@bike4mind/cli 0.20.0 → 0.20.1

package/dist/{AgentHistoryStore-C8uUKjjC.mjs → AgentHistoryStore-BQiATPsQ.mjs}

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { $ as WORK_ITEM_STATUSES, A as HttpStatus, At as usdToCredits, B as REASONING_EFFORT_INCOMPATIBLE_WITH_TOOLS_MODELS, Ct as mapMimeTypeToArtifactType, D as FORMAT_PROMPT_TEMPLATE, Dt as settingsMap, E as FIXED_TEMPERATURE_MODELS, Et as secureParameters, F as NO_TEMPERATURE_MODELS, Ft as isNearLimit, G as SupportedFabFileMimeTypes, H as REFUSAL_FALLBACK_MODELS, I as NotFoundError, It as parseRateLimitHeaders, J as UnauthorizedError, K as TTS_MAX_INPUT_CHARS, L as OllamaEmbeddingModel, M as InternalServerError, Mt as withRetry, N as MODEL_INFO_FIELD_GROUP_OF, Nt as buildRateLimitLogEntry, O as ForbiddenError, Ot as toModelInfo, P as ModelBackend, Pt as extractSnippetMeta, Q as VoyageAIEmbeddingModel, R as OpenAIEmbeddingModel, S as CorruptedFileError, St as isZodError, T as FIELD_GROUP_OF, Tt as resolveHistoryFetchLimit, U as RESPONSES_API_TOOL_MODELS, V as REASONING_SUPPORTED_MODELS, W as SpeechToTextModels, X as VIDEO_SIZE_CONSTRAINTS, Y as UnprocessableEntityError, Z as VideoModels, _ as BadRequestError, _t as isRenderableModelType, at as getQuestErrorCode, bt as isUnlimitedHistory, ct as isFieldGroup, dt as isImageAttachment, et as applyModelPriceCatalog, ft as isImageServeable, g as BFL_SAFETY_TOLERANCE, gt as isPlaceholderApiKey, h as BEDROCK_NO_PROMPT_CACHING_MODELS, ht as isModelDeprecated, it as getMcpProviderMetadata, j as ImageModels, jt as usdToCreditsStochastic, k as HTTPError, kt as toModelRecord, lt as isGPTImage2Model, m as ApiKeyType, mt as isModelAccessible, n as logger, nt as dayjsConfig_default, ot as getRetryAfterMs, p as ARTIFACT_ATTRS_PATTERN, pt as isMediaModelType, q as TooManyRequestsError, rt as defaultEmbeddingModelForEnv, st as isAudioMimeType, tt as calculateRetryDelay, ut as isGPTImageModel, v as BedrockEmbeddingModel, vt as isRetryableError, w as DEFAULT_UNKNOWN_CONTEXT_WINDOW, wt as obfuscateApiKey, x as ChatModels, xt as isUserInitiatedAbort, y as CONTEXT_WINDOW_SAFETY_BUFFER_TOKENS, yt as isSupportedFabFileMimeType, z as PermissionDeniedError } from "./ConfigStore-DD3DcC3-.mjs";
+import { $ as WORK_ITEM_STATUSES, A as HttpStatus, At as usdToCredits, B as REASONING_EFFORT_INCOMPATIBLE_WITH_TOOLS_MODELS, Ct as mapMimeTypeToArtifactType, D as FORMAT_PROMPT_TEMPLATE, Dt as settingsMap, E as FIXED_TEMPERATURE_MODELS, Et as secureParameters, F as NO_TEMPERATURE_MODELS, Ft as isNearLimit, G as SupportedFabFileMimeTypes, H as REFUSAL_FALLBACK_MODELS, I as NotFoundError, It as parseRateLimitHeaders, J as UnauthorizedError, K as TTS_MAX_INPUT_CHARS, L as OllamaEmbeddingModel, M as InternalServerError, Mt as withRetry, N as MODEL_INFO_FIELD_GROUP_OF, Nt as buildRateLimitLogEntry, O as ForbiddenError, Ot as toModelInfo, P as ModelBackend, Pt as extractSnippetMeta, Q as VoyageAIEmbeddingModel, R as OpenAIEmbeddingModel, S as CorruptedFileError, St as isZodError, T as FIELD_GROUP_OF, Tt as resolveHistoryFetchLimit, U as RESPONSES_API_TOOL_MODELS, V as REASONING_SUPPORTED_MODELS, W as SpeechToTextModels, X as VIDEO_SIZE_CONSTRAINTS, Y as UnprocessableEntityError, Z as VideoModels, _ as BadRequestError, _t as isRenderableModelType, at as getQuestErrorCode, bt as isUnlimitedHistory, ct as isFieldGroup, dt as isImageAttachment, et as applyModelPriceCatalog, ft as isImageServeable, g as BFL_SAFETY_TOLERANCE, gt as isPlaceholderApiKey, h as BEDROCK_NO_PROMPT_CACHING_MODELS, ht as isModelDeprecated, it as getMcpProviderMetadata, j as ImageModels, jt as usdToCreditsStochastic, k as HTTPError, kt as toModelRecord, lt as isGPTImage2Model, m as ApiKeyType, mt as isModelAccessible, n as logger, nt as dayjsConfig_default, ot as getRetryAfterMs, p as ARTIFACT_ATTRS_PATTERN, pt as isMediaModelType, q as TooManyRequestsError, rt as defaultEmbeddingModelForEnv, st as isAudioMimeType, tt as calculateRetryDelay, ut as isGPTImageModel, v as BedrockEmbeddingModel, vt as isRetryableError, w as DEFAULT_UNKNOWN_CONTEXT_WINDOW, wt as obfuscateApiKey, x as ChatModels, xt as isUserInitiatedAbort, y as CONTEXT_WINDOW_SAFETY_BUFFER_TOKENS, yt as isSupportedFabFileMimeType, z as PermissionDeniedError } from "./ConfigStore-CNfbeaJf.mjs";
