om-memory-system 4.23.0-next.144 → 4.25.0

package/dist/config.js

@@ -10,6 +10,10 @@ import { getAutoCaptureProviderStatus } from "./services/ai/live-model-choice.js
 import { parseBackfillModel } from "./importer/backfill-model.js";
 import { parseIgnoredDirectories, parseImportPathMaps } from "./importer/import-path-maps.js";
 import { checkedMemoryLimit, MEMORY_LIMITS } from "./utils/memory-limits.js";
+import { defaultSimilarityThreshold } from "./utils/threshold-default.js";
+import { RERANKER_TIMEOUT_DEFAULT_MS } from "./services/request-timeouts.js";
+import { checkedRerankerSettings } from "./utils/reranker-settings.js";
+import { checkedKeywordBoost, checkedKeywordMode, } from "./utils/keyword-boost.js";
 import { resolveDefaultStoragePath, runLegacyStoreMigration, legacyMigrationPaths, } from "./services/legacy-migration.js";
 const OMMS_CONFIG_DIR = join(homedir(), ".config", "omms");
 const DATA_DIR = join(homedir(), ".omms");
@@ -38,8 +42,9 @@ const DEFAULTS = {
     embeddingModel: "Xenova/nomic-embed-text-v1",
     embeddingDimensions: 768,
     embeddingUseTaskPrefixes: false,
-    similarityThreshold: 0.6,
     maxMemories: 10,
+    rerankerEnabled: false,
+    rerankerTimeoutMs: RERANKER_TIMEOUT_DEFAULT_MS,
     retrievalMaxTokens: 2000,
     maxProfileItems: 5,
     injectProfile: true,
@@ -161,384 +166,8 @@ function assertProjectRemoteProviderConfigIsSafe(projectConfig) {
             `Move them to the global config at ${CONFIG_FILES[0]}.`);
     }
 }
-export const CONFIG_TEMPLATE = `{
-  // ============================================
-  // omms (Opinionated Modular Memory System) Configuration
-  // ============================================
-  
-  // Storage location for the vector database. Leave unset to use the
-  // default ~/.omms/data (a legacy ~/.opencode-mem/data store is migrated
-  // there automatically on first start; see docs/omms-migration.md).
-  // "storagePath": "~/.omms/data",
-
-  "userEmailOverride": "",
-  "userNameOverride": "",
-  
-  // ============================================
-  // Embedding Model (for similarity search)
-  // ============================================
-  // Local = Hugging Face / ONNX via @huggingface/transformers (not Apple MLX).
-  // Remote = set embeddingApiUrl (OpenAI-compatible /embeddings); embeddingApiKey is optional.
-  
-  // Default: Nomic Embed v1 (768 dimensions, 8192 context, multilingual)
-  "embeddingModel": "Xenova/nomic-embed-text-v1",
-
-  // Opt-in Nomic task prefixes (search_document: / search_query:). After enabling,
-  // re-index existing memories so store and query vectors stay aligned.
-  // "embeddingUseTaskPrefixes": true,
-  
-  // Auto-detected dimensions (no need to set manually)
-  // "embeddingDimensions": 768,
-  
-  // Other recommended local models:
-  // "embeddingModel": "Xenova/jina-embeddings-v2-base-en",  // 768 dims, English-only, 8192 context
-  // "embeddingModel": "Xenova/jina-embeddings-v2-small-en", // 512 dims, faster, 8192 context
-  // "embeddingModel": "Xenova/all-MiniLM-L6-v2",            // 384 dims, very fast, 512 context
-  // "embeddingModel": "Xenova/all-mpnet-base-v2",           // 768 dims, good quality, 512 context
-  
-  // Optional: OpenAI-compatible API for embeddings (URL required; key only if the server needs one)
-  // "embeddingApiUrl": "https://api.openai.com/v1",
-  // "embeddingApiKey": "env://OPENAI_API_KEY",  // or "sk-..." / "file:///path/to/key"
-  // "embeddingModel": "text-embedding-3-small",  // 1536 dims, auto-detected
- 
-  // ============================================
-  // Web Server Settings
-  // ============================================
-  
-  // Start a background import of past chats on each host's next start.
