApi>): Model<TApi>
 	};
 	applyCatalogAssignments(model, policy.catalog);
 	applyCatalogCorrections(model, policy.catalog);
-	// Configured lifetimes replace catalog lifetimes rather than merging with them.
+	// Configured kinds and lifetimes replace catalog values rather than merging with them.
+	if (spec.kindConfig !== undefined) model.kind = spec.kindConfig;
 	if (spec.promptCacheConfig !== undefined) model.promptCache = { ...spec.promptCacheConfig };
 	return model;
 }

package/src/compat/auth-ids.ts

@@ -17,7 +17,9 @@ export type AuthProviderId =
 	| "charm-hyper"
 	| "cline-pass"
 	| "cloudflare-ai-gateway"
+	| "cloudflare-workers-ai"
 	| "commandcode"
+	| "coralbricks"
 	| "coreweave"
 	| "cursor"
 	| "deepinfra"
@@ -109,7 +111,9 @@ export type LoginProviderId =
 	| "charm-hyper"
 	| "cline-pass"
 	| "cloudflare-ai-gateway"
+	| "cloudflare-workers-ai"
 	| "commandcode"
+	| "coralbricks"
 	| "coreweave"
 	| "cursor"
 	| "deepinfra"

package/src/compat/axes.ts

@@ -140,6 +140,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"qwen-preserve-thinking": wire("qwenPreserveThinking", ["openai"]),
 	"reject-root-object-union": wire("rejectRootObjectUnion", OAI),
 	"retry-without-strict-on-grammar-error": wire("retryWithoutStrictOnGrammarError", OAI),
+	"session-header": wire("sessionHeader", OAI, "scalar", ["x-litellm-session-id"]),
 	"reasoning-content-field": wire("reasoningContentField", OAI, "scalar", [
 		"reasoning_content",
 		"reasoning",
@@ -169,6 +170,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"requires-thinking-as-text": wire("requiresThinkingAsText", ["openai"]),
 	"requires-tool-result-name": wire("requiresToolResultName", ["openai"]),
 	"strict-responses-pairing": wire("strictResponsesPairing", ["openai-responses"]),
+	"connection-bound-native-history": wire("connectionBoundNativeHistory", ["openai-responses"]),
 	"stateful-responses": wire("statefulResponses", ["openai-responses"]),
 	"requires-reasoning-off-juice-instruction": wire("requiresReasoningOffJuiceInstruction", ["openai-responses"]),
 	"supports-all-turns-reasoning-context": wire("supportsAllTurnsReasoningContext", ["openai-responses"]),
@@ -229,12 +231,20 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"allow-anthropic-header-overrides": wire("allowAnthropicHeaderOverrides", ["anthropic"]),
 	"disable-adaptive-thinking": wire("disableAdaptiveThinking", ["anthropic"]),
 	"disable-strict-tools": wire("disableStrictTools", ["anthropic"]),
-	"disabled-thinking": wire("disabledThinking", ["anthropic"], "scalar", ["omit", "disabled", "adaptive"]),
+	"between-tools-effort": wire("betweenToolsEffort", ["anthropic"], "scalar", ["low", "medium", "high"]),
+	"disabled-thinking": wire("disabledThinking", ["anthropic", "bedrock"], "scalar", [
+		"omit",
+		"disabled",
+		"adaptive",
+		"between-tools",
+	]),
 	"effort-beta": wire("effortBeta", ["anthropic"]),
 	"escape-builtin-tool-names": wire("escapeBuiltinToolNames", ["anthropic"]),
 	"fast-mode": wire("fastMode", ["anthropic"]),
 	"first-party-provider": wire("firstPartyProvider", ["anthropic"]),
 	"inject-claude-code-instruction": wire("injectClaudeCodeInstruction", ["anthropic"]),
+	"max-image-dimension": wire("maxImageDimension", ["anthropic"]),
+	"max-image-payload-bytes": wire("maxImagePayloadBytes", ["anthropic"]),
 	"official-endpoint": wire("officialEndpoint", ["anthropic", "openai-responses"]),
 	"replay-unsigned-thinking": wire("replayUnsignedThinking", ["anthropic"]),
 	"requires-thinking-enabled": wire("requiresThinkingEnabled", ["anthropic"]),
@@ -250,7 +260,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"supports-per-message-effort": wire("supportsPerMessageEffort", ["anthropic"]),
 	"supports-server-compaction": wire("supportsServerCompaction", ["anthropic"]),
 	"supports-between-tools-thinking": wire("supportsBetweenToolsThinking", ["anthropic"]),
-	"supports-thinking-binding-controls": wire("supportsThinkingBindingControls", ["anthropic"]),
+	"supports-thinking-binding-controls": wire("supportsThinkingBindingControls", ["anthropic", "bedrock"]),
 	"supports-turn-scoped-system": wire("supportsTurnScopedSystem", ["anthropic"]),
 