 import { n as isPathAllowed, t as assertPathAllowed } from "./pathValidation-D8tjkQXE-1HwvsuYT.mjs";
 import { n as isTerminalShellStatus, t as getShellSessionManager } from "./ShellSessionManager-6o8KZzl1-vrbPAUTq.mjs";
 import { execFile, execFileSync, spawn } from "child_process";
@@ -2044,7 +2044,7 @@ const webSearchTool = {
 	})
 };
 //#endregion
-//#region ../../b4m-core/services/dist/toolGenerators-CBGJMey9.mjs
+//#region ../../b4m-core/services/dist/toolGenerators-BVUZiwqE.mjs
 const diceRoll = async (parameters) => {
 	if (!parameters?.sides || !parameters?.times) throw new Error("Tool dice roll: Missing required parameters");
 	return sum(times(parameters.times, () => random(1, parameters.sides))).toString();
@@ -2731,7 +2731,7 @@ const askUserQuestionTool = {
 * re-export them without pulling the full tool graph. `index.ts` re-exports them
 * so the server barrel's public API is unchanged.
 */
-const generateTools = (userId, user, logger, { db, retrievalFilter, kbScope }, storage, imageGenerateStorage, statusUpdate, onStart, onFinish, llm, config, model, imageProcessorLambdaName, tools, allowedDirectories, entitlementKeys = [], sessionId, codeMinifier, availableModels, onToolLlmUsage) => {
+const generateTools = (userId, user, logger, { db, retrievalFilter, kbScope, inlinedAttachmentIds, fullyInlinedAttachmentIds }, storage, imageGenerateStorage, statusUpdate, onStart, onFinish, llm, config, model, imageProcessorLambdaName, tools, allowedDirectories, entitlementKeys = [], sessionId, codeMinifier, availableModels, onToolLlmUsage) => {
 	const context = {
 		userId,
 		user,
@@ -2750,6 +2750,8 @@ const generateTools = (userId, user, logger, { db, retrievalFilter, kbScope }, s
 		entitlementKeys,
 		retrievalFilter,
 		kbScope,
+		inlinedAttachmentIds,
+		fullyInlinedAttachmentIds,
 		codeMinifier,
 		availableModels,
 		onToolLlmUsage
@@ -7000,6 +7002,41 @@ const ChunkSchema = z$1.object({
 const isEmbeddingModel = (model, modelEnum) => {
 	return Object.values(modelEnum).includes(model);
 };
+/**
+* The embedding model's context window in tokens - the HARD ceiling an embedding call accepts.
+* Extracted from SmartChunker so a caller (the chunk queue handler, #1662) can compute the same
+* effective passage limit for owner-altitude policy resolution and cross-lake conflict reporting
+* without constructing a chunker. Throws on an unsupported model, exactly as the chunker does.
+*/
+function embeddingModelContextWindow(model) {
+	if (isEmbeddingModel(model, OpenAIEmbeddingModel)) return OPENAI_EMBEDDING_MODEL_MAP[model].contextWindow;
+	if (isEmbeddingModel(model, VoyageAIEmbeddingModel)) return VOYAGEAI_EMBEDDING_MODEL_MAP[model].contextWindow;
+	if (isEmbeddingModel(model, BedrockEmbeddingModel)) return BEDROCK_EMBEDDING_MODEL_MAP[model].contextWindow;
+	if (isEmbeddingModel(model, OllamaEmbeddingModel)) return OLLAMA_EMBEDDING_MODEL_MAP[model].contextWindow;
+	throw new Error(`Unsupported embedding model: ${model}`);
+}
+/**
+* The buffer subtracted from the model window to absorb cross-provider tokenizer differences
+* (char/4 approximations can undercount by ~8-10% vs tiktoken). A value < 1 is a percent of the
+* window (floored to >= 32 tokens); a value >= 1 is an absolute token count.
+*/
+function embeddingWindowBuffer(maxTokens, bufferPercentOrValue = .2) {
+	return bufferPercentOrValue < 1 ? Math.max(Math.floor(maxTokens * bufferPercentOrValue), 32) : Math.floor(bufferPercentOrValue);
+}
+/**
+* The effective per-chunk token limit the chunker will actually use: the SOFT passage target
+* (retrieval granularity) hard-capped to the buffered model window (an oversized chunk fails the
+* embedding call). THE single source of truth for that clamp - SmartChunker's constructor and the
+* chunk handler's conflict/observability logic (#1662) both derive from it, so a resolved policy
+* value can never drift from the granularity it actually produces. An omitted/invalid target falls
+* back to DEFAULT_PASSAGE_TOKEN_TARGET; a supplied one is floored to MIN_PASSAGE_TOKEN_TARGET.
