@elizaos/ui 2.0.0-alpha.21 → 2.0.0-alpha.210

package/packages/typescript/src/services/message.js

@@ -0,0 +1,3844 @@
+import { v4 } from "uuid";
+import { parseActionParams } from "../actions";
+import { createUniqueUuid } from "../entities";
+import { formatTaskCompletionStatus, getTaskCompletionCacheKey, } from "../features/advanced-capabilities/evaluators/task-completion";
+import { logger } from "../logger";
+import { imageDescriptionTemplate, messageHandlerTemplate, multiStepDecisionTemplate, multiStepSummaryTemplate, postActionDecisionTemplate, shouldRespondTemplate, } from "../prompts";
+import { isExplicitSelfModificationRequest } from "../should-respond";
+import { OPTIMIZED_PROMPT_SERVICE, } from "./optimized-prompt";
+import { resolveOptimizedPrompt } from "./optimized-prompt-resolver";
+import { getModelStreamChunkDeliveryDepth, runWithStreamingContext, } from "../streaming-context";
+import { runWithTrajectoryContext, setTrajectoryPurpose, } from "../trajectory-context";
+import { EventType } from "../types/events";
+import { ModelType } from "../types/model";
+import { incomingPipelineHookContext, modelStreamChunkPipelineHookContext, outgoingPipelineHookContext, parallelWithShouldRespondPipelineHookContext, preShouldRespondPipelineHookContext, } from "../types/pipeline-hooks";
+import { asUUID, ChannelType, ContentType } from "../types/primitives";
+import { composePromptFromState, getLocalServerUrl, parseBooleanFromText, parseJSONObjectFromText, parseKeyValueXml, truncateToCompleteSentence, } from "../utils";
+import { AVAILABLE_CONTEXTS_STATE_KEY, attachAvailableContexts, CONTEXT_ROUTING_STATE_KEY, mergeContextRouting, parseContextRoutingMetadata, setContextRoutingMetadata, } from "../utils/context-routing";
+import { createStreamingContext, MarkableExtractor, ResponseStreamExtractor, } from "../utils/streaming";
+import { extractFirstSentence, hasFirstSentence, } from "../utils/text-splitting";
+/**
+ * Reserved XML response keys that are NOT action names.
+ * Used when scanning parsedXml for standalone action param blocks.
+ */
+export const RESERVED_XML_KEYS = new Set([
+    "actions",
+    "thought",
+    "text",
+    "simple",
+    "providers",
+]);
+const PLANNER_CONTROL_ACTIONS = new Set(["REPLY", "RESPOND", "IGNORE", "STOP"].map(normalizeActionIdentifier));
+function escapeRegex(value) {
+    return value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+function textContainsAgentName(text, names) {
+    if (!text) {
+        return false;
+    }
+    return names.some((name) => {
+        const candidate = name?.trim();
+        if (!candidate) {
+            return false;
+        }
+        const pattern = new RegExp(`(^|[^\\p{L}\\p{N}])${escapeRegex(candidate)}(?=$|[^\\p{L}\\p{N}])`, "iu");
+        return pattern.test(text);
+    });
+}
+function textContainsUserTag(text) {
+    if (!text) {
+        return false;
+    }
+    return /<@!?[^>]+>|@\w+/u.test(text);
+}
+const DEFAULT_DUAL_PRESSURE_THRESHOLD = 20;
+const ALLOWED_CLASSIFIER_ACTIONS = new Set([
+    "REPLY",
+    "RESPOND",
+    "IGNORE",
+    "STOP",
+]);
+function resolveDualPressureThreshold(runtime) {
+    const raw = runtime.getSetting("DUAL_PRESSURE_THRESHOLD");
+    const value = Number.parseInt(String(raw ?? ""), 10);
+    if (Number.isFinite(value) && value >= 1 && value <= 100) {
+        return value;
+    }
+    return DEFAULT_DUAL_PRESSURE_THRESHOLD;
+}
+function parseOptionalPressureInt(value) {
+    if (typeof value === "number" && Number.isInteger(value)) {
+        return value >= 0 && value <= 100 ? value : null;
+    }
+    if (typeof value === "string" && value.trim() !== "") {
+        const parsed = Number.parseInt(value, 10);
+        return Number.isFinite(parsed) && parsed >= 0 && parsed <= 100
+            ? parsed
+            : null;
+    }
+    return null;
+}
+function applyDualPressureToClassifierAction(runtime, responseObject, rawAction) {
+    const threshold = resolveDualPressureThreshold(runtime);
+    const actionUpper = rawAction.trim().toUpperCase();
+    const speakRaw = responseObject?.speak_up ?? responseObject?.speakUp;
+    const holdRaw = responseObject?.hold_back ?? responseObject?.holdBack;
+    const speakUp = parseOptionalPressureInt(speakRaw);
+    const holdBack = parseOptionalPressureInt(holdRaw);
+    if (speakUp === null || holdBack === null) {
+        runtime.logger.warn({
+            src: "service:message",
+            action: actionUpper,
+            speakUp: speakRaw,
+            holdBack: holdRaw,
+        }, "Classifier response missing valid dual-pressure scores; treating as IGNORE");
+        return { pressure: null, finalActionUpper: "IGNORE" };
+    }
+    const net = speakUp - holdBack;
+    const pressure = { speakUp, holdBack, net };
+    if (actionUpper === "STOP") {
+        return { pressure, finalActionUpper: "STOP" };
+    }
+    const isEngage = actionUpper === "REPLY" || actionUpper === "RESPOND";
+    if (net <= -threshold && isEngage) {
+        runtime.logger.warn({
+            src: "service:message",
+            net,
+            threshold,
+            originalAction: actionUpper,
+            speakUp,
+            holdBack,
+        }, "Dual pressure: net below threshold but model chose engage; clamping to IGNORE");
+        return { pressure, finalActionUpper: "IGNORE" };
+    }
+    if (net >= threshold && actionUpper === "IGNORE") {
+        runtime.logger.warn({
+            src: "service:message",
+            net,
+            threshold,
+            speakUp,
+            holdBack,
+        }, "Dual pressure: high net but IGNORE chosen; allowing model decision");
+    }
+    return { pressure, finalActionUpper: actionUpper };
+}
+/**
+ * Extract action params from standalone XML blocks in a parsedXml object.
+ *
+ * When the LLM outputs `<actions>REPLY,START_CODING_TASK</actions>` alongside
+ * `<START_CODING_TASK><repo>...</repo></START_CODING_TASK>`, the XML parser
+ * puts the action block as a top-level key on parsedXml. This function finds
+ * those keys and assembles them into the legacy flat params format that
+ * `parseActionParams` consumes.
+ *
+ * Returns the assembled params string, or empty string if none found.
+ */
+export function extractStandaloneActionParams(actionNames, parsedXml) {
+    const fragments = [];
+    for (const actionName of actionNames) {
+        const upperName = actionName.toUpperCase();
+        const matchingKey = Object.keys(parsedXml).find((k) => k.toUpperCase() === upperName);
+        if (matchingKey &&
+            !RESERVED_XML_KEYS.has(matchingKey.toLowerCase()) &&
+            typeof parsedXml[matchingKey] === "string" &&
+            parsedXml[matchingKey].includes("<")) {
+            fragments.push(`<${upperName}>${parsedXml[matchingKey]}</${upperName}>`);
+        }
+    }
+    return fragments.join("\n");
+}
+function unwrapPlannerIdentifier(value) {
+    const trimmed = value.trim().replace(/^["'`]+|["'`]+$/g, "");
+    if (!trimmed) {
+        return "";
+    }
+    const nameMatch = trimmed.match(/^<name\b[^>]*>([\s\S]*?)<\/name>$/i);
+    if (nameMatch) {
+        return nameMatch[1].trim();
+    }
+    const actionMatch = trimmed.match(/^<action\b[^>]*>([\s\S]*?)<\/action>$/i);
+    if (!actionMatch) {
+        return trimmed;
+    }
+    const inner = actionMatch[1].trim();
+    if (!inner) {
+        return "";
+    }
+    const nestedNameMatch = inner.match(/<name\b[^>]*>([\s\S]*?)<\/name>/i);
+    if (nestedNameMatch) {
+        return nestedNameMatch[1].trim();
+    }
+    return /<[A-Za-z][^>]*>/.test(inner) ? trimmed : inner;
+}
+export function extractPlannerActionNames(parsedXml) {
+    return (() => {
+        if (typeof parsedXml.actions === "string") {
+            const actionsXml = parsedXml.actions;
+            if (/<action\b[^>]*>/i.test(actionsXml)) {
+                const actionEntries = [];
+                for (const match of actionsXml.matchAll(/<action\b[^>]*>([\s\S]*?)<\/action>/gi)) {
+                    const inner = match[1];
+                    const nameMatch = inner.match(/<name\b[^>]*>([\s\S]*?)<\/name>/i);
+                    const paramsMatch = inner.match(/<params\b[^>]*>([\s\S]*?)<\/params>/i);
+                    const name = unwrapPlannerIdentifier(nameMatch ? nameMatch[1] : match[0]);
+                    const paramsXml = paramsMatch ? paramsMatch[1].trim() : undefined;
+                    if (name) {
+                        actionEntries.push({ name, paramsXml });
+                    }
+                }
+                if (actionEntries.length > 0) {
+                    const inlineParamsXml = actionEntries
+                        .filter((entry) => entry.paramsXml)
+                        .map((entry) => `<${entry.name.toUpperCase()}>${entry.paramsXml}</${entry.name.toUpperCase()}>`)
+                        .join("\n");
+                    if (inlineParamsXml &&
+                        (!parsedXml.params || parsedXml.params === "")) {
+                        parsedXml.params = inlineParamsXml;
+                    }
+                    return actionEntries.map((entry) => entry.name);
+                }
+            }
+            const commaSplitActions = actionsXml
+                .split(",")
+                .map((action) => unwrapPlannerIdentifier(String(action)))
+                .filter((action) => action.length > 0);
+            if (!parsedXml.params || parsedXml.params === "") {
+                const assembled = extractStandaloneActionParams(commaSplitActions, parsedXml);
+                if (assembled) {
+                    parsedXml.params = assembled;
+                }
+            }
+            return commaSplitActions;
+        }
+        if (Array.isArray(parsedXml.actions)) {
+            return parsedXml.actions
+                .map((action) => unwrapPlannerIdentifier(String(action)))
+                .filter((action) => action.length > 0);
+        }
+        return [];
+    })();
+}
+function normalizePlannerActions(parsedXml, runtime) {
+    const normalizedActions = extractPlannerActionNames(parsedXml);
+    const finalActions = !runtime.isActionPlanningEnabled() && normalizedActions.length > 1
+        ? [normalizedActions[0]]
+        : normalizedActions;
+    const actionLookup = buildRuntimeActionLookup(runtime);
+    const validActions = finalActions.flatMap((actionName) => {
+        const normalized = normalizeActionIdentifier(actionName);
+        if (!normalized) {
+            return [];
+        }
+        if (PLANNER_CONTROL_ACTIONS.has(normalized)) {
+            return [actionName];
+        }
+        const resolvedAction = resolveRuntimeAction(actionLookup, actionName);
+        if (resolvedAction) {
+            return [resolvedAction.name];
+        }
+        const aliasedActionName = PLANNER_ACTION_ALIASES.get(normalized);
+        if (aliasedActionName) {
+            const resolvedAlias = resolveRuntimeAction(actionLookup, aliasedActionName);
+            if (resolvedAlias) {
+                runtime.logger.info({
+                    src: "service:message",
+                    actionName,
+                    aliasedActionName: resolvedAlias.name,
+                }, "Repaired planner action alias");
+                return [resolvedAlias.name];
+            }
+        }
+        runtime.logger.warn({
+            src: "service:message",
+            actionName,
+        }, "Dropping unknown planner action");
+        return [];
+    });
+    if (validActions.length > 0) {
+        return validActions;
+    }
+    const replyText = typeof parsedXml.text === "string" ? parsedXml.text.trim() : "";
+    return replyText.length > 0 ? ["REPLY"] : ["IGNORE"];
+}
+function normalizePlannerProviders(parsedXml, runtime) {
+    const rawProviders = parsedXml.providers;
+    const providerNames = (() => {
+        if (typeof rawProviders === "string") {
+            const trimmedProviders = rawProviders.trim();
+            if ((trimmedProviders.startsWith("[") && trimmedProviders.endsWith("]")) ||
+                (trimmedProviders.startsWith("{") && trimmedProviders.endsWith("}"))) {
+                try {
+                    const parsedJson = JSON.parse(trimmedProviders);
+                    if (Array.isArray(parsedJson)) {
+                        return parsedJson
+                            .map((providerName) => String(providerName).trim())
+                            .filter((providerName) => providerName.length > 0);
+                    }
+                    if (typeof parsedJson === "object" &&
+                        parsedJson !== null &&
+                        Array.isArray(parsedJson.providers)) {
+                        return (parsedJson.providers
+                            .map((providerName) => String(providerName).trim())
+                            .filter((providerName) => providerName.length > 0));
+                    }
+                }
+                catch {
+                    // Fall through to XML/comma parsing below.
+                }
+            }
+            if (rawProviders.includes("<provider>") ||
+                rawProviders.includes("<provider ")) {
+                const providers = Array.from(rawProviders.matchAll(/<provider>([\s\S]*?)<\/provider>/g))
+                    .map((match) => match[1]?.trim() ?? "")
+                    .filter((providerName) => providerName.length > 0);
+                if (providers.length > 0) {
+                    return providers;
+                }
+            }
+            return rawProviders
+                .split(/[\n,;]/)
+                .map((providerName) => providerName.replace(/^[\s"'[\](){}]+|[\s"'[\](){}]+$/g, ""))
+                .map((providerName) => providerName.trim())
+                .filter((providerName) => providerName.length > 0);
+        }
+        if (Array.isArray(rawProviders)) {
+            return rawProviders
+                .map((providerName) => String(providerName).trim())
+                .filter((providerName) => providerName.length > 0);
+        }
+        return [];
+    })();
+    if (!runtime) {
+        return providerNames;
+    }
+    const providerLookup = new Map();
+    for (const provider of runtime.providers ?? []) {
+        const normalized = normalizeActionIdentifier(provider.name);
+        if (!normalized || providerLookup.has(normalized)) {
+            continue;
+        }
+        providerLookup.set(normalized, provider.name);
+    }
+    const normalizedProviders = providerNames
+        .map((providerName) => {
+        const normalizedProviderName = normalizeActionIdentifier(providerName);
+        const canonicalProvider = providerLookup.get(normalizedProviderName) ??
+            (() => {
+                const aliasedProvider = PLANNER_PROVIDER_ALIASES.get(normalizedProviderName);
+                if (!aliasedProvider) {
+                    return undefined;
+                }
+                return providerLookup.get(normalizeActionIdentifier(aliasedProvider));
+            })();
+        if (canonicalProvider) {
+            return canonicalProvider;
+        }
+        runtime.logger.warn({
+            src: "service:message",
+            providerName,
+        }, "Dropping unknown planner provider");
+        return "";
+    })
+        .filter((providerName) => providerName.length > 0);
+    if (normalizedProviders.length === 0) {
+        return normalizedProviders;
+    }
+    const providerDefinitions = new Map((runtime.providers ?? []).map((provider) => [
+        normalizeActionIdentifier(provider.name),
+        provider,
+    ]));
+    const expandedProviders = [...normalizedProviders];
+    const seenProviders = new Set(expandedProviders.map((providerName) => normalizeActionIdentifier(providerName)));
+    for (let index = 0; index < expandedProviders.length; index += 1) {
+        const providerName = expandedProviders[index];
+        const providerDefinition = providerDefinitions.get(normalizeActionIdentifier(providerName));
+        const companionProviders = providerDefinition?.companionProviders ?? [];
+        for (const companionProvider of companionProviders) {
+            const canonicalCompanion = providerLookup.get(normalizeActionIdentifier(companionProvider));
+            if (!canonicalCompanion) {
+                runtime.logger.warn({
+                    src: "service:message",
+                    providerName,
+                    companionProvider,
+                }, "Dropping unknown companion provider");
+                continue;
+            }
+            const normalizedCompanion = normalizeActionIdentifier(canonicalCompanion);
+            if (seenProviders.has(normalizedCompanion)) {
+                continue;
+            }
+            seenProviders.add(normalizedCompanion);
+            expandedProviders.push(canonicalCompanion);
+        }
+    }
+    return expandedProviders;
+}
+const CORE_RESPONSE_STATE_PROVIDERS = [
+    "ENTITIES",
+    "CHARACTER",
+    "RECENT_MESSAGES",
+    "ACTIONS",
+    "PROVIDERS",
+];
+const STRUCTURED_RESPONSE_STATE_PROVIDERS = ["ACTIONS", "PROVIDERS"];
+const FOCUSED_PROVIDER_REPLY_STATE_PROVIDERS = ["CHARACTER", "RECENT_MESSAGES"];
+function composeResponseState(runtime, message, skipCache = false) {
+    return runtime.composeState(message, CORE_RESPONSE_STATE_PROVIDERS, true, skipCache);
+}
+function composeStructuredResponseState(runtime, message, skipCache = false) {
+    return runtime.composeState(message, STRUCTURED_RESPONSE_STATE_PROVIDERS, false, skipCache);
+}
+function composeProviderGroundedResponseState(runtime, message, providers, skipCache = false) {
+    return runtime.composeState(message, [...CORE_RESPONSE_STATE_PROVIDERS, ...providers], false, skipCache);
+}
+function composeFocusedProviderReplyState(runtime, message, providers, skipCache = false) {
+    return runtime.composeState(message, [...FOCUSED_PROVIDER_REPLY_STATE_PROVIDERS, ...providers], true, skipCache);
+}
+/**
+ * Escape Handlebars syntax in a string to prevent template injection.
+ *
+ * WHY: When embedding LLM-generated text into continuation prompts, the text
+ * goes through Handlebars.compile(). If the LLM output contains {{variable}},
+ * Handlebars will try to substitute it with state values, corrupting the prompt.
+ *
+ * This function escapes {{ to \\{{ so Handlebars outputs literal {{.
+ *
+ * @param text - Text that may contain Handlebars-like syntax
+ * @returns Text with {{ escaped to prevent interpretation
+ */
+function escapeHandlebars(text) {
+    // Single-pass replacement to avoid double-escaping triple braces.