 	// ── wire: bedrock-converse-stream ──
@@ -261,7 +271,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	// ── wire: devin-agent ──
 	"model-router": wire("modelRouter", ["devin"]),
 	"supports-parallel-tool-calls": wire("supportsParallelToolCalls", ["devin"]),
-	"trust-explicit-thinking-only": wire("trustExplicitThinkingOnly", ["devin"]),
+	"trust-explicit-thinking-only": wire("trustExplicitThinkingOnly", [...OAI, "devin"]),
 
 	// ── wire: google APIs ──
 	"antigravity-claude-tool-mode": wire("antigravityClaudeToolMode", ["google"]),
@@ -395,11 +405,35 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	/** How the model line bills an input image; shape and formulas in `./image-tokenization`. */
 	"image-tokenization": { key: "imageTokenization", set: "catalog", shape: "object" },
 	"image-model": { key: "imageModel", set: "catalog", shape: "scalar" },
+	"inline-image-byte-budget": { key: "inlineImageByteBudget", set: "catalog", shape: "scalar" },
+	/**
+	 * Most questions one System One judgment request may carry; the judge
+	 * splits larger requests. Unassigned: the endpoint takes any batch size.
+	 */
+	"judgment-questions-max": { key: "judgmentQuestionsMax", set: "catalog", shape: "scalar" },
 	"limits-patch": { key: "limitsPatch", set: "catalog", shape: "object" },
 	"long-context-cost": { key: "longContext", set: "catalog", shape: "object" },
 	"prompt-cache": { key: "promptCache", set: "catalog", shape: "object" },
+	/**
+	 * Prompt-cache lookback in block positions: how far back from a cache
+	 * breakpoint the provider looks for an earlier request's cache entry.
+	 * Unassigned: no known lookback bound.
+	 */
+	"prompt-cache-lookback": { key: "promptCacheLookback", set: "catalog", shape: "scalar" },
 	"long-usage-limit-fallback": { key: "longUsageLimitFallback", set: "catalog", shape: "scalar" },
 	"max-context-window": { key: "maxContextWindow", set: "catalog", shape: "scalar" },
+	/**
+	 * Every host of the model returns its full plaintext trace wherever the
+	 * wire does not mark reasoning as a summary (`ThinkingContent.summary ===
+	 * false` vouches for nothing more), and its identity (class + family +
+	 * revision) pins one set of weights, so a turn's reasoning may replay
+	 * natively on another host serving the same model. Marked summaries never
+	 * carry; do not declare it where a host serves summaries unmarked.
+	 * Declared per family in `classes/*.kdl`; a host whose reasoning must not be
+	 * carried in or out sets `#false` in its `providers/*.kdl`. Read through
+	 * `compat/reasoning-carry.ts`.
+	 */
+	"portable-reasoning": { key: "portableReasoning", set: "catalog", shape: "scalar", values: [true, false] },
 	"pricing-status": {
 		key: "pricingStatus",
 		set: "catalog",

package/src/compat/cascade.ts

@@ -288,7 +288,12 @@ function collect(winners: WinnerTable, pick: (rule: CompiledRule) => Record<stri
 	return out;
 }
 