+*/
+function effectiveChunkTokenLimit(opts) {
+	const maxTokens = embeddingModelContextWindow(opts.model);
+	const hardLimit = maxTokens - embeddingWindowBuffer(maxTokens, opts.bufferPercentOrValue ?? .2);
+	const { passageTokenTarget } = opts;
+	return Math.min(hardLimit, passageTokenTarget !== void 0 && Number.isFinite(passageTokenTarget) && passageTokenTarget > 0 ? Math.max(Math.floor(passageTokenTarget), 64) : 512);
+}
 var SmartChunker = class {
 	logger;
 	model;
@@ -7023,25 +7060,20 @@ var SmartChunker = class {
 			passageTokenTarget: void 0
 		} : options ?? {};
 		this.model = model;
-		if (isEmbeddingModel(model, OpenAIEmbeddingModel)) this.maxTokens = OPENAI_EMBEDDING_MODEL_MAP[model].contextWindow;
-		else if (isEmbeddingModel(model, VoyageAIEmbeddingModel)) this.maxTokens = VOYAGEAI_EMBEDDING_MODEL_MAP[model].contextWindow;
-		else if (isEmbeddingModel(model, BedrockEmbeddingModel)) this.maxTokens = BEDROCK_EMBEDDING_MODEL_MAP[model].contextWindow;
-		else if (isEmbeddingModel(model, OllamaEmbeddingModel)) this.maxTokens = OLLAMA_EMBEDDING_MODEL_MAP[model].contextWindow;
-		else throw new Error(`Unsupported embedding model: ${model}`);
+		this.maxTokens = embeddingModelContextWindow(model);
 		this.bufferPercentOrValue = bufferPercentOrValue ?? .2;
-		let buffer;
-		if (this.bufferPercentOrValue < 1) buffer = Math.max(Math.floor(this.maxTokens * this.bufferPercentOrValue), 32);
-		else buffer = Math.floor(this.bufferPercentOrValue);
-		const hardLimit = this.maxTokens - buffer;
-		const target = passageTokenTarget !== void 0 && Number.isFinite(passageTokenTarget) && passageTokenTarget > 0 ? Math.max(Math.floor(passageTokenTarget), 64) : 512;
-		this.chunkTokenLimit = Math.min(hardLimit, target);
+		this.chunkTokenLimit = effectiveChunkTokenLimit({
+			model,
+			passageTokenTarget,
+			bufferPercentOrValue: this.bufferPercentOrValue
+		});
 		this.storage = storage;
 		this.logger.updateMetadata({
 			model,
 			maxTokens: this.maxTokens,
 			chunkTokenLimit: this.chunkTokenLimit,
 			bufferPercentOrValue: this.bufferPercentOrValue,
-			passageTokenTarget: target
+			passageTokenTarget
 		});
 	}
 	async initializeEncoder() {
@@ -28418,7 +28450,125 @@ var AdminSettingsCache = class AdminSettingsCache {
 		this.logger.info("🛑 Admin settings cache shutdown complete");
 	}
 };
+/** Address of one cached override, shared by the cache and its callers so lookups are consistent. */
+function scopedOverrideKey(scopeLevel, scopeId, settingName) {
+	return JSON.stringify([
+		scopeLevel,
+		scopeId,
+		settingName
+	]);
+}
+function scopeKey(scopeLevel, scopeId) {
+	return JSON.stringify([scopeLevel, scopeId]);
+}
+/**
+* In-memory cache for scoped setting OVERRIDES, sibling to `AdminSettingsCache` (which caches the
+* flat platform table). Unlike the platform cache it cannot hold one "all settings" blob - overrides
+* are per (rung, setting) and can be many (a value per lake), so each address is cached individually
+* and negatively (a confirmed absence is cached too, so an un-overridden setting does not re-query
+* every resolve).
+*
+* Entries are nested one level per rung address so a whole rung invalidates in one delete (the shape
+* a future scoped-override writer needs). The cache key carries the scope rung, which is the whole
+* reason this is separate from `AdminSettingsCache`: that one keys by bare `settingName`, so scoped
+* values would collide there. See #1660.
+*/
+var ScopedSettingsCache = class ScopedSettingsCache {
+	cache = /* @__PURE__ */ new Map();
+	logger;
+	maxScopes = 5e3;
+	static DEFAULT_TTL = 3e5;
+	static DEVELOPMENT_TTL = 3e4;
+	constructor(logger) {
+		this.logger = logger;
+	}
+	getTTL() {
+		return process.env.NODE_ENV === "development" ? ScopedSettingsCache.DEVELOPMENT_TTL : ScopedSettingsCache.DEFAULT_TTL;
+	}
+	isValid(entry) {
+		return Date.now() - entry.timestamp < entry.ttl;
+	}
+	/**
+	* Resolve override values for the given rungs and setting names, reading through the cache. Returns
+	* a map keyed by `scopedOverrideKey(level,id,name)` -> value|null for EVERY (scope, name) pair
+	* requested, so the resolver can look up any rung deterministically. One DB query covers all misses.
