@openclaw/amazon-bedrock-provider 2026.9.2 → 2026.9.4

package/dist/bedrock-options.js

@@ -1,4 +1,18 @@
+import { resolveClaudeModelIdentity } from "openclaw/plugin-sdk/provider-model-shared";
 //#region extensions/amazon-bedrock/bedrock-options.ts
+function resolveBedrockPromptCachePolicy(model) {
+	const modelId = model.id.trim().toLowerCase().replace(/^arn:aws(?:-cn|-us-gov)?:bedrock:[^:]+:[^:]*:(?:foundation-model|inference-profile)\//, "").replace(/^(?:us|eu|apac|jp|global)\./, "");
+	if (/^amazon\.nova-(?:micro|lite|pro|premier|2-lite)-v1:0$/.test(modelId)) return "nova";
+	if (supportsBedrockClaudePromptCaching(model.id, model.name) || supportsBedrockClaudePromptCaching(resolveClaudeModelIdentity(model), model.name)) return "claude";
+}
+function resolveBedrockCachePoint(model, retention) {
+	const policy = resolveBedrockPromptCachePolicy(model);
+	if (!policy || retention === "none") return;
+	return {
+		type: "default",
+		...policy === "claude" && retention === "long" ? { ttl: "1h" } : {}
+	};
+}
 function getModelMatchCandidates(modelId, modelName) {
 	return (modelName ? [modelId, modelName] : [modelId]).flatMap((value) => {
 		const lower = value.toLowerCase();
@@ -6,7 +20,7 @@ function getModelMatchCandidates(modelId, modelName) {
 	});
 }
 /** Return whether a Bedrock model is known to support Anthropic prompt caching. */
-function supportsBedrockPromptCaching(modelId, modelName) {
+function supportsBedrockClaudePromptCaching(modelId, modelName) {
 	const candidates = getModelMatchCandidates(modelId, modelName);
 	if (!candidates.some((s) => s.includes("claude"))) {
 		if (typeof process !== "undefined" && process.env.AWS_BEDROCK_FORCE_CACHE === "1") return true;
@@ -19,4 +33,4 @@ function supportsBedrockPromptCaching(modelId, modelName) {
 	return false;
 }
 //#endregion
-export { supportsBedrockPromptCaching };
+export { resolveBedrockCachePoint, resolveBedrockPromptCachePolicy, supportsBedrockClaudePromptCaching };

package/dist/config-compat.js

@@ -1,4 +1,5 @@
 import { isRecord } from "openclaw/plugin-sdk/string-coerce-runtime";
+import { mergeMissing } from "openclaw/plugin-sdk/runtime-doctor-migrations";
 //#region extensions/amazon-bedrock/config-compat.ts
 /**
 * Legacy config migration for Amazon Bedrock discovery settings. It moves
@@ -6,14 +7,6 @@ import { isRecord } from "openclaw/plugin-sdk/string-coerce-runtime";
 */
 const LEGACY_PATH = "models.bedrockDiscovery";
 const TARGET_PATH = "plugins.entries.amazon-bedrock.config.discovery";
-const BLOCKED_OBJECT_KEYS = /* @__PURE__ */ new Set([
-	"__proto__",
-	"prototype",
-	"constructor"
-]);
-function isBlockedObjectKey(key) {
-	return BLOCKED_OBJECT_KEYS.has(key);
-}
 function getRecord(value) {
 	return isRecord(value) ? value : null;
 }
@@ -24,17 +17,6 @@ function ensureRecord(root, key) {
 	root[key] = next;
 	return next;
 }
-function mergeMissing(target, source) {
-	for (const [key, value] of Object.entries(source)) {
-		if (value === void 0 || isBlockedObjectKey(key)) continue;
-		const existing = target[key];
-		if (existing === void 0) {
-			target[key] = value;
-			continue;
-		}
-		if (isRecord(existing) && isRecord(value)) mergeMissing(existing, value);
-	}
-}
 function cloneRecord(value) {
 	return { ...value };
 }