-const resolveCache = new LRUCache<string, ResolvedAxes>({ max: 512 });
+/**
+ * Sized above the bundled catalog (~5.6k targets) plus discovered rows so a
+ * catalog-wide buildModel pass (discovery, then merge rebuilding the same ids)
+ * hits instead of thrashing. Entries share rule values: a few hundred bytes each.
+ */
+const resolveCache = new LRUCache<string, ResolvedAxes>({ max: 16384 });
 
 function keyPart(value: string | undefined): string {
 	return value === undefined ? "-1:" : `${value.length}:${value}`;

package/src/compat/judgment-questions.ts

@@ -0,0 +1,12 @@
+import type { Model } from "../types";
+import { resolveCatalogPolicy } from "./catalog-policy";
+
+/**
+ * Most questions one System One judgment request to this model may carry
+ * (`judgment-questions-max` axis), or `undefined` when the endpoint takes any
+ * batch size.
+ */
+export function resolveJudgmentQuestionsMax(model: Model): number | undefined {
+	const value = resolveCatalogPolicy(model).judgmentQuestionsMax;
+	return typeof value === "number" && Number.isInteger(value) && value > 0 ? value : undefined;
+}

package/src/compat/prompt-cache-lookback.ts

@@ -0,0 +1,23 @@
+import { toModelSpec } from "../provider-models/bundled-references";
+import type { Model } from "../types";
+import { resolveCatalogAxes } from "./resolve";
+
+const lookbackCache = new Map<string, number | null>();
+const LOOKBACK_CACHE_MAX = 8192;
+
+/**
+ * Block positions the provider's prompt cache checks back from a breakpoint for
+ * an earlier request's entry (`prompt-cache-lookback` axis), or `undefined` when
+ * the model has no known lookback bound.
+ */
+export function resolvePromptCacheLookback(model: Model): number | undefined {
+	const key = `${model.provider} ${model.id} ${model.api}`;
+	let lookback = lookbackCache.get(key);
+	if (lookback === undefined) {
+		const value = resolveCatalogAxes(toModelSpec(model)).promptCacheLookback;
+		lookback = typeof value === "number" && Number.isInteger(value) && value > 0 ? value : null;
+		if (lookbackCache.size >= LOOKBACK_CACHE_MAX) lookbackCache.clear();
+		lookbackCache.set(key, lookback);
+	}
+	return lookback ?? undefined;
+}

package/src/compat/provider-ids.d.ts

@@ -1,2 +1,2 @@
 /** Every chat-model provider with a catalog entry (`default-model` in its provider rules). */
-export type KnownProvider = "abliteration" | "aiand" | "aimlapi" | "alibaba-coding-plan" | "alibaba-token-plan" | "amazon-bedrock" | "anthropic" | "apple" | "azure" | "baseten" | "bedrock-mantle" | "cerebras" | "charm-hyper" | "cline-pass" | "cloudflare-ai-gateway" | "commandcode" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "factory-droid" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google" | "google-antigravity" | "google-gemini-cli" | "google-vertex" | "groq" | "helmcode" | "huggingface" | "kilo" | "kimi-code" | "litellm" | "lm-studio" | "local" | "meta" | "minimax" | "minimax-code" | "minimax-code-cn" | "mistral" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai" | "openai-codex" | "opencode-go" | "opencode-zen" | "openrouter" | "qianfan" | "qwen-portal" | "sakana" | "siliconflow" | "siliconflow-cn" | "singularityapi-dev" | "singularityapi-tech" | "snowflake" | "stepfun" | "synthetic" | "together" | "typesafe" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "web" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zenmux" | "zhipu-coding-plan";
+export type KnownProvider = "abliteration" | "aiand" | "aimlapi" | "alibaba-coding-plan" | "alibaba-token-plan" | "amazon-bedrock" | "anthropic" | "apple" | "azure" | "baseten" | "bedrock-mantle" | "cerebras" | "charm-hyper" | "cline-pass" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "commandcode" | "coralbricks" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "factory-droid" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google" | "google-antigravity" | "google-gemini-cli" | "google-vertex" | "groq" | "helmcode" | "huggingface" | "kilo" | "kimi-code" | "litellm" | "lm-studio" | "local" | "meta" | "minimax" | "minimax-code" | "minimax-code-cn" | "mistral" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai" | "openai-codex" | "opencode-go" | "opencode-zen" | "openrouter" | "qianfan" | "qwen-portal" | "sakana" | "siliconflow" | "siliconflow-cn" | "singularityapi-dev" | "singularityapi-tech" | "snowflake" | "stepfun" | "synthetic" | "together" | "typesafe" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "web" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zenmux" | "zhipu-coding-plan";

