@alfe.ai/gateway 0.9.13 → 0.9.15

package/dist/health.js

@@ -4533,6 +4533,7 @@ _enum(MOONSHOT_MODELS);
 _enum(QWEN_MODELS);
 string().trim().min(1);
 AnthropicModel.Opus5, AnthropicModel.Opus48, AnthropicModel.Opus47, AnthropicModel.Opus46, AnthropicModel.Sonnet5, AnthropicModel.Sonnet46, AnthropicModel.Haiku45, OpenAIModel.GPT4o, OpenAIModel.GPT4oMini, OpenAIModel.O3, OpenAIModel.GPT41, OpenAIModel.GPT41Mini, OpenAIModel.GPT41Nano, OpenAIModel.GPT54, OpenAIModel.GPT54Mini, OpenAIModel.GPT54Nano, OpenAIModel.GPT54Pro, OpenAIModel.GPT55, OpenAIModel.GPT55Pro, OpenAIModel.GPT56Sol, OpenAIModel.GPT56Terra, OpenAIModel.GPT56Luna, OpenAIModel.O3Mini, OpenAIModel.O4Mini, OpenAIModel.TextEmbedding3Small, OpenAIModel.TextEmbedding3Large, DeepSeekModel.Chat, DeepSeekModel.Reasoner, DeepSeekModel.V4Flash, DeepSeekModel.V4Pro, GoogleModel.Gemini35Flash, GoogleModel.Gemini31Pro, GoogleModel.Gemini31FlashLite, GoogleModel.Gemini25Pro, GoogleModel.Gemini25Flash, GoogleModel.Gemini25FlashLite, GoogleModel.Gemini20Flash, MiniMaxModel.M3, MiniMaxModel.M27, MiniMaxModel.M27HighSpeed, MiniMaxModel.M25, MiniMaxModel.M21, MiniMaxModel.M2, MistralModel.Large, MistralModel.Medium, MistralModel.Small, MistralModel.Codestral, MistralModel.Ministral8b, MistralModel.Ministral3b, MistralModel.MagistralMedium, MistralModel.MagistralSmall, MistralModel.DevstralMedium, XAIModel.Grok45, XAIModel.Grok43, XAIModel.Grok4, XAIModel.Grok41Fast, ZhipuModel.GLM52, ZhipuModel.GLM51, ZhipuModel.GLM46, ZhipuModel.GLM45, ZhipuModel.GLM45Air, MoonshotModel.K3, MoonshotModel.K26, MoonshotModel.K27Code, MoonshotModel.K27CodeHighSpeed, QwenModel.Qwen37Max, QwenModel.Qwen37Plus, QwenModel.Qwen36Flash, QwenModel.Qwen35Flash;
+({ Qwen3_8B: "alfe-qwen3-8b" }).Qwen3_8B;
 AnthropicModel.Opus5, AnthropicModel.Opus48, AnthropicModel.Opus47, AnthropicModel.Opus46, AnthropicModel.Sonnet5, AnthropicModel.Sonnet46, AnthropicModel.Haiku45, OpenAIModel.GPT4o, OpenAIModel.GPT4oMini, OpenAIModel.O3, OpenAIModel.GPT41, OpenAIModel.GPT41Mini, OpenAIModel.GPT41Nano, OpenAIModel.GPT54, OpenAIModel.GPT54Mini, OpenAIModel.GPT54Nano, OpenAIModel.GPT54Pro, OpenAIModel.GPT55, OpenAIModel.GPT55Pro, OpenAIModel.GPT56Sol, OpenAIModel.GPT56Terra, OpenAIModel.GPT56Luna, OpenAIModel.O3Mini, OpenAIModel.O4Mini, OpenAIModel.TextEmbedding3Small, OpenAIModel.TextEmbedding3Large, DeepSeekModel.Chat, DeepSeekModel.Reasoner, DeepSeekModel.V4Flash, DeepSeekModel.V4Pro, GoogleModel.Gemini35Flash, GoogleModel.Gemini31Pro, GoogleModel.Gemini31FlashLite, GoogleModel.Gemini25Pro, GoogleModel.Gemini25Flash, GoogleModel.Gemini25FlashLite, GoogleModel.Gemini20Flash, MiniMaxModel.M3, MiniMaxModel.M27, MiniMaxModel.M27HighSpeed, MiniMaxModel.M25, MiniMaxModel.M21, MiniMaxModel.M2, MistralModel.Large, MistralModel.Medium, MistralModel.Small, MistralModel.Codestral, MistralModel.Ministral8b, MistralModel.Ministral3b, MistralModel.MagistralMedium, MistralModel.MagistralSmall, MistralModel.DevstralMedium, XAIModel.Grok45, XAIModel.Grok43, XAIModel.Grok4, XAIModel.Grok41Fast, ZhipuModel.GLM52, ZhipuModel.GLM51, ZhipuModel.GLM46, ZhipuModel.GLM45, ZhipuModel.GLM45Air, MoonshotModel.K3, MoonshotModel.K26, MoonshotModel.K27Code, MoonshotModel.K27CodeHighSpeed, QwenModel.Qwen37Max, QwenModel.Qwen37Plus, QwenModel.Qwen36Flash, QwenModel.Qwen35Flash;
 enumValues({
 	Standard: "standard",
