@openclaw/amazon-bedrock-provider 2026.10.5-beta.3 → 2026.10.10-beta.1

package/dist/api.js

@@ -1,2 +1 @@
-import { discoverBedrockModels, resolveImplicitBedrockProvider } from "./discovery.js";
-export { discoverBedrockModels, resolveImplicitBedrockProvider };
+export { discoverBedrockModels, resolveImplicitBedrockProvider } from "./discovery.js";

package/dist/bedrock-options.js

@@ -1,31 +1,29 @@
 import { resolveClaudeModelIdentity } from "openclaw/plugin-sdk/provider-model-shared";
 //#region extensions/amazon-bedrock/bedrock-options.ts
+/** Nova requires explicit opt-in; other models retain the existing env/default policy. */
+function resolveBedrockCacheRetention(policy, cacheRetention) {
+	if (cacheRetention) return cacheRetention;
+	if (policy === "nova") return "none";
+	if (typeof process !== "undefined" && process.env.OPENCLAW_CACHE_RETENTION === "long") return "long";
+	return "short";
+}
 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$/.test(modelId)) return "nova";
-	if (supportsBedrockClaudePromptCaching(model.id, model.name) || supportsBedrockClaudePromptCaching(resolveClaudeModelIdentity(model), model.name)) return "claude";
+	return resolveBedrockClaudeCachePolicy(resolveClaudeModelIdentity(model), model.name);
 }
-function resolveBedrockCachePoint(model, retention) {
-	const policy = resolveBedrockPromptCachePolicy(model);
+function resolveBedrockCachePoint(policy, retention) {
 	if (!policy || retention === "none") return;
 	return {
 		type: "default",
 		...policy === "claude" && retention === "long" ? { ttl: "1h" } : {}
 	};
 }
-function getModelMatchCandidates(modelId, modelName) {
-	return (modelName ? [modelId, modelName] : [modelId]).flatMap((value) => {
-		const lower = value.toLowerCase();
-		return [lower, lower.replace(/[\s_.:]+/g, "-")];
-	});
-}
-function supportsBedrockClaudePromptCaching(modelId, modelName) {
-	const candidates = getModelMatchCandidates(modelId, modelName);
-	if (!candidates.some((s) => s.includes("claude"))) {
-		if (typeof process !== "undefined" && process.env.AWS_BEDROCK_FORCE_CACHE === "1") return true;
-		return false;
-	}
-	return candidates.some((candidate) => candidate.includes("-4-") || /claude-(?:fable-5|mythos-5|opus-5|sonnet-5|3-7-sonnet|3-5-haiku)/.test(candidate));
+function resolveBedrockClaudeCachePolicy(modelId, modelName) {
+	const candidates = [modelId, modelName].map((id) => resolveClaudeModelIdentity({ id }));
+	if (!candidates.some((s) => s.includes("claude"))) return typeof process !== "undefined" && process.env.AWS_BEDROCK_FORCE_CACHE === "1" ? "claude-short" : void 0;
+	if (candidates.some((candidate) => /^claude-(?:haiku-(?:4-5|5-5)|sonnet-(?:4-[56]|5(?:-5)?)|opus-(?:4-[5-8]|5(?:-5)?)|(?:fable|mythos)-5(?:-1)?)(?:$|-v\d|-\d{8}(?:-|$))/.test(candidate))) return "claude";
+	return candidates.some((candidate) => candidate.includes("-4-") || /^claude-(?:fable-5|mythos-5|opus-5|sonnet-5|haiku-5|3-7-sonnet|3-5-haiku)/.test(candidate)) ? "claude-short" : void 0;
 }
 //#endregion
-export { resolveBedrockCachePoint, resolveBedrockPromptCachePolicy, supportsBedrockClaudePromptCaching };
+export { resolveBedrockCachePoint, resolveBedrockCacheRetention, resolveBedrockPromptCachePolicy };

package/dist/config-compat.js

@@ -1,2 +1 @@
-import { t as migrateAmazonBedrockLegacyConfig } from "./.setup/config-compat-HlsW2HRS.mjs";
-export { migrateAmazonBedrockLegacyConfig };
+export { t as migrateAmazonBedrockLegacyConfig } from "./.setup/config-compat-HlsW2HRS.mjs";

