@symerian/symi 2.6.0 → 2.6.1

package/dist/{status-DHJLMwQN.js → status--iNVOTMO.js}

@@ -1,6 +1,6 @@
 import { u as resolveGatewayPort } from "./paths-Cqn-zk3M.js";
 import { B as theme, k as info, z as isRich } from "./utils-B-0b9bGM.js";
-import { B as summarizeRestartSentinel, Jt as sha256HexPrefix, _t as formatUsageReportLines, a as buildChannelSummary, gt as loadProviderUsageSummary, h as listAgentsForGateway, m as classifySessionKey, o as buildChannelAccountSnapshot, pr as lookupContextTokens, s as formatChannelAllowFrom, wr as peekSystemEvents, x as resolveSessionModelRef, z as readRestartSentinel } from "./reply-D40cmAci.js";
+import { B as summarizeRestartSentinel, Jt as sha256HexPrefix, _t as formatUsageReportLines, a as buildChannelSummary, gt as loadProviderUsageSummary, h as listAgentsForGateway, m as classifySessionKey, o as buildChannelAccountSnapshot, pr as lookupContextTokens, s as formatChannelAllowFrom, wr as peekSystemEvents, x as resolveSessionModelRef, z as readRestartSentinel } from "./reply-DYnTEYoa.js";
 import { S as parseAgentSessionKey } from "./session-key-DCt45XZa.js";
 import { t as resolveSymiPackageRoot } from "./symi-root-CrGJbkzf.js";
 import { n as runExec } from "./exec-CWkblSrI.js";
@@ -28,7 +28,7 @@ import { t as formatRuntimeStatusWithDetails } from "./runtime-status-CR9445g5.j
 import { t as readLastGatewayErrorLine } from "./diagnostics-BAMlsVVX.js";
 import { t as renderTable } from "./table-BTgkRafz.js";
 import { a as resolveGatewayProbeAuth$1, c as probeGateway, t as runSecurityAudit } from "./audit-CrITRV6w.js";
-import { a as resolveHeartbeatSummaryForAgent, t as formatHealthChannelLines } from "./health-DK6rAOhC.js";
+import { a as resolveHeartbeatSummaryForAgent, t as formatHealthChannelLines } from "./health-BjwDRAdd.js";
 import { g as resolveUpdateChannelDisplay, i as formatGitInstallLabel, m as normalizeUpdateChannel, t as checkUpdateStatus } from "./update-check-CtckACbb.js";
 import { t as resolveNodeService } from "./node-service-Cxz4e-Qd.js";
 import { n as redactSecrets, t as formatGatewayAuthUsed } from "./format-yQZNwAF2.js";

package/dist/{status-CtNKWuzg.js → status-B1_iHrOg.js}

@@ -9,7 +9,7 @@ import { c as resolveAgentWorkspaceDir } from "./agent-scope-D-jRCY0d.js";
 import { t as buildWorkspaceSkillStatus } from "./skills-status-DKXJ-tbi.js";
 import { H as VERSION, i as loadConfig, o as readConfigFileSnapshot } from "./config-DHBLS1Hl.js";
 import { n as callGateway, t as buildGatewayConnectionDetails } from "./call-BcE47FtD.js";
-import { En as peekSystemEvents, St as summarizeRestartSentinel, cn as buildChannelAccountSnapshot, in as formatUsageReportLines, ln as formatChannelAllowFrom, nt as sha256HexPrefix, rn as loadProviderUsageSummary, sn as buildChannelSummary, xt as readRestartSentinel } from "./subagent-registry-Cb5e_x99.js";
+import { En as peekSystemEvents, St as summarizeRestartSentinel, cn as buildChannelAccountSnapshot, in as formatUsageReportLines, ln as formatChannelAllowFrom, nt as sha256HexPrefix, rn as loadProviderUsageSummary, sn as buildChannelSummary, xt as readRestartSentinel } from "./subagent-registry-CXrOOgPW.js";
 import { F as resolveMainSessionKey, j as resolveFreshSessionTotalTokens, o as loadSessionStore } from "./sessions-CJXnZVjR.js";
 import { n as listChannelPlugins } from "./plugins-CwSlLxM8.js";
 import { o as getTailnetHostname, s as readTailscaleStatusJson } from "./tailscale-CbbvYNVw.js";
@@ -29,7 +29,7 @@ import { t as formatRuntimeStatusWithDetails } from "./runtime-status-hFVEC3wO.j
 import { t as readLastGatewayErrorLine } from "./diagnostics-CS1ov_hH.js";
 import { t as renderTable } from "./table-D01d2GuY.js";
 import { a as resolveGatewayProbeAuth$1, c as probeGateway, t as runSecurityAudit } from "./audit-Byo5jCLN.js";
-import { o as resolveHeartbeatSummaryForAgent, t as formatHealthChannelLines } from "./health-BpHgCv-u.js";
+import { o as resolveHeartbeatSummaryForAgent, t as formatHealthChannelLines } from "./health-C5XJPwpt.js";
 import { g as resolveUpdateChannelDisplay, i as formatGitInstallLabel, m as normalizeUpdateChannel, t as checkUpdateStatus } from "./update-check-ZdimP1aU.js";
 import { t as resolveNodeService } from "./node-service-fcZExd22.js";
 import { n as redactSecrets, t as formatGatewayAuthUsed } from "./format-DVLB9DNB.js";