@@ -5397,6 +5398,24 @@ function deriveGatewayWsUrl(apiEndpoint) {
 		return "wss://gateway.alfe.ai/ws";
 	}
 }
+function validatePrivateManagedWsUrl(value, service) {
+	let url;
+	try {
+		url = new URL(value);
+	} catch {
+		throw new Error(`${service} WebSocket URL must be a valid regional service URL`);
+	}
+	if (url.protocol !== "ws:" || !url.hostname.endsWith(".local") && !url.hostname.endsWith(".sst") || url.pathname !== "/ws" || url.username !== "" || url.password !== "" || url.search !== "" || url.hash !== "") throw new Error(`${service} WebSocket URL must be a canonical VPC-local regional service URL`);
+	return url.toString();
+}
+function _resolveManagedGatewayWsUrl(apiEndpoint) {
+	const override = process.env.ALFE_GATEWAY_WS_URL?.trim();
+	if (process.env.ALFE_PRIVATE_DATA_PLANE === "true") {
+		if (!override) throw new Error("ALFE_GATEWAY_WS_URL is required in AWS-private managed mode");
+		return validatePrivateManagedWsUrl(override, "Gateway");
+	}
+	return override && override.length > 0 ? override : deriveGatewayWsUrl(apiEndpoint);
+}
 /**
 * Parse [runtimes.*] sections from config.toml.
 * Returns a map of runtime name → RuntimeConfig.
@@ -5471,7 +5490,7 @@ async function loadManagedConfig() {
 	const apiEndpoint = process.env.ALFE_API_ENDPOINT ?? getEndpointFromToken(apiKey);
 	logger$1.debug({ apiEndpoint }, "Environment loaded, resolving identity...");
 	const identity = await resolveAgentIdentity(apiKey, apiEndpoint);
-	const gatewayWsUrl = deriveGatewayWsUrl(apiEndpoint);
+	const gatewayWsUrl = _resolveManagedGatewayWsUrl(apiEndpoint);
 	logger$1.debug({
 		agentId: identity.agentId,
 		orgId: identity.orgId,
@@ -6950,6 +6969,15 @@ const SYSTEMD_SERVICE = "alfe-gateway";
 const MANAGED_ENV_VARS = [
 	"ALFE_MANAGED",
 	"ALFE_API_KEY",
+	"ALFE_API_ENDPOINT",
+	"ALFE_API_URL",
+	"ALFE_AI_PROXY_URL",
+	"ALFE_CHAT_WS_URL",
+	"ALFE_GATEWAY_WS_URL",
+	"ALFE_SYNC_RELAY_WS_URL",
+	"ALFE_PRIVATE_DATA_PLANE",
+	"ALFE_PRIVATE_DATA_PLANE_REGION",
+	"ALFE_PRIVATE_DATA_PLANE_RELEASE",
 	"LOG_LEVEL",
 	"ALFE_CLI_VERSION"
 ];
@@ -7992,12 +8020,18 @@ async function startAiProxy(apiKey) {
 	try {
 		const { createProxyServer, DEFAULT_AI_PROXY_PORT } = await import("@alfe.ai/ai-proxy-local");
 		const { getAiServiceUrlFromToken } = await import("@alfe.ai/config");
-		const proxyUrl = getAiServiceUrlFromToken(apiKey);
+		const privateDataPlane = process.env.ALFE_PRIVATE_DATA_PLANE === "true";
+		const proxyUrlOverride = process.env.ALFE_AI_PROXY_URL?.trim();
+		const configuredProxyUrl = proxyUrlOverride && proxyUrlOverride.length > 0 ? proxyUrlOverride : void 0;
+		if (privateDataPlane && configuredProxyUrl === void 0) throw new Error("AWS-private AI proxy target is not configured");
+		const proxyUrl = configuredProxyUrl ?? getAiServiceUrlFromToken(apiKey);
 		const port = DEFAULT_AI_PROXY_PORT ?? 18193;
 		handle.server = createProxyServer({
 			port,
 			apiKey,
-			proxyUrl
+			proxyUrl,
+			allowPrivateAwsTarget: privateDataPlane,
+			requirePrivateAwsTarget: privateDataPlane
 		});
 		const server = handle.server;
 		await new Promise((resolve, reject) => {
@@ -23450,6 +23484,51 @@ function createLogger(component, additionalData) {
 		component
 	});
 }
+/**
+* Strict profile for the HIGH-VOLUME, LOW-SIGNAL families: MCP
+* `tool-result-error` and the plugin `[ERROR] alfe-tool … result-error:` line.