+    return text.replace(/\{\{\{|\{\{/g, (match) => `\\${match}`);
+}
+function sanitizeAttachmentsForStorage(attachments) {
+    if (!attachments?.length) {
+        return attachments;
+    }
+    return attachments.map((attachment) => {
+        const { _data: _discardData, _mimeType: _discardMimeType, ...rest } = attachment;
+        return rest;
+    });
+}
+function resolvePromptAttachments(attachments) {
+    if (!attachments?.length) {
+        return undefined;
+    }
+    const resolved = attachments.flatMap((attachment) => {
+        const withInlineData = attachment;
+        if (typeof withInlineData._data === "string" &&
+            withInlineData._data.trim() &&
+            typeof withInlineData._mimeType === "string" &&
+            withInlineData._mimeType.trim()) {
+            return [
+                {
+                    data: withInlineData._data,
+                    mediaType: withInlineData._mimeType,
+                    filename: attachment.title,
+                },
+            ];
+        }
+        const dataUrlMatch = attachment.url.match(/^data:([^;,]+);base64,(.+)$/i);
+        if (dataUrlMatch) {
+            return [
+                {
+                    data: dataUrlMatch[2],
+                    mediaType: dataUrlMatch[1],
+                    filename: attachment.title,
+                },
+            ];
+        }
+        return [];
+    });
+    return resolved.length > 0 ? resolved : undefined;
+}
+function normalizeShouldRespondModelType(value) {
+    if (typeof value !== "string") {
+        return "response-handler";
+    }
+    const normalized = value.trim().toLowerCase();
+    switch (normalized) {
+        case "nano":
+        case "text_nano":
+            return "nano";
+        case "small":
+        case "text_small":
+            return "small";
+        case "large":
+        case "text_large":
+            return "large";
+        case "mega":
+        case "text_mega":
+            return "mega";
+        case "response-handler":
+        case "response_handler":
+        case "responsehandler":
+            return "response-handler";
+        case "response_handler_model":
+            return "response-handler";
+        default:
+            return "response-handler";
+    }
+}
+function resolveShouldRespondModelType(model) {
+    switch (normalizeShouldRespondModelType(model)) {
+        case "nano":
+            return ModelType.TEXT_NANO;
+        case "small":
+            return ModelType.TEXT_SMALL;
+        case "large":
+            return ModelType.TEXT_LARGE;
+        case "mega":
+            return ModelType.TEXT_MEGA;
+        default:
+            return ModelType.RESPONSE_HANDLER;
+    }
+}
+/**
+ * Tracks the latest response ID per agent+room to handle message superseding
+ */
+const latestResponseIds = new Map();
+function clearLatestResponseId(agentId, roomId, responseId) {
+    const agentMap = latestResponseIds.get(agentId);
+    if (!agentMap) {
+        return;
+    }
+    if (agentMap.get(roomId) !== responseId) {
+        return;
+    }
+    agentMap.delete(roomId);
+    if (agentMap.size === 0) {
+        latestResponseIds.delete(agentId);
+    }
+}
+export function isSimpleReplyResponse(responseContent) {
+    return !!(responseContent?.actions &&
+        responseContent.actions.length === 1 &&
+        typeof responseContent.actions[0] === "string" &&
+        responseContent.actions[0].toUpperCase() === "REPLY");
+}
+function isStopResponse(responseContent) {
+    return !!(responseContent?.actions &&
+        responseContent.actions.length === 1 &&
+        typeof responseContent.actions[0] === "string" &&
+        responseContent.actions[0].toUpperCase() === "STOP");
+}
+function normalizeActionIdentifier(actionName) {
+    return unwrapPlannerIdentifier(actionName).toUpperCase().replace(/_/g, "");
+}
+const PLANNER_ACTION_ALIASES = new Map([
+    ["BULK_RESCHEDULE_MEETINGS", "PROPOSE_MEETING_TIMES"],
+    ["SCHEDULE_MEETING", "CALENDAR_ACTION"],
+    ["RESCHEDULE_MEETINGS", "CALENDAR_ACTION"],
+    ["CREATE_EVENT", "CALENDAR_ACTION"],
+    ["CREATE_RECURRING_EVENT", "CALENDAR_ACTION"],
+    ["CALENDAR_CREATE_RECURRING_EVENT", "CALENDAR_ACTION"],
+    ["SCHEDULE_RECURRING_EVENT", "CALENDAR_ACTION"],
+    ["SCHEDULE_RECURRING_MEETING", "CALENDAR_ACTION"],
+    ["SCHEDULE_RECURRING", "CALENDAR_ACTION"],
+    ["BOOK_TRAVEL", "CALL_EXTERNAL"],
+    ["CAPTURE_TRAVEL_PREFERENCES", "UPDATE_OWNER_PROFILE"],
+    ["CAPTURE_BOOKING_PREFERENCES", "UPDATE_OWNER_PROFILE"],
+    ["SET_PREFERENCES", "UPDATE_OWNER_PROFILE"],
+    ["SET_TRAVEL_PREFERENCES", "UPDATE_OWNER_PROFILE"],
+    ["CREATE_FOLLOWUP", "INBOX"],
+    ["GET_PENDING_ASSETS", "INBOX"],
+    ["GET_PENDING_ITEMS", "INBOX"],
+    ["PROPOSE_GROUP_CHAT_HANDOFF", "INBOX"],
+    ["SET_MULTI_DEVICE_MEETING_REMINDER", "PUBLISH_DEVICE_INTENT"],
+    ["HANDLE_CANCELLATION_FEE", "PUBLISH_DEVICE_INTENT"],
+    ["GET_ID_STATUS", "PUBLISH_DEVICE_INTENT"],
+    ["REQUEST_UPLOAD", "LIFEOPS_COMPUTER_USE"],
+].map(([from, to]) => [
+    normalizeActionIdentifier(from),
+    to,
+]));
+const PLANNER_PROVIDER_ALIASES = new Map([
+    ["DOCUMENT_LOOKUP", "ATTACHMENTS"],
+    ["INBOX_TRIAGE", "inboxTriage"],
+].map(([from, to]) => [normalizeActionIdentifier(from), to]));
+const PROVIDER_FOLLOWUP_PASSIVE_ACTIONS = new Set(["REPLY", "RESPOND", "NONE"].map(normalizeActionIdentifier));
+function shouldRunProviderFollowup(responseContent) {
+    if (!responseContent?.providers?.length) {
+        return false;
+    }
+    const normalizedActions = (responseContent.actions ?? [])
+        .map((actionName) => typeof actionName === "string"
+        ? normalizeActionIdentifier(actionName)
+        : "")
+        .filter((actionName) => actionName.length > 0);
+    if (normalizedActions.length === 0) {
+        return true;
+    }
+    return normalizedActions.every((actionName) => PROVIDER_FOLLOWUP_PASSIVE_ACTIONS.has(actionName));
+}
+function buildProviderFollowupPrompt(basePrompt) {
+    return `${basePrompt}
+
+[PROVIDER FOLLOW-UP]
+The requested providers have already been executed, and their grounded results are now present in context above.
+Use those provider results to produce the final reply and/or action plan for this turn.
+Do not ask for the same providers again.
+If the provider results fully answer the user, reply directly.
+If KNOWLEDGE contains a direct answer, prefer that grounded answer even when AVAILABLE_DOCUMENTS lists multiple files.
+Do not ask "which file?" when the grounded KNOWLEDGE result already resolves the request.`;
+}
+function shouldAttemptProviderRescue(responseContent) {
+    if (!responseContent) {
+        return false;
+    }
+    if ((responseContent.providers?.length ?? 0) > 0) {
+        return false;
+    }
+    const normalizedActions = (responseContent.actions ?? [])
+        .map((actionName) => typeof actionName === "string"
+        ? normalizeActionIdentifier(actionName)
+        : "")
+        .filter((actionName) => actionName.length > 0);
+    if (normalizedActions.length === 0) {
+        return true;
+    }
+    return normalizedActions.every((actionName) => PROVIDER_FOLLOWUP_PASSIVE_ACTIONS.has(actionName));
+}
+function buildProviderSelectionPrompt(draftReply) {
+    const trimmedDraftReply = draftReply?.trim() ?? "";
+    const draftReplySection = trimmedDraftReply.length > 0
+        ? `draft_reply:\n${trimmedDraftReply.replace(/<\/response>/gi, "<\\/response>")}\n\n`
+        : "";
+    const draftReplyRules = trimmedDraftReply.length > 0
+        ? [
+            "- if the draft reply asks the user to resend, restate, or clarify information that may already exist in provider context, choose the relevant providers instead of sending the draft reply as-is",
+            '- when the recent conversation already identifies a prior upload or knowledge-base question, prefer grounded provider lookup over asking "which file?" again',
+        ]
+        : [];
+    return `task: Decide whether any providers should be called before sending the assistant's reply.
+
+available provider catalog:
+{{providers}}
+
+recent conversation:
+{{recentMessages}}
+
+${draftReplySection}rules[${4 + draftReplyRules.length}]:
+- choose providers only when they can supply grounded information needed before the assistant replies
+- uploaded files, documents, prior uploads, and knowledge-base questions should use the relevant providers before asking the user to resend the material
+- if the user asks about an uploaded file or document and AVAILABLE_DOCUMENTS is available, prefer AVAILABLE_DOCUMENTS together with KNOWLEDGE before sending any clarification reply
+- return an empty providers field when no provider lookup is needed
+- do not include actions, text, or thought in the output
+${draftReplyRules.join("\n")}
+
+output:
+Return JSON or XML containing only provider names. No prose before or after it. No <think>.
+
+Examples:
+- user asks: "what is the qa codeword from the uploaded file?"
+  draft reply: "Which file are you referring to?"
+  output: {"providers":["AVAILABLE_DOCUMENTS","KNOWLEDGE"]}
+- user asks: "what is the qa codeword from the uploaded file?"
+  draft reply: "I don't have the file in my context. Which file contains the QA codeword?"
+  output: {"providers":["AVAILABLE_DOCUMENTS","KNOWLEDGE"]}
+- user asks: "thanks, that's all"
+  draft reply: "Glad to help."
+  output: {"providers":[]}`;
+}
+async function recoverProvidersForTurn(args) {
+    const prompt = composePromptFromState({
+        state: args.state,
+        template: buildProviderSelectionPrompt(args.draftReply),
+    });
+    try {
+        const result = await args.runtime.useModel(ModelType.TEXT_LARGE, {
+            prompt,
+            ...(args.attachments ? { attachments: args.attachments } : {}),
+        });
+        const rawResponse = typeof result === "string" ? result : "";
+        const parsed = parseKeyValueXml(rawResponse) ??
+            parseJSONObjectFromText(rawResponse);
+        const normalizedProviders = normalizePlannerProviders(parsed ?? { providers: rawResponse }, args.runtime);
+        if (normalizedProviders.length > 0) {
+            return normalizedProviders;
+        }
+        const shouldUseKnowledge = await shouldUseKnowledgeProviders(args.runtime, args.state, args.attachments);
+        return shouldUseKnowledge ? ["AVAILABLE_DOCUMENTS", "KNOWLEDGE"] : [];
+    }
+    catch (error) {
+        args.runtime.logger.warn({
+            src: "service:message",
+            error: error instanceof Error ? error.message : String(error),
+        }, "Provider rescue model call failed");
+        return [];
+    }
+}
+function buildGroundedFallbackReplyPrompt() {
+    return `task: Write the next assistant reply using grounded context.
+
+grounded context:
+{{providers}}
+
+recent conversation:
+{{recentMessages}}
+
+rules[5]:
+- answer directly from grounded context when it fully answers the user
+- do not ask the user to resend, rename, or specify a file if grounded document or knowledge context already answers the request
+- do not say you cannot access the file when grounded context is already present above
+- if KNOWLEDGE contains a direct answer, prefer that grounded answer even when AVAILABLE_DOCUMENTS lists multiple files
+- if grounded context is still insufficient, say exactly what is missing
+- return only the reply text
+
+output:
+Plain text only. No XML, JSON, TOON, bullets, or <think>.`;
+}
+function buildKnowledgeProviderDecisionPrompt() {
+    return `task: Decide whether the assistant should consult uploaded-document or knowledge providers before replying.
+
+available provider catalog:
+{{providers}}
+
+recent conversation:
+{{recentMessages}}
+
+rules[5]:
+- return true when the user is asking about an uploaded file, document, prior upload, or knowledge-base content
+- return true when the answer is likely already stored in uploaded documents or semantic knowledge search
+- when AVAILABLE_DOCUMENTS or KNOWLEDGE is available and the user refers to an uploaded file or prior upload, return true
+- return false for generic chat, thanks, or requests that clearly do not depend on uploaded or knowledge-base content
+- return only the structured output, with no prose
+
+output:
+Return JSON or XML only.
+
+Examples:
+- user asks: "what is the qa codeword from the uploaded file?" -> {"useKnowledgeProviders":true}
+- user asks: "thanks, that's all" -> {"useKnowledgeProviders":false}`;
+}
+async function shouldUseKnowledgeProviders(runtime, state, attachments) {
+    const prompt = composePromptFromState({
+        state,
+        template: buildKnowledgeProviderDecisionPrompt(),
+    });
+    try {
+        const result = await runtime.useModel(ModelType.TEXT_LARGE, {
+            prompt,
+            ...(attachments ? { attachments } : {}),
+        });
+        const rawResponse = typeof result === "string" ? result : "";
+        const parsed = parseKeyValueXml(rawResponse) ??
+            parseJSONObjectFromText(rawResponse);
+        const value = parsed?.useKnowledgeProviders ??
+            parsed?.use_knowledge_providers ??
+            rawResponse;
+        if (typeof value === "boolean") {
+            return value;
+        }
+        if (typeof value === "string") {
+            return value.trim().toLowerCase() === "true";
+        }
+        return false;
+    }
+    catch (error) {
+        runtime.logger.warn({
+            src: "service:message",
+            error: error instanceof Error ? error.message : String(error),
+        }, "Knowledge provider decision model call failed");
+        return false;
+    }
+}
+function buildRuntimeActionLookup(runtime) {
+    const actionMap = new Map();
+    for (const action of runtime.actions ?? []) {
+        const identifiers = [action.name, ...(action.similes ?? [])];
+        for (const identifier of identifiers) {
+            const normalized = normalizeActionIdentifier(identifier);
+            if (!normalized || actionMap.has(normalized)) {
+                continue;
+            }
+            actionMap.set(normalized, action);
+        }
+    }
+    return actionMap;
+}
+function resolveRuntimeAction(actionLookup, actionName) {
+    const normalized = normalizeActionIdentifier(actionName);
+    if (!normalized) {
+        return undefined;
+    }
+    return actionLookup.get(normalized);
+}
+const TERMINAL_ACTION_IDENTIFIERS = new Set([
+    "REPLY",
+    "IGNORE",
+    "STOP",
+    "CREATE_TASK",
+    "START_CODING_TASK",
+    "CODE_TASK",
+    "SPAWN_AGENT",
+    "SPAWN_CODING_AGENT",
+].map(normalizeActionIdentifier));
+function shouldContinueAfterActions(runtime, responseContent) {
+    // Async background/task actions handle their own follow-up and should not
+    // trigger an immediate continuation loop from the main message service.
+    const actionLookup = buildRuntimeActionLookup(runtime);
+    return !!responseContent?.actions?.some((action) => {
+        if (typeof action !== "string")
+            return false;
+        const resolvedAction = resolveRuntimeAction(actionLookup, action);
+        if (resolvedAction?.suppressPostActionContinuation) {
+            return false;
+        }
+        const canonicalAction = resolvedAction?.name ?? action;
+        return !TERMINAL_ACTION_IDENTIFIERS.has(normalizeActionIdentifier(canonicalAction));
+    });
+}
+function suppressesPostActionContinuation(runtime, responseContent) {
+    if (!responseContent?.actions?.length) {
+        return false;
+    }
+    const actionLookup = buildRuntimeActionLookup(runtime);
+    return responseContent.actions.some((action) => {
+        if (typeof action !== "string")
+            return false;
+        return (resolveRuntimeAction(actionLookup, action)
+            ?.suppressPostActionContinuation === true);
+    });
+}
+/**
+ * True when the planner's `text` field should be surfaced to the user as a
+ * preamble before action handlers run in actions-mode dispatch. The goal:
+ * the user sees "checking your inbox" rather than silence while INBOX/GMAIL
+ * do their work.
+ *
+ * Skipped when the first action is REPLY (the REPLY handler generates its own
+ * text), IGNORE (no user-visible response), or STOP (terminal). Also skipped
+ * when `text` is empty.