package/dist/{status-BSMEjz4q.js → status-BQcdARV4.js}

@@ -1,7 +1,7 @@
 import { o as createSubsystemLogger } from "./entry.js";
 import { D as resolveDefaultAgentWorkspaceDir, c as resolveAgentWorkspaceDir, l as resolveDefaultAgentId } from "./agent-scope-D-jRCY0d.js";
 import { i as loadConfig } from "./config-DHBLS1Hl.js";
-import { m as loadSymiPlugins, p as createPluginLoaderLogger } from "./subagent-registry-Cb5e_x99.js";
+import { m as loadSymiPlugins, p as createPluginLoaderLogger } from "./subagent-registry-CXrOOgPW.js";
 
 //#region src/plugins/status.ts
 const log = createSubsystemLogger("plugins");

package/dist/{status-Kv_hsY8N.js → status-DiX0DAtH.js}

@@ -1,4 +1,4 @@
-import { dt as createPluginLoaderLogger, ft as loadSymiPlugins } from "./reply-D40cmAci.js";
+import { dt as createPluginLoaderLogger, ft as loadSymiPlugins } from "./reply-DYnTEYoa.js";
 import { t as createSubsystemLogger } from "./subsystem-D9vIQve0.js";
 import { D as resolveDefaultAgentWorkspaceDir, c as resolveAgentWorkspaceDir, l as resolveDefaultAgentId } from "./agent-scope-CgUHAtCo.js";
 import { i as loadConfig } from "./config-CHwyw6l5.js";

package/dist/{subagent-registry-Cb5e_x99.js → subagent-registry-CXrOOgPW.js}

@@ -32327,7 +32327,8 @@ const BUILTIN_PROFILES = [
 				"<end_of_turn>"
 			],
 			suppressMonologue: true,
-			hasStructuredThinking: false
+			hasStructuredThinking: false,
+			validateToolArgs: true
 		},
 		promptAdditions: [
 			"CRITICAL OUTPUT RULES:",
@@ -32337,7 +32338,8 @@ const BUILTIN_PROFILES = [
 			"- Do not write verification checklists, self-assessments, or numbered assessment lists after your answer.",
 			"- Do not write 'Verification:', 'Requirements:', 'Status Check:', or 'Everything is correct'.",
 			"- If you need to verify your work, do so before writing your response, not after.",
-			"- Send brief progress updates on long tasks so the user knows you are active."
+			"- Send brief progress updates on long tasks so the user knows you are active.",
+			"- When calling tools, provide arguments as clean values only. Never include <|, |>, or XML-like tags in tool argument values. For exec/bash commands, provide only valid shell syntax."
 		],
 		ui: {
 			badge: "Local",
@@ -32361,7 +32363,8 @@ const BUILTIN_PROFILES = [
 		filters: {
 			stripPatterns: [],
 			suppressMonologue: false,
-			hasStructuredThinking: true
+			hasStructuredThinking: true,
+			validateToolArgs: false
 		},
 		promptAdditions: [],
 		ui: {
@@ -32395,7 +32398,8 @@ const BUILTIN_PROFILES = [
 				"^\\s*<bos>\\s*"
 			],
 			suppressMonologue: true,
-			hasStructuredThinking: false
+			hasStructuredThinking: false,
+			validateToolArgs: true
 		},
 		promptAdditions: [
 			"CRITICAL OUTPUT RULES:",
@@ -32406,7 +32410,8 @@ const BUILTIN_PROFILES = [
 			"- Do not write 'Verification:', 'Requirements:', 'Status Check:', or 'Everything is correct'.",
 			"- If you need to verify your work, do so before writing your response, not after.",
 			"- Keep planning and reasoning internal. Only speak when you have a result or need user input.",
-			"- When working on multi-step tasks, send a brief one-line status update so the user knows you are active."
+			"- When working on multi-step tasks, send a brief one-line status update so the user knows you are active.",
+			"- When calling tools, provide arguments as clean values only. Never include <|, |>, or XML-like tags in tool argument values. For exec/bash commands, provide only valid shell syntax."
 		],
 		ui: {
 			badge: "CoreWeave",
@@ -32434,7 +32439,8 @@ const FALLBACK_PROFILE = {
 	filters: {
 		stripPatterns: [],
 		suppressMonologue: false,
-		hasStructuredThinking: false
+		hasStructuredThinking: false,
+		validateToolArgs: false
 	},
 	promptAdditions: [],
 	ui: {
@@ -32861,6 +32867,19 @@ function stripModelTokensAggressive(text, modelId) {
 	}
 	return result;
 }
+/**
+* Clean residual fragments left after aggressive token stripping.
+* After `<|` and `|>` are removed, orphaned `<`, `"`, `|` can remain
+* at string boundaries (e.g., `<<|"|uname -a` → `<"uname -a`).
+*
+* Conservative: only trims leading junk before what looks like a command,
+* and trailing junk. Does NOT touch content in the middle of the string.
+* Preserves heredoc syntax (`<<EOF`, `<<"EOF"`) via negative lookahead.
+*/
+function cleanResidualTokenFragments(text) {
+	if (!text) return text;
+	return text.replace(/^[<"|]+(?!<[A-Z_])(?=[a-zA-Z0-9/~$.(])/gm, "").replace(/[<"|]+$/gm, "").trim();
+}
 