package/dist/discovery.js

@@ -2,13 +2,11 @@ import { refreshAwsSharedConfigCacheForBedrock } from "./aws-credential-refresh.
 import { loadBedrockControlPlaneSdk, runBedrockControlPlaneRequest } from "./control-plane.js";
 import { resolveBedrockConfigApiKey } from "./discovery-shared.js";
 import { resolveBedrockNativeThinkingLevelMap } from "./thinking-policy.js";
+import { LiveModelCatalogHttpError } from "openclaw/plugin-sdk/provider-catalog-live-runtime";
 import { resolveClaudeFable5ModelIdentity, resolveClaudeModelIdentity, resolveClaudeMythos5ModelIdentity, resolveClaudeOpus5ModelIdentity, resolveClaudeSonnet5ModelIdentity, supportsClaudeAdaptiveThinking } from "openclaw/plugin-sdk/provider-model-shared";
-import { normalizeLowercaseStringOrEmpty, normalizeOptionalLowercaseString, normalizeOptionalString } from "openclaw/plugin-sdk/string-coerce-runtime";
-import { createSubsystemLogger } from "openclaw/plugin-sdk/core";
-import { formatErrorMessage } from "openclaw/plugin-sdk/error-runtime";
+import { asOptionalRecord, normalizeLowercaseStringOrEmpty, normalizeOptionalLowercaseString, normalizeOptionalString } from "openclaw/plugin-sdk/string-coerce-runtime";
 import { isFutureDateTimestampMs, resolveExpiresAtMsFromDurationSeconds } from "openclaw/plugin-sdk/number-runtime";
 //#region extensions/amazon-bedrock/discovery.ts
-const log = createSubsystemLogger("bedrock-discovery");
 const DEFAULT_REFRESH_INTERVAL_SECONDS = 3600;
 const DEFAULT_CONTEXT_WINDOW = 32e3;
 const DEFAULT_MAX_TOKENS = 4096;
@@ -120,15 +118,11 @@ const DEFAULT_COST = {
 	cacheWrite: 0
 };
 const discoveryCache = /* @__PURE__ */ new Map();
-let hasLoggedBedrockError = false;
 function normalizeProviderFilter(filter) {
 	if (!filter || filter.length === 0) return [];
 	const normalized = new Set(filter.map((entry) => normalizeOptionalLowercaseString(entry)).filter((entry) => Boolean(entry)));
 	return Array.from(normalized).toSorted();
 }
-function buildCacheKey(params) {
-	return JSON.stringify(params);
-}
 function includesTextModalities(modalities) {
 	return (modalities ?? []).some((entry) => normalizeOptionalLowercaseString(entry) === "text");
 }
@@ -214,27 +208,21 @@ function resolveBaseModelId(profile) {
 }
 /**
 * Fetch raw inference profile summaries from the Bedrock control plane.
-* Handles pagination. Best-effort: silently returns empty array if IAM lacks
-* bedrock:ListInferenceProfiles permission.
+* All pages must succeed before discovery can cache a complete catalog.
 */
 async function fetchInferenceProfileSummaries(client, createListInferenceProfilesCommand) {
-	try {
-		const profiles = [];
-		let nextToken;
-		do {
-			const command = createListInferenceProfilesCommand({ nextToken });
-			const response = await runBedrockControlPlaneRequest({
-				operation: "Bedrock ListInferenceProfiles",
-				send: (options) => client.send(command, options)
-			});
-			for (const summary of response.inferenceProfileSummaries ?? []) profiles.push(summary);
-			nextToken = response.nextToken;
-		} while (nextToken);
-		return profiles;
-	} catch (error) {
-		log.debug?.("Skipping inference profile discovery", { error: formatErrorMessage(error) });
-		return [];
-	}
+	const profiles = [];
+	let nextToken;
+	do {
+		const command = createListInferenceProfilesCommand({ nextToken });
+		const response = await runBedrockControlPlaneRequest({
+			operation: "Bedrock ListInferenceProfiles",
+			send: (options) => client.send(command, options)
+		});
+		for (const summary of response.inferenceProfileSummaries ?? []) profiles.push(summary);
+		nextToken = response.nextToken;
+	} while (nextToken);
+	return profiles;
 }
 /**
 * Convert raw inference profile summaries into model definitions.
@@ -279,14 +267,15 @@ function resolveInferenceProfiles(profiles, defaults, providerFilter, foundation
 	}
 	return discovered;
 }
-/** Discover Bedrock models and inference profiles for one region/config. */
+/** Public discovery is advisory by default; catalog owners opt into strict acquisition. */
 async function discoverBedrockModels(params) {
 	const refreshIntervalSeconds = Math.max(0, Math.floor(params.config?.refreshInterval ?? DEFAULT_REFRESH_INTERVAL_SECONDS));
 	const providerFilter = normalizeProviderFilter(params.config?.providerFilter);