package/src/compat/provider-ids.ts

@@ -17,7 +17,9 @@ export type KnownProvider =
 	| "charm-hyper"
 	| "cline-pass"
 	| "cloudflare-ai-gateway"
+	| "cloudflare-workers-ai"
 	| "commandcode"
+	| "coralbricks"
 	| "coreweave"
 	| "cursor"
 	| "deepinfra"

package/src/compat/reasoning-carry.ts

@@ -0,0 +1,64 @@
+/**
+ * Reasoning carry: whether reasoning a model produced on one host may replay
+ * natively on another host serving the same model.
+ *
+ * The model-side half is declared with the `portable-reasoning` catalog axis
+ * per family in `rules/classes/*.kdl`: the family's stored reasoning is its own
+ * full plaintext trace, and its identity pins one set of weights. A host whose
+ * reasoning must neither be carried in nor out overrides it to `#false` in its
+ * `rules/providers/*.kdl`. Whether the target host has a native slot for
+ * earlier reasoning is derived by the request encoders in `@linxiraos/pi-ai`,
+ * not declared here.
+ *
+ * Resolved through the cascade on demand (like delegation bias) because
+ * bundled rows are frozen by the generator; results are memoized per
+ * deployment.
+ */
+import { resolveCascade } from "./cascade";
+import { classifyModel } from "./taxonomy";
+import type { ModelIdentity } from "./types";
+
+/** `null` records a deployment that does not carry reasoning. */
+const portableIdentityMemo = new Map<string, ModelIdentity | null>();
+
+/**
+ * The identity a deployment's reasoning is carried under, or `undefined` when
+ * the deployment's model family is not declared portable there. Class-only
+ * identities never qualify: they name a lineage, not a set of weights.
+ */
+export function portableReasoningIdentity(provider: string, api: string, modelId: string): ModelIdentity | undefined {
+	const key = `${provider}\0${api}\0${modelId}`;
+	const memo = portableIdentityMemo.get(key);
+	if (memo !== undefined) return memo ?? undefined;
+	const identity = classifyModel(provider, modelId, { lenient: true });
+	let portable: ModelIdentity | null = null;
+	if (identity.family !== undefined || identity.revision !== undefined) {
+		const declared = resolveCascade({
+			provider,
+			api,
+			class: identity.class,
+			model: modelId,
+			reasoning: true,
+			...(identity.family !== undefined && { family: identity.family }),
+			...(identity.revision !== undefined && { revision: identity.revision }),
+		}).catalog.portableReasoning;
+		if (declared === true) portable = identity;
+	}
+	portableIdentityMemo.set(key, portable);
+	return portable ?? undefined;
+}
+
+/**
+ * Whether reasoning recorded on `source` may replay natively on `target`:
+ * both deployments declare the family portable and classify to the same
+ * class, family, and revision.
+ */
+export function carriesReasoning(
+	source: { provider: string; api: string; model: string },
+	target: { provider: string; api: string; id: string },
+): boolean {
+	const to = portableReasoningIdentity(target.provider, target.api, target.id);
+	if (!to) return false;
+	const from = portableReasoningIdentity(source.provider, source.api, source.model);
+	return from !== undefined && from.class === to.class && from.family === to.family && from.revision === to.revision;
+}

