@gajae-code/ai 0.18.11 → 0.19.0

package/src/bedrock-claude-cache-policy.d.ts

@@ -10,7 +10,7 @@ export type BedrockClaudeGeneration = {
  *  - family-first (3.x era): anthropic.claude-3-5-haiku-20241022-v1:0
  *  - kind-first (4+ era):    anthropic.claude-opus-4-20250514-v1:0,
  *                            anthropic.claude-haiku-4-5-20251001-v1:0
- * Cross-region profiles (us./eu./au./jp./apac./global. prefixes) and inference-profile
+ * Cross-region profiles (us./eu./au./jp./apac./global./in. prefixes) and inference-profile
  * ARNs keep the canonical model id as their final path segment.
  */
 export declare function parseBedrockClaudeGeneration(id: string): BedrockClaudeGeneration | undefined;

package/src/bedrock-claude-cache-policy.ts

@@ -20,13 +20,13 @@ function extractBedrockModelId(id: string): string | undefined {
  *  - family-first (3.x era): anthropic.claude-3-5-haiku-20241022-v1:0
  *  - kind-first (4+ era):    anthropic.claude-opus-4-20250514-v1:0,
  *                            anthropic.claude-haiku-4-5-20251001-v1:0
- * Cross-region profiles (us./eu./au./jp./apac./global. prefixes) and inference-profile
+ * Cross-region profiles (us./eu./au./jp./apac./global./in. prefixes) and inference-profile
  * ARNs keep the canonical model id as their final path segment.
  */
 export function parseBedrockClaudeGeneration(id: string): BedrockClaudeGeneration | undefined {
 	const modelId = extractBedrockModelId(id);
 	if (modelId === undefined) return undefined;
-	const prefix = "(?:(?:us|eu|au|jp|apac|global)\\.)?anthropic\\.claude-";
+	const prefix = "(?:(?:us|eu|au|jp|apac|global|in)\\.)?anthropic\\.claude-";
 	const component = "(?:0|[1-9]\\d?)";
 	const suffix = "(?:-(?:[a-z][a-z0-9]*(?::[a-z0-9]+)?|\\d{8}))*";
 	const familyFirst = new RegExp(`^${prefix}([1-9]\\d?)(?:-(${component}))?-([a-z][a-z0-9]*)${suffix}$`).exec(modelId);

package/src/model-cache.ts

@@ -6,7 +6,7 @@ import { Database } from "bun:sqlite";
 import { getModelDbPath } from "@gajae-code/utils/dirs";
 import type { Api, Model } from "./types";
 
-const CACHE_SCHEMA_VERSION = 5;
+const CACHE_SCHEMA_VERSION = 6;
 
 interface CacheRow {
 	provider_id: string;
@@ -16,6 +16,7 @@ interface CacheRow {
 	static_fingerprint: string;
 	dynamic_model_ids: string | null;
 	dynamic_model_provenance: string | null;
+	discovery_user_agent: string | null;
 	models: string;
 }
 
@@ -38,6 +39,8 @@ interface CacheEntry<TApi extends Api = Api> {
 	/** IDs returned by the authoritative dynamic provider catalog, when retained. */
 	dynamicModelIds: string[] | undefined;
 	dynamicModelProvenance: string | undefined;
+	/** User-Agent used during discovery; invalidates cache if it changes. */
+	discoveryUserAgent: string | undefined;
 }
 
 let sharedDb: Database | null = null;
@@ -63,6 +66,7 @@ function getDb(dbPath?: string): Database {
 			static_fingerprint TEXT NOT NULL DEFAULT '',
 			dynamic_model_ids TEXT,
 			dynamic_model_provenance TEXT,
+			discovery_user_agent TEXT,
 			models TEXT NOT NULL
 		)
 	`);
@@ -84,7 +88,13 @@ export function closeModelCache(dbPath?: string): boolean {
 }
 
 function migrateCacheSchema(db: Database): void {
-	const columns = db.prepare("PRAGMA table_info(model_cache)").all() as TableInfoRow[];
+	const tableInfo = db.query("PRAGMA table_info(model_cache)");
+	let columns: TableInfoRow[];
+	try {
+		columns = tableInfo.all() as TableInfoRow[];
+	} finally {
+		tableInfo.finalize();
+	}
 	if (!columns.some(column => column.name === "static_fingerprint")) {
 		db.run("ALTER TABLE model_cache ADD COLUMN static_fingerprint TEXT NOT NULL DEFAULT ''");
 	}
@@ -94,7 +104,10 @@ function migrateCacheSchema(db: Database): void {
 	if (!columns.some(column => column.name === "dynamic_model_provenance")) {
 		db.run("ALTER TABLE model_cache ADD COLUMN dynamic_model_provenance TEXT");
 	}
-	db.run("UPDATE model_cache SET version = ? WHERE version IN (2, 3, 4)", [CACHE_SCHEMA_VERSION]);
+	if (!columns.some(column => column.name === "discovery_user_agent")) {
+		db.run("ALTER TABLE model_cache ADD COLUMN discovery_user_agent TEXT");
+	}
+	db.run("UPDATE model_cache SET version = ? WHERE version IN (2, 3, 4, 5)", [CACHE_SCHEMA_VERSION]);
 }
 