+	*/
+	async getOverrides(scopes, settingNames, db) {
+		const out = /* @__PURE__ */ new Map();
+		if (scopes.length === 0 || settingNames.length === 0) return out;
+		const missingScopeRefs = /* @__PURE__ */ new Map();
+		const missingNames = /* @__PURE__ */ new Set();
+		for (const scope of scopes) {
+			const inner = this.cache.get(scopeKey(scope.scopeLevel, scope.scopeId));
+			for (const name of settingNames) {
+				const outKey = scopedOverrideKey(scope.scopeLevel, scope.scopeId, name);
+				const cached = inner?.get(name);
+				if (cached && this.isValid(cached)) out.set(outKey, cached.value);
+				else {
+					if (cached) inner.delete(name);
+					out.set(outKey, null);
+					missingScopeRefs.set(scopeKey(scope.scopeLevel, scope.scopeId), scope);
+					missingNames.add(name);
+				}
+			}
+		}
+		if (missingScopeRefs.size === 0) return out;
+		const rows = await db.scopedSettings.findOverrides(Array.from(missingScopeRefs.values()), Array.from(missingNames));
+		const ttl = this.getTTL();
+		const now = Date.now();
+		const found = /* @__PURE__ */ new Set();
+		for (const row of rows) {
+			out.set(scopedOverrideKey(row.scopeLevel, row.scopeId, row.settingName), row.settingValue);
+			this.put(row.scopeLevel, row.scopeId, row.settingName, row.settingValue, now, ttl);
+			found.add(scopedOverrideKey(row.scopeLevel, row.scopeId, row.settingName));
+		}
+		for (const scope of missingScopeRefs.values()) for (const name of missingNames) if (!found.has(scopedOverrideKey(scope.scopeLevel, scope.scopeId, name))) this.put(scope.scopeLevel, scope.scopeId, name, null, now, ttl);
+		if (this.cache.size > this.maxScopes) this.evictOldestScopes();
+		return out;
+	}
+	put(scopeLevel, scopeId, name, value, now, ttl) {
+		const sk = scopeKey(scopeLevel, scopeId);
+		let inner = this.cache.get(sk);
+		if (!inner) {
+			inner = /* @__PURE__ */ new Map();
+			this.cache.set(sk, inner);
+		}
+		inner.set(name, {
+			value,
+			timestamp: now,
+			ttl
+		});
+	}
+	evictOldestScopes() {
+		const freshness = (inner) => Math.max(...Array.from(inner.values()).map((e) => e.timestamp));
+		const scopesByAge = Array.from(this.cache.entries()).sort((a, b) => freshness(a[1]) - freshness(b[1]));
+		const toRemove = this.cache.size - this.maxScopes;
+		for (let i = 0; i < toRemove; i++) this.cache.delete(scopesByAge[i][0]);
+		this.logger.warn(`ScopedSettingsCache evicted ${toRemove} oldest rungs (size limit)`);
+	}
+	/** Invalidate every cached value for one rung address (call this from a future scoped-override writer). */
+	invalidateScope(scopeLevel, scopeId) {
+		this.cache.delete(scopeKey(scopeLevel, scopeId));
+	}
+	invalidateAll() {
+		this.cache.clear();
+	}
+	/** Snapshot for monitoring, parity with AdminSettingsCache.getStats(). */
+	getStats() {
+		let entries = 0;
+		for (const inner of this.cache.values()) entries += inner.size;
+		return {
+			scopes: this.cache.size,
+			entries
+		};
+	}
+	/** Graceful shutdown - no timers to stop (unlike AdminSettingsCache), just drop the map. */
+	shutdown() {
+		this.cache.clear();
+	}
+};
 let globalSettingsCache = null;
+let globalScopedSettingsCache = null;
 /**
 * Get or create the global settings cache instance
 */
@@ -28426,6 +28576,32 @@ function getSettingsCache(logger) {
 	if (!globalSettingsCache) globalSettingsCache = new AdminSettingsCache(logger || new Logger());
 	return globalSettingsCache;
 }
+function getScopedSettingsCache(logger) {
+	if (!globalScopedSettingsCache) globalScopedSettingsCache = new ScopedSettingsCache(logger || new Logger());
+	return globalScopedSettingsCache;
+}
+/**
+* Read org/owner/lake OVERRIDE values for a set of rungs and setting names, through the scoped cache.
+* Returns a map keyed by `scopedOverrideKey(level,id,name)` -> value|null (see ScopedSettingsCache).
+* Platform values are NOT here - the resolver layers these over the platform read. Passing no rungs
+* (a platform-altitude resolve) returns an empty map and touches neither cache nor DB.
+*/
+async function getScopedOverrides(scopes, settingNames, db, options) {
+	return getScopedSettingsCache(options?.logger).getOverrides(scopes, settingNames, db);
+}
+/**
+* Invalidate cached overrides for one rung address, or all of them. Call from a scoped-override writer.
+*
+* In-process only: this clears the calling process's map. With negative caching and the 5-minute TTL,
+* a newly-written override can stay invisible on other Lambda/container instances for up to one TTL.
+* That is acceptable while no writer ships (this PR), but a future write path needs a cross-instance
+* invalidation story (shared cache or short TTL) rather than relying on this alone.