package/src/compat/request-size.ts

@@ -0,0 +1,25 @@
+import { toModelSpec } from "../provider-models/bundled-references";
+import type { Model } from "../types";
+import { resolveCatalogAxes } from "./resolve";
+
+const imageBudgetCache = new Map<string, number | null>();
+const IMAGE_BUDGET_CACHE_MAX = 8192;
+
+/**
+ * Maximum encoded inline-image bytes retained per request for a deployment with a documented body-size limit.
+ * The axis is a host contract: a rerouted `baseUrl` (proxy, gateway) imposes its own limit, so the budget
+ * applies only while the model targets its official endpoint.
+ */
+export function resolveInlineImageByteBudget(model: Model): number | undefined {
+	const compat = model.compat;
+	if (!compat || !("officialEndpoint" in compat) || !compat.officialEndpoint) return undefined;
+	const key = `${model.provider} ${model.id} ${model.api}`;
+	let budget = imageBudgetCache.get(key);
+	if (budget === undefined) {
+		const value = resolveCatalogAxes(toModelSpec(model)).inlineImageByteBudget;
+		budget = typeof value === "number" && Number.isFinite(value) && value > 0 ? value : null;
+		if (imageBudgetCache.size >= IMAGE_BUDGET_CACHE_MAX) imageBudgetCache.clear();
+		imageBudgetCache.set(key, budget);
+	}
+	return budget ?? undefined;
+}

package/src/compat/resolve.ts

@@ -107,11 +107,12 @@ class IdentityFacts {
 		return this.is("kimi") && this.family("k2.7-code", "k3");
 	}
 
-	/** Adaptive-thinking Claude generation floor (Opus ≥ min; Sonnet/Fable/Mythos ≥ 5). */
+	/** Adaptive-thinking Claude generation floor (Opus ≥ min; Sonnet/Fable/Mythos ≥ 5; Haiku ≥ 5.5). */
 	anthropicAdaptiveGenAtLeast(opusMin: string): boolean {
 		if (!this.is("anthropic")) return false;
 		if (this.family("opus")) return this.revGte(opusMin);
 		if (this.family("sonnet", "fable", "mythos")) return this.revGte("5");
+		if (this.family("haiku")) return this.revGte("5.5");
 		return false;
 	}
 }
@@ -502,6 +503,7 @@ function detectOpenAICompat(
 		supportsDeveloperRole: isOpenAIHost || isAzureHost,
 		supportsMultipleSystemMessages: supportsMultipleSystemMessagesDefault,
 		supportsReasoningEffort: !isGrok && !d.isXiaomiMimo && (!(d.isZai || d.isZhipu) || supportsZaiReasoningEffort),
+		trustExplicitThinkingOnly: undefined,
 		// API-conditional: this completions-only Copilot exclusion cannot be a
 		// provider rule without changing Copilot Responses rows.
 		supportsReasoningParams: provider !== "github-copilot",
@@ -731,6 +733,7 @@ function resolveOpenAIResponsesPolicy(
 			hostMatchesUrl(baseUrl, "openrouter") ||
 			hostMatchesUrl(baseUrl, "deepseekFamily"),
 		supportsReasoningEffort: !isXaiHost,
+		trustExplicitThinkingOnly: undefined,
 		supportsLongPromptCacheRetention: isOpenAIUrl,
 		supportsPromptCacheBreakpoints,
 		promptCacheBreakpointTtl: supportsPromptCacheBreakpoints ? "30m" : undefined,
@@ -792,6 +795,7 @@ function resolveOpenAIResponsesPolicy(
 			(LOCAL_OPENAI_COMPAT_PROVIDERS[backendProvider] === true || hasLocalLoopbackBaseUrl(baseUrl)),
 		supportsObfuscationOptOut: isOpenAIUrl || provider === "openai",
 		storeResponses: false,
+		connectionBoundNativeHistory: false,
 		officialEndpoint: isOfficialOpenAIEndpoint(provider, baseUrl),
 		harmonyLeakMitigation: false,
 		rejectRootObjectUnion: false,
@@ -854,6 +858,7 @@ function pickResponsesOnly(compat: ResolvedOpenAIResponsesCompat): ResponsesOnly
 		supportsImageDetailOriginal: compat.supportsImageDetailOriginal,
 		supportsObfuscationOptOut: compat.supportsObfuscationOptOut,
 		storeResponses: compat.storeResponses,
+		connectionBoundNativeHistory: compat.connectionBoundNativeHistory,
 		supportsAllTurnsReasoningContext: compat.supportsAllTurnsReasoningContext,
 		supportsConfigurationUpdate: compat.supportsConfigurationUpdate,
 		supportsSteering: compat.supportsSteering,
@@ -910,6 +915,11 @@ function resolveAnthropicPolicy(
 		escapeBuiltinToolNames: false,
 		injectClaudeCodeInstruction: true,
 		stripImageInput: false,
+		// Unset means "canonical Anthropic image contract"; the request builder
+		// owns those numbers. The keys must exist for a KDL rule or a
+		// user-authored `compat` block to override them.
+		maxImageDimension: undefined,
+		maxImagePayloadBytes: undefined,
 		thinkingLoopGuard: undefined,
 		// Present as keys so models.yml `compat` can set them; unset unless a rule assigns them.
 		stripThinkingHistory: undefined,
@@ -932,6 +942,9 @@ function resolveBedrockPolicy(spec: ModelSpec<"bedrock-converse-stream">, axes:
 		promptCacheMinimumTokens: 0,
 		promptCacheMaximumCheckpoints: 0,
 		supportsForcedToolChoice: true,
+		// Opt-out: every prefix-bound Claude line on Converse takes the controls,
+		// including custom Bedrock-wire providers; KDL or models.yml sets `false`.
+		supportsThinkingBindingControls: true,
 	};
 	// Reasoning capability is a mechanism gate; adaptive-lineage duration is rule-owned.
 	compat.streamIdleTimeoutMs = compatReasoning(spec, axes) ? BEDROCK_REASONING_STREAM_IDLE_TIMEOUT_MS : undefined;
@@ -1009,7 +1022,8 @@ function defaultThinkingMode<TApi extends Api>(spec: ModelSpec<TApi>, facts: Ide
 				return "anthropic-budget-effort";
 			}
 			if (facts.is("anthropic")) {
-				if (facts.revGte("4.6") && !facts.family("haiku")) return "anthropic-adaptive";
+				// Haiku stays on budget thinking until 5.5, its first adaptive generation.
+				if (facts.revGte(facts.family("haiku") ? "5.5" : "4.6")) return "anthropic-adaptive";
 				if (facts.family("opus") && facts.revGte("4.5")) return "anthropic-budget-effort";
 			}
 			return "budget";