 	const defaultContextWindow = resolveDefaultContextWindow(params.config);
 	const defaultMaxTokens = resolveDefaultMaxTokens(params.config);
-	const cacheKey = buildCacheKey({
+	const cacheKey = JSON.stringify({
 		region: params.region,
+		discoveryMode: params.discoveryMode,
 		providerFilter,
 		refreshIntervalSeconds,
 		defaultContextWindow,
@@ -295,11 +284,8 @@ async function discoverBedrockModels(params) {
 	const now = params.now?.() ?? Date.now();
 	if (refreshIntervalSeconds > 0) {
 		const cached = discoveryCache.get(cacheKey);
-		if (cached && isFutureDateTimestampMs(cached.expiresAt, { nowMs: now })) {
-			if (cached.value) return cached.value;
-			if (cached.inFlight) return cached.inFlight;
-		}
-		if (cached) discoveryCache.delete(cacheKey);
+		if (cached && isFutureDateTimestampMs(cached.expiresAt, { nowMs: now })) return cached.result;
+		discoveryCache.delete(cacheKey);
 	}
 	const sdk = await loadBedrockControlPlaneSdk();
 	const clientFactory = params.clientFactory ?? ((region) => sdk.createClient(region));
@@ -311,7 +297,11 @@ async function discoverBedrockModels(params) {
 			const [foundationResponse, profileSummaries] = await Promise.all([runBedrockControlPlaneRequest({
 				operation: "Bedrock ListFoundationModels",
 				send: (options) => client.send(foundationCommand, options)
-			}), fetchInferenceProfileSummaries(client, (input) => sdk.createListInferenceProfilesCommand(input))]);
+			}), fetchInferenceProfileSummaries(client, (input) => sdk.createListInferenceProfilesCommand(input)).catch((error) => {
+				if (params.discoveryMode === "strict") throw error;
+				discoveryCache.delete(cacheKey);
+				return [];
+			})]);
 			const discovered = [];
 			const seenIds = /* @__PURE__ */ new Set();
 			const foundationModels = /* @__PURE__ */ new Map();
@@ -343,37 +333,28 @@ async function discoverBedrockModels(params) {
 				if (aGlobal !== bGlobal) return aGlobal - bGlobal;
 				return a.name.localeCompare(b.name);
 			});
+		} catch (error) {
+			const status = asOptionalRecord(asOptionalRecord(error)?.$metadata)?.httpStatusCode;
+			if (typeof status === "number") throw new LiveModelCatalogHttpError("amazon-bedrock", status);
+			throw error;
 		} finally {
 			client.destroy();
 		}
-	})();
+	})().catch((error) => {
+		discoveryCache.delete(cacheKey);
+		if (params.discoveryMode === "strict") throw error;
+		return [];
+	});
 	if (refreshIntervalSeconds > 0) {
 		const expiresAt = resolveExpiresAtMsFromDurationSeconds(refreshIntervalSeconds, { nowMs: now });
 		if (expiresAt !== void 0) discoveryCache.set(cacheKey, {
 			expiresAt,
-			inFlight: discoveryPromise
+			result: discoveryPromise
 		});
 	}
-	try {
-		const value = await discoveryPromise;
-		if (refreshIntervalSeconds > 0) {
-			const expiresAt = resolveExpiresAtMsFromDurationSeconds(refreshIntervalSeconds, { nowMs: now });
-			if (expiresAt !== void 0) discoveryCache.set(cacheKey, {
-				expiresAt,
-				value
-			});
-		}
-		return value;
-	} catch (error) {
-		if (refreshIntervalSeconds > 0) discoveryCache.delete(cacheKey);
-		if (!hasLoggedBedrockError) {
-			hasLoggedBedrockError = true;
-			log.warn("Failed to discover Bedrock models", { error: formatErrorMessage(error) });
-		}
-		return [];
-	}
+	return discoveryPromise;
 }
-/** Resolve the implicit Bedrock provider config from env, plugin config, and discovery. */
+/** Public resolution keeps advisory null results; strict catalog callers retain acquired empties. */
 async function resolveImplicitBedrockProvider(params) {
 	const env = params.env ?? process.env;
 	const discoveryConfig = params.pluginConfig?.discovery;
@@ -384,10 +365,11 @@ async function resolveImplicitBedrockProvider(params) {
 	const region = discoveryConfig?.region ?? normalizeOptionalString(env.AWS_REGION) ?? normalizeOptionalString(env.AWS_DEFAULT_REGION) ?? "us-east-1";
 	const models = await discoverBedrockModels({
 		region,
+		discoveryMode: params.discoveryMode,
 		config: discoveryConfig,
 		clientFactory: params.clientFactory
 	});
-	if (models.length === 0) return null;
+	if (models.length === 0 && params.discoveryMode !== "strict") return null;
 	return {
 		baseUrl: `https://bedrock-runtime.${region}.amazonaws.com`,
 		api: "bedrock-converse-stream",