+*/
+function invalidateScopedSettingsCache(scope) {
+	if (!globalScopedSettingsCache) return;
+	if (scope) globalScopedSettingsCache.invalidateScope(scope.scopeLevel, scope.scopeId);
+	else globalScopedSettingsCache.invalidateAll();
+}
 function getSettingsValue(key, settings, defaultValue) {
 	const settingConfig = settingsMap[key];
 	if (!settingConfig) {
@@ -28478,7 +28654,7 @@ async function getSettingsByNames(settingNames, db, options) {
 	const allSettings = await getSettingsCache(logger).getSettingsMap(db);
 	const result = {};
 	settingNames.forEach((name) => {
-		result[name] = allSettings[name] || null;
+		result[name] = allSettings[name] ?? null;
 	});
 	return result;
 }
@@ -28503,6 +28679,10 @@ function shutdownSettingsCache() {
 		globalSettingsCache.shutdown();
 		globalSettingsCache = null;
 	}
+	if (globalScopedSettingsCache) {
+		globalScopedSettingsCache.shutdown();
+		globalScopedSettingsCache = null;
+	}
 }
 /**
 * Get cache statistics for monitoring
@@ -28860,6 +29040,12 @@ const EMBEDDING_TOKEN_LIMITS = {
 */
 const KNOWLEDGE_FILE_TOKEN_ALLOCATION = .7;
 /**
+* The two always-on blocks this function injects itself, downstream of the caller's own
+* systemPromptDetails assembly - so a caller reporting per-source token telemetry (see
+* ChatCompletionProcess) cannot see them without reading this list back off the return value.
+*/
+const BUILDER_INJECTED_BLOCK_IDS = ["formatPrompt", "imagePrompt"];
+/**
 * Rounds the final safety pass may spend shrinking the payload. It has to re-measure between rounds
 * (see the pass for why one shot overshoots), and each round costs two real tokenizer calls, so this
 * bounds the work. Converges in one or two rounds in practice.
@@ -29194,6 +29380,7 @@ async function fetchAndProcessPreviousMessages(session, historyCount = null, { d
 			});
 		}
 	}
+	const priorToolNames = chatHistoryItems.flatMap((item) => (item.promptMeta?.functionCalls ?? []).map((fc) => fc.name)).filter((name) => Boolean(name));
 	return [
 		convertedMessages,
 		chatHistoryItems.length,
@@ -29203,7 +29390,8 @@ async function fetchAndProcessPreviousMessages(session, historyCount = null, { d
 			itemCount: chatHistoryItems.length,
 			oldestIncludedQuestId,
 			excludedOlderQuestCount,
-			recentGeneratedImages
+			recentGeneratedImages,
+			priorToolNames
 		}
 	];
 }
@@ -29495,12 +29683,16 @@ const noopResize = async (imageBuffer) => imageBuffer;
 async function processFabFilesServer(embeddingFactory, fabFiles, userPrompt, attachedContentTokenBudget, modelInfo, sendStatusUpdate, { logger, storage, db, resizeImageForModel = noopResize }, progressCallback) {
 	if (!fabFiles || fabFiles.length === 0) return {
 		userMessages: [],
-		errorMessages: []
+		errorMessages: [],
+		deliveredFileIds: [],
+		fullyDeliveredFileIds: []
 	};
 	const fileProcessingStartTime = Date.now();
 	let systemContent = "";
 	const userMessages = [];
 	const errorMessages = [];
+	const deliveredFileIds = /* @__PURE__ */ new Set();
+	const fullyDeliveredFileIds = /* @__PURE__ */ new Set();
 	const contextFiles = [];
 	const supportsVision = modelInfo?.supportsVision ?? false;
 	if (fabFiles.length > 0) sendStatusUpdate("Munching attached files...");
@@ -29524,6 +29716,8 @@ async function processFabFilesServer(embeddingFactory, fabFiles, userPrompt, att
 	const maxTokens = Math.max(1, Math.floor(attachedContentTokenBudget / textFileCount));
 	const fileContentCache = /* @__PURE__ */ new Map();
 	const processFileInParallel = async (file) => {
+		let delivered = false;
+		let fullyDelivered = false;
 		try {
 			if (isAudioMimeType(file.mimeType)) {
 				logger.warn(`[processFabFilesServer] Skipping audio file ${file.fileName} — audio is not attachable to an LLM.`);
@@ -29552,6 +29746,8 @@ async function processFabFilesServer(embeddingFactory, fabFiles, userPrompt, att
 							type: "text",
 							text: `Image URL: ${fileUrl}\nFile: "${file.fileName}" (fabFileId: ${file.id})\nWhen referencing this file, use the exact filename "${file.fileName}" — do not rename based on image content.`
 						});
+						delivered = true;
+						fullyDelivered = true;
 						break;
 					}
 					case ModelBackend.Anthropic:
@@ -29586,6 +29782,8 @@ async function processFabFilesServer(embeddingFactory, fabFiles, userPrompt, att
 								type: "text",
 								text: `Image URL: ${fileUrl}\nFile: "${file.fileName}" (fabFileId: ${file.id})\nWhen referencing this file, use the exact filename "${file.fileName}" — do not rename based on image content.`
 							});
+							delivered = true;
+							fullyDelivered = true;
 						} else if (modelInfo.id.startsWith("moonshot")) {
 							const moonshotBuffer = await resizeImageForModel(await storage.download(file.filePath), void 0, logger);
 							const { mime: moonshotMimeType } = await getFileType(moonshotBuffer, file.fileName, file.mimeType);