 //#endregion
 //#region src/agents/output-normalizer.ts
@@ -32946,10 +32965,14 @@ function detectRepetition(text, minBlock) {
 	const searchStart = sampleStart + minBlock;
 	return text.indexOf(sample, searchStart) >= 0;
 }
+function sanitizeStringValue(value, modelId) {
+	const stripped = stripModelTokensAggressive(value, modelId);
+	return isGemmaModel$1(modelId) ? cleanResidualTokenFragments(stripped) : stripped;
+}
 function sanitizeArgs(args, modelId) {
 	const result = {};
-	for (const [key, value] of Object.entries(args)) if (typeof value === "string") result[key] = stripModelTokensAggressive(value, modelId);
-	else if (Array.isArray(value)) result[key] = value.map((item) => typeof item === "string" ? stripModelTokensAggressive(item, modelId) : item && typeof item === "object" ? sanitizeArgs(item, modelId) : item);
+	for (const [key, value] of Object.entries(args)) if (typeof value === "string") result[key] = sanitizeStringValue(value, modelId);
+	else if (Array.isArray(value)) result[key] = value.map((item) => typeof item === "string" ? sanitizeStringValue(item, modelId) : item && typeof item === "object" ? sanitizeArgs(item, modelId) : item);
 	else if (value && typeof value === "object") result[key] = sanitizeArgs(value, modelId);
 	else result[key] = value;
 	return result;
@@ -35243,7 +35266,7 @@ async function recordLoopOutcome(args) {
 	if (!args.ctx?.sessionKey) return;
 	try {
 		const { getDiagnosticSessionState } = await import("./diagnostic-session-state-CIjIGxEE.js").then((n) => n.n);
-		const { recordToolCallOutcome } = await import("./tool-loop-detection-BU3fbtCd.js");
+		const { recordToolCallOutcome } = await import("./tool-loop-detection-D7qjFnRh.js");
 		recordToolCallOutcome(getDiagnosticSessionState({
 			sessionKey: args.ctx.sessionKey,
 			sessionId: args.ctx?.agentId
@@ -35265,7 +35288,7 @@ async function runBeforeToolCallHook(args) {
 	if (args.ctx?.sessionKey) {
 		const { getDiagnosticSessionState } = await import("./diagnostic-session-state-CIjIGxEE.js").then((n) => n.n);
 		const { logToolLoopAction } = await import("./diagnostic-Ci8Xsc_Y.js").then((n) => n.n);
-		const { detectToolCallLoop, recordToolCall } = await import("./tool-loop-detection-BU3fbtCd.js");
+		const { detectToolCallLoop, recordToolCall } = await import("./tool-loop-detection-D7qjFnRh.js");
 		const sessionState = getDiagnosticSessionState({
 			sessionKey: args.ctx.sessionKey,
 			sessionId: args.ctx?.agentId
@@ -35536,6 +35559,88 @@ function toClientToolDefinitions(tools, onClientToolCall, hookContext) {
 	});
 }
 