+ */
+export function shouldEmitPlannerPreamble(runtime, responseContent) {
+    if (!responseContent)
+        return false;
+    const text = typeof responseContent.text === "string" ? responseContent.text.trim() : "";
+    if (text.length === 0)
+        return false;
+    const firstAction = typeof responseContent.actions?.[0] === "string"
+        ? normalizeActionIdentifier(responseContent.actions[0])
+        : "";
+    if (firstAction.length === 0)
+        return false;
+    const resolvedAction = (runtime.actions ?? []).find((action) => normalizeActionIdentifier(action.name) === firstAction &&
+        action.suppressPostActionContinuation === true);
+    if (resolvedAction) {
+        return false;
+    }
+    return (firstAction !== normalizeActionIdentifier("REPLY") &&
+        firstAction !== normalizeActionIdentifier("IGNORE") &&
+        firstAction !== normalizeActionIdentifier("STOP"));
+}
+function callbackTextPreview(content) {
+    if (!content || typeof content !== "object") {
+        return "";
+    }
+    const text = typeof content.text === "string" ? content.text.trim() : "";
+    if (!text) {
+        return "";
+    }
+    return text.replace(/\s+/g, " ").slice(0, 200);
+}
+function summarizeAttachmentKeyPart(url) {
+    const trimmed = url.trim();
+    if (trimmed.length <= 256) {
+        return trimmed;
+    }
+    return `${trimmed.slice(0, 128)}...(${trimmed.length})`;
+}
+function callbackDeliveryKey(content) {
+    if (!content || typeof content !== "object") {
+        return "";
+    }
+    const text = typeof content.text === "string"
+        ? content.text.replace(/\s+/g, " ").trim()
+        : "";
+    const attachmentKeys = Array.isArray(content.attachments)
+        ? content.attachments
+            .map((attachment) => {
+            if (!attachment || typeof attachment !== "object") {
+                return "";
+            }
+            const url = typeof attachment.url === "string"
+                ? summarizeAttachmentKeyPart(attachment.url)
+                : "";
+            const title = typeof attachment.title === "string" ? attachment.title.trim() : "";
+            const contentType = typeof attachment.contentType === "string"
+                ? attachment.contentType
+                : "";
+            if (!url && !title && !contentType) {
+                return "";
+            }
+            return `${contentType}:${title}:${url}`;
+        })
+            .filter((key) => key.length > 0)
+            .sort()
+        : [];
+    if (!text && attachmentKeys.length === 0) {
+        return "";
+    }
+    return JSON.stringify({
+        text,
+        attachments: attachmentKeys,
+    });
+}
+function getLatestVisibleReplyText(responseContent, actionResults) {
+    for (let index = actionResults.length - 1; index >= 0; index--) {
+        const result = actionResults[index];
+        const actionName = typeof result?.data?.actionName === "string"
+            ? result.data.actionName
+            : "";
+        if (normalizeActionIdentifier(actionName) !== "REPLY") {
+            continue;
+        }
+        if (typeof result.text === "string" && result.text.trim().length > 0) {
+            return result.text.trim();
+        }
+    }
+    const responseText = typeof responseContent?.text === "string"
+        ? responseContent.text.trim()
+        : "";
+    return responseText;
+}
+function isLikelyClarifyingQuestion(text) {
+    const normalized = text.trim();
+    if (!normalized) {
+        return false;
+    }
+    if (/[?؟]\s*$/.test(normalized)) {
+        return true;
+    }
+    const firstSentence = extractFirstSentence(normalized)
+        .first.trim()
+        .toLowerCase();
+    return /^(what|which|when|where|who|whom|whose|why|how|can you|could you|would you|will you|do you|did you|are you|is it|should i|should we)\b/.test(firstSentence);
+}
+function shouldWaitForUserAfterIncompleteReflection(responseContent, actionResults) {
+    const latestVisibleReply = getLatestVisibleReplyText(responseContent, actionResults);
+    if (!isLikelyClarifyingQuestion(latestVisibleReply)) {
+        return false;
+    }
+    if (actionResults.length === 0) {
+        return isSimpleReplyResponse(responseContent);
+    }
+    return actionResults.every((result) => {
+        const actionName = typeof result?.data?.actionName === "string"
+            ? result.data.actionName
+            : "";
+        return normalizeActionIdentifier(actionName) === "REPLY";
+    });
+}
+function formatActionResultsForPrompt(actionResults) {
+    if (actionResults.length === 0) {
+        return "No action results available.";
+    }
+    return [
+        "# Action Results",
+        ...actionResults.map((result, index) => {
+            const actionNameValue = result.data?.actionName;
+            const actionName = typeof actionNameValue === "string"
+                ? actionNameValue
+                : "Unknown Action";
+            const lines = [
+                `${index + 1}. ${actionName} - ${result.success === false ? "failed" : "succeeded"}`,
+            ];
+            if (typeof result.text === "string" && result.text.trim()) {
+                lines.push(`Output: ${result.text.trim().slice(0, 2000)}`);
+            }
+            if (result.error) {
+                const errorText = result.error instanceof Error
+                    ? result.error.message
+                    : String(result.error);
+                lines.push(`Error: ${errorText.slice(0, 1000)}`);
+            }
+            return lines.join("\n");
+        }),
+    ].join("\n\n");
+}
+function withActionResults(state, actionResults) {
+    return {
+        ...state,
+        values: {
+            ...state.values,
+            actionResults: formatActionResultsForPrompt(actionResults),
+        },
+        data: {
+            ...state.data,
+            actionResults,
+        },
+    };
+}
+function withTaskCompletion(state, taskCompletion) {
+    if (!taskCompletion) {
+        return state;
+    }
+    return {
+        ...state,
+        values: {
+            ...state.values,
+            taskCompletionStatus: formatTaskCompletionStatus(taskCompletion),
+            taskCompleted: taskCompletion.completed,
+            taskCompletionAssessed: taskCompletion.assessed,
+            taskCompletionReason: taskCompletion.reason,
+        },
+        data: {
+            ...state.data,
+            taskCompletion,
+        },
+    };
+}
+function getStructuredOutputFailure(state) {
+    const candidate = state.data?.structuredOutputFailure;
+    if (!candidate || typeof candidate !== "object") {
+        return null;
+    }
+    return candidate;
+}
+function summarizeStructuredOutputFailure(failure) {
+    if (!failure) {
+        return "Structured output parsing failed, but no additional diagnostics were recorded.";
+    }
+    const parts = [
+        `Kind: ${failure.kind}`,
+        `Model: ${failure.model}`,
+        `Format: ${failure.format}`,
+        `Attempts: ${failure.attempts}/${failure.maxRetries + 1}`,
+    ];
+    if (failure.key) {
+        parts.push(`Key: ${failure.key}`);
+    }
+    if (failure.parseError) {
+        parts.push(`Error: ${failure.parseError}`);
+    }
+    if (failure.issues && failure.issues.length > 0) {
+        parts.push(`Issues: ${failure.issues.join(" | ")}`);
+    }
+    if (failure.responsePreview) {
+        parts.push(`Response Preview:\n${failure.responsePreview}`);
+    }
+    return parts.join("\n");
+}
+function summarizeActionResultsForUser(actionResults) {
+    if (actionResults.length === 0) {
+        return "";
+    }
+    const summary = actionResults
+        .slice(-3)
+        .map((result) => {
+        const actionName = typeof result.data?.actionName === "string"
+            ? result.data.actionName
+            : "unknown action";
+        return `${actionName} (${result.success === false ? "failed" : "succeeded"})`;
+    })
+        .join(", ");
+    return `Completed action state before the error: ${summary}.`;
+}
+function withContextRoutingValues(state, contextRoutingStateValues) {
+    if (!contextRoutingStateValues) {
+        return state;
+    }
+    const mergedStateValues = {
+        ...state.values,
+    };
+    if (contextRoutingStateValues[AVAILABLE_CONTEXTS_STATE_KEY] !== undefined) {
+        mergedStateValues[AVAILABLE_CONTEXTS_STATE_KEY] = contextRoutingStateValues[AVAILABLE_CONTEXTS_STATE_KEY];
+    }
+    if (contextRoutingStateValues[CONTEXT_ROUTING_STATE_KEY] !== undefined) {
+        mergedStateValues[CONTEXT_ROUTING_STATE_KEY] = contextRoutingStateValues[CONTEXT_ROUTING_STATE_KEY];
+    }
+    return {
+        ...state,
+        values: mergedStateValues,
+    };
+}
+async function composeContinuationDecisionState(runtime, message, contextRoutingStateValues) {
+    // Continuation prompts run after the runtime has already persisted an
+    // assistant reply and/or action_result memories. Refresh RECENT_MESSAGES so
+    // the follow-up planner does not reuse stale conversation history cached on
+    // the original user turn.
+    return withContextRoutingValues(await runtime.composeState(message, ["RECENT_MESSAGES", "ACTIONS"], false, false), contextRoutingStateValues);
+}
+function withoutProviders(state, providerNamesToOmit) {
+    if (providerNamesToOmit.length === 0) {
+        return state;
+    }
+    const omittedProviderNames = new Set(providerNamesToOmit.map((providerName) => providerName.trim().toUpperCase()));
+    const providerResults = typeof state.data?.providers === "object" && state.data?.providers !== null
+        ? state.data.providers
+        : {};
+    const providerOrder = Array.isArray(state.data?.providerOrder)
+        ? state.data.providerOrder
+        : Object.keys(providerResults);
+    const filteredProviderOrder = providerOrder.filter((providerName) => !omittedProviderNames.has(providerName.toUpperCase()));
+    const filteredProviderResults = Object.fromEntries(Object.entries(providerResults).filter(([providerName]) => !omittedProviderNames.has(providerName.trim().toUpperCase())));
+    const filteredProvidersText = filteredProviderOrder
+        .map((providerName) => filteredProviderResults[providerName]?.text)
+        .filter((text) => typeof text === "string" && text.trim() !== "")
+        .join("\n");
+    return {
+        ...state,
+        values: {
+            ...state.values,
+            providers: filteredProvidersText,
+        },
+        data: {
+            ...state.data,
+            providerOrder: filteredProviderOrder,
+            providers: filteredProviderResults,
+        },
+        text: filteredProvidersText,
+    };
+}
+function buildShouldRespondCharacterText(providerResult) {
+    if (!providerResult) {
+        return "";
+    }
+    const values = typeof providerResult.values === "object" && providerResult.values !== null
+        ? providerResult.values
+        : {};
+    const bio = typeof values.bio === "string" ? values.bio : "";
+    const directions = typeof values.directions === "string" ? values.directions : "";
+    const system = typeof values.system === "string" ? values.system : "";
+    const classifierText = [bio, directions, system]
+        .filter((section) => section.trim().length > 0)
+        .join("\n\n");
+    return (classifierText ||
+        (typeof providerResult.text === "string" ? providerResult.text : ""));
+}
+function prepareShouldRespondState(state) {
+    const stateWithoutActions = withoutProviders(state, ["ACTIONS"]);
+    const providerResults = typeof stateWithoutActions.data?.providers === "object" &&
+        stateWithoutActions.data?.providers !== null
+        ? {
+            ...stateWithoutActions.data.providers,
+        }
+        : null;
+    if (!providerResults?.CHARACTER) {
+        return stateWithoutActions;
+    }
+    providerResults.CHARACTER = {
+        ...providerResults.CHARACTER,
+        text: buildShouldRespondCharacterText(providerResults.CHARACTER),
+    };
+    const providerOrder = Array.isArray(stateWithoutActions.data?.providerOrder)
+        ? stateWithoutActions.data.providerOrder
+        : Object.keys(providerResults);
+    const providersText = providerOrder
+        .map((providerName) => providerResults[providerName]?.text)
+        .filter((text) => typeof text === "string" && text.trim() !== "")
+        .join("\n");
+    return {
+        ...stateWithoutActions,
+        values: {
+            ...stateWithoutActions.values,
+            providers: providersText,
+        },
+        data: {
+            ...stateWithoutActions.data,
+            providers: providerResults,
+        },
+        text: providersText,
+    };
+}
+function isBenchmarkMode(state) {
+    const benchmarkFlag = state.values?.benchmark_has_context;
+    if (typeof benchmarkFlag === "boolean") {
+        return benchmarkFlag;
+    }
+    if (typeof benchmarkFlag === "string") {
+        return parseBooleanFromText(benchmarkFlag);
+    }
+    return false;
+}
+/**
+ * Default implementation of the MessageService interface.
+ * This service handles the complete message processing pipeline including:
+ * - Message validation and memory creation
+ * - Smart response decision (shouldRespond)
+ * - Single-shot or multi-step processing strategies
+ * - Action execution and evaluation
+ * - Attachment processing
+ * - Message deletion and channel clearing
+ *
+ * This is the standard message handler used by elizaOS and can be replaced
+ * with custom implementations via the IMessageService interface.
+ */
+export class DefaultMessageService {
+    /**
+     * Main message handling entry point
+     */
+    async handleMessage(runtime, message, callback, options) {
+        const source = typeof message.content?.source === "string" &&
+            message.content.source.trim() !== ""
+            ? message.content.source
+            : "messageService";
+        let trajectoryStepId = typeof message.metadata === "object" &&
+            message.metadata !== null &&
+            "trajectoryStepId" in message.metadata
+            ? message.metadata.trajectoryStepId
+            : undefined;
+        if (!(typeof trajectoryStepId === "string" && trajectoryStepId.trim() !== "")) {
+            try {
+                await runtime.emitEvent(EventType.MESSAGE_RECEIVED, {
+                    runtime,
+                    message,
+                    callback,
+                    source,
+                });
+            }
+            catch (error) {
+                runtime.logger.warn({
+                    src: "service:message",
+                    agentId: runtime.agentId,
+                    entityId: message.entityId,
+                    roomId: message.roomId,
+                    error: error instanceof Error ? error.message : String(error),
+                }, "Failed to emit MESSAGE_RECEIVED before handling message");
+            }
+            trajectoryStepId =
+                typeof message.metadata === "object" &&
+                    message.metadata !== null &&
+                    "trajectoryStepId" in message.metadata
+                    ? message.metadata.trajectoryStepId
+                    : undefined;
+        }
+        return await runWithTrajectoryContext(typeof trajectoryStepId === "string" && trajectoryStepId.trim() !== ""
+            ? {
+                trajectoryStepId: trajectoryStepId.trim(),
+                runId: runtime.getCurrentRunId?.(),
+                roomId: message.roomId,
+                messageId: message.id,
+            }
+            : undefined, async () => {
+            // Determine shouldRespondModel from options or runtime settings
+            const shouldRespondModelSetting = runtime.getSetting("SHOULD_RESPOND_MODEL");
+            const resolvedShouldRespondModel = normalizeShouldRespondModelType(options?.shouldRespondModel ?? shouldRespondModelSetting);
+            // Single ID used for tracking, streaming, and the final message (before opts / chunk wrapper).
+            const responseId = asUUID(v4());
+            // WHY voice detection wraps onStreamChunk here instead of using a
+            // separate ResponseStreamExtractor + AsyncLocalStorage context:
+            //
+            // Previously handleMessage created a second XML extractor
+            // (ResponseStreamExtractor) and injected it via runWithStreamingContext.
+            // Both extractors received the same raw LLM tokens in useModel and
+            // emitted independently, causing the dual-extractor garbling bug —
+            // consumers saw overlapping deltas that produced unintelligible TTS.
+            //
+            // The fix: a single extractor (ValidationStreamExtractor in
+            // dynamicPromptExecFromState) now provides `accumulated` — the full
+            // extracted text — via the third StreamChunkCallback argument. Voice
+            // detection wraps the caller's callback to intercept accumulated text
+            // for first-sentence detection, then forwards to the original. This
+            // keeps voice logic in handleMessage (encapsulation) without adding a
+            // second extraction pipeline.
+            //
+            // The `streamTextFallback` path exists for action handlers or other
+            // call sites that don't provide `accumulated` (raw token streams).
+            let firstSentenceSent = false;
+            let streamTextFallback = "";
+            const userOnStreamChunk = options?.onStreamChunk;
+            const wrappedOnStreamChunk = userOnStreamChunk
+                ? async (chunk, messageId, accumulated) => {
+                    let streamText;
+                    // If we have accumulated text, also sync streamTextFallback so the
+                    // fallback path has accurate state if the stream source later changes.
+                    if (accumulated !== undefined) {
+                        streamTextFallback = accumulated;
+                        streamText = accumulated;
+                    }
+                    else {
+                        streamTextFallback += chunk;
+                        streamText = streamTextFallback;
+                    }
+                    // Skip when this callback is invoked from `useModel`'s stream loop:
+                    // `source: "use_model"` already ran for the same raw chunk (Node ALS).
+                    if (getModelStreamChunkDeliveryDepth() === 0) {
+                        await runtime.applyPipelineHooks("model_stream_chunk", modelStreamChunkPipelineHookContext({
+                            source: "message_service",
+                            chunk,
+                            messageId,
+                            roomId: message.roomId,
+                            runId: runtime.getCurrentRunId(),
+                            responseId,
+                            accumulated,
+                        }));
+                    }
+                    // Only run first-sentence TTS detection when `accumulated` is present.
+                    // Raw-token streams (no accumulated) may contain XML markup or partial
+                    // structured output that would garble hasFirstSentence() and TTS.
+                    if (!firstSentenceSent &&
+                        accumulated !== undefined &&
+                        hasFirstSentence(streamText)) {
+                        const { first } = extractFirstSentence(streamText);
+                        if (first.length > 5) {
+                            firstSentenceSent = true;
+                            (async () => {
+                                try {
+                                    const voiceSettings = runtime.character.settings
+                                        ?.voice;
+                                    const model = voiceSettings?.model || "en_US-male-medium";
+                                    const voiceId = voiceSettings?.url ||
+                                        voiceSettings?.voiceId ||
+                                        "nova";
+                                    let audioBuffer = null;
+                                    const params = {
+                                        text: first,
+                                        voice: voiceId,
+                                        model: model,
+                                    };
+                                    const result = runtime.getModel(ModelType.TEXT_TO_SPEECH)
+                                        ? await runtime.useModel(ModelType.TEXT_TO_SPEECH, params)
+                                        : undefined;
+                                    if (result instanceof ArrayBuffer ||
+                                        Object.prototype.toString.call(result) ===
+                                            "[object ArrayBuffer]") {
+                                        audioBuffer = Buffer.from(result);
+                                    }
+                                    else if (Buffer.isBuffer(result)) {
+                                        audioBuffer = result;
+                                    }
+                                    else if (result instanceof Uint8Array) {
+                                        audioBuffer = Buffer.from(result);
+                                    }
+                                    if (audioBuffer && callback) {
+                                        const audioBase64 = audioBuffer.toString("base64");
+                                        await callback({
+                                            text: "",
+                                            attachments: [
+                                                {
+                                                    id: v4(),
+                                                    url: `data:audio/wav;base64,${audioBase64}`,
+                                                    title: "Voice Response",
+                                                    source: "voice-cache",
+                                                    description: "Voice response for first sentence",
+                                                    text: first,
+                                                    contentType: ContentType.AUDIO,
+                                                },
+                                            ],
+                                            source: "voice",
+                                        });
+                                    }
+                                }
+                                catch (error) {
+                                    runtime.logger.error({ error }, "Error generating voice for first sentence");
+                                }
+                            })();
+                        }
+                    }
+                    await userOnStreamChunk(chunk, messageId, accumulated);
+                }
+                : undefined;
+            const opts = {
+                maxRetries: options?.maxRetries ?? 3,
+                timeoutDuration: options?.timeoutDuration ?? 60 * 60 * 1000, // 1 hour
+                useMultiStep: options?.useMultiStep ??
+                    parseBooleanFromText(String(runtime.getSetting("USE_MULTI_STEP") ?? "")),
+                maxMultiStepIterations: options?.maxMultiStepIterations ??
+                    parseInt(String(runtime.getSetting("MAX_MULTISTEP_ITERATIONS") ?? "6"), 10),
+                continueAfterActions: options?.continueAfterActions ??
+                    parseBooleanFromText(String(runtime.getSetting("CONTINUE_AFTER_ACTIONS") ?? "true")),
+                onStreamChunk: wrappedOnStreamChunk,
+                keepExistingResponses: options?.keepExistingResponses ??