 export function readModelCache<TApi extends Api>(
@@ -102,13 +115,24 @@ export function readModelCache<TApi extends Api>(
 	ttlMs: number,
 	now: () => number,
 	dbPath?: string,
+	expectedDiscoveryUserAgent?: string,
 ): CacheEntry<TApi> | null {
 	try {
 		const db = getDb(dbPath);
-		const row = db.query<CacheRow, [string]>("SELECT * FROM model_cache WHERE provider_id = ?").get(providerId);
+		const select = db.query<CacheRow, [string]>("SELECT * FROM model_cache WHERE provider_id = ?");
+		let row: CacheRow | null;
+		try {
+			row = select.get(providerId);
+		} finally {
+			select.finalize();
+		}
 		if (!row || row.version !== CACHE_SCHEMA_VERSION) {
 			return null;
 		}
+		// Invalidate cache if discovery User-Agent has changed
+		if (expectedDiscoveryUserAgent !== undefined && row.discovery_user_agent !== expectedDiscoveryUserAgent) {
+			return null;
+		}
 		const models = JSON.parse(row.models) as Model<TApi>[];
 		const ageMs = now() - row.updated_at;
 		const fresh = Number.isFinite(ageMs) && ageMs >= 0 && ageMs <= ttlMs;
@@ -120,6 +144,7 @@ export function readModelCache<TApi extends Api>(
 			staticFingerprint: row.static_fingerprint ?? "",
 			dynamicModelIds: row.dynamic_model_ids === null ? undefined : (JSON.parse(row.dynamic_model_ids) as string[]),
 			dynamicModelProvenance: row.dynamic_model_provenance ?? undefined,
+			discoveryUserAgent: row.discovery_user_agent ?? undefined,
 		};
 	} catch {
 		return null;
@@ -135,12 +160,13 @@ export function writeModelCache<TApi extends Api>(
 	dbPath?: string,
 	dynamicModelIds?: readonly string[],
 	dynamicModelProvenance?: string,
+	discoveryUserAgent?: string,
 ): void {
 	try {
 		const db = getDb(dbPath);
 		db.run(
-			`INSERT OR REPLACE INTO model_cache (provider_id, version, updated_at, authoritative, static_fingerprint, dynamic_model_ids, dynamic_model_provenance, models)
-			 VALUES (?, ?, ?, ?, ?, ?, ?, ?)`,
+			`INSERT OR REPLACE INTO model_cache (provider_id, version, updated_at, authoritative, static_fingerprint, dynamic_model_ids, dynamic_model_provenance, discovery_user_agent, models)
+			 VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)`,
 			[
 				providerId,
 				CACHE_SCHEMA_VERSION,
@@ -149,6 +175,7 @@ export function writeModelCache<TApi extends Api>(
 				staticFingerprint,
 				dynamicModelIds === undefined ? null : JSON.stringify(dynamicModelIds),
 				dynamicModelProvenance ?? null,
+				discoveryUserAgent ?? null,
 				JSON.stringify(models),
 			],
 		);
@@ -166,11 +193,12 @@ export function insertModelCacheIfAbsent<TApi extends Api>(
 	dbPath?: string,
 	dynamicModelIds?: readonly string[],
 	dynamicModelProvenance?: string,
+	discoveryUserAgent?: string,
 ): boolean {
 	try {
 		const result = getDb(dbPath).run(
-			`INSERT INTO model_cache (provider_id, version, updated_at, authoritative, static_fingerprint, dynamic_model_ids, dynamic_model_provenance, models)
-			 VALUES (?, ?, ?, ?, ?, ?, ?, ?)
+			`INSERT INTO model_cache (provider_id, version, updated_at, authoritative, static_fingerprint, dynamic_model_ids, dynamic_model_provenance, discovery_user_agent, models)
+			 VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
 			 ON CONFLICT(provider_id) DO NOTHING`,
 			[
 				providerId,
@@ -180,6 +208,7 @@ export function insertModelCacheIfAbsent<TApi extends Api>(
 				staticFingerprint,
 				dynamicModelIds === undefined ? null : JSON.stringify(dynamicModelIds),
 				dynamicModelProvenance ?? null,
+				discoveryUserAgent ?? null,
 				JSON.stringify(models),
 			],
 		);
@@ -195,6 +224,7 @@ export function updateModelCacheIfUnchanged<TApi extends Api>(
 	expectedDynamicModelIds: readonly string[] | undefined,
 	expectedDynamicModelProvenance: string | undefined,
 	expectedModels: readonly Model<TApi>[],
+	expectedDiscoveryUserAgent: string | undefined,
 	updatedAt: number,
 	models: Model<TApi>[],
 	authoritative: boolean,
@@ -202,30 +232,35 @@ export function updateModelCacheIfUnchanged<TApi extends Api>(
 	dbPath?: string,
 	dynamicModelIds?: readonly string[],
 	dynamicModelProvenance?: string,
+	discoveryUserAgent?: string,
 ): boolean {
 	try {
 		const expectedIds = expectedDynamicModelIds === undefined ? null : JSON.stringify(expectedDynamicModelIds);
 		const provenance = expectedDynamicModelProvenance ?? null;
+		const expectedUserAgent = expectedDiscoveryUserAgent ?? null;
 		const nextIds = dynamicModelIds === undefined ? null : JSON.stringify(dynamicModelIds);
 		const nextProvenance = dynamicModelProvenance ?? null;
 		const expectedModelsJson = JSON.stringify(expectedModels);
 		const result = getDb(dbPath).run(
 			`UPDATE model_cache
 			 SET updated_at = ?, authoritative = ?, static_fingerprint = ?, dynamic_model_ids = ?,
-			     dynamic_model_provenance = ?, models = ?
+			     dynamic_model_provenance = ?, discovery_user_agent = ?, models = ?
 			 WHERE provider_id = ? AND updated_at = ?
-			   AND dynamic_model_ids IS ? AND dynamic_model_provenance IS ? AND models = ?`,
+			   AND dynamic_model_ids IS ? AND dynamic_model_provenance IS ?
+			   AND discovery_user_agent IS ? AND models = ?`,
 			[
 				updatedAt,
 				authoritative ? 1 : 0,
 				staticFingerprint,
 				nextIds,
 				nextProvenance,
+				discoveryUserAgent ?? null,
 				JSON.stringify(models),
 				providerId,
 				expectedUpdatedAt,
 				expectedIds,
 				provenance,
+				expectedUserAgent,
 				expectedModelsJson,
 			],
 		);