+//#endregion
+//#region src/agents/pi-tools.validate.ts
+/**
+* Tool argument validation — validates that exec/bash commands contain
+* valid shell syntax after model token sanitization.
+*
+* Gated on ModelProfile.filters.validateToolArgs — only active for
+* models known to leak control tokens into structured output (Gemma, Ollama).
+*
+* @module
+*/
+/**
+* Validate that a command string contains extractable valid shell syntax.
+* Returns the cleaned command if valid, or an error reason if not.
+*
+* Does NOT attempt to parse the full shell grammar — just checks for
+* obvious corruption (residual control tokens, empty commands).
+* Preserves heredocs, redirects, pipes, and other legitimate shell syntax.
+*/
+function validateShellCommand(command) {
+	if (!command || !command.trim()) return {
+		valid: false,
+		cleaned: "",
+		reason: "empty command"
+	};
+	let cleaned = cleanResidualTokenFragments(command);
+	if (/(?:^|[^<])<\|/.test(cleaned) || /\|>/.test(cleaned)) {
+		cleaned = cleaned.replace(/<\|[^|>]*(?:\|>)?/g, "").trim();
+		cleaned = cleanResidualTokenFragments(cleaned);
+	}
+	if (!cleaned || !cleaned.trim()) return {
+		valid: false,
+		cleaned: "",
+		reason: "command is empty after removing control tokens"
+	};
+	if (!/[a-zA-Z0-9_/.~-]/.test(cleaned)) return {
+		valid: false,
+		cleaned,
+		reason: "no recognizable shell command found"
+	};
+	return {
+		valid: true,
+		cleaned: cleaned.trim()
+	};
+}
+
+//#endregion
+//#region src/agents/pi-tools.validate-wrapper.ts
+const EXEC_TOOL_NAMES = new Set(["exec", "bash"]);
+/**
+* Wrap tools with argument validation. Only exec/bash tools are validated.
+* Other tools pass through unchanged.
+*
+* Call this AFTER tools are created and the model profile is resolved.
+*/
+function wrapToolsWithArgValidation(tools, profile) {
+	if (!profile.filters?.validateToolArgs) return tools;
+	return tools.map((tool) => wrapToolWithArgValidation(tool));
+}
+/**
+* Wrap a single tool with argument validation for exec/bash commands.
+*/
+function wrapToolWithArgValidation(tool) {
+	const toolName = (tool.name ?? "").toLowerCase().trim();
+	if (!EXEC_TOOL_NAMES.has(toolName)) return tool;
+	const originalExecute = tool.execute;
+	return {
+		...tool,
+		execute(toolCallId, params, ...rest) {
+			const command = params.command ?? params.cmd;
+			if (typeof command !== "string") return originalExecute.call(tool, toolCallId, params, ...rest);
+			const result = validateShellCommand(command);
+			if (!result.valid) return Promise.resolve({ output: `Error: ${result.reason}. The command appears to contain model control tokens. Please provide a clean shell command without any <|, |>, or XML-like tags. Only valid shell syntax is accepted.` });
+			if (result.cleaned !== command) return originalExecute.call(tool, toolCallId, {
+				...params,
+				command: result.cleaned
+			}, ...rest);
+			return originalExecute.call(tool, toolCallId, params, ...rest);
+		}
+	};
+}
+
 //#endregion
 //#region src/agents/plan-mode.ts
 /**
@@ -39488,10 +39593,11 @@ async function runEmbeddedAttempt(params) {
 				await resourceLoader.reload();
 			}
 			const hookRunner = getGlobalHookRunner();
-			const { builtInTools, customTools } = splitSdkTools({
+			const { builtInTools: rawBuiltInTools, customTools } = splitSdkTools({
 				tools,
 				sandboxEnabled: !!sandbox?.enabled
 			});
+			const builtInTools = wrapToolsWithArgValidation(rawBuiltInTools, modelProfile);
 			let clientToolCallDetected = null;
 			const clientToolLoopDetection = resolveToolLoopDetectionConfig({
 				cfg: params.config,
@@ -52836,7 +52942,7 @@ function isVoiceChannelType(type) {
 function createDefaultDeps() {
 	return {
 		sendMessageWhatsApp: async (...args) => {
-			const { sendMessageWhatsApp } = await import("./web-Czp0JS6-.js");
+			const { sendMessageWhatsApp } = await import("./web-Bqrgp43v.js");
 			return await sendMessageWhatsApp(...args);
 		},
 		sendMessageTelegram: async (...args) => {
@@ -68150,7 +68256,7 @@ function loadWebLoginQr() {
 	return webLoginQrPromise;
 }
 function loadWebChannel() {
-	webChannelPromise ??= import("./web-Czp0JS6-.js");
+	webChannelPromise ??= import("./web-Bqrgp43v.js");
 	return webChannelPromise;
 }
 function loadWhatsAppActions() {