package/dist/memory-embedding-adapter.js

@@ -1,5 +1,5 @@
 import { DEFAULT_BEDROCK_EMBEDDING_MODEL, createBedrockEmbeddingProvider, hasAwsCredentials } from "./embedding-provider.js";
-import { isMissingEmbeddingApiKeyError } from "openclaw/plugin-sdk/memory-core-host-engine-embeddings";
+import { isMissingEmbeddingApiKeyError } from "openclaw/plugin-sdk/embedding-provider-adapter";
 //#region extensions/amazon-bedrock/memory-embedding-adapter.ts
 /**
 * Memory embedding adapter for Amazon Bedrock. It exposes Bedrock embeddings to

package/dist/register.sync.runtime.js

@@ -1,13 +1,15 @@
 import { refreshAwsSharedConfigCacheForBedrock } from "./aws-credential-refresh.js";
-import { supportsBedrockPromptCaching } from "./bedrock-options.js";
+import { resolveBedrockPromptCachePolicy, supportsBedrockClaudePromptCaching } from "./bedrock-options.js";
 import { loadBedrockControlPlaneSdk, runBedrockControlPlaneRequest } from "./control-plane.js";
 import { resolveBedrockConfigApiKey } from "./discovery-shared.js";
 import { bedrockMemoryEmbeddingProviderAdapter } from "./memory-embedding-adapter.js";
 import { isLatestAdaptiveBedrockModelRef, isOpus47OrNewerBedrockModelRef, resolveBedrockClaudeThinkingProfile, resolveBedrockNativeThinkingLevelMap, supportsBedrockNativeMaxEffort } from "./thinking-policy.js";
 import { streamSimpleBedrock } from "./stream.runtime.js";
 import { resolvePluginConfigObject } from "openclaw/plugin-sdk/plugin-config-runtime";
+import { runLiveProviderCatalog } from "openclaw/plugin-sdk/provider-catalog-live-runtime";
 import { buildProviderReplayFamilyHooks, normalizeProviderId, resolveClaudeFable5ModelIdentity, resolveClaudeModelIdentity, resolveClaudeMythos5ModelIdentity, resolveClaudeOpus5ModelIdentity, resolveClaudeSonnet5ModelIdentity } from "openclaw/plugin-sdk/provider-model-shared";
 import { createPayloadPatchStreamWrapper } from "openclaw/plugin-sdk/provider-stream-shared";
+import { splitSystemPromptCacheBoundary } from "openclaw/plugin-sdk/provider-transport-runtime";
 //#region extensions/amazon-bedrock/register.sync.runtime.ts
 function normalizeBedrockResolvedModel({ modelId, model }) {
 	const thinkingLevelMap = resolveBedrockNativeThinkingLevelMap(modelId, model.params);
@@ -87,9 +89,6 @@ function createGuardrailWrapStreamFn(innerWrapStreamFn, guardrailConfig) {
 		});
 	};
 }
-function sharedRuntimeWouldInjectCachePoints(modelId) {
-	return supportsBedrockPromptCaching(modelId);
-}
 /**
 * Detect Bedrock application inference profile ARNs — these are the only IDs
 * where model-name-based checks fail because the ARN is opaque.
@@ -101,21 +100,6 @@ function isBedrockAppInferenceProfile(modelId) {
 	return BEDROCK_APP_INFERENCE_PROFILE_RE.test(modelId);
 }
 /**
-* The shared runtime's `supportsPromptCaching` checks `model.id` for specific Claude
-* model name patterns, which fails for application inference profile ARNs (opaque
-* IDs that may not contain the model name). When OpenClaw's `isAnthropicBedrockModel`
-* identifies the model but the shared runtime won't inject cache points, we do it via onPayload.
-*
-* Gated to application inference profile ARNs only — regular Claude model IDs and
-* system-defined inference profiles (us.anthropic.claude-*) are left to the shared runtime.
-*/
-function needsCachePointInjection(modelId) {
-	if (!isBedrockAppInferenceProfile(modelId)) return false;
-	if (sharedRuntimeWouldInjectCachePoints(modelId)) return false;
-	if (isAnthropicBedrockModel(modelId)) return true;
-	return false;
-}
-/**
 * Extract the region from a Bedrock ARN.
 * e.g. "arn:aws:bedrock:us-east-1:123:application-inference-profile/abc" → "us-east-1"
 */
@@ -123,13 +107,6 @@ function extractRegionFromArn(arn) {
 	const parts = arn.split(":");
 	return parts.length >= 4 && parts[3] ? parts[3] : void 0;
 }
-/**
-* Check if a resolved foundation model ARN supports prompt caching using the
-* same matcher OpenClaw uses for direct model IDs.
-*/
-function resolvedModelSupportsCaching(modelArn) {
-	return supportsBedrockPromptCaching(modelArn);
-}
 const appProfileTraitsCache = /* @__PURE__ */ new Map();
 async function resolveAppProfileTraits(modelId, fallbackRegion, signal) {
 	const cached = appProfileTraitsCache.get(modelId);
@@ -150,7 +127,7 @@ async function resolveAppProfileTraits(modelId, fallbackRegion, signal) {
 		})).models ?? [];
 		const modelArns = models.map((model) => model.modelArn ?? "");
 		const traits = {
-			cacheEligible: models.length > 0 && modelArns.every((modelArn) => resolvedModelSupportsCaching(modelArn)),
+			cacheEligible: models.length > 0 && modelArns.every((modelArn) => supportsBedrockClaudePromptCaching(modelArn)),
 			omitTemperature: modelArns.some(isOpus47OrNewerBedrockModelRef)
 		};
 		appProfileTraitsCache.set(modelId, traits);
@@ -178,11 +155,15 @@ function makeCachePoint(cacheRetention) {
 * Inject Bedrock Converse cache points into the payload when the shared runtime skipped them
 * because it didn't recognize the model ID (application inference profiles).
 */
-function injectBedrockCachePoints(payload, cacheRetention) {