@@ -29609,6 +29807,8 @@ async function processFabFilesServer(embeddingFactory, fabFiles, userPrompt, att
 								type: "text",
 								text: `Image URL: ${fileUrl}\nFile: "${file.fileName}" (fabFileId: ${file.id})\nWhen referencing this file, use the exact filename "${file.fileName}" — do not rename based on image content.`
 							});
+							delivered = true;
+							fullyDelivered = true;
 						} else logger.warn(`Vision support for the model ${modelInfo.id} is not implemented. Skipping image processing.`);
 						break;
 					case ModelBackend.Ollama: {
@@ -29634,6 +29834,8 @@ async function processFabFilesServer(embeddingFactory, fabFiles, userPrompt, att
 							type: "text",
 							text: `Image URL: ${fileUrl}\nFile: "${file.fileName}" (fabFileId: ${file.id})\nWhen referencing this file, use the exact filename "${file.fileName}". Do not rename based on image content.`
 						});
+						delivered = true;
+						fullyDelivered = true;
 						break;
 					}
 					default: logger.error(`Unsupported backend for model ${modelInfo.id} backend ${modelInfo?.backend ?? "undefined"}`);
@@ -29680,7 +29882,9 @@ async function processFabFilesServer(embeddingFactory, fabFiles, userPrompt, att
 					}
 					if (truncatedResults.length > 0) {
 						deliveredViaCosine = true;
+						delivered = true;
 						const deliveredEveryChunk = !scanTruncated && totalChunks === truncatedResults.length;
+						fullyDelivered = deliveredEveryChunk && !anyChunkCut;
 						let notice = "";
 						if (!deliveredEveryChunk) notice = excerptNotice(file.fileName);
 						else if (anyChunkCut) notice = CONTENT_TRUNCATION_NOTICE;
@@ -29707,6 +29911,7 @@ async function processFabFilesServer(embeddingFactory, fabFiles, userPrompt, att
 					const finalMaxFileSize = maxTokens > 0 ? maxSizeBasedonMaxTokens : MAX_FILE_SIZE;
 					logger.log(`[processFabFilesServer] Final max file size: ${finalMaxFileSize}`);
 					sendStatusUpdate("Adding file content to prompt...");
+					fullyDelivered = fabContent.length <= finalMaxFileSize;
 					if (fabContent.length > finalMaxFileSize) {
 						await sendStatusUpdate("File is too large, truncating...");
 						const originalFileSize = fabContent.length;
@@ -29717,6 +29922,7 @@ async function processFabFilesServer(embeddingFactory, fabFiles, userPrompt, att
 							content: errorMsg
 						});
 					} else errorMsg = null;
+					delivered = true;
 					if (file.system) systemContent += fabContent;
 					else contextFiles.push({
 						fileName: file.fileName,
@@ -29746,6 +29952,8 @@ async function processFabFilesServer(embeddingFactory, fabFiles, userPrompt, att
 					}
 				}
 			}
+			if (delivered) deliveredFileIds.add(file.id);
+			if (fullyDelivered) fullyDeliveredFileIds.add(file.id);
 		} catch (error) {
 			logger.updateMetadata({ fileId: file.id });
 			logger.error(`🕐 [processFabFilesServer] Error processing file ${file.fileName}: ${error}`);
@@ -29784,7 +29992,9 @@ async function processFabFilesServer(embeddingFactory, fabFiles, userPrompt, att
 	logger.info(`📁 File processing completed in ${fileProcessingTime}ms for ${fabFiles.length} files`);
 	return {
 		userMessages,
-		errorMessages
+		errorMessages,
+		deliveredFileIds: Array.from(deliveredFileIds),
+		fullyDeliveredFileIds: Array.from(fullyDeliveredFileIds)
 	};
 }
 /**
@@ -29950,7 +30160,11 @@ const processMessages = (messages, tokenBudget, { truncationNotice } = {}) => {
 async function buildAndSortMessages(previousMessages, fabMessages, userPrompt, maxInputTokens, settings, historyCount = 0, logger, tokenizer, options = { verbose: false }) {
 	if (!(maxInputTokens > 0)) {
 		logger.error(`Invalid maxInputTokens: ${maxInputTokens}. Must be greater than 0.`);
-		return [];
+		return {
+			messages: [],
+			messageTruncation: null,
+			injectedBlocks: []
+		};
 	}
 	if (process.env.VERBOSE_CHAT_CONTEXT !== "false") {
 		if (options.verbose) {
@@ -29990,18 +30204,28 @@ async function buildAndSortMessages(previousMessages, fabMessages, userPrompt, m
 	const systemMessages = [];
 	let systemTokenCount = 0;
 	const builderInjectedPriorities = /* @__PURE__ */ new Map();
-	if (!options.skipAdminPromptTemplates) {
+	const injectedBlockGateResults = /* @__PURE__ */ new Map();
+	if (options.skipAdminPromptTemplates) {
+		injectedBlockGateResults.set("formatPrompt", { reason: "mode_skipped" });
+		injectedBlockGateResults.set("imagePrompt", { reason: "mode_skipped" });
+	} else {
 		if (getSettingsValue("UseFormatPrompt", settings)) {
 			const formatPromptTemplate = settings.FormatPromptTemplate;
 			const withFormatPrompt = includeHardcodedSystemMessage(fabMessages, formatPromptTemplate);
-			if (withFormatPrompt.length > fabMessages.length) builderInjectedPriorities.set(withFormatPrompt[0], 60);
+			if (withFormatPrompt.length > fabMessages.length) {
+				builderInjectedPriorities.set(withFormatPrompt[0], 60);
+				injectedBlockGateResults.set("formatPrompt", { message: withFormatPrompt[0] });