+* These are the NORMAL model-facing error path — a model omitting a required
+* argument produces one every turn — so a single agent can hold an issue open
+* indefinitely at the default budget.
+*
+* We keep sending them (continued visibility is the point); we just send far
+* fewer. Numbers and why:
+*
+* - `dedupeWindowMs: 6h` — a repeating fingerprint reports at most 4×/day per
+*   agent instead of the default 144×/day. This is the lever that actually
+*   collapses the observed burn, because the noisy issues repeat one message
+*   shape over and over.
+* - `errorCapturesMax: 2` per `errorCapturesWindowMs: 1h` — the family ceiling
+*   drops from 720 events/day/agent (5 per 10 min) to 48. An hourly window,
+*   rather than folding the budget into the 6h dedupe, keeps a NEWLY-failing
+*   tool visible within the hour instead of behind a 6h blackout.
+* - `dedupeMapMaxKeys: 200` — the dedupe map must now retain keys 36× longer,
+*   so 50 slots would evict still-live entries on a wide tool surface and let
+*   them re-fire. 200 covers every tool an agent realistically exposes.
+*
+* These are PER-DAEMON-PROCESS (i.e. per agent VM) budgets; fleet-wide volume
+* scales with agent count and is unbounded from here.
+*
+* Suppression is not silent, but it is COARSE: for the error family,
+* `errorsSuppressed` is one counter per throttle INSTANCE, incremented across
+* every fingerprint and reset on each allowed capture (see
+* {@link CaptureThrottle.allowErrorCapture}). The tag therefore reads "error
+* events this instance suppressed since its last allowed error capture",
+* landing on whichever issue happened to be the one let through — it is NOT
+* that issue's own occurrence count. Use it as a family noise-level signal,
+* not as per-issue volume.
+*
+* The crash family is accounted SEPARATELY via `crashesSuppressed`
+* ({@link CaptureThrottle.allowCrashCapture}). A suppressed error never
+* contributes to the next allowed crash capture, or vice versa, even on a
+* shared instance.
+*/
+const STRICT_ERROR_CAPTURE_PROFILE = Object.freeze({
+	errorCapturesMax: 2,
+	errorCapturesWindowMs: 60 * 6e4,
+	dedupeWindowMs: 360 * 6e4,
+	dedupeMapMaxKeys: 200
+});
 const ANSI_RE = /\x1b\[[0-9;]*[A-Za-z]/g;
 function stripAnsi(line) {
 	return line.replace(ANSI_RE, "");
@@ -23561,8 +23640,32 @@ var ErrorLineDetector = class {
 	}
 };
 /**
-* Bounds Sentry quota per RuntimeProcess. Worst case per agent:
+* Exact shape emitted by `installToolErrorCapture` in
+* `@alfe.ai/agent-api-client` for a tool that RETURNED an error result:
+*
+*   [ERROR] alfe-tool plugin=<plugin> tool=<name> result-error: <msg>
+*
+* `plugin` / `tool` are `safeToken`-sanitised there (`[A-Za-z0-9_.@/-]`), so
+* they never contain whitespace. Deliberately anchored and `result-error`-only:
+* the sibling `thrown:` kind is a genuine unhandled exception and keeps the
+* default budget, as does any other `[ERROR]` line or runtime stack trace.
+*/
+const ALFE_TOOL_RESULT_ERROR_LINE = /^\[ERROR\] alfe-tool plugin=[A-Za-z0-9_.@/-]+ tool=[A-Za-z0-9_.@/-]+ result-error:/;
+/**
+* True for the plugin tool-`result-error` line above — the runtime-side twin of
+* MCP's `tool-result-error`, i.e. the NORMAL model-facing failure path rather
+* than a runtime fault. Callers route these to a strict
+* {@link STRICT_ERROR_CAPTURE_PROFILE} throttle so genuine crashes and stderr
+* keep the default budget.
+*/
+function isPluginToolResultErrorLine(line) {
+	return ALFE_TOOL_RESULT_ERROR_LINE.test(stripAnsi(line));
+}
+/**
+* Bounds Sentry quota per instance. Worst case at the default limits:
 * ≤ `ERROR_CAPTURES_MAX` output events + ≤ 1 crash event per 10-minute window.
+* Pass {@link CaptureThrottleOptions} (e.g. {@link STRICT_ERROR_CAPTURE_PROFILE})
+* to give a noisier family its own, tighter budget.