-	if (!cacheRetention || cacheRetention === "none") return;
+function injectBedrockCachePoints(payload, cacheRetention, context, model) {
+	if (!cacheRetention || cacheRetention === "none" || resolveBedrockPromptCachePolicy(model)) return;
 	const point = makeCachePoint(cacheRetention);
 	const system = payload.system;
-	if (Array.isArray(system) && system.length > 0 && !hasCachePoint(system)) system.push(point);
+	if (Array.isArray(system) && system.length > 0 && !hasCachePoint(system)) {
+		const split = context.systemPrompt && splitSystemPromptCacheBoundary(context.systemPrompt);
+		if (!split || split.stablePrefix) system.splice(split ? 1 : system.length, 0, point);
+	}
+	if (context.messages.some((message) => message.role === "user" && message.runtimeContextCarrier)) return;
 	const messages = payload.messages;
 	if (Array.isArray(messages) && messages.length > 0) {
 		for (const msg of messages.toReversed()) if (msg.role === "user" && Array.isArray(msg.content)) {
@@ -221,7 +202,7 @@ function registerAmazonBedrockPlugin(api) {
 			id: modelId,
 			params: model?.params
 		};
-		if (isAnthropicBedrockModel(modelId) || resolveClaudeModelIdentity(modelRef).startsWith("claude-")) return streamFn;
+		if (resolveBedrockPromptCachePolicy(modelRef) === "nova" || isAnthropicBedrockModel(modelId) || resolveClaudeModelIdentity(modelRef).startsWith("claude-")) return streamFn;
 		if (isBedrockAppInferenceProfile(modelId)) return streamFn;
 		return createBedrockNoCacheWrapper(streamFn);
 	};
@@ -282,18 +263,22 @@ function registerAmazonBedrockPlugin(api) {
 		auth: [],
 		catalog: {
 			order: "simple",
-			run: async (ctx) => {
-				const { resolveImplicitBedrockProvider } = await import("./discovery.js");
-				const implicit = await resolveImplicitBedrockProvider({
-					pluginConfig: resolveCurrentPluginConfig(ctx.config),
-					env: ctx.env
-				});
-				if (!implicit) return null;
-				return { provider: implicit };
-			}
+			run: (ctx) => runLiveProviderCatalog({
+				providerId,
+				run: async () => {
+					const { resolveImplicitBedrockProvider } = await import("./discovery.js");
+					const implicit = await resolveImplicitBedrockProvider({
+						discoveryMode: "strict",
+						pluginConfig: resolveCurrentPluginConfig(ctx.config),
+						env: ctx.env
+					});
+					return implicit ? { provider: implicit } : null;
+				}
+			})
 		},
 		resolveConfigApiKey: ({ env }) => resolveBedrockConfigApiKey(env),
 		normalizeResolvedModel: normalizeBedrockResolvedModel,
+		supportsSystemPromptCacheBoundary: true,
 		createStreamFn: ({ model }) => model.api === "bedrock-converse-stream" ? bedrockStreamFn : void 0,
 		...anthropicByModelReplayHooks,
 		wrapStreamFn: ({ modelId, config, model, streamFn, thinkingLevel, extraParams }) => {
@@ -326,11 +311,11 @@ function registerAmazonBedrockPlugin(api) {
 				} else wrapped = createBedrockServiceTierWrapper(wrapped, serviceTier);
 			}
 			const region = resolveBedrockRegion(config) ?? extractRegionFromBaseUrl(model?.baseUrl) ?? currentPluginConfig?.discovery?.region;
-			const mayNeedCacheInjection = isBedrockAppInferenceProfile(modelId) && !sharedRuntimeWouldInjectCachePoints(modelId);
+			const mayNeedCacheInjection = isBedrockAppInferenceProfile(modelId) && !supportsBedrockClaudePromptCaching(modelId);
 			const shouldOmitTemperature = opus47OrNewer || fable5 || isLatestAdaptiveBedrockModelRef(modelId, model?.params);
 			const shouldPatchMaxThinking = supportsNativeMax && thinkingLevel === "max";
 			const shouldPatchPayload = shouldOmitTemperature || shouldPatchMaxThinking;
-			const heuristicMatch = needsCachePointInjection(modelId);
+			const heuristicMatch = isAnthropicBedrockModel(modelId);
 			if (!region && !mayNeedCacheInjection && !shouldOmitTemperature && !shouldPatchMaxThinking) return createAwsCredentialRefreshStreamWrapper(wrapped);
 			const underlying = wrapped ?? streamFn;
 			if (!underlying) return wrapped;
@@ -356,7 +341,7 @@ function registerAmazonBedrockPlugin(api) {
 						onPayload: async (payload, payloadModel) => {
 							if (payload && typeof payload === "object") {
 								const payloadRecord = payload;
-								injectBedrockCachePoints(payloadRecord, cacheRetention);
+								injectBedrockCachePoints(payloadRecord, cacheRetention, context, streamModel);
 								if (shouldPatchMaxThinking) patchMaxThinkingEffort(payloadRecord);
 								if (shouldOmitTemperature) omitUnsupportedClaudePayloadTemperature(payloadRecord);
 								else if (mayNeedTemperatureTrait) {
@@ -373,7 +358,7 @@ function registerAmazonBedrockPlugin(api) {
 						const traits = await resolveAppProfileTraits(modelId, region, merged.signal);
 						if (payload && typeof payload === "object") {
 							const payloadRecord = payload;
-							if (traits.cacheEligible) injectBedrockCachePoints(payloadRecord, cacheRetention);
+							if (traits.cacheEligible) injectBedrockCachePoints(payloadRecord, cacheRetention, context, streamModel);
 							if (shouldPatchMaxThinking) patchMaxThinkingEffort(payloadRecord);
 							if (traits.omitTemperature) omitUnsupportedClaudePayloadTemperature(payloadRecord);
 						}