package/src/model-manager.ts

@@ -60,6 +60,8 @@ export interface ModelManagerOptions<TApi extends Api = Api, TModelsDevPayload =
 	canPublishCache?: () => boolean;
 	/** Credential-and-endpoint identity required to reuse dynamic catalog IDs. */
 	cacheDynamicModelProvenance?: string;
+	/** Discovery User-Agent (e.g., for Antigravity): cache invalidates on change. */
+	discoveryUserAgent?: string;
 }
 
 /**
@@ -146,7 +148,13 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
 	const provenance = options.cacheDynamicModelProvenance;
 	const now = options.now ?? Date.now;
 	const ttlMs = options.cacheTtlMs ?? DEFAULT_CACHE_TTL_MS;
-	const legacyCache = readModelCache<TApi>(options.providerId, ttlMs, now, options.cacheDbPath);
+	const legacyCache = readModelCache<TApi>(
+		options.providerId,
+		ttlMs,
+		now,
+		options.cacheDbPath,
+		options.discoveryUserAgent,
+	);
 	if (
 		strategy !== "online-if-uncached" ||
 		typeof options.fetchDynamicModels !== "function" ||
@@ -185,7 +193,7 @@ async function resolveProviderModelsUncoalesced<TApi extends Api = Api, TModelsD
 	const staticModels = passModelList<TApi>(
 		options.staticModels ?? getBundledModels(options.providerId as GeneratedProvider),
 	);
-	const cache = readModelCache<TApi>(options.providerId, ttlMs, now, dbPath);
+	const cache = readModelCache<TApi>(options.providerId, ttlMs, now, dbPath, options.discoveryUserAgent);
 	const dynamicFetcher = options.fetchDynamicModels;
 	const hasDynamicFetcher = typeof dynamicFetcher === "function";
 	const cacheDynamicModelIdsCurrent =
@@ -262,6 +270,7 @@ async function resolveProviderModelsUncoalesced<TApi extends Api = Api, TModelsD
 				dbPath,
 				cache.dynamicModelIds,
 				cache.dynamicModelProvenance,
+				options.discoveryUserAgent,
 			);
 		}
 		return {
@@ -315,6 +324,7 @@ async function resolveProviderModelsUncoalesced<TApi extends Api = Api, TModelsD
 					dbPath,
 					dynamicModels.map(model => model.id),
 					options.cacheDynamicModelProvenance,
+					options.discoveryUserAgent,
 				);
 			}
 		} else {
@@ -322,10 +332,9 @@ async function resolveProviderModelsUncoalesced<TApi extends Api = Api, TModelsD
 			// stale state remains visible while retry backoff still applies. Re-read
 			// the current row and trust only that row's provenance and dynamic IDs,
 			// not the initial cacheProvenanceMismatch snapshot. A concurrent writer
-			// may have replaced the provider row; if the latest row is unbound or
-			// belongs to another context, use no latest fallback and do not
-			// overwrite or downgrade it.
-			const latestCache = readModelCache<TApi>(options.providerId, ttlMs, now, dbPath);
+			// may have replaced the provider row. This reread filters mismatched
+			// discovery User-Agents, while the checks below enforce dynamic provenance.
+			const latestCache = readModelCache<TApi>(options.providerId, ttlMs, now, dbPath, options.discoveryUserAgent);
 			const latestCacheMatchesCurrentContext =
 				options.cacheDynamicModelProvenance === undefined
 					? latestCache !== null && latestCache.dynamicModelIds === undefined