+                    parseBooleanFromText(String(runtime.getSetting("BASIC_CAPABILITIES_KEEP_RESP") ?? "")),
+                shouldRespondModel: resolvedShouldRespondModel,
+            };
+            let visibleCallbackCount = 0;
+            let firstVisibleCallbackPreview = "";
+            const deliveredCallbackKeys = new Set();
+            const instrumentedCallback = callback
+                ? async (content, actionName) => {
+                    const deliveryKey = callbackDeliveryKey(content);
+                    const preview = callbackTextPreview(content);
+                    if (deliveryKey && deliveredCallbackKeys.has(deliveryKey)) {
+                        runtime.logger.warn({
+                            src: "service:message",
+                            agentId: runtime.agentId,
+                            messageId: message.id,
+                            roomId: message.roomId,
+                            action: typeof content?.action ===
+                                "string"
+                                ? String(content.action)
+                                : actionName,
+                            source: typeof content.source === "string"
+                                ? content.source
+                                : undefined,
+                            preview: preview ||
+                                (Array.isArray(content.attachments) &&
+                                    content.attachments.length > 0
+                                    ? `[attachments:${content.attachments.length}]`
+                                    : ""),
+                        }, "Suppressing duplicate visible callback reply emitted for a single turn");
+                        return [];
+                    }
+                    if (deliveryKey) {
+                        deliveredCallbackKeys.add(deliveryKey);
+                    }
+                    if (preview) {
+                        visibleCallbackCount += 1;
+                        if (visibleCallbackCount === 1) {
+                            firstVisibleCallbackPreview = preview;
+                        }
+                        else {
+                            runtime.logger.warn({
+                                src: "service:message",
+                                agentId: runtime.agentId,
+                                messageId: message.id,
+                                roomId: message.roomId,
+                                callbackCount: visibleCallbackCount,
+                                action: typeof content?.action ===
+                                    "string"
+                                    ? String(content.action)
+                                    : actionName,
+                                source: typeof content.source === "string"
+                                    ? content.source
+                                    : undefined,
+                                firstPreview: firstVisibleCallbackPreview,
+                                currentPreview: preview,
+                            }, "Multiple visible callback replies emitted for a single turn");
+                        }
+                    }
+                    return actionName === undefined
+                        ? callback(content)
+                        : callback(content, actionName);
+                }
+                : undefined;
+            // Set up timeout monitoring
+            let timeoutId;
+            try {
+                runtime.logger.info({
+                    src: "service:message",
+                    agentId: runtime.agentId,
+                    entityId: message.entityId,
+                    roomId: message.roomId,
+                }, "Message received");
+                // Track this response ID - ensure map exists for this agent
+                let agentResponses = latestResponseIds.get(runtime.agentId);
+                if (!agentResponses) {
+                    agentResponses = new Map();
+                    latestResponseIds.set(runtime.agentId, agentResponses);
+                }
+                const previousResponseId = agentResponses.get(message.roomId);
+                if (previousResponseId) {
+                    logger.debug({
+                        src: "service:message",
+                        roomId: message.roomId,
+                        previousResponseId,
+                        responseId,
+                    }, "Updating response ID");
+                }
+                agentResponses.set(message.roomId, responseId);
+                // Start run tracking with roomId for proper log association
+                const runId = runtime.startRun(message.roomId);
+                if (!runId) {
+                    runtime.logger.error("Failed to start run tracking");
+                    return {
+                        didRespond: false,
+                        responseContent: null,
+                        responseMessages: [],
+                        state: { values: {}, data: {}, text: "" },
+                        mode: "none",
+                    };
+                }
+                const startTime = Date.now();
+                // Emit run started event
+                await runtime.emitEvent(EventType.RUN_STARTED, {
+                    runtime,
+                    source: "messageHandler",
+                    runId,
+                    messageId: message.id,
+                    roomId: message.roomId,
+                    entityId: message.entityId,
+                    startTime,
+                    status: "started",
+                });
+                const timeoutPromise = new Promise((_, reject) => {
+                    timeoutId = setTimeout(async () => {
+                        await runtime.emitEvent(EventType.RUN_TIMEOUT, {
+                            runtime,
+                            source: "messageHandler",
+                            runId,
+                            messageId: message.id,
+                            roomId: message.roomId,
+                            entityId: message.entityId,
+                            startTime,
+                            status: "timeout",
+                            endTime: Date.now(),
+                            duration: Date.now() - startTime,
+                            error: "Run exceeded timeout",
+                        });
+                        reject(new Error("Run exceeded timeout"));
+                    }, opts.timeoutDuration);
+                });
+                // Wrap processing with streaming context for automatic streaming in useModel calls
+                // Use ResponseStreamExtractor to filter XML and only stream <text> (if REPLY) or <message>
+                let streamingContext;
+                // Voice handling state
+                let firstSentenceSent = false;
+                let firstSentenceText = "";
+                if (opts.onStreamChunk) {
+                    const extractor = new ResponseStreamExtractor();
+                    const onStreamChunk = opts.onStreamChunk;
+                    let streamText = "";
+                    streamingContext = {
+                        onStreamChunk: async (chunk, msgId) => {
+                            if (extractor.done)
+                                return;
+                            const textToStream = extractor.push(chunk);
+                            if (textToStream) {
+                                streamText += textToStream;
+                                // Check for first sentence to send to voice
+                                if (!firstSentenceSent && hasFirstSentence(streamText)) {
+                                    const { first } = extractFirstSentence(streamText);
+                                    firstSentenceText = first;
+                                    if (first.length > 5) {
+                                        // Minimal length check
+                                        firstSentenceSent = true;
+                                        // Process voice in background
+                                        (async () => {
+                                            try {
+                                                const voiceSettings = runtime.character.settings
+                                                    ?.voice;
+                                                const model = voiceSettings?.model || "en_US-male-medium";
+                                                const voiceId = voiceSettings?.url ||
+                                                    voiceSettings?.voiceId ||
+                                                    "nova";
+                                                let audioBuffer = null;
+                                                const params = {
+                                                    text: first,
+                                                    voice: voiceId,
+                                                    model: model,
+                                                };
+                                                const result = runtime.getModel(ModelType.TEXT_TO_SPEECH)
+                                                    ? await runtime.useModel(ModelType.TEXT_TO_SPEECH, params)
+                                                    : undefined;
+                                                if (result instanceof ArrayBuffer ||
+                                                    Object.prototype.toString.call(result) ===
+                                                        "[object ArrayBuffer]") {
+                                                    audioBuffer = Buffer.from(result);
+                                                }
+                                                else if (Buffer.isBuffer(result)) {
+                                                    audioBuffer = result;
+                                                }
+                                                else if (result instanceof Uint8Array) {
+                                                    audioBuffer = Buffer.from(result);
+                                                }
+                                                if (audioBuffer && instrumentedCallback) {
+                                                    const audioBase64 = audioBuffer.toString("base64");
+                                                    await instrumentedCallback({
+                                                        text: "",
+                                                        attachments: [
+                                                            {
+                                                                id: v4(),
+                                                                url: `data:audio/wav;base64,${audioBase64}`,
+                                                                title: "Voice Response",
+                                                                source: "voice-cache",
+                                                                description: "Voice response for first sentence",
+                                                                text: first,
+                                                                contentType: ContentType.AUDIO,
+                                                            },
+                                                        ],
+                                                        source: "voice",
+                                                    });
+                                                }
+                                            }
+                                            catch (error) {
+                                                runtime.logger.error({ error }, "Error generating voice for first sentence");
+                                            }
+                                        })();
+                                    }
+                                }
+                                await onStreamChunk(textToStream, msgId);
+                            }
+                        },
+                        messageId: responseId,
+                    };
+                }
+                const processingPromise = runWithStreamingContext(streamingContext, () => this.processMessage(runtime, message, instrumentedCallback, responseId, runId, startTime, opts));
+                const result = await Promise.race([
+                    processingPromise,
+                    timeoutPromise,
+                ]);
+                // Clean up timeout
+                clearTimeout(timeoutId);
+                // Voice: Handle the rest of the message
+                if (firstSentenceSent && result.responseContent?.text) {
+                    const fullText = result.responseContent.text;
+                    const rest = fullText.replace(firstSentenceText, "").trim();
+                    if (rest.length > 0) {
+                        // Generate voice for rest
+                        // (Async immediately)
+                        (async () => {
+                            try {
+                                const voiceSettings = runtime.character.settings?.voice;
+                                const model = voiceSettings?.model || "en_US-male-medium";
+                                const voiceId = voiceSettings?.url || voiceSettings?.voiceId || "nova";
+                                let audioBuffer = null;
+                                const params = {
+                                    text: rest,
+                                    voice: voiceId,
+                                    model: model,
+                                };
+                                const result = runtime.getModel(ModelType.TEXT_TO_SPEECH)
+                                    ? await runtime.useModel(ModelType.TEXT_TO_SPEECH, params)
+                                    : undefined;
+                                if (result instanceof ArrayBuffer ||
+                                    Object.prototype.toString.call(result) ===
+                                        "[object ArrayBuffer]") {
+                                    audioBuffer = Buffer.from(result);
+                                }
+                                else if (Buffer.isBuffer(result)) {
+                                    audioBuffer = result;
+                                }
+                                else if (result instanceof Uint8Array) {
+                                    audioBuffer = Buffer.from(result);
+                                }
+                                if (audioBuffer && instrumentedCallback) {
+                                    const audioBase64 = audioBuffer.toString("base64");
+                                    await instrumentedCallback({
+                                        text: "",
+                                        attachments: [
+                                            {
+                                                id: v4(),
+                                                url: `data:audio/wav;base64,${audioBase64}`,
+                                                title: "Voice Response",
+                                                source: "voice",
+                                                description: "Voice response for remaining text",
+                                                text: rest,
+                                                contentType: ContentType.AUDIO,
+                                            },
+                                        ],
+                                        source: "voice",
+                                    });
+                                }
+                            }
+                            catch (error) {
+                                runtime.logger.error({ error }, "Error generating voice for remaining text");
+                            }
+                        })();
+                    }
+                }
+                return result;
+            }
+            finally {
+                clearTimeout(timeoutId);
+                // Ensure latestResponseIds is cleaned up even if processMessage
+                // threw before reaching its own cleanup at the end of the method.
+                clearLatestResponseId(runtime.agentId, message.roomId, responseId);
+            }
+        });
+    }
+    /**
+     * Internal message processing implementation
+     */
+    async processMessage(runtime, message, callback, responseId, runId, startTime, opts) {
+        const agentResponses = latestResponseIds.get(runtime.agentId);
+        if (!agentResponses)
+            throw new Error("Agent responses map not found");
+        // Skip messages from self (unless it's an autonomous message)
+        const isAutonomousMessage = message.content?.metadata &&
+            typeof message.content.metadata === "object" &&
+            message.content.metadata.isAutonomous ===
+                true;
+        if (message.entityId === runtime.agentId && !isAutonomousMessage) {
+            runtime.logger.debug({ src: "service:message", agentId: runtime.agentId }, "Skipping message from self");
+            await this.emitRunEnded(runtime, runId, message, startTime, "self");
+            return {
+                didRespond: false,
+                responseContent: null,
+                responseMessages: [],
+                state: { values: {}, data: {}, text: "" },
+                mode: "none",
+            };
+        }
+        runtime.logger.debug({
+            src: "service:message",
+            messagePreview: truncateToCompleteSentence(message.content.text || "", 50),
+        }, "Processing message");
+        // ── Save the incoming message to memory ────────────────────────────
+        runtime.logger.debug({ src: "service:message" }, "Saving message to memory");
+        let memoryToQueue;
+        if (message.id) {
+            const existingMemory = await runtime.getMemoryById(message.id);
+            if (existingMemory) {
+                runtime.logger.debug({ src: "service:message" }, "Memory already exists, skipping creation");
+                memoryToQueue = existingMemory;
+            }
+            else {
+                const createdMemoryId = await runtime.createMemory(message, "messages");
+                memoryToQueue = { ...message, id: createdMemoryId };
+            }
+            await runtime.queueEmbeddingGeneration(memoryToQueue, "high");
+        }
+        else {
+            const memoryId = await runtime.createMemory(message, "messages");
+            message.id = memoryId;
+            memoryToQueue = { ...message, id: memoryId };
+            await runtime.queueEmbeddingGeneration(memoryToQueue, "normal");
+        }
+        // Check if LLM is off by default
+        const agentUserState = await runtime.getParticipantUserState(message.roomId, runtime.agentId);
+        const defLllmOff = parseBooleanFromText(String(runtime.getSetting("BASIC_CAPABILITIES_DEFLLMOFF") || ""));
+        if (defLllmOff && agentUserState === null) {
+            runtime.logger.debug({ src: "service:message" }, "LLM is off by default");
+            await this.emitRunEnded(runtime, runId, message, startTime, "off");
+            return {
+                didRespond: false,
+                responseContent: null,
+                responseMessages: [],
+                state: { values: {}, data: {}, text: "" },
+                mode: "none",
+            };
+        }
+        // Check if room is muted
+        const agentName = runtime.character.name ?? "agent";
+        const mentionContext = message.content.mentionContext;
+        const explicitlyAddressesAgent = mentionContext?.isMention === true ||
+            mentionContext?.isReply === true ||
+            textContainsAgentName(message.content.text, [
+                runtime.character.name,
+                runtime.character.username,
+            ]);
+        if (agentUserState === "MUTED" &&
+            message.content.text &&
+            !explicitlyAddressesAgent &&
+            !message.content.text.toLowerCase().includes(agentName.toLowerCase())) {
+            runtime.logger.debug({ src: "service:message", roomId: message.roomId }, "Ignoring muted room");
+            await this.emitRunEnded(runtime, runId, message, startTime, "muted");
+            return {
+                didRespond: false,
+                responseContent: null,
+                responseMessages: [],
+                state: { values: {}, data: {}, text: "" },
+                mode: "none",
+            };
+        }
+        // Room context for shouldRespond (fetch before compose so providers see
+        // post-attachment and post-incoming-hook message state).
+        const room = await runtime.getRoom(message.roomId);
+        // Process attachments before state composition / incoming hooks
+        if (message.content.attachments && message.content.attachments.length > 0) {
+            message.content.attachments = await this.processAttachments(runtime, message.content.attachments);
+            if (message.id) {
+                await runtime.updateMemory({
+                    id: message.id,
+                    content: {
+                        ...message.content,
+                        attachments: sanitizeAttachmentsForStorage(message.content.attachments),
+                    },
+                });
+            }
+        }
+        const preIncomingHookText = typeof message.content?.text === "string" ? message.content.text : "";
+        await runtime.applyPipelineHooks("incoming_before_compose", incomingPipelineHookContext(message, {
+            roomId: message.roomId,
+            responseId,
+            runId,
+        }));
+        const postIncomingHookText = typeof message.content?.text === "string" ? message.content.text : "";
+        if (message.id && postIncomingHookText !== preIncomingHookText) {
+            await runtime.updateMemory({
+                id: message.id,
+                content: message.content,
+            });
+            await runtime.queueEmbeddingGeneration({ ...message, id: message.id }, "normal");
+        }
+        const promptAttachments = resolvePromptAttachments(message.content.attachments);
+        // Compose initial state (after incoming hooks so providers/actions text matches this turn)
+        let state = await composeResponseState(runtime, message);
+        state = attachAvailableContexts(state, runtime);
+        const metadata = typeof message.content.metadata === "object" &&
+            message.content.metadata !== null
+            ? message.content.metadata
+            : null;
+        const isAutonomous = metadata?.isAutonomous === true;
+        const autonomyMode = typeof metadata?.autonomyMode === "string" ? metadata.autonomyMode : null;
+        await runtime.applyPipelineHooks("pre_should_respond", preShouldRespondPipelineHookContext(message, {
+            roomId: message.roomId,
+            responseId,
+            runId,
+            state,
+            isAutonomous,
+        }));
+        let shouldRespondToMessage = true;
+        let terminalDecision = null;
+        let routedDecision = null;
+        let dualPressureLog = null;
+        let shouldRespondClassifierAction = null;
+        const parallelJoin = {};
+        const setTranslatedUserText = (text) => {
+            parallelJoin.translatedUserText = text;
+        };
+        const parallelHookCtx = parallelWithShouldRespondPipelineHookContext({
+            roomId: message.roomId,
+            responseId,
+            runId,
+            message,
+            state,
+            room: room ?? undefined,
+            mentionContext,
+            isAutonomous,
+            setTranslatedUserText,
+        });
+        if (isAutonomous) {
+            runtime.logger.debug({ src: "service:message", autonomyMode }, "Autonomy message bypassing shouldRespond checks");
+            shouldRespondToMessage = true;
+            await runtime.applyPipelineHooks("parallel_with_should_respond", parallelHookCtx);
+        }
+        else {
+            const [classifyOutcome] = await Promise.all([
+                this.runNonAutonomousShouldRespondClassify(runtime, message, state, room ?? undefined, mentionContext, opts, promptAttachments),
+                runtime.applyPipelineHooks("parallel_with_should_respond", parallelHookCtx),
+            ]);
+            shouldRespondToMessage = classifyOutcome.shouldRespondToMessage;
+            terminalDecision = classifyOutcome.terminalDecision;
+            routedDecision = classifyOutcome.routedDecision;
+            dualPressureLog = classifyOutcome.dualPressureLog;
+            shouldRespondClassifierAction =
+                classifyOutcome.shouldRespondClassifierAction;
+            state = classifyOutcome.state;
+        }
+        const joinedTranslation = typeof parallelJoin.translatedUserText === "string"
+            ? parallelJoin.translatedUserText
+            : undefined;
+        if (joinedTranslation !== undefined &&
+            joinedTranslation !== message.content.text) {
+            message.content.text = joinedTranslation;
+            if (message.id) {
+                await runtime.updateMemory({
+                    id: message.id,
+                    content: message.content,
+                });
+                await runtime.queueEmbeddingGeneration({ ...message, id: message.id }, "normal");
+            }
+            if (message.id) {
+                runtime.stateCache.delete(message.id);
+                runtime.stateCache.delete(`${message.id}_action_results`);
+            }
+            state = await composeResponseState(runtime, message);
+            state = attachAvailableContexts(state, runtime);
+        }
+        let responseContent = null;
+        let responseMessages = [];
+        let mode = "none";
+        // Holds a deferred simple-mode reply that will be flushed after
+        // evaluators + reflection have had a chance to override it. Declared
+        // out here so the post-evaluation flush at the bottom of handleMessage
+        // can see the same variable that the simple-mode branch sets.
+        let pendingSimpleEmit = null;
+        // Track memory IDs created for the simple-mode reply so we can clean
+        // them up if reflection overrides the deferred emit (Greptile P1 fix).
+        let pendingSimpleMemoryIds = [];
+        if (shouldRespondToMessage) {
+            const resolvedRouting = mergeContextRouting(state, message);
+            let executionState = state;
+            if (routedDecision) {
+                executionState = withContextRoutingValues(await runtime.composeState(message, ["ACTIONS", "PROVIDERS"], false, false), {
+                    [AVAILABLE_CONTEXTS_STATE_KEY]: state.values?.[AVAILABLE_CONTEXTS_STATE_KEY],
+                    [CONTEXT_ROUTING_STATE_KEY]: resolvedRouting,
+                });
+            }
+            const result = opts.useMultiStep
+                ? await this.runMultiStepCore(runtime, message, executionState, callback, opts, responseId, promptAttachments, {
+                    precomposedState: executionState,
+                })
+                : await this.runSingleShotCore(runtime, message, executionState, opts, responseId, promptAttachments, {
+                    precomposedState: executionState,
+                });
+            responseContent = result.responseContent;
+            responseMessages = result.responseMessages;
+            state = result.state;
+            mode = result.mode;
+            // Race check before we send anything
+            const currentResponseId = agentResponses.get(message.roomId);
+            if (currentResponseId !== responseId && !opts.keepExistingResponses) {
+                runtime.logger.info({
+                    src: "service:message",
+                    agentId: runtime.agentId,
+                    roomId: message.roomId,
+                }, "Response discarded - newer message being processed");
+                return {
+                    didRespond: false,
+                    responseContent: null,
+                    responseMessages: [],
+                    state,
+                    mode: "none",
+                };
+            }
+            if (responseContent && message.id) {
+                responseContent.inReplyTo = createUniqueUuid(runtime, message.id);
+            }
+            const providerStateValues = {
+                [AVAILABLE_CONTEXTS_STATE_KEY]: state.values?.[AVAILABLE_CONTEXTS_STATE_KEY],
+                [CONTEXT_ROUTING_STATE_KEY]: state.values?.[CONTEXT_ROUTING_STATE_KEY],
+            };
+            if (responseContent?.providers && responseContent.providers.length > 0) {
+                state = withContextRoutingValues(await composeProviderGroundedResponseState(runtime, message, responseContent.providers), providerStateValues);
+            }
+            if (responseContent && shouldRunProviderFollowup(responseContent)) {
+                const providerFollowupState = responseContent.providers && responseContent.providers.length > 0
+                    ? withContextRoutingValues(await composeFocusedProviderReplyState(runtime, message, responseContent.providers), providerStateValues)
+                    : state;
+                runtime.logger.info({
+                    src: "service:message",
+                    providers: responseContent.providers ?? [],
+                    actions: responseContent.actions ?? [],
+                }, "Running provider follow-up pass");
+                const providerContinuation = await this.runSingleShotCore(runtime, message, providerFollowupState, opts, responseId, promptAttachments, {
+                    precomposedState: providerFollowupState,
+                    failureStage: "answering from requested provider results",
+                    providerFollowup: true,
+                });
+                responseContent = providerContinuation.responseContent;
+                responseMessages = providerContinuation.responseMessages;
+                state = providerContinuation.state;
+                mode = providerContinuation.mode;
+                if (responseContent && message.id) {
+                    responseContent.inReplyTo = createUniqueUuid(runtime, message.id);
+                }
+                runtime.logger.info({
+                    src: "service:message",
+                    finalActions: responseContent?.actions ?? [],
+                    finalProviders: responseContent?.providers ?? [],
+                    hasText: typeof responseContent?.text === "string" &&
+                        responseContent.text.length > 0,
+                }, "Provider follow-up pass completed");
+                if (responseContent?.providers && responseContent.providers.length > 0) {
+                    state = withContextRoutingValues(await runtime.composeState(message, responseContent.providers, false, false), providerStateValues);
+                }
+            }
+            // Save response memory to database.