package/dist/{tool-loop-detection-BgbtzUGc.js → tool-loop-detection-C7TCF2V2.js}

@@ -9,6 +9,8 @@ const TOOL_CALL_HISTORY_SIZE = 30;
 const WARNING_THRESHOLD = 10;
 const CRITICAL_THRESHOLD = 20;
 const GLOBAL_CIRCUIT_BREAKER_THRESHOLD = 30;
+const CONSECUTIVE_FAILURE_WARNING_THRESHOLD = 3;
+const CONSECUTIVE_FAILURE_CRITICAL_THRESHOLD = 5;
 const DEFAULT_LOOP_DETECTION_CONFIG = {
 	enabled: false,
 	historySize: TOOL_CALL_HISTORY_SIZE,
@@ -18,8 +20,11 @@ const DEFAULT_LOOP_DETECTION_CONFIG = {
 	detectors: {
 		genericRepeat: true,
 		knownPollNoProgress: true,
-		pingPong: true
-	}
+		pingPong: true,
+		consecutiveFailure: true
+	},
+	consecutiveFailureWarningThreshold: CONSECUTIVE_FAILURE_WARNING_THRESHOLD,
+	consecutiveFailureCriticalThreshold: CONSECUTIVE_FAILURE_CRITICAL_THRESHOLD
 };
 function asPositiveInt(value, fallback) {
 	if (typeof value !== "number" || !Number.isInteger(value) || value <= 0) return fallback;
@@ -37,10 +42,13 @@ function resolveLoopDetectionConfig(config) {
 		warningThreshold,
 		criticalThreshold,
 		globalCircuitBreakerThreshold,
+		consecutiveFailureWarningThreshold: asPositiveInt(config?.consecutiveFailureWarningThreshold, DEFAULT_LOOP_DETECTION_CONFIG.consecutiveFailureWarningThreshold),
+		consecutiveFailureCriticalThreshold: asPositiveInt(config?.consecutiveFailureCriticalThreshold, DEFAULT_LOOP_DETECTION_CONFIG.consecutiveFailureCriticalThreshold),
 		detectors: {
 			genericRepeat: config?.detectors?.genericRepeat ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.genericRepeat,
 			knownPollNoProgress: config?.detectors?.knownPollNoProgress ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.knownPollNoProgress,
-			pingPong: config?.detectors?.pingPong ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.pingPong
+			pingPong: config?.detectors?.pingPong ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.pingPong,
+			consecutiveFailure: (config?.detectors)?.consecutiveFailure ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.consecutiveFailure
 		}
 	};
 }
@@ -218,6 +226,22 @@ function canonicalPairKey(signatureA, signatureB) {
 	return [signatureA, signatureB].toSorted().join("|");
 }
 /**
+* Count consecutive error results for a specific tool, walking backward from
+* the most recent call. A success or a call to a different tool resets the streak.
+* Only counts entries that have been resolved (have a resultHash or isError flag).
+*/
+function getConsecutiveFailureStreak(history, toolName) {
+	let count = 0;
+	for (let i = history.length - 1; i >= 0; i--) {
+		const entry = history[i];
+		if (entry.toolName !== toolName) break;
+		if (entry.resultHash === void 0 && entry.isError === void 0) break;
+		if (!entry.isError) break;
+		count++;
+	}
+	return count;
+}
+/**
 * Detect if an agent is stuck in a repetitive tool call loop.
 * Checks if the same tool+params combination has been called excessively.
 */
@@ -230,6 +254,31 @@ function detectToolCallLoop(state, toolName, params, config) {
 	const noProgressStreak = noProgress.count;
 	const knownPollTool = isKnownPollToolCall(toolName, params);
 	const pingPong = getPingPongStreak(history, currentHash);
+	if (resolvedConfig.detectors.consecutiveFailure) {
+		const failureStreak = getConsecutiveFailureStreak(history, toolName);
+		if (failureStreak >= resolvedConfig.consecutiveFailureCriticalThreshold) {
+			log.error(`Consecutive failure circuit breaker: ${toolName} failed ${failureStreak} times in a row`);
+			return {
+				stuck: true,
+				level: "critical",
+				detector: "consecutive_failure",
+				count: failureStreak,
+				message: `CRITICAL: ${toolName} has failed ${failureStreak} consecutive times. Stop calling this tool. Explain to the user what you were trying to do and why it failed.`,
+				warningKey: `consecutive-failure:${toolName}`
+			};
+		}
+		if (failureStreak >= resolvedConfig.consecutiveFailureWarningThreshold) {
+			log.warn(`Consecutive failure warning: ${toolName} failed ${failureStreak} times in a row`);
+			return {
+				stuck: true,
+				level: "warning",
+				detector: "consecutive_failure",
+				count: failureStreak,
+				message: `WARNING: You have failed to call ${toolName} ${failureStreak} times consecutively. Stop retrying and explain to the user what you were trying to do. The tool arguments may contain invalid syntax from model control tokens.`,
+				warningKey: `consecutive-failure:${toolName}`
+			};
+		}
+	}
 	if (noProgressStreak >= resolvedConfig.globalCircuitBreakerThreshold) {
 		log.error(`Global circuit breaker triggered: ${toolName} repeated ${noProgressStreak} times with no progress`);
 		return {
@@ -334,6 +383,7 @@ function recordToolCallOutcome(state, params) {
 		if (call.toolName !== params.toolName || call.argsHash !== argsHash) continue;
 		if (call.resultHash !== void 0) continue;
 		call.resultHash = resultHash;
+		call.isError = params.error !== void 0;
 		matched = true;
 		break;
 	}