package/dist/stream.runtime.js

@@ -1,14 +1,14 @@
-import { supportsBedrockPromptCaching } from "./bedrock-options.js";
+import { resolveBedrockCachePoint, resolveBedrockPromptCachePolicy } from "./bedrock-options.js";
 import { supportsBedrockNativeMaxEffort } from "./thinking-policy.js";
-import { requiresClaudeMandatoryAdaptiveThinking, resolveClaudeFable5ModelIdentity, resolveClaudeModelIdentity, resolveClaudeMythos5ModelIdentity, resolveClaudeOpus5ModelIdentity, resolveClaudeSonnet5ModelIdentity, supportsClaudeAdaptiveThinking, supportsClaudeNativeXhighEffort } from "openclaw/plugin-sdk/provider-model-shared";
+import { bindsClaudeThinkingPrefix, requiresClaudeMandatoryAdaptiveThinking, resolveClaudeFable5ModelIdentity, resolveClaudeModelIdentity, resolveClaudeMythos5ModelIdentity, resolveClaudeOpus5ModelIdentity, resolveClaudeSonnet5ModelIdentity, supportsClaudeAdaptiveThinking, supportsClaudeNativeXhighEffort } from "openclaw/plugin-sdk/provider-model-shared";
 import { applyAnthropicRefusal, createDeferredEventBuffer, notifyLlmRequestActivity } from "openclaw/plugin-sdk/provider-stream-shared";
+import { describeToolResultMediaPlaceholder, failTransportStream, finalizeTerminalToolCallArguments, notifyProviderHttpMetadata, splitSystemPromptCacheBoundary, stripSystemPromptCacheBoundary } from "openclaw/plugin-sdk/provider-transport-runtime";
 import { isRecord, normalizeOptionalString } from "openclaw/plugin-sdk/string-coerce-runtime";
-import { BedrockRuntimeClient, BedrockRuntimeServiceException, CachePointType, CacheTTL, ConversationRole, ConverseStreamCommand, ImageFormat, StopReason, ToolResultStatus } from "@aws-sdk/client-bedrock-runtime";
+import { BedrockRuntimeClient, BedrockRuntimeServiceException, CacheTTL, ConversationRole, ConverseStreamCommand, ImageFormat, StopReason, ToolResultStatus } from "@aws-sdk/client-bedrock-runtime";
 import { NodeHttpHandler } from "@smithy/node-http-handler";
 import { expectDefined } from "openclaw/plugin-sdk/expect-runtime";
-import { AssistantMessageEventStream, adjustMaxTokensForThinking, buildBaseOptions, calculateCost, clampReasoning, createHttpProxyAgentsForTarget, parseStreamingJson, sanitizeSurrogates, transformMessages } from "openclaw/plugin-sdk/llm";
+import { AssistantMessageEventStream, adjustMaxTokensForThinking, buildBaseOptions, calculateCost, clampReasoning, createHttpProxyAgentsForTarget, createToolArgumentPreviewSchedule, parseStreamingJson, sanitizeSurrogates, transformMessages } from "openclaw/plugin-sdk/llm";
 import { canonicalizeBase64 } from "openclaw/plugin-sdk/media-runtime";
-import { describeToolResultMediaPlaceholder, failTransportStream, finalizeTerminalToolCallArguments, notifyProviderHttpMetadata } from "openclaw/plugin-sdk/provider-transport-runtime";
 //#region extensions/amazon-bedrock/stream.runtime.ts
 /**
 * Amazon Bedrock Converse streaming runtime. It maps OpenClaw messages/tools,
@@ -73,6 +73,7 @@ const streamBedrock = (model, context, options = {}) => {
 		};
 		const blocks = output.content;
 		const pendingToolCallEnds = [];
+		const toolArgumentPreviewSchedules = /* @__PURE__ */ new WeakMap();
 		const redactedReasoningChunks = /* @__PURE__ */ new Map();
 		const fable5 = usesClaudeFable5BedrockContract(model);
 		const refusalBuffer = usesClaudeStreamingRefusalBedrockContract(model) ? createDeferredEventBuffer(stream) : void 0;
@@ -105,14 +106,15 @@ const streamBedrock = (model, context, options = {}) => {
 		let client;
 		try {
 			client = new BedrockRuntimeClient(config);
-			const cacheRetention = resolveCacheRetention(options.cacheRetention);
+			const cacheRetention = resolveCacheRetention(model, options.cacheRetention);
+			const cachePoint = resolveBedrockCachePoint(model, cacheRetention);
 			const additionalModelRequestFields = buildAdditionalModelRequestFields(model, options);
 			const thinking = additionalModelRequestFields?.thinking;
 			const sendsAdaptiveThinking = thinking !== null && typeof thinking === "object" && thinking.type === "adaptive";
 			let commandInput = {
 				modelId: model.id,
-				messages: convertMessages(context, model, cacheRetention),
-				system: buildSystemPrompt(context.systemPrompt, model, cacheRetention),
+				messages: convertMessages(context, model, cachePoint),
+				system: buildSystemPrompt(context.systemPrompt, cacheRetention, cachePoint),
 				inferenceConfig: {
 					...options.maxTokens !== void 0 && { maxTokens: options.maxTokens },
 					...options.temperature !== void 0 && !sendsAdaptiveThinking && { temperature: options.temperature }
@@ -151,8 +153,8 @@ const streamBedrock = (model, context, options = {}) => {
 						type: "start",
 						partial: output
 					});
-				} else if (item.contentBlockStart) handleContentBlockStart(item.contentBlockStart, blocks, output, eventSink);
-				else if (item.contentBlockDelta) handleContentBlockDelta(item.contentBlockDelta, blocks, output, eventSink, redactedReasoningChunks);
+				} else if (item.contentBlockStart) handleContentBlockStart(item.contentBlockStart, blocks, output, eventSink, toolArgumentPreviewSchedules);
+				else if (item.contentBlockDelta) handleContentBlockDelta(item.contentBlockDelta, blocks, output, eventSink, redactedReasoningChunks, toolArgumentPreviewSchedules);
 				else if (item.contentBlockStop) handleContentBlockStop(item.contentBlockStop, blocks, output, eventSink, redactedReasoningChunks, pendingToolCallEnds);
 				else if (item.messageStop) {
 					sawMessageStop = true;
@@ -287,7 +289,7 @@ function resolveSimpleBedrockOptions(model, options) {
 		thinkingBudgets: options.thinkingBudgets
 	};
 }
-function handleContentBlockStart(event, blocks, output, stream) {