+			} else injectedBlockGateResults.set("formatPrompt", { reason: "not_triggered" });
 			fabMessages = withFormatPrompt;
-		}
+		} else injectedBlockGateResults.set("formatPrompt", { reason: "setting_disabled" });
 		if (getSettingsValue("UseImagePrompt", settings)) {
 			const withImagePrompt = includeImagePromptSystemMessage(fabMessages, userPromptContent, options.imageGenerationAvailable ?? false);
-			if (withImagePrompt.length > fabMessages.length) builderInjectedPriorities.set(withImagePrompt[0], 50);
+			if (withImagePrompt.length > fabMessages.length) {
+				builderInjectedPriorities.set(withImagePrompt[0], 50);
+				injectedBlockGateResults.set("imagePrompt", { message: withImagePrompt[0] });
+			} else injectedBlockGateResults.set("imagePrompt", { reason: "not_triggered" });
 			fabMessages = withImagePrompt;
-		}
+		} else injectedBlockGateResults.set("imagePrompt", { reason: "setting_disabled" });
 	}
 	const imageMessages = [];
 	const nonImageMessages = [];
@@ -30067,6 +30291,21 @@ async function buildAndSortMessages(previousMessages, fabMessages, userPrompt, m
 		admittedSystemMessages.add(message);
 	}
 	systemMessages.push(...systemCandidates.filter((message) => admittedSystemMessages.has(message)));
+	const injectedBlocks = BUILDER_INJECTED_BLOCK_IDS.map((id) => {
+		const gateResult = injectedBlockGateResults.get(id);
+		if (!gateResult?.message) return {
+			id,
+			injected: false,
+			delivered: false,
+			...gateResult?.reason ? { reason: gateResult.reason } : {}
+		};
+		return {
+			id,
+			injected: true,
+			delivered: admittedSystemMessages.has(gateResult.message),
+			...typeof gateResult.message.content === "string" ? { content: gateResult.message.content } : {}
+		};
+	});
 	if (systemCandidates.length > 0 && systemMessages.length === 0) logger.warn(`No system instructions fit the budget: ${systemCandidates.length} message(s) dropped, cap ${systemTokenCap} of ${preSystemBudget} est. tokens (${Math.round(65)}% after the attached-content reserve of ${contentReserve}). Smallest was ${Math.min(...systemCandidates.map(systemMessageTokens))} est. tokens.`);
 	tokenBudget -= systemTokenCount;
 	const historyMessages = isUnlimitedHistory(historyCount) ? previousMessages : historyCount > 0 ? previousMessages.slice(-historyCount * 2) : [];
@@ -30197,16 +30436,16 @@ async function buildAndSortMessages(previousMessages, fabMessages, userPrompt, m
 		];
 	};
 	const messages = assemble(processedContentMessages, processedPreviousMessages);
-	const recordDebugInfo = (finalContentMessages) => {
+	const buildDebugInfo = (finalContentMessages) => {
 		const historyWindowed = previousMessages.length > historyMessages.length;
 		const budgetTruncated = allRemovedMessages.length > 0 || contentSqueezed || historyCutMidMessage;
-		buildAndSortMessages.lastDebugInfo = { messageTruncation: {
+		return {
 			wasTruncated: budgetTruncated,
 			originalMessageCount: originalTotalMessageCount,
 			truncatedMessageCount: processedPreviousMessages.length + finalContentMessages.length,
 			truncationMethod: budgetTruncated ? "token-budget" : historyWindowed ? "history-limit" : void 0,
 			removedMessages: allRemovedMessages.length > 0 ? allRemovedMessages : void 0
-		} };
+		};
 	};
 	const finalTokenCount = await calculateTotalTokenLength(messages, {
 		estimateOnly: false,
@@ -30251,8 +30490,11 @@ async function buildAndSortMessages(previousMessages, fabMessages, userPrompt, m
 			else break;
 		}
 		if (currentTokenCount > maxInputTokens) logger.warn(`Final safety pass could not bring the payload under maxInputTokens (${currentTokenCount} > ${maxInputTokens}). Remaining: ${systemMessages.length} system, ${processedPreviousMessages.length} history, ${reducedContentMessages.length} content, ${imageMessages.length} image message(s) plus the user prompt` + (lastRatio === null ? ", with nothing shrinkable to measure." : `, at ${lastRatio.toFixed(2)} real tokens per estimated token.`));
-		recordDebugInfo(reducedContentMessages);
-		return ensureToolPairingIntegrity(assemble(reducedContentMessages, processedPreviousMessages), logger);
+		return {
+			messages: ensureToolPairingIntegrity(assemble(reducedContentMessages, processedPreviousMessages), logger),
+			messageTruncation: buildDebugInfo(reducedContentMessages),
+			injectedBlocks
+		};
 	}
 	if (process.env.VERBOSE_MESSAGE_BUILDING === "true") {
 		logger.log("=== Verbose Message Building Log ===");
@@ -30265,18 +30507,16 @@ async function buildAndSortMessages(previousMessages, fabMessages, userPrompt, m
 		logger.log(`\nTotal messages: ${messages.length}`);
 		logger.log("=== End of Verbose Message Building Log ===");
 	}
-	recordDebugInfo(processedContentMessages);
-	return ensureToolPairingIntegrity(messages, logger);
-}
-/**
-* Returns the debug info populated by the most recent buildAndSortMessages call.