-  "autoBackfill": true,
-  // "piBackfillModel": "inherit", // or "external", or "provider/model"
-  // "opencodeBackfillModel": "inherit", // or "external", or "provider/model"
-  // Directory maps for history recorded in moved or deleted directories.
-  // "importPathMaps": [{ "from": "~/code/app-feat-x", "to": "~/code/app" }],
-  // Unresolved directories to hide on the Directory maps page. They stay unimported.
-  // "importIgnoredDirectories": ["/private/tmp/scratch"],
-  // Claude Code's folder, when it is not ~/.claude (like CLAUDE_CONFIG_DIR).
-  // "claudeConfigDir": "~/.claude-work",
-
-  // Register the web app to start when you log in.
-  "webServerAutoStart": true,
-  // Enable web UI for managing memories (accessible at http://localhost:4747)
-  "webServerEnabled": true,
-  
-  // Port for web UI server
-  "webServerPort": 4747,
-  
-  // Host address for web UI (use 127.0.0.1 for local only, 0.0.0.0 for network access)
-  "webServerHost": "127.0.0.1",
-
-  // Optional HTTP Basic Auth for the web UI. Accepts literal, env://, or file:// secrets.
-  // "webServerAuthPassword": "",
-  // "webServerAuthUsername": "",
-
-  // Deprecated: imported once into the API token table, then no longer read.
-  // Create API tokens on the Settings page (Keys and access) instead.
-  // "webServerApiToken": "env://OMMS_WEB_TOKEN",
-  
-  // ============================================
-  // Database Settings
-  // ============================================
-  
-  // Maximum vectors per database shard (auto-creates new shard when limit reached)
-  "maxVectorsPerShard": 50000,
-  
-  // Automatically delete old memories based on retention period
-  "autoCleanupEnabled": true,
-  
-  // Days to keep memories before auto-cleanup (only if autoCleanupEnabled is true)
-  "autoCleanupRetentionDays": 30,
-  
-  // Automatically detect and remove duplicate memories
-  "deduplicationEnabled": true,
-  
-   // Similarity threshold (0-1) for detecting duplicates (higher = stricter)
-   "deduplicationSimilarityThreshold": 0.90,
-   
-  // ============================================
-  // Memory Scope Settings
-  // ============================================
-
-  // Default scope for memory list/search queries
-  // "project" keeps queries within the current project, "all-projects" searches across all project shards
-  "memory": {
-    "defaultScope": "project"
-  },
-
-  // ============================================
-  // Model for auto-capture and profile learning (same rule in OpenCode and Pi)
-  // ============================================
-
-   // Which model summarises your work, in this order:
-   //   1. The host model below: opencodeProvider/opencodeModel in OpenCode,
-   //      piProvider/piModel in Pi. Set the model to "inherit" to follow
-   //      whatever model the session is using, or to "external" to send
-   //      every call to the external API below.
-   //   2. If no host model is set: the external API (memoryModel/memoryApiUrl/
-   //      memoryApiKey further down).
-   //   3. If neither is set: the session's own model.
-   // If the host model fails and the external API is configured, the external
-   // API is used instead.
-   //
-   // Host models go through OpenCode's or Pi's own sign-in (OAuth like Claude
-   // Pro/Max, GitHub Copilot, ChatGPT, or your own key there), so no separate API
-   // key is needed here.
-   //
-   // Examples (OpenCode names from 'opencode providers list'; Pi names from its model list):
-   //   "opencodeProvider": "anthropic",    "opencodeModel": "claude-haiku-4-5-20251001"
-   //   "opencodeProvider": "openai",       "opencodeModel": "inherit"
-   //   "piProvider": "openai-codex",       "piModel": "gpt-5.6-luna"
-   //
-   // "opencodeProvider": "anthropic",
-   // "opencodeModel": "claude-haiku-4-5-20251001",
-   // "piProvider": "openai-codex",
-   // "piModel": "gpt-5.6-luna",
-
-   // ============================================
-   // Auto-Capture Settings
-   // ============================================
-  
-  // Auto-capture runs in the background without blocking your main session.