+            // - simple mode: persists after hooks in the branch below.
+            // - actions mode: do NOT persist the initial LLM text here.
+            //   The action callbacks produce the real user-facing messages;
+            //   saving the planner text now would emit a premature reply that
+            //   may be contradicted once the action completes or fails.
+            // - other non-simple modes (e.g. "none"): persist immediately.
+            if (responseMessages.length > 0 &&
+                mode !== "simple" &&
+                mode !== "actions") {
+                for (const responseMemory of responseMessages) {
+                    // Update the content in case inReplyTo was added
+                    if (responseContent) {
+                        responseMemory.content = responseContent;
+                    }
+                    runtime.logger.debug({ src: "service:message", memoryId: responseMemory.id }, "Saving response to memory");
+                    await runtime.createMemory(responseMemory, "messages");
+                    await this.emitMessageSent(runtime, responseMemory, message.content.source ?? "messageHandler");
+                }
+            }
+            if (responseContent) {
+                if (mode === "simple") {
+                    // Log provider usage for simple responses
+                    if (responseContent.providers &&
+                        responseContent.providers.length > 0) {
+                        runtime.logger.debug({
+                            src: "service:message",
+                            providers: responseContent.providers,
+                        }, "Simple response used providers");
+                    }
+                    // WHY order: hooks → createMemory → deferred callback matches wire + DB.
+                    await runtime.applyPipelineHooks("outgoing_before_deliver", outgoingPipelineHookContext(responseContent, {
+                        source: "simple",
+                        roomId: message.roomId,
+                        message,
+                        responseId: responseContent.responseId ?? responseMessages[0]?.id,
+                    }));
+                    if (responseMessages.length > 0) {
+                        for (const responseMemory of responseMessages) {
+                            if (responseContent) {
+                                responseMemory.content = responseContent;
+                            }
+                            runtime.logger.debug({ src: "service:message", memoryId: responseMemory.id }, "Saving response to memory");
+                            await runtime.createMemory(responseMemory, "messages");
+                            await this.emitMessageSent(runtime, responseMemory, message.content.source ?? "messageHandler");
+                            if (responseMemory.id) {
+                                pendingSimpleMemoryIds.push(responseMemory.id);
+                            }
+                        }
+                    }
+                    pendingSimpleEmit = responseContent;
+                }
+                else if (mode === "actions") {
+                    // Surface the planner's text before action handlers run, so the
+                    // user sees the agent's plan rather than silence. The full
+                    // responseContent is already persisted as a memory above.
+                    if (callback &&
+                        !isBenchmarkMode(state) &&
+                        shouldEmitPlannerPreamble(runtime, responseContent)) {
+                        await callback({
+                            ...responseContent,
+                            actions: [],
+                        });
+                    }
+                    // Pass onStreamChunk to processActions so each action can manage its own streaming context
+                    await runtime.processActions(message, responseMessages, state, async (content) => {
+                        runtime.logger.debug({ src: "service:message", content }, "Action callback");
+                        if (responseContent) {
+                            responseContent.actionCallbacks = content;
+                        }
+                        if (callback) {
+                            return callback(content);
+                        }
+                        return [];
+                    }, { onStreamChunk: opts.onStreamChunk });
+                    if (opts.continueAfterActions &&
+                        message.id &&
+                        shouldContinueAfterActions(runtime, responseContent) &&
+                        !suppressesPostActionContinuation(runtime, responseContent)) {
+                        const continuation = await this.runPostActionContinuation(runtime, message, state, callback, opts, runtime.getActionResults(message.id));
+                        if (continuation.responseMessages.length > 0) {
+                            responseMessages = [
+                                ...responseMessages,
+                                ...continuation.responseMessages,
+                            ];
+                        }
+                        if (continuation.responseContent) {
+                            responseContent = continuation.responseContent;
+                            mode = continuation.mode;
+                        }
+                        state = continuation.state;
+                    }
+                }
+            }
+        }
+        else {
+            // Agent decided not to respond
+            runtime.logger.debug({ src: "service:message" }, "Agent decided not to respond");
+            // Check if we still have the latest response ID
+            const currentResponseId = agentResponses.get(message.roomId);
+            if (currentResponseId !== responseId && !opts.keepExistingResponses) {
+                runtime.logger.info({
+                    src: "service:message",
+                    agentId: runtime.agentId,
+                    roomId: message.roomId,
+                }, "Ignore response discarded - newer message being processed");
+                await this.emitRunEnded(runtime, runId, message, startTime, "replaced");
+                return {
+                    didRespond: false,
+                    responseContent: null,
+                    responseMessages: [],
+                    state,
+                    mode: "none",
+                };
+            }
+            if (!message.id) {
+                runtime.logger.error({ src: "service:message", agentId: runtime.agentId }, "Message ID is missing, cannot create ignore response");
+                await this.emitRunEnded(runtime, runId, message, startTime, "noMessageId");
+                return {
+                    didRespond: false,
+                    responseContent: null,
+                    responseMessages: [],
+                    state,
+                    mode: "none",
+                };
+            }
+            // Construct a minimal content object indicating the terminal decision
+            const terminalAction = terminalDecision ?? "IGNORE";
+            const terminalContent = {
+                thought: terminalAction === "STOP"
+                    ? "Agent decided to stop and end the run."
+                    : "Agent decided not to respond to this message.",
+                actions: [terminalAction],
+                simple: true,
+                inReplyTo: createUniqueUuid(runtime, message.id),
+            };
+            await runtime.applyPipelineHooks("outgoing_before_deliver", outgoingPipelineHookContext(terminalContent, {
+                source: "excluded",
+                roomId: message.roomId,
+                message,
+            }));
+            const terminalMemory = {
+                id: asUUID(v4()),
+                entityId: runtime.agentId,
+                agentId: runtime.agentId,
+                content: terminalContent,
+                roomId: message.roomId,
+                createdAt: Date.now(),
+            };
+            await runtime.createMemory(terminalMemory, "messages");
+            await this.emitMessageSent(runtime, terminalMemory, message.content.source ?? "messageHandler");
+            runtime.logger.debug({ src: "service:message", memoryId: terminalMemory.id }, "Saved terminal response to memory");
+            if (callback) {
+                await callback(terminalContent);
+            }
+        }
+        // Clean up the response ID
+        clearLatestResponseId(runtime.agentId, message.roomId, responseId);
+        // Run evaluators before ending the turn because reflection can now mark
+        // the task incomplete and trigger another continuation/action pass.
+        const runEvaluate = () => runtime.evaluate(message, state, shouldRespondToMessage && !isStopResponse(responseContent), async (content) => {
+            runtime.logger.debug({ src: "service:message", content }, "Evaluate callback");
+            if (responseContent) {
+                responseContent.evalCallbacks = content;
+            }
+            if (callback) {
+                await runtime.applyPipelineHooks("outgoing_before_deliver", outgoingPipelineHookContext(content, {
+                    source: "evaluate",
+                    roomId: message.roomId,
+                    message,
+                    responseId: content.responseId,
+                }));
+                return callback(content);
+            }
+            return [];
+        }, responseMessages);
+        await runEvaluate();
+        if (opts.continueAfterActions &&
+            message.id &&
+            !isBenchmarkMode(state)) {
+            const taskCompletion = await runtime.getCache(getTaskCompletionCacheKey(message.id));
+            await runtime.deleteCache(getTaskCompletionCacheKey(message.id));
+            if (taskCompletion?.assessed &&
+                !taskCompletion.completed &&
+                // Honor `suppressPostActionContinuation` here too. The flag's
+                // contract per Action.suppressPostActionContinuation is "stop after
+                // this action — don't run any continuation LLM turn." Without this
+                // guard, an action that already emitted a complete user-facing
+                // reply (e.g. CALENDAR_ACTION) will get a second visible callback
+                // when the reflection evaluator marks the task as incomplete and
+                // triggers another LLM/processActions pass.
+                !suppressesPostActionContinuation(runtime, responseContent)) {
+                const directReplyText = typeof responseContent?.text === "string"
+                    ? responseContent.text.trim()
+                    : "";
+                let latestActionResults = [];
+                const shouldWaitForUser = isSimpleReplyResponse(responseContent) && directReplyText.length > 0
+                    ? isLikelyClarifyingQuestion(directReplyText)
+                    : (() => {
+                        latestActionResults = runtime.getActionResults(message.id);
+                        return shouldWaitForUserAfterIncompleteReflection(responseContent, latestActionResults);
+                    })();
+                if (shouldWaitForUser) {
+                    runtime.logger.debug({
+                        src: "service:message",
+                        messageId: message.id,
+                        taskCompletionReason: taskCompletion.reason,
+                        replyPreview: getLatestVisibleReplyText(responseContent, latestActionResults).slice(0, 200),
+                    }, "Skipping reflection continuation because the agent is waiting for user input");
+                }
+                else {
+                    const continuation = await this.runReflectionTaskContinuation(runtime, message, state, callback, opts, taskCompletion);
+                    if (continuation.responseMessages.length > 0) {
+                        responseMessages = [
+                            ...responseMessages,
+                            ...continuation.responseMessages,
+                        ];
+                    }
+                    if (continuation.responseContent) {
+                        responseContent = continuation.responseContent;
+                        mode = continuation.mode;
+                    }
+                    // Reflection produced a continuation (may or may not have
+                    // responseContent — e.g. actions that set results but the
+                    // helper returned early). Drop the deferred chatty REPLY
+                    // either way: emitting both would show two contradictory
+                    // messages, and even when responseContent is null the
+                    // continuation's action callbacks already went to the user.
+                    if (pendingSimpleEmit &&
+                        (continuation.responseContent ||
+                            continuation.responseMessages.length > 0)) {
+                        // Clean up orphaned memories that were persisted before
+                        // we knew reflection would override (Greptile P1 fix).
+                        for (const memId of pendingSimpleMemoryIds) {
+                            await runtime.deleteMemory(memId);
+                        }
+                        pendingSimpleMemoryIds = [];
+                        pendingSimpleEmit = null;
+                    }
+                    state = continuation.state;
+                }
+            }
+        }
+        // Flush the deferred simple-mode reply now that reflection has had its
+        // chance to override. If reflection produced its own response, this is
+        // already null and the original chatty REPLY is dropped.
+        if (pendingSimpleEmit && callback) {
+            await callback(pendingSimpleEmit);
+        }
+        const didRespond = responseMessages.length > 0 && !isStopResponse(responseContent);
+        // Collect metadata for logging
+        let entityName = "noname";
+        if (message.metadata &&
+            "entityName" in message.metadata &&
+            typeof message.metadata.entityName === "string") {
+            entityName = message.metadata.entityName;
+        }
+        const isDM = message.content && message.content.channelType === ChannelType.DM;
+        let roomName = entityName;
+        if (!isDM) {
+            const roomDatas = await runtime.getRoomsByIds([message.roomId]);
+            if (roomDatas?.length) {
+                const roomData = roomDatas[0];
+                if (roomData.name) {
+                    roomName = roomData.name;
+                }
+                if (roomData.worldId) {
+                    const worldData = await runtime.getWorld(roomData.worldId);
+                    if (worldData) {
+                        roomName = `${worldData.name}-${roomName}`;
+                    }
+                }
+            }
+        }
+        const date = new Date();
+        // Extract available actions from provider data
+        const stateData = state.data;
+        const stateDataProviders = stateData?.providers;
+        const actionsProvider = stateDataProviders?.ACTIONS;
+        const actionsProviderData = actionsProvider?.data;
+        const actionsData = actionsProviderData && "actionsData" in actionsProviderData
+            ? actionsProviderData.actionsData
+            : undefined;
+        const availableActions = actionsData?.map((a) => a.name) ?? [];
+        const _logData = {
+            at: date.toString(),
+            timestamp: Math.floor(date.getTime() / 1000),
+            messageId: message.id,
+            userEntityId: message.entityId,
+            input: message.content.text,
+            thought: responseContent?.thought,
+            simple: responseContent?.simple,
+            availableActions,
+            actions: responseContent?.actions,
+            providers: responseContent?.providers,
+            irt: responseContent?.inReplyTo,
+            output: responseContent?.text,
+            entityName,
+            source: message.content.source,
+            channelType: message.content.channelType,
+            roomName,
+        };
+        // Emit run ended event
+        await runtime.emitEvent(EventType.RUN_ENDED, {
+            runtime,
+            source: "messageHandler",
+            runId,
+            messageId: message.id,
+            roomId: message.roomId,
+            entityId: message.entityId,
+            startTime,
+            status: "completed",
+            endTime: Date.now(),
+            duration: Date.now() - startTime,
+        });
+        return {
+            didRespond,
+            responseContent,
+            responseMessages,
+            state,
+            mode,
+            ...(dualPressureLog !== null || shouldRespondClassifierAction !== null
+                ? {
+                    dualPressure: dualPressureLog,
+                    shouldRespondClassifierAction,
+                }
+                : {}),
+        };
+    }
+    async runNonAutonomousShouldRespondClassify(runtime, message, state, room, mentionContext, opts, promptAttachments) {
+        let shouldRespondToMessage = true;
+        let terminalDecision = null;
+        let routedDecision = null;
+        let dualPressureLog = null;
+        let shouldRespondClassifierAction = null;
+        let workingState = state;
+        const checkShouldRespondEnabled = runtime.isCheckShouldRespondEnabled();
+        const responseDecision = this.shouldRespond(runtime, message, room, mentionContext);
+        runtime.logger.debug({ src: "service:message", responseDecision, checkShouldRespondEnabled }, "Response decision");
+        if (!checkShouldRespondEnabled) {
+            runtime.logger.debug({ src: "service:message" }, "checkShouldRespond disabled, always responding (ChatGPT mode)");
+            shouldRespondToMessage = true;
+        }
+        else if (responseDecision.skipEvaluation) {
+            runtime.logger.debug({
+                src: "service:message",
+                agentName: runtime.character.name ?? "Agent",
+                reason: responseDecision.reason,
+            }, "Skipping LLM evaluation");
+            routedDecision = parseContextRoutingMetadata(responseDecision);
+            setContextRoutingMetadata(message, routedDecision);
+            shouldRespondToMessage = responseDecision.shouldRespond;
+        }
+        else {
+            workingState = {
+                ...workingState,
+                values: {
+                    ...workingState.values,
+                    dualPressureThreshold: resolveDualPressureThreshold(runtime),
+                },
+            };
+            const shouldRespondState = prepareShouldRespondState(workingState);
+            const optimizedPromptService = runtime.getService(OPTIMIZED_PROMPT_SERVICE);
+            const baselineShouldRespond = runtime.character.templates?.shouldRespondTemplate ||
+                shouldRespondTemplate;
+            const resolvedShouldRespondTemplate = resolveOptimizedPrompt(optimizedPromptService, "should_respond", baselineShouldRespond);
+            const _shouldRespondPrompt = composePromptFromState({
+                state: shouldRespondState,
+                template: resolvedShouldRespondTemplate,
+            });
+            runtime.logger.debug({
+                src: "service:message",
+                agentName: runtime.character.name ?? "Agent",
+                reason: responseDecision.reason,
+                model: opts.shouldRespondModel,
+            }, "Using LLM evaluation");
+            setTrajectoryPurpose("should_respond");
+            const responseObject = await runtime.dynamicPromptExecFromState({
+                state: shouldRespondState,
+                params: {
+                    prompt: resolvedShouldRespondTemplate,
+                    ...(promptAttachments ? { attachments: promptAttachments } : {}),
+                },
+                schema: [
+                    {
+                        field: "name",
+                        description: "The name of the agent responding",
+                        validateField: false,
+                        streamField: false,
+                    },
+                    {
+                        field: "reasoning",
+                        description: "Your reasoning for this decision",
+                        validateField: false,
+                        streamField: false,
+                    },
+                    {
+                        field: "speak_up",
+                        description: "Integer 0-100 pressure TO engage",
+                        validateField: false,
+                        streamField: false,
+                    },
+                    {
+                        field: "hold_back",
+                        description: "Integer 0-100 pressure to STAY QUIET",
+                        validateField: false,
+                        streamField: false,
+                    },
+                    {
+                        field: "action",
+                        description: "REPLY | RESPOND | IGNORE | STOP (REPLY and RESPOND both mean engage)",
+                        validateField: false,
+                        streamField: false,
+                    },
+                    {
+                        field: "primaryContext",
+                        description: "Primary domain context from available_contexts (e.g., wallet, knowledge)",
+                        validateField: false,
+                        streamField: false,
+                    },
+                    {
+                        field: "secondaryContexts",
+                        description: "Optional comma-separated additional domain contexts",
+                        validateField: false,
+                        streamField: false,
+                    },
+                    {
+                        field: "evidenceTurnIds",
+                        description: "Optional comma-separated message IDs that influenced this decision",
+                        validateField: false,
+                        streamField: false,
+                    },
+                ],
+                options: {
+                    contextCheckLevel: 0,
+                    maxRetries: Math.max(1, Math.min(opts.maxRetries, 2)),
+                    retryBackoff: {
+                        initialMs: 500,
+                        multiplier: 2,
+                        maxMs: 2000,
+                    },
+                    modelType: resolveShouldRespondModelType(opts.shouldRespondModel),
+                    preferredEncapsulation: "toon",
+                },
+            });
+            runtime.logger.debug({ src: "service:message", responseObject }, "Parsed evaluation result");
+            const rawAction = typeof responseObject?.action === "string" ? responseObject.action : "";
+            const actionUpper = rawAction.trim().toUpperCase();
+            const hasValidClassifierAction = actionUpper.length > 0 && ALLOWED_CLASSIFIER_ACTIONS.has(actionUpper);
+            routedDecision = parseContextRoutingMetadata(responseObject);
+            setContextRoutingMetadata(message, routedDecision);
+            if (!hasValidClassifierAction) {
+                runtime.logger.warn({
+                    src: "service:message",
+                    action: responseObject?.action,
+                }, "Classifier response missing valid action; treating as IGNORE");
+                terminalDecision = "IGNORE";
+                shouldRespondToMessage = false;
+            }
+            else {
+                const dual = applyDualPressureToClassifierAction(runtime, responseObject, rawAction);
+                dualPressureLog = dual.pressure;
+                shouldRespondClassifierAction = dual.finalActionUpper;
+                if (dual.finalActionUpper === "IGNORE" ||
+                    dual.finalActionUpper === "STOP") {
+                    terminalDecision = dual.finalActionUpper;
+                }
+                shouldRespondToMessage =
+                    dual.finalActionUpper === "REPLY" ||
+                        dual.finalActionUpper === "RESPOND";
+            }
+        }
+        return {
+            shouldRespondToMessage,
+            terminalDecision,
+            routedDecision,
+            dualPressureLog,
+            shouldRespondClassifierAction,
+            state: workingState,
+        };
+    }
+    /**
+     * Determines whether the agent should respond to a message.