package/dist/memory-embedding-adapter.js

@@ -1,5 +1,4 @@
 import { DEFAULT_BEDROCK_EMBEDDING_MODEL, createBedrockEmbeddingProvider, hasAwsCredentials } from "./embedding-provider.js";
-import { isMissingEmbeddingApiKeyError } from "openclaw/plugin-sdk/embedding-provider-adapter";
 //#region extensions/amazon-bedrock/memory-embedding-adapter.ts
 const bedrockMemoryEmbeddingProviderAdapter = {
 	id: "bedrock",
@@ -7,8 +6,6 @@ const bedrockMemoryEmbeddingProviderAdapter = {
 	transport: "remote",
 	authProviderId: "amazon-bedrock",
 	autoSelectPriority: 60,
-	allowExplicitWhenConfiguredAuto: true,
-	shouldContinueAutoSelection: isMissingEmbeddingApiKeyError,
 	create: async (options) => {
 		if (!await hasAwsCredentials()) throw new Error("No API key found for provider \"bedrock\". AWS credentials are not available. Set AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY, AWS_PROFILE, or AWS_BEARER_TOKEN_BEDROCK, configure an EC2/ECS/EKS role, or set memory.search.provider to another provider.");
 		const { provider, client } = await createBedrockEmbeddingProvider({