package/dist/{tool-loop-detection-BU3fbtCd.js → tool-loop-detection-D7qjFnRh.js}

@@ -7,6 +7,8 @@ const TOOL_CALL_HISTORY_SIZE = 30;
 const WARNING_THRESHOLD = 10;
 const CRITICAL_THRESHOLD = 20;
 const GLOBAL_CIRCUIT_BREAKER_THRESHOLD = 30;
+const CONSECUTIVE_FAILURE_WARNING_THRESHOLD = 3;
+const CONSECUTIVE_FAILURE_CRITICAL_THRESHOLD = 5;
 const DEFAULT_LOOP_DETECTION_CONFIG = {
 	enabled: false,
 	historySize: TOOL_CALL_HISTORY_SIZE,
@@ -16,8 +18,11 @@ const DEFAULT_LOOP_DETECTION_CONFIG = {
 	detectors: {
 		genericRepeat: true,
 		knownPollNoProgress: true,
-		pingPong: true
-	}
+		pingPong: true,
+		consecutiveFailure: true
+	},
+	consecutiveFailureWarningThreshold: CONSECUTIVE_FAILURE_WARNING_THRESHOLD,
+	consecutiveFailureCriticalThreshold: CONSECUTIVE_FAILURE_CRITICAL_THRESHOLD
 };
 function asPositiveInt(value, fallback) {
 	if (typeof value !== "number" || !Number.isInteger(value) || value <= 0) return fallback;
@@ -35,10 +40,13 @@ function resolveLoopDetectionConfig(config) {
 		warningThreshold,
 		criticalThreshold,
 		globalCircuitBreakerThreshold,
+		consecutiveFailureWarningThreshold: asPositiveInt(config?.consecutiveFailureWarningThreshold, DEFAULT_LOOP_DETECTION_CONFIG.consecutiveFailureWarningThreshold),
+		consecutiveFailureCriticalThreshold: asPositiveInt(config?.consecutiveFailureCriticalThreshold, DEFAULT_LOOP_DETECTION_CONFIG.consecutiveFailureCriticalThreshold),
 		detectors: {
 			genericRepeat: config?.detectors?.genericRepeat ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.genericRepeat,
 			knownPollNoProgress: config?.detectors?.knownPollNoProgress ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.knownPollNoProgress,
-			pingPong: config?.detectors?.pingPong ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.pingPong
+			pingPong: config?.detectors?.pingPong ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.pingPong,
+			consecutiveFailure: (config?.detectors)?.consecutiveFailure ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.consecutiveFailure
 		}
 	};
 }
@@ -216,6 +224,22 @@ function canonicalPairKey(signatureA, signatureB) {
 	return [signatureA, signatureB].toSorted().join("|");
 }
 /**
+* Count consecutive error results for a specific tool, walking backward from
+* the most recent call. A success or a call to a different tool resets the streak.
+* Only counts entries that have been resolved (have a resultHash or isError flag).
+*/
+function getConsecutiveFailureStreak(history, toolName) {
+	let count = 0;
+	for (let i = history.length - 1; i >= 0; i--) {
+		const entry = history[i];
+		if (entry.toolName !== toolName) break;
+		if (entry.resultHash === void 0 && entry.isError === void 0) break;
+		if (!entry.isError) break;
+		count++;
+	}
+	return count;
+}
+/**
 * Detect if an agent is stuck in a repetitive tool call loop.
 * Checks if the same tool+params combination has been called excessively.
 */
@@ -228,6 +252,31 @@ function detectToolCallLoop(state, toolName, params, config) {
 	const noProgressStreak = noProgress.count;
 	const knownPollTool = isKnownPollToolCall(toolName, params);
 	const pingPong = getPingPongStreak(history, currentHash);
+	if (resolvedConfig.detectors.consecutiveFailure) {
+		const failureStreak = getConsecutiveFailureStreak(history, toolName);
+		if (failureStreak >= resolvedConfig.consecutiveFailureCriticalThreshold) {
+			log.error(`Consecutive failure circuit breaker: ${toolName} failed ${failureStreak} times in a row`);
+			return {
+				stuck: true,
+				level: "critical",
+				detector: "consecutive_failure",
+				count: failureStreak,
+				message: `CRITICAL: ${toolName} has failed ${failureStreak} consecutive times. Stop calling this tool. Explain to the user what you were trying to do and why it failed.`,
+				warningKey: `consecutive-failure:${toolName}`
+			};
+		}
+		if (failureStreak >= resolvedConfig.consecutiveFailureWarningThreshold) {
+			log.warn(`Consecutive failure warning: ${toolName} failed ${failureStreak} times in a row`);
+			return {
+				stuck: true,
+				level: "warning",
+				detector: "consecutive_failure",
+				count: failureStreak,
+				message: `WARNING: You have failed to call ${toolName} ${failureStreak} times consecutively. Stop retrying and explain to the user what you were trying to do. The tool arguments may contain invalid syntax from model control tokens.`,
+				warningKey: `consecutive-failure:${toolName}`
+			};
+		}
+	}
 	if (noProgressStreak >= resolvedConfig.globalCircuitBreakerThreshold) {
 		log.error(`Global circuit breaker triggered: ${toolName} repeated ${noProgressStreak} times with no progress`);
 		return {
@@ -332,6 +381,7 @@ function recordToolCallOutcome(state, params) {
 		if (call.toolName !== params.toolName || call.argsHash !== argsHash) continue;
 		if (call.resultHash !== void 0) continue;
 		call.resultHash = resultHash;
+		call.isError = params.error !== void 0;
 		matched = true;
 		break;
 	}