package/src/compat/rules/README.md

@@ -213,10 +213,16 @@ The three value shapes are:
 - **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.
-One object axis carries a computed form: `long-context-cost` accepts either the absolute rates (`input-threshold` + `input`/`output`/`cache-read`/`cache-write`) or `input-threshold` + `multiplier` (with optional `input-threshold-inclusive`), which derives the tier from the row's live base price at build time so the rule tracks upstream list-price updates (xAI's SuperGrok 200K tier). Rows without a token price carry no tier.
+One object axis carries a computed form: `long-context-cost` accepts either the absolute rates (`input-threshold` + `input`/`output`/`cache-read`/`cache-write`) or `input-threshold` + `multiplier` (with optional `input-threshold-inclusive`), which derives the tier from the row's base price at build time, after `cost-patch` and `cost-fallback`, so the rule tracks upstream list-price updates (xAI's SuperGrok 200K tier). Rows without a token price carry no tier. An empty `long-context-cost {}` adds no tier; a more specific rule uses it to opt a host out of a broader rule's band (GitHub Copilot's Haiku 5.5 rows, whose long tier is the `-1m` sibling).
 
 `context-window-authoritative #true` preserves a host's supplied context window through runtime model selection instead of applying inferred expansion or reference-price-tier caps. Explicit user context overrides still apply afterward. It applies to rows materialized through `buildModel` (discovery and regenerated bundles). See [`providers/factory-droid.kdl`](providers/factory-droid.kdl).
 