-  // Note: Ollama may not support tool calling. Use OpenAI, Anthropic, or Groq for best results.
-  
-  "autoCaptureEnabled": true,
-  
-  // Provider type: "openai-chat" | "openai-responses" | "anthropic" | "minimax" | "orcarouter" | "google-gemini"
-  // Note: "openai-chat" is a generic OpenAI API-compatible mode.
-  // Any service that follows the OpenAI Chat Completions API can use it via custom "memoryApiUrl".
-  "memoryProvider": "openai-chat",
-  
-  // External API (step 2 above, and the fallback when a host model fails).
-  // Uncomment all 3 lines and replace memoryApiKey before use:
-  // "memoryModel": "gpt-4o-mini",
-  // "memoryApiUrl": "https://api.openai.com/v1",
-  // "memoryApiKey": "sk-...",
-
-  // API Key Formats:
-  // Direct value:        "sk-..."
-  // From file:           "file://~/.config/litellm-key.txt"
-  // From env variable:   "env://LITELLM_API_KEY"
-  
-  // Examples for different providers:
-  // Any OpenAI-compatible endpoint can use the "openai-chat" provider pattern below.
-  // Common examples: DeepSeek, Qwen (via Alibaba Cloud ModelStudio),
-  // Zhipu GLM (BigModel platform), and Kimi (Moonshot AI platform).
-
-  // OpenAI Chat Completion (default, backward compatible):
-  //   "memoryProvider": "openai-chat"
-  //   "memoryModel": "gpt-4o-mini"
-  //   "memoryApiUrl": "https://api.openai.com/v1"
-  //   "memoryApiKey": "sk-..."
-
-  // DeepSeek (OpenAI-compatible example):
-  //   "memoryProvider": "openai-chat"
-  //   "memoryModel": "deepseek-chat"
-  //   "memoryApiUrl": "https://api.deepseek.com/v1"
-  //   "memoryApiKey": "sk-..."
-  
-  // OpenAI Responses API (recommended, with session support):
-  //   "memoryProvider": "openai-responses"
-  //   "memoryModel": "gpt-4o"
-  //   "memoryApiUrl": "https://api.openai.com/v1"
-  //   "memoryApiKey": "sk-..."
-  
-  // Anthropic (with session support):
-  //   "memoryProvider": "anthropic"
-  //   "memoryModel": "claude-3-5-haiku-20241022"
-  //   "memoryApiUrl": "https://api.anthropic.com/v1"
-  //   "memoryApiKey": "sk-ant-..."
-
-  // MiniMax (Anthropic Messages-compatible endpoint, with session support):
-  //   "memoryProvider": "minimax"
-  //   "memoryModel": "MiniMax-M3"
-  //   "memoryApiUrl": "https://api.minimax.io"        // global endpoint
-  //   "memoryApiKey": "<MiniMax API key>"
-  //   // China endpoint: "memoryApiUrl": "https://api.minimaxi.com"
-  //   // Optional adaptive thinking for MiniMax-M3:
-  //   "memoryExtraParams": { "thinking": { "type": "adaptive" } }
-
-  // OrcaRouter (OpenAI-compatible gateway, namespaced model IDs, with session support):
-  //   "memoryProvider": "orcarouter"
-  //   "memoryApiKey": "<OrcaRouter API key>"
-  //   // memoryApiUrl and memoryModel are optional — they default to
-  //   // https://api.orcarouter.ai/v1 and "orcarouter/auto" (a routing alias).
-  //   // OrcaRouter rejects bare model names, so if you set memoryModel, use a
-  //   // namespaced ID such as "openai/gpt-5.5" or "deepseek/deepseek-v4-flash".
-  //   "memoryModel": "openai/gpt-5.5"
-
-  // Groq (OpenAI-compatible, use openai-chat provider):
-  //   "memoryProvider": "openai-chat"
-  //   "memoryModel": "llama-3.3-70b-versatile"
-  //   "memoryApiUrl": "https://api.groq.com/openai/v1"
-  //   "memoryApiKey": "gsk_..."