+function handleContentBlockStart(event, blocks, output, stream, toolArgumentPreviewSchedules) {
 	const index = event.contentBlockIndex;
 	const start = event.start;
 	if (start?.toolUse) {
@@ -300,6 +302,7 @@ function handleContentBlockStart(event, blocks, output, stream) {
 			partialJson: "",
 			index
 		};
+		toolArgumentPreviewSchedules.set(block, createToolArgumentPreviewSchedule());
 		output.content.push(block);
 		stream.push({
 			type: "toolcall_start",
@@ -308,7 +311,7 @@ function handleContentBlockStart(event, blocks, output, stream) {
 		});
 	}
 }
-function handleContentBlockDelta(event, blocks, output, stream, redactedReasoningChunks) {
+function handleContentBlockDelta(event, blocks, output, stream, redactedReasoningChunks, toolArgumentPreviewSchedules) {
 	const contentBlockIndex = event.contentBlockIndex;
 	const delta = event.delta;
 	let index = blocks.findIndex((b) => b.index === contentBlockIndex);
@@ -340,7 +343,7 @@ function handleContentBlockDelta(event, blocks, output, stream, redactedReasonin
 		}
 	} else if (delta?.toolUse && block?.type === "toolCall") {
 		block.partialJson = (block.partialJson || "") + (delta.toolUse.input || "");
-		block.arguments = parseStreamingJson(block.partialJson);
+		if (toolArgumentPreviewSchedules.get(block)?.(block.partialJson.length)) block.arguments = parseStreamingJson(block.partialJson);
 		stream.push({
 			type: "toolcall_delta",
 			contentIndex: index,
@@ -500,10 +503,11 @@ function mapThinkingLevelToEffort(model, level) {
 }
 /**
 * Resolve cache retention preference.
-* Defaults to "short" and uses OPENCLAW_CACHE_RETENTION for backward compatibility.
+* Nova requires explicit opt-in; other models retain the existing env/default policy.
 */
-function resolveCacheRetention(cacheRetention) {
+function resolveCacheRetention(model, cacheRetention) {
 	if (cacheRetention) return cacheRetention;
+	if (resolveBedrockPromptCachePolicy(model) === "nova") return "none";
 	if (typeof process !== "undefined" && process.env.OPENCLAW_CACHE_RETENTION === "long") return "long";
 	return "short";
 }
@@ -519,9 +523,6 @@ function isAnthropicClaudeModel(model) {
 	const name = model.name?.toLowerCase() ?? "";
 	return id.includes("anthropic.claude") || id.includes("anthropic/claude") || name.includes("anthropic.claude") || name.includes("anthropic/claude") || name.includes("claude");
 }
-function supportsPromptCaching(model) {
-	return usesClaudeFable5BedrockContract(model) || supportsBedrockPromptCaching(model.id, model.name) || supportsBedrockPromptCaching(resolveClaudeModelIdentity(model), model.name);
-}
 /**
 * Check if the model supports thinking signatures in reasoningContent.
 * Only Anthropic Claude models support the signature field.
@@ -533,14 +534,15 @@ function supportsPromptCaching(model) {
 function supportsThinkingSignature(model) {
 	return isAnthropicClaudeModel(model);
 }
-function buildSystemPrompt(systemPrompt, model, cacheRetention) {
+function buildSystemPrompt(systemPrompt, cacheRetention, cachePoint) {
 	if (!systemPrompt) return;
-	const blocks = [{ text: sanitizeSurrogates(systemPrompt) }];
-	if (cacheRetention !== "none" && supportsPromptCaching(model)) blocks.push({ cachePoint: {
-		type: CachePointType.DEFAULT,
-		...cacheRetention === "long" ? { ttl: CacheTTL.ONE_HOUR } : {}
-	} });
-	return blocks;
+	if (cacheRetention === "none") return [{ text: sanitizeSurrogates(stripSystemPromptCacheBoundary(systemPrompt)) }];
+	const split = splitSystemPromptCacheBoundary(systemPrompt);
+	const stablePrefix = split?.stablePrefix ?? systemPrompt;
+	const blocks = stablePrefix ? [{ text: sanitizeSurrogates(stripSystemPromptCacheBoundary(stablePrefix)) }] : [];
+	if (stablePrefix && cachePoint) blocks.push({ cachePoint });
+	if (split?.dynamicSuffix) blocks.push({ text: sanitizeSurrogates(stripSystemPromptCacheBoundary(split.dynamicSuffix)) });
+	return blocks.length > 0 ? blocks : void 0;
 }
 function normalizeToolCallId(id) {
 	const sanitized = id.replace(/[^a-zA-Z0-9_-]/g, "_");
@@ -561,7 +563,7 @@ function createBedrockToolResult(message) {
 		status: message.isError ? ToolResultStatus.ERROR : ToolResultStatus.SUCCESS
 	} };
 }
-function convertMessages(context, model, cacheRetention) {
+function convertMessages(context, model, cachePoint) {
 	const result = [];
 	let firstVolatileMessageIndex;
 	const transformedMessages = transformMessages(context.messages, model, normalizeToolCallId);
@@ -581,7 +583,7 @@ function convertMessages(context, model, cacheRetention) {
 					default: continue;
 				}
 				if (content.length === 0) continue;
-				if (m.runtimeContextCarrier === true && firstVolatileMessageIndex === void 0) firstVolatileMessageIndex = result.length;
+				if (m.runtimeContextCarrier === true && !bindsClaudeThinkingPrefix(model) && firstVolatileMessageIndex === void 0) firstVolatileMessageIndex = result.length;
 				result.push({
 					role: ConversationRole.USER,
 					content
@@ -654,12 +656,9 @@ function convertMessages(context, model, cacheRetention) {
 			default: continue;
 		}
 	}
-	if (cacheRetention !== "none" && supportsPromptCaching(model) && result.at(-1)?.role === ConversationRole.USER) {
+	if (cachePoint && result.at(-1)?.role === ConversationRole.USER) {
 		const cacheAnchor = result.findLast((message, index) => message.role === ConversationRole.USER && (firstVolatileMessageIndex === void 0 || index < firstVolatileMessageIndex));
-		if (cacheAnchor?.content) cacheAnchor.content.push({ cachePoint: {
-			type: CachePointType.DEFAULT,
-			...cacheRetention === "long" ? { ttl: CacheTTL.ONE_HOUR } : {}
-		} });
+		if (cacheAnchor?.content) cacheAnchor.content.push({ cachePoint });
 	}
 	return result;
 }