package/dist/register.sync.runtime.js

@@ -1,4 +1,4 @@
-import { resolveBedrockPromptCachePolicy, supportsBedrockClaudePromptCaching } from "./bedrock-options.js";
+import { resolveBedrockCachePoint, resolveBedrockPromptCachePolicy } from "./bedrock-options.js";
 import { loadBedrockControlPlaneSdk, runBedrockControlPlaneRequest } from "./control-plane.js";
 import { bedrockMemoryEmbeddingProviderAdapter } from "./memory-embedding-adapter.js";
 import { isLatestAdaptiveBedrockModelRef, isOpus47OrNewerBedrockModelRef, resolveBedrockClaudeThinkingProfile, resolveBedrockNativeThinkingLevelMap, supportsBedrockNativeMaxEffort } from "./thinking-policy.js";
@@ -106,14 +106,14 @@ async function resolveAppProfileTraits(modelId, fallbackRegion, signal) {
 		const controlPlaneClient = sdk.createClient(region);
 		client = controlPlaneClient;
 		const command = sdk.createGetInferenceProfileCommand({ inferenceProfileIdentifier: modelId });
-		const models = (await runBedrockControlPlaneRequest({
+		const modelArns = ((await runBedrockControlPlaneRequest({
 			operation: "Bedrock GetInferenceProfile",
 			signal,
 			send: (options) => controlPlaneClient.send(command, options)
-		})).models ?? [];
-		const modelArns = models.map((model) => model.modelArn ?? "");
+		})).models ?? []).map((model) => model.modelArn ?? "");
+		const cachePolicies = modelArns.map((id) => resolveBedrockPromptCachePolicy({ id }));
 		const traits = {
-			cacheEligible: models.length > 0 && modelArns.every((modelArn) => supportsBedrockClaudePromptCaching(modelArn)),
+			cachePolicy: cachePolicies.length > 0 && cachePolicies.every((policy) => policy === "claude" || policy === "claude-short") ? cachePolicies.includes("claude-short") ? "claude-short" : "claude" : void 0,
 			omitTemperature: modelArns.some(isOpus47OrNewerBedrockModelRef)
 		};
 		appProfileTraitsCache.set(modelId, traits);
@@ -121,7 +121,7 @@ async function resolveAppProfileTraits(modelId, fallbackRegion, signal) {
 	} catch {
 		signal?.throwIfAborted();
 		return {
-			cacheEligible: isAnthropicBedrockModel(modelId),
+			cachePolicy: isAnthropicBedrockModel(modelId) ? "claude-short" : void 0,
 			omitTemperature: isOpus47OrNewerBedrockModelRef(modelId)
 		};
 	} finally {
@@ -135,12 +135,9 @@ function hasCachePoint(blocks) {
 * 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, context, model) {
-	if (!cacheRetention || cacheRetention === "none" || resolveBedrockPromptCachePolicy(model)) return;
-	const point = { cachePoint: {
-		type: "default",
-		...cacheRetention === "long" ? { ttl: "1h" } : {}
-	} };
+function injectBedrockCachePoints(payload, cachePoint, context) {
+	if (!cachePoint) return;
+	const point = { cachePoint };
 	const system = payload.system;
 	if (Array.isArray(system) && system.length > 0 && !hasCachePoint(system)) {
 		const split = context.systemPrompt && splitSystemPromptCacheBoundary(context.systemPrompt);
@@ -149,8 +146,14 @@ function injectBedrockCachePoints(payload, cacheRetention, context, model) {
 	if (context.messages.some(hasRuntimeContextMarker)) return;
 	const messages = payload.messages;
 	if (Array.isArray(messages)) {
-		const userContent = messages.toReversed().find((msg) => msg.role === "user" && Array.isArray(msg.content))?.content;
-		if (userContent && !hasCachePoint(userContent)) userContent.push(point);
+		let remaining = 2;
+		for (let index = messages.length - 1; index >= 0; index--) {
+			const message = messages[index];
+			if (message?.role === "user" && Array.isArray(message.content)) {
+				if (!hasCachePoint(message.content)) message.content.push(point);
+				if (--remaining === 0) break;
+			}
+		}
 	}
 }
 function patchMaxThinkingEffort(payload) {
@@ -252,7 +255,7 @@ function registerAmazonBedrockPlugin(api) {
 				} else wrapped = createBedrockServiceTierWrapper(wrapped, serviceTier);
 			}
 			const region = resolveBedrockRegion(config, provider) ?? extractRegionFromBaseUrl(model?.baseUrl) ?? currentPluginConfig?.discovery?.region;
-			const mayNeedCacheInjection = isBedrockAppInferenceProfile(modelId) && !supportsBedrockClaudePromptCaching(modelId);
+			const mayNeedCacheInjection = isBedrockAppInferenceProfile(modelId) && !resolveBedrockPromptCachePolicy(modelRef);
 			const shouldOmitTemperature = opus47OrNewer || fable5 || isLatestAdaptiveBedrockModelRef(modelId, model?.params);
 			const shouldPatchMaxThinking = supportsNativeMax && thinkingLevel === "max";
 			const shouldPatchPayload = shouldOmitTemperature || shouldPatchMaxThinking;
@@ -273,16 +276,16 @@ function registerAmazonBedrockPlugin(api) {
 						return originalOnPayload?.(payload, payloadModel);
 					} } : {}
 				});
-				const cacheRetention = typeof merged.cacheRetention === "string" ? merged.cacheRetention : "short";
+				const cacheRetention = merged.cacheRetention ?? "short";
 				return underlying(streamModel, context, {
 					...merged,
 					onPayload: async (payload, payloadModel) => {
-						const traits = heuristicMatch ? void 0 : await resolveAppProfileTraits(modelId, region, merged.signal);
+						const traits = await resolveAppProfileTraits(modelId, region, merged.signal);
 						if (payload && typeof payload === "object") {
 							const payloadRecord = payload;
-							if (heuristicMatch || traits?.cacheEligible) injectBedrockCachePoints(payloadRecord, cacheRetention, context, streamModel);
+							injectBedrockCachePoints(payloadRecord, resolveBedrockCachePoint(traits.cachePolicy, cacheRetention), context);
 							if (shouldPatchMaxThinking) patchMaxThinkingEffort(payloadRecord);
-							if (heuristicMatch ? shouldOmitTemperature || "temperature" in merged && (await resolveAppProfileTraits(modelId, region, merged.signal)).omitTemperature : traits?.omitTemperature) omitUnsupportedClaudePayloadTemperature(payloadRecord);
+							if (shouldOmitTemperature || traits.omitTemperature) omitUnsupportedClaudePayloadTemperature(payloadRecord);
 						}
 						return originalOnPayload?.(payload, payloadModel);
 					}