-  
-  // Maximum iterations for multi-turn AI analysis (for openai-responses, anthropic, and minimax)
-  "autoCaptureMaxIterations": 5,
-   
-  // Timeout per iteration in milliseconds (30 seconds default)
-  "autoCaptureIterationTimeout": 30000,
-
-  // Tries for one turn's capture, on Pi and OpenCode, 2 and 4 seconds apart. After the
-  // last one fails, a turn the model could not reach goes to the capture retry queue.
-  "autoCaptureMaxRetries": 3,
-
-  // Maximum UTF-8 bytes for the auto-capture markdown context sent to the summary model.
-  // Prevents HTTP 400 context overflows on models with ~131K token windows (e.g. Groq Llama).
-  // Rough guide: tokens ≈ bytes / 4 for mixed code/prose.
-  "autoCaptureMaxContextBytes": 131072,
-   
-  // Days to keep AI session history before cleanup
-  "aiSessionRetentionDays": 7,
-
-  // Capture diagnostics: every capture attempt always writes one metadata line
-  // (model, stop reason, sizes, outcome) to ~/.omms/omms.log, with no
-  // conversation text. Set captureTrace to true to also write each attempt's
-  // prompt and raw reply to ~/.omms/traces/ for debugging. Traces can contain
-  // conversation content: <private> text and common API key formats are
-  // redacted, files are readable only by you, and they are deleted after
-  // captureTraceRetentionDays. Only this global file can turn tracing on.
-  // "captureTrace": false,
-  // "captureTraceRetentionDays": 7,
-  // "captureAttemptRetentionDays": 30,
-
-  // Capture retry queue: when a live capture fails because the model cannot be
-  // reached (network error, timeout, HTTP 408, 429 or 5xx), OMMS keeps a copy of
-  // the turn in ~/.omms/data and tries it again later. The copy has <private>
-  // text and common API key formats removed. Turns are deleted when the retry
-  // ends or after captureRetryRetentionHours (0 to 720). 0 turns the queue off.
-  // Only this global file can set it.
-  // "captureRetryRetentionHours": 72,
-
-  // Temperature for AI API requests (set to false to omit parameter for models that don't support it)
-  // Some reasoning models (like o1, o3, gpt-5) don't support temperature parameter
-  // Set to false and add "memoryTemperature": false in config when using such models
-  "memoryTemperature": 0.3,
-
-  // Extra parameters to include in API request body
-  // Useful for local inference servers (e.g. llama-server with --jinja) that support
-  // additional parameters like disabling thinking/reasoning mode
-  // Example for Qwen3 models: { "enable_thinking": false }
-  // "memoryExtraParams": {},
-
-  // Language for auto-capture summaries (default: "auto" for auto-detection)
-  // Options: "auto", "en", "id", "zh", "ja", "es", "fr", "de", "ru", "pt", "ar", "ko"
-  // "autoCaptureLanguage": "auto",
-
-  // ============================================
-  // Toast Notifications
-  // ============================================
-
-  // Show toast when memory is auto-captured
-  "showAutoCaptureToasts": true,
-
-  // Show toast when user profile is updated
-  "showUserProfileToasts": true,
-
-  // Show toast for error messages
-  "showErrorToasts": true,
-
-  // ============================================
-  // User Profile System
-  // ============================================
-
-  // Analyze user prompts every N prompts to build/update your user profile
-  // When N uncaptured prompts accumulate, AI will analyze them to identify:
-  // - User preferences (code style, communication style, tool preferences)
-  // - User patterns (recurring topics, problem domains, technical interests)
-  // - User workflows (development habits, sequences, learning style)
-  // - Skill level (overall and per-domain assessment)
-  "userProfileAnalysisInterval": 10,
-
-  // Days before inactive items (all types) are eligible for removal
-  "userProfileStaleDays": 2,
-
-  // Days before an inactive workflow is eligible for removal. A workflow seen
-  // userProfileMinEvidenceForRetention times or more is kept.