+`inline-image-byte-budget` caps retained base64 image characters for a deployment's request body. It leaves room for tool schemas and text beneath the provider's body-size limit; the coding agent reads it through `resolveInlineImageByteBudget` before sending live history. The budget applies only while `compat.officialEndpoint` holds: a custom `baseUrl` imposes its own body limit.
+
+`judgment-questions-max` is the most questions one System One judgment request may carry (razorback16/openjev: 256; Cloudflare Clef on Workers AI and through OpenRouter Decisions: 64; unidentified `*openjev*` deployments: 16). The coding agent reads it through `resolveJudgmentQuestionsMax` and `NativeJudge` splits larger requests into parallel chunks; unassigned means the endpoint takes any batch size. [`providers/judgment-questions.kdl`](providers/judgment-questions.kdl) lists the servers verified uncapped or not isolatable.
+
+`prompt-cache-lookback` is how many block positions back from a cache breakpoint the model's prompt cache looks for an earlier request's entry (Claude: 20, counting the breakpoint; a run of consecutive `tool_use` or `tool_result` blocks is one position). The coding agent reads it through `resolvePromptCacheLookback` and keeps warm-cache tool-result pruning inside the window; unassigned means no known bound.
+
 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`.
 
 ### Time-based pricing
@@ -468,11 +474,15 @@ provider "muse-code" {
 
 Only `discovery` enrolls a provider in `generate-models.ts`; providers without it are never fetched at generation time (see the `charm-hyper` entry for why a live gateway deliberately omits it).
 
+`automatic-default #false` keeps a provider available for explicit selection but excludes it from startup fallback and automatic model presets. The default is `#true`; `apple` opts out because its on-device context cannot accommodate the standard coding-agent prompt in many projects.
+
+`kind-apis { <kind> "<api>" }` maps each non-chat kind (`image`, `tts`, `stt`, `embedding`, `rerank`, `video`) to the API discovery assigns rows of that kind. A runner API must sit under the kind it serves (`RUNNER_API_KINDS` in `src/types.ts`); chat APIs, which serve hosted image generation, and multi-kind `local-inference` may back any kind.
+
 ### Seed rows
 
 A `seed` _defines_ bundled rows for providers whose catalog cannot be discovered at generation time — credential-scoped rosters, unauthenticated regens, or models ahead of upstream catalogs. Every other stratum patches rows; this one authors them. Runtime model managers hand the rows to `staticModels` through `seedModels(provider)`; the generator bundles them per the seed's `bundle` policy. Values are literal — a seed never derives from another provider's row, and pricing is never borrowed.
 
-`seed` properties: `api` and `base-url` are per-row defaults (a `model` may override either with the same property names); `bundle` defaults to `always`; `precedence="seed"` is optional. `model` takes the wire id positionally, requires `name=`, and its body MUST declare `reasoning`, `input` (`"text"` and/or `"image"`), `cost` (all four per-million rates), and `limits` (`context=` / `max-tokens=`, an omitted limit is `null`); `supports-tools #true` is optional. Any other directive is an axis from the cascade vocabulary: thinking axes become the row's explicit `thinking` (then `thinking-mode` and `thinking-efforts` are both required), wire axes become its explicit `compat` and must apply to the row's API, and catalog axes are rejected because they stay rule-owned in the cascade block. Explicit `thinking`/`compat` on a seed row win over the cascade exactly as they do for any authored spec.
+`seed` properties: `api` and `base-url` are per-row defaults (a `model` may override either with the same property names); `bundle` defaults to `always`; `precedence="seed"` is optional. `model` takes the row id positionally, requires `name=`, and its body MUST declare `reasoning`, `input` (`"text"` and/or `"image"`), `cost` (all four per-million rates), and `limits` (`context=` / `max-tokens=`, an omitted limit is `null`); `supports-tools #true` is optional, and `request-model-id "<wire id>"` sets the row's `requestModelId` when the upstream id differs from the row id (an OpenAI Decisions judge row beside the same-id chat model). Any other directive is an axis from the cascade vocabulary: thinking axes become the row's explicit `thinking` (then `thinking-mode` and `thinking-efforts` are both required), wire axes become its explicit `compat` and must apply to the row's API, and catalog axes are rejected because they stay rule-owned in the cascade block. Explicit `thinking`/`compat` on a seed row win over the cascade exactly as they do for any authored spec.
 