@@ -353,6 +362,7 @@ async function resolveProviderModelsUncoalesced<TApi extends Api = Api, TModelsD
 						latestCache.dynamicModelIds,
 						latestCache.dynamicModelProvenance,
 						latestCache.models,
+						latestCache.discoveryUserAgent,
 						now(),
 						snapshotModels,
 						false,
@@ -360,6 +370,7 @@ async function resolveProviderModelsUncoalesced<TApi extends Api = Api, TModelsD
 						dbPath,
 						options.cacheDynamicModelProvenance === undefined ? undefined : [],
 						options.cacheDynamicModelProvenance,
+						options.discoveryUserAgent,
 					);
 					if (!updated) cacheModels = [];
 				} else if (latestCache == null) {
@@ -372,6 +383,7 @@ async function resolveProviderModelsUncoalesced<TApi extends Api = Api, TModelsD
 						dbPath,
 						options.cacheDynamicModelProvenance === undefined ? undefined : [],
 						options.cacheDynamicModelProvenance,
+						options.discoveryUserAgent,
 					);
 					if (!inserted) cacheModels = [];
 				}

package/src/model-pricing.d.ts

@@ -1,3 +1,11 @@
 import type { Api, Model, ModelCost } from "./types";
 export declare function getOpenAIModelCost<TApi extends Api>(model: Model<TApi>, inputTokens: number): ModelCost | undefined;
+export declare function getAnthropicModelCost<TApi extends Api>(model: Model<TApi>, inputTokens: number): ModelCost | undefined;
+/**
+ * $/million rate for 1-hour cache writes, or undefined when the model has a single
+ * cache-write rate. Cost calculation applies it to `usage.cttl.ephemeral1h` and bills
+ * the remaining cache-write tokens at `ModelCost.cacheWrite`.
+ */
+export declare function getAnthropicCacheWriteOneHourRate<TApi extends Api>(model: Model<TApi>, inputTokens: number): number | undefined;
+export declare function applyAnthropicModelPricing<TApi extends Api>(model: Model<TApi>): void;
 export declare function applyOpenAIModelPricing<TApi extends Api>(model: Model<TApi>): void;

package/src/model-pricing.ts

@@ -6,6 +6,17 @@ interface TieredPricing {
 }
 
 const LONG_CONTEXT_THRESHOLD = 272_000;
+// Anthropic Haiku 5.5's direct Messages API charges five times the base rates
+// when prompt input exceeds 100K tokens, including cached input.
+// https://platform.claude.com/docs/en/models/haiku-5-5/overview#pricing
+const ANTHROPIC_HAIKU_5_5_LONG_CONTEXT_PRICING: LongContextPricing = {
+	threshold: 100_000,
+	cost: { input: 0.5, output: 2.5, cacheRead: 0.05, cacheWrite: 0.625 },
+};
+// `ModelCost.cacheWrite` is the 5-minute cache-write rate. Haiku 5.5 charges 1-hour
+// writes at twice the base input rate in each prompt-length tier ($0.20 up to 100K
+// tokens, $1 above), and the provider defaults to `ttl: "1h"`.
+const ANTHROPIC_HAIKU_5_5_CACHE_WRITE_1H = { standard: 0.2, longContext: 1 };
 
 const GPT_5_6_SOL_PRICING: TieredPricing = {
 	cost: { input: 5, output: 30, cacheRead: 0.5, cacheWrite: 6.25 },
@@ -95,6 +106,82 @@ export function getOpenAIModelCost<TApi extends Api>(model: Model<TApi>, inputTo
 	return inputTokens > pricing.longContextPricing.threshold ? pricing.longContextPricing.cost : pricing.cost;
 }
 