 */
 var CaptureThrottle = class {
 	windowStart = 0;
@@ -23572,16 +23675,24 @@ var CaptureThrottle = class {
 	lastCrashCaptureAt = 0;
 	hasCapturedCrash = false;
 	crashesSuppressed = 0;
-	constructor(stableUptimeMs) {
+	errorCapturesMax;
+	errorCapturesWindowMs;
+	dedupeWindowMs;
+	dedupeMapMaxKeys;
+	constructor(stableUptimeMs, options = {}) {
 		this.stableUptimeMs = stableUptimeMs;
+		this.errorCapturesMax = options.errorCapturesMax ?? 5;
+		this.errorCapturesWindowMs = options.errorCapturesWindowMs ?? 6e5;
+		this.dedupeWindowMs = options.dedupeWindowMs ?? 6e5;
+		this.dedupeMapMaxKeys = options.dedupeMapMaxKeys ?? 50;
 	}
 	allowErrorCapture(fingerprintKey, now = Date.now()) {
-		if (now - this.windowStart >= 6e5) {
+		if (now - this.windowStart >= this.errorCapturesWindowMs) {
 			this.windowStart = now;
 			this.windowCount = 0;
 		}
 		const lastSent = this.lastSentByKey.get(fingerprintKey);
-		if (lastSent !== void 0 && now - lastSent < 6e5 || this.windowCount >= 5) {
+		if (lastSent !== void 0 && now - lastSent < this.dedupeWindowMs || this.windowCount >= this.errorCapturesMax) {
 			this.errorsSuppressed++;
 			return {
 				allow: false,
@@ -23590,7 +23701,7 @@ var CaptureThrottle = class {
 		}
 		this.windowCount++;
 		this.lastSentByKey.set(fingerprintKey, now);
-		if (this.lastSentByKey.size > 50) {
+		if (this.lastSentByKey.size > this.dedupeMapMaxKeys) {
 			const oldest = this.lastSentByKey.keys().next().value;
 			if (oldest !== void 0) this.lastSentByKey.delete(oldest);
 		}
@@ -23653,6 +23764,7 @@ var RuntimeProcess = class {
 	ringBuffer = new OutputRingBuffer();
 	detector = new ErrorLineDetector();
 	throttle = new CaptureThrottle(STABLE_UPTIME_MS);
+	toolResultErrorThrottle = new CaptureThrottle(STABLE_UPTIME_MS, STRICT_ERROR_CAPTURE_PROFILE);
 	detectorFlushTimer = null;
 	turnActivityProbe = null;
 	constructor(options) {
@@ -23837,7 +23949,7 @@ var RuntimeProcess = class {
 	/** Report a completed error block to Sentry, throttle permitting. */
 	emitErrorBlock(block) {
 		if (!block) return;
-		const { allow, suppressedCount } = this.throttle.allowErrorCapture(block.fingerprintKey);
+		const { allow, suppressedCount } = (isPluginToolResultErrorLine(block.lines[0]) ? this.toolResultErrorThrottle : this.throttle).allowErrorCapture(block.fingerprintKey);
 		if (!allow) {
 			log$3.debug({
 				runtime: this.options.runtime,
@@ -24563,7 +24675,7 @@ const log = createLogger("McpErrorCapture");
 /** Quiet-period flush for a pending multi-line stderr block. */
 const MCP_STDERR_FLUSH_DEBOUNCE_MS = 1500;
 function createMcpErrorHooks() {
-	const resultErrorThrottle = new CaptureThrottle(6e4);
+	const resultErrorThrottle = new CaptureThrottle(6e4, STRICT_ERROR_CAPTURE_PROFILE);
 	const errorThrottle = new CaptureThrottle(6e4);
 	const crashThrottle = new CaptureThrottle(6e4);
 	const throttleFor = (kind) => kind === "tool-result-error" ? resultErrorThrottle : kind === "server-crash" ? crashThrottle : errorThrottle;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@alfe.ai/gateway",
-  "version": "0.9.13",
+  "version": "0.9.15",
   "description": "Alfe local gateway daemon — persistent control plane for agent integrations",
   "type": "module",
   "bin": {
@@ -23,9 +23,9 @@
     "pino-roll": "^1.2.0",
     "smol-toml": ">=1.6.1",
     "ws": "^8.18.0",
-    "@alfe.ai/agent-api-client": "^0.17.0",
-    "@alfe.ai/ai-proxy-local": "^0.0.16",
+    "@alfe.ai/ai-proxy-local": "^0.0.17",
     "@alfe.ai/config": "^0.4.1",
+    "@alfe.ai/agent-api-client": "^0.17.1",
     "@alfe.ai/integration-manifest": "^0.4.0",
     "@alfe.ai/integrations": "^0.6.4",
     "@alfe.ai/mcp-bundler": "^0.4.1"