package/dist/{tool-loop-detection-B6j1r-Wk.js → tool-loop-detection-DPVtQOfM.js}

@@ -10,6 +10,8 @@ const TOOL_CALL_HISTORY_SIZE = 30;
 const WARNING_THRESHOLD = 10;
 const CRITICAL_THRESHOLD = 20;
 const GLOBAL_CIRCUIT_BREAKER_THRESHOLD = 30;
+const CONSECUTIVE_FAILURE_WARNING_THRESHOLD = 3;
+const CONSECUTIVE_FAILURE_CRITICAL_THRESHOLD = 5;
 const DEFAULT_LOOP_DETECTION_CONFIG = {
 	enabled: false,
 	historySize: TOOL_CALL_HISTORY_SIZE,
@@ -19,8 +21,11 @@ const DEFAULT_LOOP_DETECTION_CONFIG = {
 	detectors: {
 		genericRepeat: true,
 		knownPollNoProgress: true,
-		pingPong: true
-	}
+		pingPong: true,
+		consecutiveFailure: true
+	},
+	consecutiveFailureWarningThreshold: CONSECUTIVE_FAILURE_WARNING_THRESHOLD,
+	consecutiveFailureCriticalThreshold: CONSECUTIVE_FAILURE_CRITICAL_THRESHOLD
 };
 function asPositiveInt(value, fallback) {
 	if (typeof value !== "number" || !Number.isInteger(value) || value <= 0) return fallback;
@@ -38,10 +43,13 @@ function resolveLoopDetectionConfig(config) {
 		warningThreshold,
 		criticalThreshold,
 		globalCircuitBreakerThreshold,
+		consecutiveFailureWarningThreshold: asPositiveInt(config?.consecutiveFailureWarningThreshold, DEFAULT_LOOP_DETECTION_CONFIG.consecutiveFailureWarningThreshold),
+		consecutiveFailureCriticalThreshold: asPositiveInt(config?.consecutiveFailureCriticalThreshold, DEFAULT_LOOP_DETECTION_CONFIG.consecutiveFailureCriticalThreshold),
 		detectors: {
 			genericRepeat: config?.detectors?.genericRepeat ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.genericRepeat,
 			knownPollNoProgress: config?.detectors?.knownPollNoProgress ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.knownPollNoProgress,
-			pingPong: config?.detectors?.pingPong ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.pingPong
+			pingPong: config?.detectors?.pingPong ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.pingPong,
+			consecutiveFailure: (config?.detectors)?.consecutiveFailure ?? DEFAULT_LOOP_DETECTION_CONFIG.detectors.consecutiveFailure
 		}
 	};
 }
@@ -219,6 +227,22 @@ function canonicalPairKey(signatureA, signatureB) {
 	return [signatureA, signatureB].toSorted().join("|");
 }
 /**
+* Count consecutive error results for a specific tool, walking backward from
+* the most recent call. A success or a call to a different tool resets the streak.
+* Only counts entries that have been resolved (have a resultHash or isError flag).
+*/
+function getConsecutiveFailureStreak(history, toolName) {