+const BEDROCK_HAIKU_5_5_ID = /^(?:(?:au|eu|global|jp|us).)?anthropic.claude-haiku-5-5$/;
+// Bedrock bills Haiku 5.5 prompts above 100K tokens at higher "Long Context" rates.
+// AWS documents the threshold but the per-token rates only on its pricing page,
+// which is not machine-readable; mirror Anthropic's published 5x multiplier on each
+// inference profile's own base rates.
+// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-haiku-5-5.html#model-card-anthropic-claude-haiku-5-5-pricing
+const BEDROCK_HAIKU_5_5_LONG_CONTEXT_MULTIPLIER = 5;
+
+function getAnthropicLongContextPricing<TApi extends Api>(model: Model<TApi>): LongContextPricing | undefined {
+	if (model.provider === "anthropic" && model.api === "anthropic-messages" && model.id === "claude-haiku-5-5") {
+		return {
+			threshold: ANTHROPIC_HAIKU_5_5_LONG_CONTEXT_PRICING.threshold,
+			cost: { ...ANTHROPIC_HAIKU_5_5_LONG_CONTEXT_PRICING.cost },
+		};
+	}
+	if (
+		model.provider === "amazon-bedrock" &&
+		model.api === "bedrock-converse-stream" &&
+		BEDROCK_HAIKU_5_5_ID.test(model.id)
+	) {
+		const m = BEDROCK_HAIKU_5_5_LONG_CONTEXT_MULTIPLIER;
+		const scale = (rate: number): number => Math.round(rate * m * 1e9) / 1e9;
+		return {
+			threshold: ANTHROPIC_HAIKU_5_5_LONG_CONTEXT_PRICING.threshold,
+			cost: {
+				input: scale(model.cost.input),
+				output: scale(model.cost.output),
+				cacheRead: scale(model.cost.cacheRead),
+				cacheWrite: scale(model.cost.cacheWrite),
+			},
+		};
+	}
+	return undefined;
+}
+
+export function getAnthropicModelCost<TApi extends Api>(
+	model: Model<TApi>,
+	inputTokens: number,
+): ModelCost | undefined {
+	const pricing = getAnthropicLongContextPricing(model);
+	if (!pricing || inputTokens <= pricing.threshold) return undefined;
+	return pricing.cost;
+}
+
+/**
+ * $/million rate for 1-hour cache writes, or undefined when the model has a single
+ * cache-write rate. Cost calculation applies it to `usage.cttl.ephemeral1h` and bills
+ * the remaining cache-write tokens at `ModelCost.cacheWrite`.
+ */
+export function getAnthropicCacheWriteOneHourRate<TApi extends Api>(
+	model: Model<TApi>,
+	inputTokens: number,
+): number | undefined {
+	const aboveThreshold = inputTokens > ANTHROPIC_HAIKU_5_5_LONG_CONTEXT_PRICING.threshold;
+	if (model.provider === "anthropic" && model.api === "anthropic-messages" && model.id === "claude-haiku-5-5") {
+		return aboveThreshold
+			? ANTHROPIC_HAIKU_5_5_CACHE_WRITE_1H.longContext
+			: ANTHROPIC_HAIKU_5_5_CACHE_WRITE_1H.standard;
+	}
+	if (
+		model.provider === "amazon-bedrock" &&
+		model.api === "bedrock-converse-stream" &&
+		BEDROCK_HAIKU_5_5_ID.test(model.id)
+	) {
+		// Same structure as direct: 1-hour writes cost twice the tier's input rate.
+		const multiplier = aboveThreshold ? BEDROCK_HAIKU_5_5_LONG_CONTEXT_MULTIPLIER : 1;
+		return Math.round(2 * model.cost.input * multiplier * 1e9) / 1e9;
+	}
+	return undefined;
+}
+
+export function applyAnthropicModelPricing<TApi extends Api>(model: Model<TApi>): void {
+	const pricing = getAnthropicLongContextPricing(model);
+	if (pricing) model.longContextPricing = pricing;
+}
+
 export function applyOpenAIModelPricing<TApi extends Api>(model: Model<TApi>): void {
 	if (model.provider !== "openai" && model.provider !== "openai-codex") {
 		return;