-  "userProfileWorkflowStaleDays": 30,
-
-  // Number of preferences shown in UI
-  "userProfileDisplayPreferences": 20,
-  
-  // Number of patterns shown in UI
-  "userProfileDisplayPatterns": 15,
-  
-  // Number of workflows shown in UI
-  "userProfileDisplayWorkflows": 10,
-  
-  // Number of preferences injected into LLM conversation context
-  // Keep this small — the strongest signals are enough; more dilute LLM attention
-  "userProfileInjectPreferences": 5,
-  
-  // Number of patterns injected into LLM conversation context
-  "userProfileInjectPatterns": 5,
-  
-  // Number of workflows injected into LLM conversation context
-  "userProfileInjectWorkflows": 3,
-  
-  // Days before preference confidence starts to decay (if not reinforced)
-  // Preferences that aren't seen again will gradually lose confidence and be removed
-  "userProfileConfidenceDecayDays": 30,
-  
-  // Number of profile versions to keep in changelog (for rollback/debugging)
-  // Older versions are automatically cleaned up
-  "userProfileChangelogRetentionCount": 5,
-
-  // Minimum evidence count for a preference/pattern to survive confidence decay
-  // Items confirmed fewer times are more likely to be pruned when confidence decays
-  "userProfileMinEvidenceForRetention": 3,
-
-  // Merge duplicate or vague profile items every N analysed prompts (about 1 model call per 100 items)
-  "userProfileAutoCleanupEnabled": true,
-  "userProfileAutoCleanupInterval": 500,
-
-  // Enable LLM validation of existing preferences against recent behavior.
-  // When enabled, each analysis round checks if top-5 preferences still match recent prompts.
-  // Experimental — disabled by default.
-  "userProfileValidationEnabled": false,
-
-  // ============================================
-  // Search Settings
-  // ============================================
-  
-  // Minimum similarity score (0-1) for memory search results
-  "similarityThreshold": 0.6,
-
-  // Maximum number of memories to return in search results.
-  // Manual searches can ask for fewer; prompt retrieval uses this ceiling.
-  "maxMemories": 10,
-
-  // ============================================
-  // Memory Context Controls
-  // ============================================
-
-  // Recent memories added at fresh session start (OpenCode V1 and Claude Code).
-  // Pi and OpenCode V2 search the store with each prompt instead.
-  "chatMessage": {
-    "enabled": true,
-    // Recent memories injected at session start.
-    "maxMemories": 3,
-    "excludeCurrentSession": true,
-    // When to inject: "first" message of a session or "always".
-    // "injectOn": "first",
-  },
-
-  // OpenCode profile-learning input limit, in UTF-8 bytes. Smaller values can
-  // omit prompts from that input. Does not limit Pi or Claude Code profile input.
-  "userProfileMaxContextBytes": 32768,
-
-  // Shared budget for automatic memory context (profile text, memories, and
-  // formatting) across all hosts, in approximate tokens: ceil(UTF-8 bytes / 4).
-  // Covers prompt retrieval, recent-memory injection, and session-memory
-  // restoration. Not an exact provider token limit or a spending limit.
-  "retrievalMaxTokens": 2000,
-
-  // ============================================
-  // Advanced Settings
-  // ============================================
-  
-  // Inject user profile into AI context (preferences, patterns, workflows)
-  "injectProfile": true
-}
-`;
+import { CONFIG_TEMPLATE } from "./config-template.js";
+export { CONFIG_TEMPLATE };
 function ensureConfigExists() {
     const configPath = join(OMMS_CONFIG_DIR, "omms.jsonc");
     const hasOmmsConfig = CONFIG_FILES.slice(0, 2).some((path) => existsSync(path));
@@ -640,8 +269,11 @@ function resolveMemoryApiKey(value) {
 }
 function buildConfig(fileConfig) {
     const memoryApiKey = resolveMemoryApiKey(fileConfig.memoryApiKey);
-    const embeddingDimensions = fileConfig.embeddingDimensions ??
-        getEmbeddingDimensions(fileConfig.embeddingModel ?? DEFAULTS.embeddingModel);
+    checkedRerankerSettings(fileConfig);
+    const embeddingModel = fileConfig.embeddingModel ?? DEFAULTS.embeddingModel;
+    // The default follows the embedder; a value present in the file always wins.