+     * Uses simple rules for obvious cases (DM, mentions) and defers to LLM for ambiguous cases.
+     */
+    shouldRespond(runtime, message, room, mentionContext) {
+        if (!room) {
+            return {
+                shouldRespond: false,
+                skipEvaluation: true,
+                reason: "no room context",
+            };
+        }
+        function normalizeEnvList(value) {
+            if (!value || typeof value !== "string")
+                return [];
+            const cleaned = value.trim().replace(/^\[|\]$/g, "");
+            return cleaned
+                .split(",")
+                .map((v) => v.trim())
+                .filter(Boolean);
+        }
+        // Channel types that always trigger a response (private channels)
+        const alwaysRespondChannels = [
+            ChannelType.DM,
+            ChannelType.VOICE_DM,
+            ChannelType.SELF,
+            ChannelType.API,
+        ];
+        // Sources that always trigger a response
+        const alwaysRespondSources = ["client_chat"];
+        // Support runtime-configurable overrides via env settings
+        const customChannels = normalizeEnvList(runtime.getSetting("ALWAYS_RESPOND_CHANNELS") ??
+            runtime.getSetting("SHOULD_RESPOND_BYPASS_TYPES"));
+        const customSources = normalizeEnvList(runtime.getSetting("ALWAYS_RESPOND_SOURCES") ??
+            runtime.getSetting("SHOULD_RESPOND_BYPASS_SOURCES"));
+        const respondChannels = new Set([
+            ...alwaysRespondChannels.map((t) => t.toString()),
+            ...customChannels,
+        ].map((s) => s.trim().toLowerCase()));
+        const respondSources = [...alwaysRespondSources, ...customSources].map((s) => s.trim().toLowerCase());
+        const roomType = room.type?.toString().toLowerCase();
+        const sourceStr = message.content.source?.toLowerCase() || "";
+        const textMentionsAgentByName = textContainsAgentName(message.content.text, [runtime.character.name, runtime.character.username]);
+        const textMentionsTaggedParticipants = textContainsUserTag(message.content.text);
+        // 1. DM/VOICE_DM/API channels: always respond (private channels)
+        if (respondChannels.has(roomType)) {
+            return {
+                shouldRespond: true,
+                skipEvaluation: true,
+                reason: `private channel: ${roomType}`,
+            };
+        }
+        // 2. Specific sources (e.g., client_chat): always respond
+        if (respondSources.some((pattern) => sourceStr.includes(pattern))) {
+            return {
+                shouldRespond: true,
+                skipEvaluation: true,
+                reason: `whitelisted source: ${sourceStr}`,
+            };
+        }
+        // 3. Platform mentions and replies: always respond
+        const hasPlatformMention = !!(mentionContext?.isMention || mentionContext?.isReply);
+        if (hasPlatformMention) {
+            const mentionType = mentionContext?.isMention ? "mention" : "reply";
+            return {
+                shouldRespond: true,
+                skipEvaluation: true,
+                reason: `platform ${mentionType}`,
+            };
+        }
+        // 4. Mixed-address messages should still reach the agent when the text
+        // explicitly names it alongside other tagged participants.
+        if (textMentionsTaggedParticipants && textMentionsAgentByName) {
+            return {
+                shouldRespond: true,
+                skipEvaluation: true,
+                reason: "text address with tagged participants",
+            };
+        }
+        // 5. Clear self-modification requests should bypass the ignore-biased
+        // classifier even in group chat, but only for narrow personality/style
+        // update phrasing to avoid broad false positives.
+        if (isExplicitSelfModificationRequest(message.content.text || "")) {
+            return {
+                shouldRespond: true,
+                skipEvaluation: true,
+                reason: "explicit self-modification request",
+                primaryContext: "social",
+                secondaryContexts: ["system"],
+            };
+        }
+        // 6. All other cases are ambiguous enough to need the classifier.
+        // Lack of a platform mention is not proof the message isn't directed
+        // at the agent in a fast-moving group conversation.
+        return {
+            shouldRespond: false,
+            skipEvaluation: false,
+            reason: textMentionsAgentByName
+                ? "agent named in text requires LLM evaluation"
+                : "needs LLM evaluation",
+            primaryContext: "general",
+        };
+    }
+    /**
+     * Processes attachments by generating descriptions for supported media types.
+     */
+    async processAttachments(runtime, attachments) {
+        if (!attachments || attachments.length === 0) {
+            return [];
+        }
+        runtime.logger.debug({ src: "service:message", count: attachments.length }, "Processing attachments");
+        const processedAttachments = await Promise.all(attachments.map(async (attachment) => {
+            const processedAttachment = { ...attachment };
+            const isRemote = /^(http|https):\/\//.test(attachment.url);
+            const url = isRemote
+                ? attachment.url
+                : getLocalServerUrl(attachment.url);
+            // Only process images that don't already have descriptions
+            if (attachment.contentType === ContentType.IMAGE &&
+                !attachment.description) {
+                // Skip image analysis when vision / image-description is explicitly
+                // disabled (e.g. the user toggled the Vision capability off).
+                const disableImageDesc = runtime.getSetting("DISABLE_IMAGE_DESCRIPTION");
+                if (disableImageDesc === true || disableImageDesc === "true") {
+                    return processedAttachment;
+                }
+                runtime.logger.debug({ src: "service:message", imageUrl: attachment.url }, "Generating image description");
+                let imageUrl = url;
+                const runtimeFetch = runtime.fetch ?? globalThis.fetch;
+                const inlineData = attachment;
+                if (typeof inlineData._data === "string" &&
+                    inlineData._data.trim() &&
+                    typeof inlineData._mimeType === "string" &&
+                    inlineData._mimeType.trim()) {
+                    imageUrl = `data:${inlineData._mimeType};base64,${inlineData._data}`;
+                }
+                else if (!isRemote) {
+                    // Convert local/internal media to base64
+                    const res = await runtimeFetch(url);
+                    if (!res.ok)
+                        throw new Error(`Failed to fetch image: ${res.statusText}`);
+                    const arrayBuffer = await res.arrayBuffer();
+                    const buffer = Buffer.from(arrayBuffer);
+                    const contentType = res.headers.get("content-type") || "application/octet-stream";
+                    imageUrl = `data:${contentType};base64,${buffer.toString("base64")}`;
+                }
+                const optimizedMediaService = runtime.getService(OPTIMIZED_PROMPT_SERVICE);
+                const resolvedImagePrompt = resolveOptimizedPrompt(optimizedMediaService, "media_description", imageDescriptionTemplate);
+                const response = await runtime.useModel(ModelType.IMAGE_DESCRIPTION, {
+                    prompt: resolvedImagePrompt,
+                    imageUrl,
+                });
+                if (typeof response === "string") {
+                    const parsedXml = parseKeyValueXml(response);
+                    if (parsedXml && (parsedXml.description || parsedXml.text)) {
+                        processedAttachment.description =
+                            (typeof parsedXml.description === "string"
+                                ? parsedXml.description
+                                : "") || "";
+                        processedAttachment.title =
+                            (typeof parsedXml.title === "string"
+                                ? parsedXml.title
+                                : "Image") || "Image";
+                        processedAttachment.text =
+                            (typeof parsedXml.text === "string" ? parsedXml.text : "") ||
+                                (typeof parsedXml.description === "string"
+                                    ? parsedXml.description
+                                    : "") ||
+                                "";
+                        runtime.logger.debug({
+                            src: "service:message",
+                            descriptionPreview: processedAttachment.description?.substring(0, 100),
+                        }, "Generated image description");
+                    }
+                    else {
+                        // Fallback: Try simple regex parsing
+                        const responseStr = response;
+                        const titleMatch = responseStr.match(/<title>([^<]+)<\/title>/);
+                        const descMatch = responseStr.match(/<description>([^<]+)<\/description>/);
+                        const textMatch = responseStr.match(/<text>([^<]+)<\/text>/);
+                        if (titleMatch || descMatch || textMatch) {
+                            processedAttachment.title = titleMatch?.[1] || "Image";
+                            processedAttachment.description = descMatch?.[1] || "";
+                            processedAttachment.text =
+                                textMatch?.[1] || descMatch?.[1] || "";
+                            runtime.logger.debug({
+                                src: "service:message",
+                                descriptionPreview: processedAttachment.description?.substring(0, 100),
+                            }, "Used fallback XML parsing for description");
+                        }
+                        else {
+                            runtime.logger.warn({ src: "service:message" }, "Failed to parse XML response for image description");
+                        }
+                    }
+                }
+                else if (response &&
+                    typeof response === "object" &&
+                    "description" in response) {
+                    // Handle object responses for backwards compatibility
+                    const objResponse = response;
+                    processedAttachment.description = objResponse.description;
+                    processedAttachment.title = objResponse.title || "Image";
+                    processedAttachment.text = objResponse.description;
+                    runtime.logger.debug({
+                        src: "service:message",
+                        descriptionPreview: processedAttachment.description?.substring(0, 100),
+                    }, "Generated image description");
+                }
+                else {
+                    runtime.logger.warn({ src: "service:message" }, "Unexpected response format for image description");
+                }
+            }
+            else if (attachment.contentType === ContentType.DOCUMENT &&
+                !attachment.text) {
+                const docFetch = runtime.fetch ?? globalThis.fetch;
+                const res = await docFetch(url);
+                if (!res.ok)
+                    throw new Error(`Failed to fetch document: ${res.statusText}`);
+                const contentType = res.headers.get("content-type") || "";
+                const isPlainText = contentType.startsWith("text/plain");
+                if (isPlainText) {
+                    runtime.logger.debug({ src: "service:message", documentUrl: attachment.url }, "Processing plain text document");
+                    const textContent = await res.text();
+                    processedAttachment.text = textContent;
+                    processedAttachment.title =
+                        processedAttachment.title || "Text File";
+                    runtime.logger.debug({
+                        src: "service:message",
+                        textPreview: processedAttachment.text?.substring(0, 100),
+                    }, "Extracted text content");
+                }
+                else {
+                    runtime.logger.warn({ src: "service:message", contentType }, "Skipping non-plain-text document");
+                }
+            }
+            else if (attachment.contentType === ContentType.AUDIO &&
+                !attachment.text) {
+                runtime.logger.debug({ src: "service:message", audioUrl: attachment.url }, "Transcribing audio attachment");
+                try {
+                    let transcriptionInput = url;
+                    const audioFetch = runtime.fetch ?? globalThis.fetch;
+                    // For local/internal URLs, fetch the audio as a buffer
+                    if (!isRemote) {
+                        const res = await audioFetch(url);
+                        if (!res.ok)
+                            throw new Error(`Failed to fetch audio: ${res.statusText}`);
+                        const arrayBuffer = await res.arrayBuffer();
+                        transcriptionInput = Buffer.from(arrayBuffer);
+                    }
+                    const transcript = await runtime.useModel(ModelType.TRANSCRIPTION, transcriptionInput);
+                    if (typeof transcript === "string" && transcript.trim()) {
+                        processedAttachment.text = transcript.trim();
+                        processedAttachment.title = processedAttachment.title || "Audio";
+                        processedAttachment.description = `Transcript: ${transcript.trim()}`;
+                        runtime.logger.debug({
+                            src: "service:message",
+                            transcriptPreview: processedAttachment.text?.substring(0, 100),
+                        }, "Transcribed audio attachment");
+                    }
+                }
+                catch (err) {
+                    runtime.logger.warn({ src: "service:message", err }, "Audio transcription failed, continuing without transcript");
+                }
+            }
+            else if (attachment.contentType === ContentType.VIDEO &&
+                !attachment.text) {
+                runtime.logger.debug({ src: "service:message", videoUrl: attachment.url }, "Transcribing video attachment");
+                try {
+                    let transcriptionInput = url;
+                    const videoFetch = runtime.fetch ?? globalThis.fetch;
+                    // For local/internal URLs, fetch the video as a buffer
+                    if (!isRemote) {
+                        const res = await videoFetch(url);
+                        if (!res.ok)
+                            throw new Error(`Failed to fetch video: ${res.statusText}`);
+                        const arrayBuffer = await res.arrayBuffer();
+                        transcriptionInput = Buffer.from(arrayBuffer);
+                    }
+                    const transcript = await runtime.useModel(ModelType.TRANSCRIPTION, transcriptionInput);
+                    if (typeof transcript === "string" && transcript.trim()) {
+                        processedAttachment.text = transcript.trim();
+                        processedAttachment.title = processedAttachment.title || "Video";
+                        processedAttachment.description = `Transcript: ${transcript.trim()}`;
+                        runtime.logger.debug({
+                            src: "service:message",
+                            transcriptPreview: processedAttachment.text?.substring(0, 100),
+                        }, "Transcribed video attachment");
+                    }
+                }
+                catch (err) {
+                    runtime.logger.warn({ src: "service:message", err }, "Video transcription failed, continuing without transcript");
+                }
+            }
+            return processedAttachment;
+        }));
+        return processedAttachments;
+    }
+    async runPostActionContinuation(runtime, message, state, callback, opts, initialActionResults) {
+        const contextRoutingStateValues = {
+            [AVAILABLE_CONTEXTS_STATE_KEY]: state.values?.[AVAILABLE_CONTEXTS_STATE_KEY],
+            [CONTEXT_ROUTING_STATE_KEY]: state.values?.[CONTEXT_ROUTING_STATE_KEY],
+        };
+        const taskCompletion = state.data?.taskCompletion;
+        if (!message.id || initialActionResults.length === 0) {
+            return {
+                responseContent: null,
+                responseMessages: [],
+                state,
+                mode: "none",
+            };
+        }
+        const traceActionResults = [...initialActionResults];
+        const responseMessages = [];
+        let accumulatedState = state;
+        let responseContent = null;
+        for (let iterationCount = 0; iterationCount < opts.maxMultiStepIterations; iterationCount++) {
+            accumulatedState = withTaskCompletion(withActionResults(await composeContinuationDecisionState(runtime, message, contextRoutingStateValues), traceActionResults), taskCompletion);
+            const continuation = await this.runSingleShotCore(runtime, message, accumulatedState, opts, asUUID(v4()), resolvePromptAttachments(message.content.attachments), {
+                prompt: runtime.character.templates?.postActionDecisionTemplate ||
+                    postActionDecisionTemplate,
+                precomposedState: accumulatedState,
+                failureStage: "preparing the follow-up reply after actions",
+            });
+            if (!continuation.responseContent) {
+                runtime.logger.debug({ src: "service:message", iteration: iterationCount + 1 }, "Post-action continuation produced no response");
+                break;
+            }
+            responseContent = continuation.responseContent;
+            if (message.id) {
+                responseContent.inReplyTo = createUniqueUuid(runtime, message.id);
+            }
+            if (responseContent.providers && responseContent.providers.length > 0) {
+                accumulatedState = withActionResults(withContextRoutingValues(await composeProviderGroundedResponseState(runtime, message, responseContent.providers), contextRoutingStateValues), traceActionResults);
+            }
+            else {
+                accumulatedState = withActionResults(continuation.state, traceActionResults);
+            }
+            accumulatedState = withTaskCompletion(accumulatedState, taskCompletion);
+            if (continuation.responseMessages.length > 0 &&
+                continuation.mode !== "simple") {
+                for (const responseMemory of continuation.responseMessages) {
+                    responseMemory.content = responseContent;
+                    await runtime.createMemory(responseMemory, "messages");
+                    await this.emitMessageSent(runtime, responseMemory, message.content.source ?? "messageHandler");
+                }
+                responseMessages.push(...continuation.responseMessages);
+            }
+            if (continuation.mode === "simple") {
+                await runtime.applyPipelineHooks("outgoing_before_deliver", outgoingPipelineHookContext(responseContent, {
+                    source: "continuation_simple",
+                    roomId: message.roomId,
+                    message,
+                    responseId: responseContent.responseId ??