package/src/model-thinking.d.ts

@@ -87,12 +87,23 @@ export declare function mapEffortToAnthropicAdaptiveEffort<TApi extends Api>(mod
  * - Thinking content is omitted by default (needs display: "summarized")
  */
 export declare function hasOpus47ApiRestrictions(modelId: string): boolean;
+/**
+ * Haiku 5+ enables adaptive thinking when the `thinking` field is omitted.
+ * An explicit user-facing Off choice must therefore send `type: "disabled"`.
+ */
+export declare function requiresAnthropicThinkingDisableWhenOff(modelId: string): boolean;
+/**
+ * Haiku 5+ and Opus 4.7+ reject sampling parameters on the Anthropic Messages API.
+ * https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide#remove-sampling-parameters
+ */
+export declare function hasAnthropicModelSamplingRestrictions(modelId: string): boolean;
 /**
  * Adaptive thinking `display` is supported starting with Anthropic Opus 4.7.
  * Older adaptive-thinking models (Opus 4.6, Sonnet 4.6+) reject the field.
  * Fable (5+) postdates Opus 4.7, accepts `display`, and defaults it to
  * "omitted" — thinking tokens are billed but no content streams back — so it
  * must opt in like Opus 4.7+ (issue #2791).
+ * Haiku 5+ also accepts `display` and defaults it to "omitted".
  *
  * Shares `hasOpus47ApiRestrictions` version parsing on purpose: the two
  * predicates describe the same API generation, and a private `claude-opus-(\d+)-(\d+)`

package/src/model-thinking.ts

@@ -5,7 +5,7 @@ import {
 	isCodexGpt56Tier,
 	isCodexProductTransport,
 } from "./context-cap-policy";
-import { applyOpenAIModelPricing } from "./model-pricing";
+import { applyAnthropicModelPricing, applyOpenAIModelPricing } from "./model-pricing";
 import {
 	isAuditedOpenAIReasoningTransport,
 	parseDirectXaiReasoningEffortGeneration,
@@ -56,6 +56,17 @@ const DEFAULT_REASONING_EFFORTS_WITH_XHIGH_AND_MAX: readonly Effort[] = [
 	Effort.XHigh,
 	Effort.Max,
 ];
+// Haiku 5.5 supports low, medium, high, xhigh, and max on both Anthropic
+// Messages and Bedrock Converse; neither contract lists minimal.
+// https://platform.claude.com/docs/en/build-with-claude/effort#effort-levels
+// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-haiku-5-5.html
+const ANTHROPIC_HAIKU_5_5_EFFORTS: readonly Effort[] = [
+	Effort.Low,
+	Effort.Medium,
+	Effort.High,
+	Effort.XHigh,
+	Effort.Max,
+];
 const GEMINI_3_PRO_EFFORTS: readonly Effort[] = [Effort.Low, Effort.High];
 const GEMINI_3_FLASH_EFFORTS: readonly Effort[] = [Effort.Minimal, Effort.Low, Effort.Medium, Effort.High];
 // Gemini 3.7 Flash dropped `minimal`; the official API returns an error for it.
@@ -95,8 +106,8 @@ type SemVer = {
 	patch: number;
 };
 
-type GeminiKind = "pro" | "flash";
-type AnthropicKind = "opus" | "sonnet" | "fable";
+type GeminiKind = "pro" | "flash" | "argon";
+type AnthropicKind = "opus" | "sonnet" | "fable" | "haiku";
 type OpenAIVariant =
 	| "astra"
 	| "base"
@@ -328,6 +339,23 @@ export function modelSupportsReasoningControl<TApi extends Api>(
 export function applyGeneratedModelPolicies(models: ApiModel<Api>[]): void {
 	for (let index = 0; index < models.length; index++) {
 		const source = models[index]!;
+		const inputCost = nonNegativeFinitePrice(source.cost.input);
+		const outputCost = nonNegativeFinitePrice(source.cost.output);
+		const cacheReadCost = nonNegativeFinitePrice(source.cost.cacheRead);
+		const cacheWriteCost = nonNegativeFinitePrice(source.cost.cacheWrite);
+		if (
+			inputCost !== source.cost.input ||
+			outputCost !== source.cost.output ||
+			cacheReadCost !== source.cost.cacheRead ||
+			cacheWriteCost !== source.cost.cacheWrite
+		) {
+			source.cost = {
+				input: inputCost,
+				output: outputCost,
+				cacheRead: cacheReadCost,
+				cacheWrite: cacheWriteCost,
+			};
+		}
 		if (source.provider === "xai" && (source.id === "grok-4.5" || source.id === "grok-4.6")) {
 			source.reasoning = true;
 		}
@@ -359,6 +387,10 @@ export function applyGeneratedModelPolicies(models: ApiModel<Api>[]): void {
 	}
 }
 