+    const similarityThreshold = checkedOwnMemoryLimit(MEMORY_LIMITS.similarityThreshold, fileConfig, "similarityThreshold", defaultSimilarityThreshold(embeddingModel));
+    const embeddingDimensions = fileConfig.embeddingDimensions ?? getEmbeddingDimensions(embeddingModel);
     const autoCaptureMaxContextBytes = checkedOwnMemoryLimit(MEMORY_LIMITS.autoCaptureMaxContextBytes, fileConfig, "autoCaptureMaxContextBytes", DEFAULTS.autoCaptureMaxContextBytes);
     const maxMemories = checkedOwnMemoryLimit(MEMORY_LIMITS.maxMemories, fileConfig, "maxMemories", DEFAULTS.maxMemories);
     const retrievalMaxTokens = checkedOwnMemoryLimit(MEMORY_LIMITS.retrievalMaxTokens, fileConfig, "retrievalMaxTokens", DEFAULTS.retrievalMaxTokens);
@@ -685,7 +317,14 @@ function buildConfig(fileConfig) {
             ? // No ambient OPENAI_API_KEY: an unset key means no Authorization header.
                 resolveSecretValue(fileConfig.embeddingApiKey)
             : undefined,
-        similarityThreshold: fileConfig.similarityThreshold ?? DEFAULTS.similarityThreshold,
+        similarityThreshold,
+        rerankerEnabled: fileConfig.rerankerEnabled ?? DEFAULTS.rerankerEnabled,
+        rerankerApiUrl: fileConfig.rerankerApiUrl?.trim() || undefined,
+        rerankerModel: fileConfig.rerankerModel?.trim() || undefined,
+        rerankerApiKey: fileConfig.rerankerApiKey,
+        rerankerTimeoutMs: fileConfig.rerankerTimeoutMs ?? DEFAULTS.rerankerTimeoutMs,
+        keywordBoost: checkedKeywordBoost(fileConfig.keywordBoost),
+        keywordMode: checkedKeywordMode(fileConfig.keywordMode),
         maxMemories,
         retrievalMaxTokens,
         maxProfileItems: fileConfig.maxProfileItems ?? DEFAULTS.maxProfileItems,
@@ -887,6 +526,14 @@ export function initConfig(directory, options = {}) {
     delete projectOverrides.piBackfillModel;
     delete projectOverrides.importPathMaps;
     delete projectOverrides.importIgnoredDirectories;
+    // The reranker decides where memory text is sent, so only the global file
+    // may set it; a checked-in project file must not redirect a user's
+    // memories to its own server.
+    delete projectOverrides.rerankerEnabled;
+    delete projectOverrides.rerankerApiUrl;
+    delete projectOverrides.rerankerModel;
+    delete projectOverrides.rerankerApiKey;
+    delete projectOverrides.rerankerTimeoutMs;
     // A checked-in project file must not point the web app at another folder.
     delete projectOverrides.claudeConfigDir;
     delete projectOverrides.webServerAutoStart;

package/dist/core/retrieval.js

@@ -26,7 +26,7 @@ export async function buildRetrievalSection(prompt, directory, sessionId, budget
         : search.results;
     if (results.length === 0)
         return null;
-    const memoryContext = await formatContextForPrompt(tags.user.userEmail || null, { results }, budget);
+    const memoryContext = await formatContextForPrompt(tags.user.userEmail || null, { results }, budget, search.queryVector);
     return memoryContext || null;
 }
 /**

package/dist/services/chat/chat-answer.d.ts

@@ -0,0 +1,56 @@
+import type { RetrievedSources } from "./chat-retrieve.js";
+import type { ChatRole, ChatSource } from "./chat-types.js";
+/** Each source is cut to this many characters in the prompt. Stored snapshots keep it whole. */
+export declare const MAX_SOURCE_CHARS = 2000;
+export declare const ANSWER_TOOL: {
+    readonly type: "function";
+    readonly function: {
+        readonly name: "answer";
+        readonly description: "Answer the question from the numbered sources";
+        readonly parameters: {
+            readonly type: "object";
+            readonly properties: {
+                readonly answer: {
+                    readonly type: "string";
+                    readonly description: "The answer, with a marker such as [S1] or [P2] after each claim";
+                };
+                readonly citations: {
+                    readonly type: "array";
+                    readonly items: {
+                        readonly type: "string";
+                    };
+                };
+                readonly insufficient: {
+                    readonly type: "boolean";
+                    readonly description: "True when the sources do not answer the question";
+                };
+            };
+            readonly required: readonly ["answer", "citations", "insufficient"];
+        };
+    };
+};
+export declare function numberSources(found: RetrievedSources): ChatSource[];
+export declare function buildAnswerPrompt(question: string, turns: Array<{
+    role: ChatRole;
+    content: string;
+}>, sources: ChatSource[]): {
+    system: string;
+    user: string;
+};
+/**
+ * Removes `[S9]`-style markers that name a source that was not sent, and lists
+ * the valid ones in order of first use. Other bracketed text is left alone.