+                        continuation.responseMessages[0]?.id,
+                }));
+                if (continuation.responseMessages.length > 0) {
+                    for (const responseMemory of continuation.responseMessages) {
+                        responseMemory.content = responseContent;
+                        await runtime.createMemory(responseMemory, "messages");
+                        await this.emitMessageSent(runtime, responseMemory, message.content.source ?? "messageHandler");
+                    }
+                    responseMessages.push(...continuation.responseMessages);
+                }
+                if (callback) {
+                    await callback(responseContent);
+                }
+                break;
+            }
+            if (continuation.mode !== "actions") {
+                break;
+            }
+            await runtime.processActions(message, continuation.responseMessages, accumulatedState, async (content) => {
+                runtime.logger.debug({ src: "service:message", content }, "Post-action callback");
+                if (responseContent) {
+                    responseContent.actionCallbacks = content;
+                }
+                if (callback) {
+                    return callback(content);
+                }
+                return [];
+            }, { onStreamChunk: opts.onStreamChunk });
+            if (!shouldContinueAfterActions(runtime, responseContent) ||
+                suppressesPostActionContinuation(runtime, responseContent)) {
+                break;
+            }
+            const latestActionResults = runtime.getActionResults(message.id);
+            if (latestActionResults.length === 0) {
+                runtime.logger.warn({ src: "service:message", iteration: iterationCount + 1 }, "Post-action continuation produced no new action results");
+                break;
+            }
+            traceActionResults.push(...latestActionResults);
+        }
+        accumulatedState = withTaskCompletion(withActionResults(accumulatedState, traceActionResults), taskCompletion);
+        return {
+            responseContent,
+            responseMessages,
+            state: accumulatedState,
+            mode: responseContent ? "simple" : "none",
+        };
+    }
+    async runReflectionTaskContinuation(runtime, message, state, callback, opts, taskCompletion) {
+        const contextRoutingStateValues = {
+            [AVAILABLE_CONTEXTS_STATE_KEY]: state.values?.[AVAILABLE_CONTEXTS_STATE_KEY],
+            [CONTEXT_ROUTING_STATE_KEY]: state.values?.[CONTEXT_ROUTING_STATE_KEY],
+        };
+        const initialActionResults = message.id
+            ? runtime.getActionResults(message.id)
+            : [];
+        let accumulatedState = withTaskCompletion(withActionResults(await composeContinuationDecisionState(runtime, message, contextRoutingStateValues), initialActionResults), taskCompletion);
+        const continuation = await this.runSingleShotCore(runtime, message, accumulatedState, opts, asUUID(v4()), resolvePromptAttachments(message.content.attachments), {
+            prompt: runtime.character.templates?.postActionDecisionTemplate ||
+                postActionDecisionTemplate,
+            precomposedState: accumulatedState,
+            failureStage: "continuing after reflection marked the task incomplete",
+        });
+        if (!continuation.responseContent) {
+            return {
+                responseContent: null,
+                responseMessages: [],
+                state: accumulatedState,
+                mode: "none",
+            };
+        }
+        const responseMessages = [];
+        const responseContent = continuation.responseContent;
+        if (message.id) {
+            responseContent.inReplyTo = createUniqueUuid(runtime, message.id);
+        }
+        if (responseContent.providers && responseContent.providers.length > 0) {
+            accumulatedState = withTaskCompletion(withActionResults(withContextRoutingValues(await composeProviderGroundedResponseState(runtime, message, responseContent.providers), contextRoutingStateValues), initialActionResults), taskCompletion);
+        }
+        else {
+            accumulatedState = withTaskCompletion(withActionResults(continuation.state, initialActionResults), taskCompletion);
+        }
+        if (continuation.responseMessages.length > 0 &&
+            continuation.mode !== "simple") {
+            for (const responseMemory of continuation.responseMessages) {
+                responseMemory.content = responseContent;
+                await runtime.createMemory(responseMemory, "messages");
+                await this.emitMessageSent(runtime, responseMemory, message.content.source ?? "messageHandler");
+            }
+            responseMessages.push(...continuation.responseMessages);
+        }
+        if (continuation.mode === "simple") {
+            await runtime.applyPipelineHooks("outgoing_before_deliver", outgoingPipelineHookContext(responseContent, {
+                source: "continuation_simple",
+                roomId: message.roomId,
+                message,
+                responseId: responseContent.responseId ?? continuation.responseMessages[0]?.id,
+            }));
+            if (continuation.responseMessages.length > 0) {
+                for (const responseMemory of continuation.responseMessages) {
+                    responseMemory.content = responseContent;
+                    await runtime.createMemory(responseMemory, "messages");
+                    await this.emitMessageSent(runtime, responseMemory, message.content.source ?? "messageHandler");
+                }
+                responseMessages.push(...continuation.responseMessages);
+            }
+            if (callback) {
+                await callback(responseContent);
+            }
+            return {
+                responseContent,
+                responseMessages,
+                state: accumulatedState,
+                mode: "simple",
+            };
+        }
+        if (continuation.mode !== "actions") {
+            return {
+                responseContent,
+                responseMessages,
+                state: accumulatedState,
+                mode: continuation.mode,
+            };
+        }
+        await runtime.processActions(message, continuation.responseMessages, accumulatedState, async (content) => {
+            runtime.logger.debug({ src: "service:message", content }, "Reflection continuation callback");
+            responseContent.actionCallbacks = content;
+            if (callback) {
+                return callback(content);
+            }
+            return [];
+        }, { onStreamChunk: opts.onStreamChunk });
+        const latestActionResults = message.id
+            ? runtime.getActionResults(message.id)
+            : [];
+        accumulatedState = withTaskCompletion(withActionResults(accumulatedState, latestActionResults.length > 0
+            ? latestActionResults
+            : initialActionResults), taskCompletion);
+        if (latestActionResults.length > 0 &&
+            shouldContinueAfterActions(runtime, responseContent) &&
+            !suppressesPostActionContinuation(runtime, responseContent)) {
+            return await this.runPostActionContinuation(runtime, message, accumulatedState, callback, opts, latestActionResults);
+        }
+        return {
+            responseContent,
+            responseMessages,
+            state: accumulatedState,
+            mode: "actions",
+        };
+    }
+    /**
+     * Single-shot strategy: one LLM call to generate response
+     * Uses dynamicPromptExecFromState for validation-aware structured output
+     */
+    async runSingleShotCore(runtime, message, state, opts, responseId, promptAttachments, overrides) {
+        state =
+            overrides?.precomposedState ??
+                (await composeStructuredResponseState(runtime, message));
+        if (!state.values?.actionNames) {
+            runtime.logger.warn({ src: "service:message" }, "actionNames data missing from state");
+        }
+        let responseContent = null;
+        // Create streaming context for retry state tracking
+        const streamingExtractor = opts.onStreamChunk
+            ? new MarkableExtractor()
+            : undefined;
+        const streamingCtx = streamingExtractor && opts.onStreamChunk
+            ? createStreamingContext(streamingExtractor, opts.onStreamChunk, responseId)
+            : undefined;
+        // Resolve the template prompt once so it's available for both the primary
+        // call and any follow-up repair prompts (e.g. parameter repair).
+        const optimizedResponseService = runtime.getService(OPTIMIZED_PROMPT_SERVICE);
+        const baselineResponseTemplate = runtime.character.templates?.messageHandlerTemplate ||
+            messageHandlerTemplate;
+        let prompt = overrides?.prompt ||
+            resolveOptimizedPrompt(optimizedResponseService, "response", baselineResponseTemplate);
+        if (overrides?.providerFollowup) {
+            prompt = buildProviderFollowupPrompt(prompt);
+        }
+        // Use dynamicPromptExecFromState for structured output with validation
+        setTrajectoryPurpose("response");
+        const parsedXml = await runtime.dynamicPromptExecFromState({
+            state,
+            params: {
+                prompt,
+                ...(promptAttachments ? { attachments: promptAttachments } : {}),
+            },
+            schema: [
+                // WHY validateField: false on non-streamed fields?
+                // At validation level 1, each field gets validation codes by default.
+                // If a non-streamed field's code is corrupted, we'd retry unnecessarily.
+                // By opting out, we reduce token overhead AND avoid false failures.
+                {
+                    field: "thought",
+                    description: "Your internal reasoning about the message and what to do",
+                    validateField: false,
+                    streamField: false,
+                },
+                {
+                    field: "actions",
+                    description: "Ordered action entries. For XML, use one or more <action><name>ACTION_NAME</name><params>...</params></action> blocks inside <actions>.",
+                    required: false,
+                    validateField: false,
+                    streamField: false,
+                },
+                {
+                    field: "providers",
+                    description: "Optional provider names to call before the final reply or action. Use an empty field when no provider lookup is needed.",
+                    required: false,
+                    validateField: false,
+                    streamField: false,
+                },
+                // WHY streamField: true? This is the user-facing output - stream it!
+                // WHY validateField default? At level 1, we want to validate text integrity
+                {
+                    field: "text",
+                    description: "The text response to send to the user",
+                    streamField: true,
+                },
+                {
+                    field: "simple",
+                    description: "Whether this is a simple response (true/false)",
+                    validateField: false,
+                    streamField: false,
+                },
+            ],
+            options: {
+                modelType: ModelType.ACTION_PLANNER,
+                preferredEncapsulation: "xml",
+                maxRetries: opts.maxRetries,
+                // Stream through the filtered context callback for real-time output
+                onStreamChunk: streamingCtx?.onStreamChunk,
+            },
+        });
+        runtime.logger.debug({ src: "service:message", parsedXml }, "Parsed Response Content");
+        if (parsedXml) {
+            // Mark streaming as complete now that we have a valid response
+            streamingExtractor?.markComplete();
+            const finalActions = normalizePlannerActions(parsedXml, runtime);
+            responseContent = {
+                ...parsedXml,
+                thought: String(parsedXml.thought || ""),
+                actions: finalActions,
+                providers: normalizePlannerProviders(parsedXml, runtime),
+                text: String(parsedXml.text || ""),
+                simple: parsedXml.simple === true || parsedXml.simple === "true",
+            };
+        }
+        else {
+            // dynamicPromptExecFromState returned null - use streamed text if available
+            const streamedText = streamingCtx?.getStreamedText?.() || "";
+            const isTextComplete = streamingCtx?.isComplete?.() ?? false;
+            if (isTextComplete && streamedText) {
+                runtime.logger.info({
+                    src: "service:message",
+                    streamedTextLength: streamedText.length,
+                    streamedTextPreview: streamedText.substring(0, 100),
+                }, "Text extraction complete - using streamed text");
+                responseContent = {
+                    thought: "Response generated via streaming",
+                    actions: ["REPLY"],
+                    providers: [],
+                    text: streamedText,
+                    simple: true,
+                };
+            }
+            else if (streamedText && !isTextComplete) {
+                // Text was cut mid-stream - attempt continuation
+                runtime.logger.debug({
+                    src: "service:message",
+                    streamedTextLength: streamedText.length,
+                    streamedTextPreview: streamedText.substring(0, 100),
+                }, "Text cut mid-stream - attempting continuation");
+                // Reset extractor for fresh streaming of continuation
+                streamingCtx?.reset?.();
+                // Build continuation prompt with full context (reuses `prompt` from outer scope)
+                const escapedStreamedText = escapeHandlebars(streamedText);
+                const continuationPrompt = `${prompt}
+
+[CONTINUATION REQUIRED]
+Your previous response was cut off. The user already received this text:
+"${escapedStreamedText}"
+
+Continue EXACTLY from where you left off. Do NOT repeat what was already said.
+Output ONLY the continuation, starting immediately after the last character above.`;
+                const continuationParsed = await runtime.dynamicPromptExecFromState({
+                    state,
+                    params: {
+                        prompt: continuationPrompt,
+                        ...(promptAttachments ? { attachments: promptAttachments } : {}),
+                    },
+                    schema: [
+                        {
+                            field: "text",
+                            description: "Continuation of response",
+                            required: true,
+                            streamField: true,
+                        },
+                    ],
+                    options: {
+                        modelType: ModelType.ACTION_PLANNER,
+                        preferredEncapsulation: "xml",
+                        contextCheckLevel: 0, // Fast mode for continuations - we trust the model
+                        onStreamChunk: streamingCtx?.onStreamChunk,
+                    },
+                });
+                const continuationText = String(continuationParsed?.text || "");
+                const fullText = streamedText + continuationText;
+                responseContent = {
+                    thought: "Response completed via continuation",
+                    actions: ["REPLY"],
+                    providers: [],
+                    text: fullText,
+                    simple: true,
+                };
+            }
+            else {
+                runtime.logger.warn({ src: "service:message" }, "dynamicPromptExecFromState returned null");
+                const groundedFallback = await this.tryGroundedFallbackReply(runtime, message, state, responseId, promptAttachments);
+                if (groundedFallback) {
+                    return groundedFallback;
+                }
+                return await this.buildStructuredFailureReply(runtime, message, state, responseId, overrides?.failureStage ?? "preparing the reply");
+            }
+        }
+        if (!responseContent) {
+            return {
+                responseContent: null,
+                responseMessages: [],
+                state,
+                mode: "none",
+            };
+        }
+        if (!overrides?.providerFollowup &&
+            shouldAttemptProviderRescue(responseContent)) {
+            const rescuedProviders = await recoverProvidersForTurn({
+                runtime,
+                state,
+                draftReply: String(responseContent.text || ""),
+                attachments: promptAttachments,
+            });
+            if (rescuedProviders.length > 0) {
+                runtime.logger.info({
+                    src: "service:message",
+                    rescuedProviders,
+                    originalActions: responseContent.actions ?? [],
+                }, "Selected providers during reply rescue pass");
+                responseContent.providers = rescuedProviders;
+            }
+        }
+        // Action parameter repair (Python parity):
+        // If the model selected actions with required parameters but omitted <params>,
+        // do a second pass asking for ONLY a <params> block.
+        const requiredByAction = new Map();
+        const actionByName = new Map();
+        for (const action of runtime.actions) {
+            const normalizedName = action.name.trim().toUpperCase();
+            if (normalizedName) {
+                actionByName.set(normalizedName, action);
+            }
+        }
+        for (const a of responseContent.actions ?? []) {
+            const actionName = typeof a === "string" ? a.trim().toUpperCase() : "";
+            if (!actionName)
+                continue;
+            const actionDef = actionByName.get(actionName);
+            const required = actionDef?.parameters?.filter((p) => p.required).map((p) => p.name) ??
+                [];
+            if (required.length > 0) {
+                requiredByAction.set(actionName, required);
+            }
+        }
+        const existingParams = parseActionParams(responseContent.params);
+        const missingRequiredParams = () => {
+            for (const [actionName, required] of requiredByAction) {
+                const params = existingParams.get(actionName);
+                if (!params)
+                    return true;
+                for (const key of required) {
+                    if (!(key in params))
+                        return true;
+                }
+            }
+            return false;
+        };
+        if (requiredByAction.size > 0 && missingRequiredParams()) {
+            const requirementLines = Array.from(requiredByAction.entries())
+                .map(([a, req]) => `- ${a}: ${req.join(", ")}`)
+                .join("\n");
+            const repairPrompt = [
+                prompt,
+                "",
+                "# Parameter Repair",
+                "You selected actions that require parameters but did not include a complete params object.",
+                "Return ONLY XML with a top-level <params> field.",
+                "Example:",
+                "<response>",
+                "  <params>",
+                "    <SEND_MESSAGE>",
+                "      <target>room-or-channel-id</target>",
+                "      <text>message body</text>",
+                "    </SEND_MESSAGE>",
+                "  </params>",
+                "</response>",
+                "",
+                "Required parameters by action:",
+                requirementLines,
+                "",
+                "Do not include thought, actions, providers, text, or any other fields.",
+            ].join("\n");
+            const repairResponse = await runtime.useModel(ModelType.TEXT_LARGE, {
+                prompt: repairPrompt,
+            });
+            const repairParsed = parseKeyValueXml(repairResponse);
+            if (repairParsed?.params) {
+                responseContent.params = repairParsed.params;
+            }
+        }
+        const benchmarkMode = isBenchmarkMode(state);
+        // Benchmark mode (Python parity): force action-based loop when benchmark context is present.
+        if (benchmarkMode) {
+            if (!responseContent.actions || responseContent.actions.length === 0) {
+                responseContent.actions = ["REPLY"];
+            }
+            if (!responseContent.providers ||
+                responseContent.providers.length === 0) {
+                responseContent.providers = ["CONTEXT_BENCH"];
+            }
+            // Suppress any direct planner answer; the REPLY action should generate final output.
+            if (responseContent.actions.some((a) => a.toUpperCase() === "REPLY")) {
+                responseContent.text = "";
+            }
+        }
+        // LLM terminal-control ambiguity handling
+        if (responseContent.actions && responseContent.actions.length > 1) {
+            const isIgnore = (a) => typeof a === "string" && a.toUpperCase() === "IGNORE";
+            const isStop = (a) => typeof a === "string" && a.toUpperCase() === "STOP";
+            const hasIgnore = responseContent.actions.some(isIgnore);
+            const hasStop = responseContent.actions.some(isStop);
+            if (hasIgnore) {
+                if (!responseContent.text || responseContent.text.trim() === "") {
+                    responseContent.actions = ["IGNORE"];
+                }
+                else {
+                    const filtered = responseContent.actions.filter((a) => !isIgnore(a));
+                    responseContent.actions = filtered.length ? filtered : ["REPLY"];
+                }
+            }
+            if (hasStop) {
+                const filtered = responseContent.actions.filter((a) => !isStop(a));
+                responseContent.actions = filtered.length ? filtered : ["STOP"];
+            }
+        }
+        // Automatically determine if response is simple
+        const isSimple = isSimpleReplyResponse(responseContent);
+        const isStop = isStopResponse(responseContent);
+        responseContent.simple = isSimple;
+        // Include message ID for streaming coordination (so broadcast uses same ID)
+        responseContent.responseId = responseId;
+        const responseMessages = [
+            {
+                id: responseId,
+                entityId: runtime.agentId,
+                agentId: runtime.agentId,
+                content: responseContent,
+                roomId: message.roomId,
+                createdAt: Date.now(),
+            },
+        ];
+        return {
+            responseContent,
+            responseMessages,
+            state,
+            mode: isStop
+                ? "none"
+                : isSimple && responseContent.text
+                    ? "simple"
+                    : "actions",
+        };
+    }
+    async tryGroundedFallbackReply(runtime, message, state, responseId, promptAttachments) {
+        let groundedState = state;
+        const selectedProviders = await recoverProvidersForTurn({
+            runtime,
+            state,
+            attachments: promptAttachments,
+        });
+        if (selectedProviders.length > 0) {
+            groundedState = await composeFocusedProviderReplyState(runtime, message, selectedProviders);
+        }
+        const prompt = composePromptFromState({
+            state: groundedState,
+            template: buildGroundedFallbackReplyPrompt(),
+        });
+        try {
+            const result = await runtime.useModel(ModelType.TEXT_SMALL, {
+                prompt,
+                ...(promptAttachments ? { attachments: promptAttachments } : {}),
+            });
+            const text = typeof result === "string" ? result.trim() : "";
+            if (!text) {
+                return null;
+            }
+            const responseContent = {
+                thought: selectedProviders.length > 0
+                    ? "Grounded fallback reply from selected providers"
+                    : "Grounded fallback reply",
+                actions: ["REPLY"],
+                providers: selectedProviders,
+                text,
+                simple: true,
+                responseId,
+            };
+            const responseMessages = [
+                {
+                    id: responseId,
+                    entityId: runtime.agentId,
+                    agentId: runtime.agentId,
+                    content: responseContent,
+                    roomId: message.roomId,
+                    createdAt: Date.now(),
+                },
+            ];
+            return {
+                responseContent,
+                responseMessages,
+                state: groundedState,
+                mode: "simple",
+            };
+        }
+        catch (error) {
+            runtime.logger.warn({
+                src: "service:message",
+                error: error instanceof Error ? error.message : String(error),
+            }, "Grounded fallback reply generation failed");
+            return null;
+        }
+    }
+    async buildStructuredFailureReply(runtime, message, state, responseId, stage) {
+        const failure = getStructuredOutputFailure(state);
+        const recentMessages = typeof state.values?.recentMessages === "string" &&
+            state.values.recentMessages.trim().length > 0
+            ? state.values.recentMessages
+            : typeof state.text === "string" && state.text.trim().length > 0
+                ? state.text
+                : typeof message.content.text === "string"
+                    ? message.content.text
+                    : "(unavailable)";
+        const actionResults = Array.isArray(state.data?.actionResults)
+            ? state.data.actionResults
+            : [];
+        const failurePrompt = [
+            "You are recovering from an internal structured-output failure while responding to a user.",
+            "Write the next user-facing reply in plain language.",
+            "",
+            "Rules:",
+            "- Explain what failed and why using only the diagnostics below.",
+            "- Mention any completed or failed actions if action results are available.",
+            "- Be transparent, concise, and avoid inventing causes.",
+            "- If the model returned malformed XML, TOON, or JSON, say that clearly.",
+            "- Suggest the most useful next step for the user.",
+            "- Return only the reply text. No XML, JSON, TOON, bullet labels, or <think>.",
+            "",
+            `Failure Stage: ${stage}`,
+            "",
+            "Structured Failure Diagnostics:",
+            summarizeStructuredOutputFailure(failure),
+            "",
+            "Recent Conversation:",
+            recentMessages,
+            "",
+            "Action Results So Far:",
+            typeof state.values?.actionResults === "string" &&
+                state.values.actionResults.trim().length > 0
+                ? state.values.actionResults
+                : "No action results available.",
+            "",
+            "Reply:",
+        ].join("\n");
+        let replyText = "";
+        for (const modelType of [
+            ModelType.TEXT_LARGE,
+            ModelType.RESPONSE_HANDLER,
+            ModelType.TEXT_SMALL,
+            ModelType.TEXT_NANO,
+        ]) {
+            try {
+                const response = await runtime.useModel(modelType, {
+                    prompt: failurePrompt,
+                });
+                if (typeof response !== "string") {
+                    continue;
+                }
+                const cleaned = response
+                    .replace(/<think>[\s\S]*?<\/think>/g, "")
+                    .trim();
+                const looksStructuredReply = cleaned.startsWith("<") ||
+                    /^TOON\b/i.test(cleaned) ||
+                    /^(thought|text)\s*:/i.test(cleaned);
+                const parsed = looksStructuredReply
+                    ? parseKeyValueXml(cleaned)
+                    : null;
+                replyText =
+                    typeof parsed?.text === "string" && parsed.text.trim().length > 0
+                        ? parsed.text.trim()
+                        : cleaned;
+                if (replyText) {
+                    break;
+                }
+            }
+            catch (error) {
+                runtime.logger.warn({
+                    src: "service:message",
+                    stage,
+                    modelType,
+                    error: error instanceof Error ? error.message : String(error),
+                }, "Structured failure reply generation failed for model");
+            }
+        }
+        if (!replyText) {
+            const failureReason = failure?.parseError ??