package/dist/stream.runtime.js

@@ -1,4 +1,4 @@
-import { resolveBedrockCachePoint, resolveBedrockPromptCachePolicy } from "./bedrock-options.js";
+import { resolveBedrockCachePoint, resolveBedrockCacheRetention, resolveBedrockPromptCachePolicy } from "./bedrock-options.js";
 import { resolveBedrockRuntimeAuth } from "./aws-credential-refresh.js";
 import { isClaude5BedrockModel, resolveBedrockClaudeThinkingProfile, supportsBedrockNativeMaxEffort } from "./thinking-policy.js";
 import { AssistantMessageEventStream, adjustMaxTokensForThinking, buildBaseOptions, calculateCost, clampReasoning, createHttpProxyAgentsForTarget, createToolArgumentPreviewSchedule, hasRuntimeContextMarker, parseStreamingJson, sanitizeSurrogates, transformMessages } from "openclaw/plugin-sdk/llm";
@@ -75,8 +75,9 @@ const streamBedrock = (model, context, options = {}) => {
 		try {
 			options.signal?.throwIfAborted();
 			client = new BedrockRuntimeClient(config);
-			const cacheRetention = resolveCacheRetention(model, options.cacheRetention);
-			const cachePoint = resolveBedrockCachePoint(model, cacheRetention);
+			const cachePolicy = resolveBedrockPromptCachePolicy(model);
+			const cacheRetention = resolveBedrockCacheRetention(cachePolicy, options.cacheRetention);
+			const cachePoint = resolveBedrockCachePoint(cachePolicy, cacheRetention);
 			const additionalModelRequestFields = buildAdditionalModelRequestFields(model, options);
 			const thinking = additionalModelRequestFields?.thinking;
 			const sendsAdaptiveThinking = thinking !== null && typeof thinking === "object" && thinking.type === "adaptive";
@@ -161,10 +162,7 @@ const streamBedrock = (model, context, options = {}) => {
 				error,
 				cleanup: () => {
 					output.content = output.content.filter((block) => block.type !== "toolCall");
-					for (const block of output.content) {
-						delete block.index;
-						delete block.partialJson;
-					}
+					for (const block of output.content) delete block.index;
 					if (refusalBuffer) {
 						refusalBuffer.discard();
 						output.content = [];
@@ -452,16 +450,6 @@ function mapThinkingLevelToEffort(model, level) {
 	}
 }
 /**
-* Resolve cache retention preference.
-* Nova requires explicit opt-in; other models retain the existing env/default policy.
-*/
-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";
-}
-/**
 * Check if the model is an Anthropic Claude model on Bedrock.
 * Checks both model ID and model name to support application inference profiles
 * whose ARNs don't contain the model name.
@@ -604,8 +592,14 @@ function convertMessages(context, model, cachePoint) {
 		}
 	}
 	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 });
+		let remaining = 2;
+		for (let index = (firstVolatileMessageIndex ?? result.length) - 1; index >= 0; index--) {
+			const message = result[index];
+			if (message?.role === ConversationRole.USER && message.content) {
+				message.content.push({ cachePoint });
+				if (--remaining === 0) break;
+			}
+		}
 	}
 	return result;
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@openclaw/amazon-bedrock-provider",
-  "version": "2026.10.5-beta.3",
+  "version": "2026.10.10-beta.1",
   "description": "OpenClaw Amazon Bedrock provider plugin with model discovery, embeddings, and guardrail support.",
   "repository": {
     "type": "git",
@@ -25,10 +25,10 @@
       "minHostVersion": ">=2026.5.12-beta.1"
     },
     "compat": {
-      "pluginApi": ">=2026.10.5-beta.3"
+      "pluginApi": ">=2026.10.10-beta.1"
     },
     "build": {
-      "openclawVersion": "2026.10.5-beta.3",
+      "openclawVersion": "2026.10.10-beta.1",
       "bundledDist": false
     },
     "release": {
@@ -47,7 +47,7 @@
     "assets/activity.svg"
   ],
   "peerDependencies": {
-    "openclaw": ">=2026.10.5-beta.3"
+    "openclaw": ">=2026.10.10-beta.1"
   },
   "peerDependenciesMeta": {
     "openclaw": {