+function nonNegativeFinitePrice(value: number): number {
+	return Number.isFinite(value) && value >= 0 ? value : 0;
+}
+
 /**
  * Link OpenAI model variants to their context promotion targets.
  *
@@ -517,12 +549,34 @@ export function hasOpus47ApiRestrictions(modelId: string): boolean {
 	return semverGte(parsed.version, "4.7") && parsed.kind === "opus";
 }
 
+function isAnthropicHaikuAdaptiveModel(modelId: string): boolean {
+	const parsed = parseAnthropicModel(getCanonicalModelId(modelId));
+	return parsed?.kind === "haiku" && semverGte(parsed.version, "5.0");
+}
+
+/**
+ * Haiku 5+ enables adaptive thinking when the `thinking` field is omitted.
+ * An explicit user-facing Off choice must therefore send `type: "disabled"`.
+ */
+export function requiresAnthropicThinkingDisableWhenOff(modelId: string): boolean {
+	return isAnthropicHaikuAdaptiveModel(modelId);
+}
+
+/**
+ * Haiku 5+ and Opus 4.7+ reject sampling parameters on the Anthropic Messages API.
+ * https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide#remove-sampling-parameters
+ */
+export function hasAnthropicModelSamplingRestrictions(modelId: string): boolean {
+	return hasOpus47ApiRestrictions(modelId) || isAnthropicHaikuAdaptiveModel(modelId);
+}
+
 /**
  * Adaptive thinking `display` is supported starting with Anthropic Opus 4.7.
  * Older adaptive-thinking models (Opus 4.6, Sonnet 4.6+) reject the field.
  * Fable (5+) postdates Opus 4.7, accepts `display`, and defaults it to
  * "omitted" — thinking tokens are billed but no content streams back — so it
  * must opt in like Opus 4.7+ (issue #2791).
+ * Haiku 5+ also accepts `display` and defaults it to "omitted".
  *
  * Shares `hasOpus47ApiRestrictions` version parsing on purpose: the two
  * predicates describe the same API generation, and a private `claude-opus-(\d+)-(\d+)`
@@ -533,7 +587,7 @@ export function hasOpus47ApiRestrictions(modelId: string): boolean {
  */
 export function supportsAnthropicAdaptiveThinkingDisplay(modelId: string): boolean {
 	if (/claude-fable-\d/.test(modelId)) return true;
-	return hasOpus47ApiRestrictions(modelId);
+	return hasOpus47ApiRestrictions(modelId) || isAnthropicHaikuAdaptiveModel(modelId);
 }
 
 function anthropicModelHasRealXHighEffort<TApi extends Api>(model: ApiModel<TApi>): boolean {
@@ -555,6 +609,7 @@ function anthropicModelHasRealXHighEffort<TApi extends Api>(model: ApiModel<TApi
 
 function applyGeneratedModelPolicy(model: ApiModel<Api>): void {
 	applyOpenAIModelPricing(model);
+	applyAnthropicModelPricing(model);
 	const copilotLimits = model.provider === "github-copilot" ? COPILOT_GENERATED_LIMITS[model.id] : undefined;
 	if (copilotLimits) {
 		model.contextWindow = copilotLimits.contextWindow;
@@ -729,6 +784,17 @@ function applyGeneratedModelPolicy(model: ApiModel<Api>): void {
 	) {
 		model.contextWindow = 1_000_000;
 	}
+	// Google Gemini 4 Argon supports up to 65,536 output tokens across provider
+	// aliases. Strip catalog provider prefixes and OpenRouter's batch suffix.
+	// https://ai.google.dev/gemini-api/docs/models/gemini-4-argon
+	const canonicalId = model.id
+		.toLowerCase()
+		.split("/")
+		.at(-1)
+		?.replace(/:batch$/, "");
+	if (canonicalId === "gemini-4-argon") {
+		model.maxTokens = Math.min(model.maxTokens, 65_536);
+	}
 }
 
 function scrubGeneratedModelName(name: string): string {
@@ -847,6 +913,19 @@ function inferDefaultEffort<TApi extends Api>(model: ApiModel<TApi>, parsedModel
 	if (model.provider === "omlx") {
 		return Effort.Medium;
 	}
+	if (
+		parsedModel.family === "anthropic" &&
+		parsedModel.kind === "haiku" &&
+		((model.provider === "anthropic" &&
+			model.api === "anthropic-messages" &&
+			semverGte(parsedModel.version, "5.0")) ||
+			(model.provider === "amazon-bedrock" &&
+				model.api === "bedrock-converse-stream" &&
+				semverGte(parsedModel.version, "5.5")))
+	) {
+		// Haiku 5.5 defaults to medium adaptive effort instead of the global high.
+		return Effort.Medium;
+	}
 	return undefined;
 }
 