-*/
-function getLastBuildDebugInfo() {
-	return buildAndSortMessages.lastDebugInfo?.messageTruncation || null;
+	return {
+		messages: ensureToolPairingIntegrity(messages, logger),
+		messageTruncation: buildDebugInfo(processedContentMessages),
+		injectedBlocks
+	};
 }
 var llm_exports = /* @__PURE__ */ __exportAll({
 	ATTACHED_CONTENT_EXTRACTION_SHARE: () => ATTACHED_CONTENT_EXTRACTION_SHARE,
 	ATTACHMENT_DELIVERED_NOTICE: () => ATTACHMENT_DELIVERED_NOTICE,
+	BUILDER_INJECTED_BLOCK_IDS: () => BUILDER_INJECTED_BLOCK_IDS,
 	EXTRACTION_SYSTEM_RESERVE_MAX_SHARE: () => EXTRACTION_SYSTEM_RESERVE_MAX_SHARE,
 	FORMAT_PROMPT_PRIORITY: () => 60,
 	IMAGE_PROMPT_PRIORITY: () => 50,
@@ -30298,7 +30538,6 @@ var llm_exports = /* @__PURE__ */ __exportAll({
 	fetchAndProcessPreviousMessages: () => fetchAndProcessPreviousMessages,
 	generateSafeEmbedding: () => generateSafeEmbedding,
 	getCachedSignedUrl: () => getCachedSignedUrl,
-	getLastBuildDebugInfo: () => getLastBuildDebugInfo,
 	includeHardcodedSystemMessage: () => includeHardcodedSystemMessage,
 	includeImagePromptSystemMessage: () => includeImagePromptSystemMessage,
 	processFabFilesServer: () => processFabFilesServer,
@@ -34541,6 +34780,7 @@ __reExport(/* @__PURE__ */ __exportAll({
 	AdminSettingsCache: () => AdminSettingsCache,
 	BEDROCK_EMBEDDING_MODEL_MAP: () => BEDROCK_EMBEDDING_MODEL_MAP,
 	BFLImageService: () => BFLImageService,
+	BUILDER_INJECTED_BLOCK_IDS: () => BUILDER_INJECTED_BLOCK_IDS,
 	BadRequestError: () => BadRequestError,
 	BaseStorage: () => BaseStorage,
 	BedrockEmbeddingService: () => BedrockEmbeddingService,
@@ -34602,6 +34842,7 @@ __reExport(/* @__PURE__ */ __exportAll({
 	RapidReplyMappingsCache: () => RapidReplyMappingsCache,
 	S3Storage: () => S3Storage,
 	SQSService: () => SQSService,
+	ScopedSettingsCache: () => ScopedSettingsCache,
 	SmartChunker: () => SmartChunker,
 	TOKEN_BUFFER_PERCENTAGE: () => TOKEN_BUFFER_PERCENTAGE,
 	TOOL_RESULT_NOT_RECORDED: () => TOOL_RESULT_NOT_RECORDED,
@@ -34659,12 +34900,12 @@ __reExport(/* @__PURE__ */ __exportAll({
 	getFileContent: () => getFileContent,
 	getFileExtension: () => getFileExtension,
 	getFileType: () => getFileType,
-	getLastBuildDebugInfo: () => getLastBuildDebugInfo,
 	getLlmWithFallback: () => getLlmWithFallback,
 	getMimeTypeByExtension: () => getMimeTypeByExtension,
 	getNotificationDeduplicator: () => getNotificationDeduplicator,
 	getProviderFromModel: () => getProviderFromModel,
 	getRetryAfterMs: () => getRetryAfterMs,
+	getScopedOverrides: () => getScopedOverrides,
 	getSettingByName: () => getSettingByName,
 	getSettingsByNames: () => getSettingsByNames,
 	getSettingsCacheStats: () => getSettingsCacheStats,
@@ -34674,6 +34915,7 @@ __reExport(/* @__PURE__ */ __exportAll({
 	includeHardcodedSystemMessage: () => includeHardcodedSystemMessage,
 	includeImagePromptSystemMessage: () => includeImagePromptSystemMessage,
 	initializeConfig: () => initializeConfig,
+	invalidateScopedSettingsCache: () => invalidateScopedSettingsCache,
 	invalidateSettingsCache: () => invalidateSettingsCache,
 	isAiEditableOfficeMime: () => isAiEditableOfficeMime,
 	isGPT5ModelWithToolSupport: () => isGPT5ModelWithToolSupport,
@@ -34708,6 +34950,7 @@ __reExport(/* @__PURE__ */ __exportAll({
 	resolveEmbeddingConfig: () => resolveEmbeddingConfig,
 	resolveSupportedMimeType: () => resolveSupportedMimeType,
 	safeInputWindow: () => safeInputWindow,
+	scopedOverrideKey: () => scopedOverrideKey,
 	secureParameters: () => secureParameters,
 	sendToConnection: () => sendToConnection,
 	shouldTriggerFallback: () => shouldTriggerFallback,

package/dist/{ApiClient-B_CQrUiF.mjs → ApiClient-BPmlalut.mjs}

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
-import { n as logger, t as ConfigStore } from "./ConfigStore-DD3DcC3-.mjs";
-import { t as version } from "./package-BqKSCbso.mjs";
+import { n as logger, t as ConfigStore } from "./ConfigStore-CNfbeaJf.mjs";
+import { t as version } from "./package-CxHSRXdp.mjs";
 import axios, { isAxiosError } from "axios";
 //#region src/auth/OAuthClient.ts
 /**