+                failure?.issues?.[0] ??
+                "the model returned output that did not match the required format";
+            replyText = [
+                `I hit an internal parsing error while ${stage}.`,
+                `Reason: ${failureReason}.`,
+                summarizeActionResultsForUser(actionResults),
+                "Please try again or ask me to retry the last step.",
+            ]
+                .filter(Boolean)
+                .join(" ");
+        }
+        replyText = truncateToCompleteSentence(replyText.trim(), 2000);
+        const responseContent = {
+            thought: `Explain the structured-output failure during ${stage}.`,
+            actions: ["REPLY"],
+            providers: [],
+            text: replyText,
+            simple: true,
+            responseId,
+        };
+        const responseMessages = [
+            {
+                id: responseId,
+                entityId: runtime.agentId,
+                agentId: runtime.agentId,
+                content: responseContent,
+                roomId: message.roomId,
+                createdAt: Date.now(),
+            },
+        ];
+        return {
+            responseContent,
+            responseMessages,
+            state,
+            mode: "simple",
+        };
+    }
+    /**
+     * Multi-step strategy: iterative action execution with final summary
+     */
+    async runMultiStepCore(runtime, message, state, callback, opts, responseId, promptAttachments, overrides) {
+        const contextRoutingStateValues = {
+            [AVAILABLE_CONTEXTS_STATE_KEY]: overrides?.precomposedState?.values?.[AVAILABLE_CONTEXTS_STATE_KEY],
+            [CONTEXT_ROUTING_STATE_KEY]: overrides?.precomposedState?.values?.[CONTEXT_ROUTING_STATE_KEY],
+        };
+        const traceActionResult = [];
+        let accumulatedState = state;
+        let iterationCount = 0;
+        while (iterationCount < opts.maxMultiStepIterations) {
+            iterationCount++;
+            runtime.logger.debug({
+                src: "service:message",
+                iteration: iterationCount,
+                maxIterations: opts.maxMultiStepIterations,
+            }, "Starting multi-step iteration");
+            accumulatedState = withContextRoutingValues((await runtime.composeState(message, ["RECENT_MESSAGES", "ACTION_STATE", "PROVIDERS"], false, false)), contextRoutingStateValues);
+            accumulatedState.data.actionResults = traceActionResult;
+            // Use dynamicPromptExecFromState for structured decision output
+            const optimizedPlannerService = runtime.getService(OPTIMIZED_PROMPT_SERVICE);
+            const baselinePlannerTemplate = runtime.character.templates?.multiStepDecisionTemplate ||
+                multiStepDecisionTemplate;
+            const resolvedPlannerTemplate = resolveOptimizedPrompt(optimizedPlannerService, "action_planner", baselinePlannerTemplate);
+            const parsedStep = await runtime.dynamicPromptExecFromState({
+                state: accumulatedState,
+                params: {
+                    prompt: resolvedPlannerTemplate,
+                    ...(promptAttachments ? { attachments: promptAttachments } : {}),
+                },
+                schema: [
+                    // Multi-step decision loop - internal reasoning, no streaming needed
+                    // WHY: This is orchestration logic, not user-facing output
+                    {
+                        field: "thought",
+                        description: "Your reasoning for the selected providers and/or action, and how this step contributes to resolving the user's request",
+                        validateField: false,
+                        streamField: false,
+                    },
+                    {
+                        field: "providers",
+                        description: "Comma-separated list of providers to call to gather necessary data",
+                        validateField: false,
+                        streamField: false,
+                    },
+                    {
+                        field: "action",
+                        description: "Name of the action to execute after providers return (can be empty if no action is needed)",
+                        validateField: false,
+                        streamField: false,
+                    },
+                    // WHY parameters: Actions need input data. Without this field in the schema,
+                    // the LLM won't be instructed to output parameters, breaking action execution.
+                    {
+                        field: "params",
+                        description: "Optional TOON parameters for the selected action. Use a `params` object keyed by action name when the action needs input.",
+                        validateField: false,
+                        streamField: false,
+                    },
+                    {
+                        field: "isFinish",
+                        description: "true if the task is fully resolved and no further steps are needed, false otherwise",
+                        validateField: false,
+                        streamField: false,
+                    },
+                ],
+                options: {
+                    modelType: ModelType.ACTION_PLANNER,
+                    preferredEncapsulation: "toon",
+                },
+            });
+            if (!parsedStep) {
+                runtime.logger.warn({ src: "service:message", iteration: iterationCount }, "Failed to parse multi-step result");
+                traceActionResult.push({
+                    data: { actionName: "parse_error" },
+                    success: false,
+                    error: "Failed to parse step result",
+                });
+                return await this.buildStructuredFailureReply(runtime, message, withActionResults(accumulatedState, traceActionResult), responseId, "planning the next multi-step action");
+            }
+            const thought = typeof parsedStep.thought === "string" ? parsedStep.thought : undefined;
+            // Handle providers as comma-separated string or array
+            let providers = [];
+            if (Array.isArray(parsedStep.providers)) {
+                providers = parsedStep.providers;
+            }
+            else if (typeof parsedStep.providers === "string") {
+                providers = parsedStep.providers
+                    .split(",")
+                    .map((p) => p.trim())
+                    .filter((p) => p.length > 0);
+            }
+            const action = typeof parsedStep.action === "string" ? parsedStep.action : undefined;
+            const isFinish = parsedStep.isFinish;
+            // Check for completion condition
+            if (isFinish === "true" || isFinish === true) {
+                runtime.logger.info({
+                    src: "service:message",
+                    agentId: runtime.agentId,
+                    iteration: iterationCount,
+                }, "Multi-step task completed");
+                if (callback) {
+                    await callback({
+                        text: "",
+                        thought: typeof thought === "string" ? thought : "",
+                    });
+                }
+                break;
+            }
+            // Validate that we have something to do
+            const providersArray = Array.isArray(providers) ? providers : [];
+            if ((!providersArray || providersArray.length === 0) && !action) {
+                runtime.logger.warn({ src: "service:message", iteration: iterationCount }, "No providers or action specified, forcing completion");
+                break;
+            }
+            // Total timeout for all providers running in parallel (configurable via PROVIDERS_TOTAL_TIMEOUT_MS env var)
+            // Since providers run in parallel, this is the max wall-clock time allowed
+            const PROVIDERS_TOTAL_TIMEOUT_MS = parseInt(String(runtime.getSetting("PROVIDERS_TOTAL_TIMEOUT_MS") || "1000"), 10);
+            // Track which providers have completed (for timeout diagnostics)
+            const completedProviders = new Set();
+            const providerByName = new Map(runtime.providers.map((provider) => [provider.name, provider]));
+            const providerPromises = [];
+            for (const name of providersArray) {
+                if (typeof name !== "string")
+                    continue;
+                providerPromises.push((async (providerName) => {
+                    const provider = providerByName.get(providerName);
+                    if (!provider) {
+                        runtime.logger.warn({ src: "service:message", providerName }, "Provider not found");
+                        completedProviders.add(providerName);
+                        return {
+                            providerName,
+                            success: false,
+                            error: `Provider not found: ${providerName}`,
+                        };
+                    }
+                    try {
+                        const providerResult = await provider.get(runtime, message, state);
+                        completedProviders.add(providerName);
+                        if (!providerResult) {
+                            runtime.logger.warn({ src: "service:message", providerName }, "Provider returned no result");
+                            return {
+                                providerName,
+                                success: false,
+                                error: "Provider returned no result",
+                            };
+                        }
+                        const success = !!providerResult.text;
+                        return {
+                            providerName,
+                            success,
+                            text: success ? providerResult.text : undefined,
+                            error: success ? undefined : "Provider returned no result",
+                        };
+                    }
+                    catch (err) {
+                        completedProviders.add(providerName);
+                        const errorMsg = err instanceof Error ? err.message : String(err);
+                        runtime.logger.error({ src: "service:message", providerName, error: errorMsg }, "Provider execution failed");
+                        return { providerName, success: false, error: errorMsg };
+                    }
+                })(name));
+            }
+            // Create timeout promise for provider execution (with cleanup)
+            let timeoutId;
+            const timeoutPromise = new Promise((resolve) => {
+                timeoutId = setTimeout(() => resolve("timeout"), PROVIDERS_TOTAL_TIMEOUT_MS);
+            });
+            // Race between all providers completing and timeout
+            const allProvidersPromise = Promise.allSettled(providerPromises);
+            const raceResult = await Promise.race([
+                allProvidersPromise,
+                timeoutPromise,
+            ]);
+            // Clear timeout if providers completed first
+            if (timeoutId !== undefined) {
+                clearTimeout(timeoutId);
+            }
+            // Check if providers took too long - abort pipeline and notify user
+            if (raceResult === "timeout") {
+                // Identify which providers were still pending when timeout hit
+                const allProviderNames = providersArray.filter((name) => typeof name === "string");
+                const pendingProviders = allProviderNames.filter((name) => !completedProviders.has(name));
+                runtime.logger.error({
+                    src: "service:message",
+                    timeoutMs: PROVIDERS_TOTAL_TIMEOUT_MS,
+                    pendingProviders,
+                    completedProviders: Array.from(completedProviders),
+                }, `Providers took too long (>${PROVIDERS_TOTAL_TIMEOUT_MS}ms) - slow providers: ${pendingProviders.join(", ")}`);
+                if (callback) {
+                    const timeoutContent = {
+                        text: "Providers took too long to respond. Please optimize your providers or use caching.",
+                        actions: [],
+                        thought: "Provider timeout - pipeline aborted",
+                    };
+                    await runtime.applyPipelineHooks("outgoing_before_deliver", outgoingPipelineHookContext(timeoutContent, {
+                        source: "simple",
+                        roomId: message.roomId,
+                        message,
+                    }));
+                    await callback(timeoutContent);
+                }
+                return {
+                    responseContent: null,
+                    responseMessages: [],
+                    state,
+                    mode: "none",
+                };
+            }
+            // Providers completed in time
+            const providerResults = raceResult;
+            // Process results and notify via callback
+            for (const result of providerResults) {
+                if (result.status === "fulfilled") {
+                    const { providerName, success, text, error } = result.value;
+                    traceActionResult.push({
+                        data: { actionName: providerName },
+                        success,
+                        text,
+                        error,
+                    });
+                    if (callback) {
+                        await callback({
+                            text: `🔎 Provider executed: ${providerName}`,
+                            actions: [providerName],
+                            thought: typeof thought === "string" ? thought : "",
+                        });
+                    }
+                }
+                else {
+                    runtime.logger.error({
+                        src: "service:message",
+                        error: result.reason || "Unknown provider failure",
+                    }, "Unexpected provider promise rejection");
+                }
+            }
+            if (action) {
+                const actionContent = {
+                    text: `🔎 Executing action: ${action}`,
+                    actions: [action],
+                    thought: thought || "",
+                };
+                if (parsedStep && typeof parsedStep.params === "string") {
+                    actionContent.params = parsedStep.params;
+                }
+                await runtime.processActions(message, [
+                    {
+                        id: v4(),
+                        entityId: runtime.agentId,
+                        roomId: message.roomId,
+                        createdAt: Date.now(),
+                        content: actionContent,
+                    },
+                ], state, async () => {
+                    return [];
+                });
+                // Get cached action results from runtime
+                const cachedState = runtime.stateCache.get(`${message.id}_action_results`);
+                const cachedStateValues = cachedState?.values;
+                const rawActionResults = cachedStateValues?.actionResults;
+                const actionResults = Array.isArray(rawActionResults)
+                    ? rawActionResults
+                    : [];
+                const result = actionResults.length > 0 ? actionResults[0] : null;
+                const success = result?.success ?? false;
+                traceActionResult.push({
+                    data: { actionName: typeof action === "string" ? action : "unknown" },
+                    success,
+                    text: result && "text" in result && typeof result.text === "string"
+                        ? result.text
+                        : undefined,
+                    values: result &&
+                        "values" in result &&
+                        typeof result.values === "object" &&
+                        result.values !== null
+                        ? result.values
+                        : undefined,
+                    error: success
+                        ? undefined
+                        : result && "text" in result && typeof result.text === "string"
+                            ? result.text
+                            : undefined,
+                });
+            }
+        }
+        if (iterationCount >= opts.maxMultiStepIterations) {
+            runtime.logger.warn({ src: "service:message", maxIterations: opts.maxMultiStepIterations }, "Reached maximum iterations, forcing completion");
+        }
+        accumulatedState = withContextRoutingValues((await runtime.composeState(message, ["RECENT_MESSAGES", "ACTION_STATE"], false, false)), contextRoutingStateValues);
+        // Use dynamicPromptExecFromState for final summary generation
+        // Stream the final summary for better UX
+        const summary = await runtime.dynamicPromptExecFromState({
+            state: accumulatedState,
+            params: {
+                prompt: runtime.character.templates?.multiStepSummaryTemplate ||
+                    multiStepSummaryTemplate,
+                ...(promptAttachments ? { attachments: promptAttachments } : {}),
+            },
+            schema: [
+                {
+                    field: "thought",
+                    description: "Your internal reasoning about the summary",
+                    validateField: false,
+                    streamField: false,
+                },
+                // WHY streamField: true? This is the final user-facing output
+                {
+                    field: "text",
+                    description: "The final summary message to send to the user",
+                    required: true,
+                    streamField: true,
+                },
+            ],
+            options: {
+                modelSize: "large",
+                preferredEncapsulation: opts.onStreamChunk ? "xml" : "toon",
+                requiredFields: ["text"],
+                // Stream the final summary to the user
+                onStreamChunk: opts.onStreamChunk,
+            },
+        });
+        let responseContent = null;
+        const summaryText = summary?.text;
+        if (typeof summaryText === "string" && summaryText) {
+            responseContent = {
+                actions: ["MULTI_STEP_SUMMARY"],
+                text: summaryText,
+                thought: (typeof summary?.thought === "string"
+                    ? summary.thought
+                    : "Final user-facing message after task completion.") ||
+                    "Final user-facing message after task completion.",
+                simple: true,
+                responseId,
+            };
+        }
+        else {
+            return await this.buildStructuredFailureReply(runtime, message, withActionResults(accumulatedState, traceActionResult), responseId, "writing the final summary");
+        }
+        const responseMessages = responseContent
+            ? [
+                {
+                    id: responseId,
+                    entityId: runtime.agentId,
+                    agentId: runtime.agentId,
+                    content: responseContent,
+                    roomId: message.roomId,
+                    createdAt: Date.now(),
+                },
+            ]
+            : [];
+        return {
+            responseContent,
+            responseMessages,
+            state: accumulatedState,
+            mode: responseContent ? "simple" : "none",
+        };
+    }
+    /**
+     * Helper to emit run ended events
+     */
+    async emitRunEnded(runtime, runId, message, startTime, status) {
+        await runtime.emitEvent(EventType.RUN_ENDED, {
+            runtime,
+            source: "messageHandler",
+            runId,
+            messageId: message.id,
+            roomId: message.roomId,
+            entityId: message.entityId,
+            startTime,
+            status: status,
+            endTime: Date.now(),
+            duration: Date.now() - startTime,
+        });
+    }
+    async emitMessageSent(runtime, message, source) {
+        await runtime.emitEvent(EventType.MESSAGE_SENT, {
+            runtime,
+            message,
+            source,
+        });
+    }
+    /**
+     * Deletes a message from the agent's memory.
+     * This method handles the actual deletion logic that was previously in event handlers.
+     *
+     * @param runtime - The agent runtime instance
+     * @param message - The message memory to delete
+     * @returns Promise resolving when deletion is complete
+     */
+    async deleteMessage(runtime, message) {
+        if (!message.id) {
+            runtime.logger.error({ src: "service:message", agentId: runtime.agentId }, "Cannot delete memory: message ID is missing");
+            return;
+        }
+        runtime.logger.info({
+            src: "service:message",
+            agentId: runtime.agentId,
+            messageId: message.id,
+            roomId: message.roomId,
+        }, "Deleting memory");
+        await runtime.deleteMemory(message.id);
+        runtime.logger.debug({ src: "service:message", messageId: message.id }, "Successfully deleted memory");
+    }
+    /**
+     * Clears all messages from a channel/room.
+     * This method handles bulk deletion of all message memories in a room.
+     *
+     * @param runtime - The agent runtime instance
+     * @param roomId - The room ID to clear messages from
+     * @param channelId - The original channel ID (for logging)
+     * @returns Promise resolving when channel is cleared
+     */
+    async clearChannel(runtime, roomId, channelId) {
+        runtime.logger.info({ src: "service:message", agentId: runtime.agentId, channelId, roomId }, "Clearing message memories from channel");
+        // Get all message memories for this room
+        const memories = await runtime.getMemoriesByRoomIds({
+            tableName: "messages",
+            roomIds: [roomId],
+        });
+        runtime.logger.debug({ src: "service:message", channelId, count: memories.length }, "Found message memories to delete");
+        // Delete each message memory
+        let deletedCount = 0;
+        for (const memory of memories) {
+            if (memory.id) {
+                try {
+                    await runtime.deleteMemory(memory.id);
+                    deletedCount++;
+                }
+                catch (error) {
+                    runtime.logger.warn({ src: "service:message", error, memoryId: memory.id }, "Failed to delete message memory");
+                }
+            }
+        }
+        runtime.logger.info({
+            src: "service:message",
+            agentId: runtime.agentId,
+            channelId,
+            deletedCount,
+            totalCount: memories.length,
+        }, "Cleared message memories from channel");
+    }
+}