+	let count = 0;
+	for (let i = history.length - 1; i >= 0; i--) {
+		const entry = history[i];
+		if (entry.toolName !== toolName) break;
+		if (entry.resultHash === void 0 && entry.isError === void 0) break;
+		if (!entry.isError) break;
+		count++;
+	}
+	return count;
+}
+/**
 * Detect if an agent is stuck in a repetitive tool call loop.
 * Checks if the same tool+params combination has been called excessively.
 */
@@ -231,6 +255,31 @@ function detectToolCallLoop(state, toolName, params, config) {
 	const noProgressStreak = noProgress.count;
 	const knownPollTool = isKnownPollToolCall(toolName, params);
 	const pingPong = getPingPongStreak(history, currentHash);
+	if (resolvedConfig.detectors.consecutiveFailure) {
+		const failureStreak = getConsecutiveFailureStreak(history, toolName);
+		if (failureStreak >= resolvedConfig.consecutiveFailureCriticalThreshold) {
+			log.error(`Consecutive failure circuit breaker: ${toolName} failed ${failureStreak} times in a row`);
+			return {
+				stuck: true,
+				level: "critical",
+				detector: "consecutive_failure",
+				count: failureStreak,
+				message: `CRITICAL: ${toolName} has failed ${failureStreak} consecutive times. Stop calling this tool. Explain to the user what you were trying to do and why it failed.`,
+				warningKey: `consecutive-failure:${toolName}`
+			};
+		}
+		if (failureStreak >= resolvedConfig.consecutiveFailureWarningThreshold) {
+			log.warn(`Consecutive failure warning: ${toolName} failed ${failureStreak} times in a row`);
+			return {
+				stuck: true,
+				level: "warning",
+				detector: "consecutive_failure",
+				count: failureStreak,
+				message: `WARNING: You have failed to call ${toolName} ${failureStreak} times consecutively. Stop retrying and explain to the user what you were trying to do. The tool arguments may contain invalid syntax from model control tokens.`,
+				warningKey: `consecutive-failure:${toolName}`
+			};
+		}
+	}
 	if (noProgressStreak >= resolvedConfig.globalCircuitBreakerThreshold) {
 		log.error(`Global circuit breaker triggered: ${toolName} repeated ${noProgressStreak} times with no progress`);
 		return {
@@ -335,6 +384,7 @@ function recordToolCallOutcome(state, params) {
 		if (call.toolName !== params.toolName || call.argsHash !== argsHash) continue;
 		if (call.resultHash !== void 0) continue;
 		call.resultHash = resultHash;
+		call.isError = params.error !== void 0;
 		matched = true;
 		break;
 	}