@@ -979,6 +1058,8 @@ function inferGeminiSupportedEfforts(model: GeminiModel): readonly Effort[] {
 	if (model.kind === "pro") {
 		return GEMINI_3_PRO_EFFORTS;
 	}
+	// Gemini 3.7+ Flash and Gemini 4 Argon accept low/medium/high only.
+	// https://ai.google.dev/gemini-api/docs/models/gemini-4-argon
 	if (semverGte(model.version, "3.7")) {
 		return GEMINI_3_7_FLASH_EFFORTS;
 	}
@@ -993,6 +1074,9 @@ function inferAnthropicSupportedEfforts<TApi extends Api>(
 		(model.api === "anthropic-messages" || model.api === "bedrock-converse-stream") &&
 		semverGte(parsedModel.version, "4.6")
 	) {
+		if (parsedModel.kind === "haiku" && semverGte(parsedModel.version, "5.5")) {
+			return ANTHROPIC_HAIKU_5_5_EFFORTS;
+		}
 		if (parsedModel.kind === "fable") {
 			// Fable exposes Anthropic's Messages-only xhigh preset; Bedrock
 			// Converse lacks it (same split as Opus 4.7+ below).
@@ -1054,11 +1138,7 @@ function inferThinkingControlMode<TApi extends Api>(
 		case "google-generative-ai":
 		case "google-gemini-cli":
 		case "google-vertex":
-			return parsedModel.family === "gemini" &&
-				semverGte(parsedModel.version, "3.0") &&
-				parsedModel.version.major === 3
-				? "google-level"
-				: "budget";
+			return parsedModel.family === "gemini" && semverGte(parsedModel.version, "3.0") ? "google-level" : "budget";
 
 		case "anthropic-messages":
 			if (parsedModel.family === "anthropic") {
@@ -1075,7 +1155,7 @@ function inferThinkingControlMode<TApi extends Api>(
 			if (parsedModel.family === "anthropic") {
 				if (
 					semverGte(parsedModel.version, "4.6") &&
-					(parsedModel.kind === "opus" || parsedModel.kind === "fable")
+					(parsedModel.kind === "opus" || parsedModel.kind === "fable" || parsedModel.kind === "haiku")
 				) {
 					return "anthropic-adaptive";
 				}
@@ -1104,7 +1184,7 @@ function parseGeminiModel(modelId: string): GeminiModel | null {
 	if (modelId.endsWith(GEMINI_SUFFIX)) {
 		modelId = modelId.slice(0, -GEMINI_SUFFIX.length);
 	}
-	const match = /gemini-(\d+(?:\.\d+){0,2})-(pro|flash)\b/.exec(modelId);
+	const match = /gemini-(\d+(?:\.\d+){0,2})-(pro|flash|argon)\b/.exec(modelId);
 	if (!match) {
 		return null;
 	}
@@ -1116,7 +1196,7 @@ function parseGeminiModel(modelId: string): GeminiModel | null {
 }
 
 function parseAnthropicModel(modelId: string): AnthropicModel | null {
-	const match = /claude-(opus|sonnet|fable)-(\d{1,2}(?:[.-]\d{1,2}){0,2})\b/.exec(modelId);
+	const match = /claude-(opus|sonnet|fable|haiku)-(\d{1,2}(?:[.-]\d{1,2}){0,2})\b/.exec(modelId);
 	if (!match) {
 		return null;
 	}
@@ -1124,6 +1204,11 @@ function parseAnthropicModel(modelId: string): AnthropicModel | null {
 	if (!version) {
 		return null;
 	}
+	// Haiku is recognized from generation 5 only. Haiku 4.5 predates the
+	// adaptive-thinking policy and keeps the generic (unparsed) budget ladder.
+	if (match[1] === "haiku" && !semverGte(version, "5.0")) {
+		return null;
+	}
 	return { family: "anthropic", kind: match[1] as AnthropicKind, version };
 }