 `bundle` decides when the generator includes the rows:
 

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" "snowflake" "litellm" "kilo" "zenmux" "opencode-zen" "opencode-go" \
+	"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"
+	"gmi-cloud" "stencil" "singularityapi-dev" "singularityapi-tech" "coralbricks"

package/src/compat/rules/auth/cloudflare-workers-ai.kdl

@@ -0,0 +1,9 @@
+// Cloudflare Workers AI runs the Clef decision models under
+// `/client/v4/accounts/{account_id}/ai/run`. Login stores the API token with
+// the account id (the shared Cloudflare credential); an env-supplied token
+// takes its account from `CLOUDFLARE_ACCOUNT_ID`.
+auth "cloudflare-workers-ai" {
+	name "Cloudflare Workers AI"
+	env "CLOUDFLARE_API_TOKEN" "CLOUDFLARE_AUTH_TOKEN"
+	login "custom" hook="cloudflare-workers-ai"
+}

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

@@ -0,0 +1,13 @@
+// `/v1/models` is key-protected — it answers `401 missing_api_key` without a
+// bearer key and `401 invalid_api_key` to a rejected one (verified
+// 2026-10-02) — so login validates the pasted key there. The probe bills
+// nothing, unlike a chat-completions ping against a prepaid balance.
+auth "coralbricks" {
+	name "CoralBricks"
+	login "api-key" {
+		auth-url "https://www.coralbricks.ai/api-keys"
+		instructions "Create or copy your API key from the CoralBricks dashboard (keys look like cb_...)"
+		prompt "Paste your CoralBricks API key" placeholder="cb_..."
+		validate "models-endpoint" label="CoralBricks" url="https://inference.coralbricks.ai/v1/models"
+	}
+}

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

@@ -4,6 +4,6 @@ auth "minimax-code-cn" {
 		auth-url "https://platform.minimaxi.com/subscribe/token-plan"
 		instructions "Subscribe to Token Plan and copy your API key"
 		prompt "Paste your MiniMax Token Plan API key" placeholder="sk-..."
-		validate "chat-completions" base-url="https://api.minimaxi.com/v1" model="MiniMax-M3"
+		validate "anthropic-messages" base-url="https://api.minimaxi.com/anthropic" model="MiniMax-M3"
 	}
 }

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

@@ -4,6 +4,6 @@ auth "minimax-code" {
 		auth-url "https://platform.minimax.io/subscribe/token-plan"
 		instructions "Subscribe to Token Plan and copy your API key"
 		prompt "Paste your MiniMax Token Plan API key" placeholder="sk-..."
-		validate "chat-completions" label="MiniMax Token Plan" base-url="https://api.minimax.io/v1" model="MiniMax-M3"
+		validate "anthropic-messages" label="MiniMax Token Plan" base-url="https://api.minimax.io/anthropic" model="MiniMax-M3"
 	}
 }

package/src/compat/rules/auth/openai-codex.kdl

@@ -9,7 +9,7 @@ auth "openai-codex" {
 		authorize-params {
 			id_token_add_organizations "true"
 			codex_cli_simplified_flow "true"
-			originator "omp"
+			originator "zeta"
 		}
 		instructions "A browser window should open. Complete login to finish."
 		// OpenAI only allowlists this exact URI; a busy port must fail rather than fall back.

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

@@ -1,5 +1,5 @@
 // stencil.so account. Not a model provider: the credential authorizes
-// omp-hosted Stencil services such as `omp stream` (live.omp.sh).
+// Zeta-hosted Stencil services such as `zetacode stream` (live.omp.sh).
 //
 // Two hosts: the Stencil API at api.stencil.so (`{base}`, STENCIL_BASE_URL)
 // and the OIDC issuer at auth.stencil.so (`{auth}`, STENCIL_AUTH_URL) which