package/dist/thinking-policy.js

@@ -1,4 +1,4 @@
-import { resolveClaudeFable5ModelIdentity, resolveClaudeModelIdentity, resolveClaudeMythos5ModelIdentity, resolveClaudeOpus5ModelIdentity, resolveClaudeSonnet5ModelIdentity } from "openclaw/plugin-sdk/provider-model-shared";
+import { resolveClaudeFable5ModelIdentity, resolveClaudeModelIdentity, resolveClaudeMythos5ModelIdentity, resolveClaudeOpus5ModelIdentity, resolveClaudeSonnet5ModelIdentity } from "openclaw/plugin-sdk/claude-model-runtime";
 //#region extensions/amazon-bedrock/thinking-policy.ts
 const BASE_CLAUDE_THINKING_LEVELS = [
 	{ id: "off" },

package/node_modules/@aws-sdk/client-bedrock/dist-cjs/index.js

@@ -86,7 +86,7 @@ const commonParams = {
     UseDualStack: { type: "builtInParams", name: "useDualstackEndpoint" },
 };
 
-var version = "3.1116.0";
+var version = "3.1120.0";
 var packageInfo = {
 	version: version};
 

package/node_modules/@aws-sdk/client-bedrock/node_modules/@aws-sdk/token-providers/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aws-sdk/token-providers",
-  "version": "3.1117.0",
+  "version": "3.1121.0",
   "description": "A collection of token providers",
   "keywords": [
     "aws",

package/node_modules/@aws-sdk/client-bedrock/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aws-sdk/client-bedrock",
-  "version": "3.1117.0",
+  "version": "3.1121.0",
   "description": "AWS SDK for JavaScript Bedrock Client for Node.js, Browser and React Native",
   "homepage": "https://github.com/aws/aws-sdk-js-v3/tree/main/clients/client-bedrock",
   "license": "Apache-2.0",
@@ -48,7 +48,7 @@
   "dependencies": {
     "@aws-sdk/core": "^3.977.9",
     "@aws-sdk/credential-provider-node": "^3.972.81",
-    "@aws-sdk/token-providers": "3.1117.0",
+    "@aws-sdk/token-providers": "3.1121.0",
     "@aws-sdk/types": "^3.974.5",
     "@smithy/core": "^3.33.3",
     "@smithy/fetch-http-handler": "^5.7.2",

package/node_modules/@aws-sdk/client-bedrock-runtime/dist-cjs/index.js

@@ -90,7 +90,7 @@ const commonParams = {
     UseDualStack: { type: "builtInParams", name: "useDualstackEndpoint" },
 };
 
-var version = "3.1116.0";
+var version = "3.1120.0";
 var packageInfo = {
 	version: version};
 

package/node_modules/@aws-sdk/client-bedrock-runtime/node_modules/@aws-sdk/token-providers/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aws-sdk/token-providers",
-  "version": "3.1117.0",
+  "version": "3.1121.0",
   "description": "A collection of token providers",
   "keywords": [
     "aws",

package/node_modules/@aws-sdk/client-bedrock-runtime/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aws-sdk/client-bedrock-runtime",
-  "version": "3.1117.0",
+  "version": "3.1121.0",
   "description": "AWS SDK for JavaScript Bedrock Runtime Client for Node.js, Browser and React Native",
   "homepage": "https://github.com/aws/aws-sdk-js-v3/tree/main/clients/client-bedrock-runtime",
   "license": "Apache-2.0",
@@ -59,7 +59,7 @@
     "@aws-sdk/eventstream-handler-node": "^3.972.34",
     "@aws-sdk/middleware-eventstream": "^3.972.29",
     "@aws-sdk/middleware-websocket": "^3.972.52",
-    "@aws-sdk/token-providers": "3.1117.0",
+    "@aws-sdk/token-providers": "3.1121.0",
     "@aws-sdk/types": "^3.974.5",
     "@smithy/core": "^3.33.3",
     "@smithy/fetch-http-handler": "^5.7.2",

package/openclaw.plugin.json

@@ -1,5 +1,6 @@
 {
   "id": "amazon-bedrock",
+  "categories": ["models", "memory"],
   "name": "Amazon Bedrock",
   "description": "OpenClaw Amazon Bedrock provider plugin with model discovery, embeddings, and guardrail support.",
   "activation": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@openclaw/amazon-bedrock-provider",
-  "version": "2026.9.2",
+  "version": "2026.9.4",
   "description": "OpenClaw Amazon Bedrock provider plugin with model discovery, embeddings, and guardrail support.",
   "repository": {
     "type": "git",
@@ -8,8 +8,8 @@
   },
   "type": "module",
   "dependencies": {
-    "@aws-sdk/client-bedrock": "3.1117.0",
-    "@aws-sdk/client-bedrock-runtime": "3.1117.0",
+    "@aws-sdk/client-bedrock": "3.1121.0",
+    "@aws-sdk/client-bedrock-runtime": "3.1121.0",
     "@aws-sdk/credential-provider-node": "3.972.81",
     "@smithy/node-http-handler": "4.11.3",
     "@smithy/shared-ini-file-loader": "4.7.2",
@@ -25,10 +25,10 @@
       "minHostVersion": ">=2026.5.12-beta.1"
     },
     "compat": {
-      "pluginApi": ">=2026.9.2"
+      "pluginApi": ">=2026.9.4"
     },
     "build": {
-      "openclawVersion": "2026.9.2",
+      "openclawVersion": "2026.9.4",
       "bundledDist": false
     },
     "release": {
@@ -45,7 +45,7 @@
     "README.md"
   ],
   "peerDependencies": {
-    "openclaw": ">=2026.9.2"
+    "openclaw": ">=2026.9.4"
   },
   "peerDependenciesMeta": {
     "openclaw": {