+ */
+export declare function validateCitations(answer: string, sentLabels: string[]): {
+    text: string;
+    cited: string[];
+};
+export interface FinalAnswer {
+    text: string;
+    cited: string[];
+    insufficient: boolean;
+    /** True when a full answer cites no valid source; the page shows a note. */
+    uncited: boolean;
+}
+/** The checked answer, or null when the reply holds no answer text. */
+export declare function finaliseAnswer(raw: unknown, sources: ChatSource[]): FinalAnswer | null;

package/dist/services/chat/chat-answer.js

@@ -0,0 +1,88 @@
+import { MAX_PLAN_TURNS, MAX_TURN_CHARS } from "./chat-plan.js";
+/** Each source is cut to this many characters in the prompt. Stored snapshots keep it whole. */
+export const MAX_SOURCE_CHARS = 2000;
+export const ANSWER_TOOL = {
+    type: "function",
+    function: {
+        name: "answer",
+        description: "Answer the question from the numbered sources",
+        parameters: {
+            type: "object",
+            properties: {
+                answer: {
+                    type: "string",
+                    description: "The answer, with a marker such as [S1] or [P2] after each claim",
+                },
+                citations: { type: "array", items: { type: "string" } },
+                insufficient: {
+                    type: "boolean",
+                    description: "True when the sources do not answer the question",
+                },
+            },
+            required: ["answer", "citations", "insufficient"],
+        },
+    },
+};
+const SYSTEM = [
+    "You answer questions about one user's saved memories of past coding sessions and their profile.",
+    "Use only the numbered sources given. Do not use outside knowledge.",
+    "Put the source marker, such as [S1] or [P2], after each claim it supports.",
+    "When the sources do not answer the question, say so briefly and set insufficient to true.",
+    "Write plain, short British English. Call the answer tool.",
+].join(" ");
+export function numberSources(found) {
+    return [
+        ...found.memories.map((source, i) => ({ ...source, label: `S${i + 1}` })),
+        ...found.profile.map((source, i) => ({ ...source, label: `P${i + 1}` })),
+    ];
+}
+function describeSource(source) {
+    const meta = source.kind === "memory"
+        ? [source.projectName, source.tags?.length ? `tags: ${source.tags.join(", ")}` : ""]
+        : [source.section, source.category];
+    const header = meta.filter(Boolean).join("; ");
+    return `[${source.label}] ${source.text.slice(0, MAX_SOURCE_CHARS)}${header ? `\n(${header})` : ""}`;
+}
+export function buildAnswerPrompt(question, turns, sources) {
+    const history = turns
+        .slice(-MAX_PLAN_TURNS)
+        .map((turn) => `${turn.role}: ${turn.content.slice(0, MAX_TURN_CHARS)}`);
+    const user = [
+        history.length ? `Earlier turns:\n${history.join("\n")}` : "",
+        `Sources:\n${sources.map(describeSource).join("\n\n")}`,
+        `Question: ${question}`,
+    ]
+        .filter(Boolean)
+        .join("\n\n");
+    return { system: SYSTEM, user };
+}
+const MARKER = /\s*\[([SP]\d+)\]/g;
+/**
+ * Removes `[S9]`-style markers that name a source that was not sent, and lists
+ * the valid ones in order of first use. Other bracketed text is left alone.
+ */
+export function validateCitations(answer, sentLabels) {
+    const sent = new Set(sentLabels);
+    const cited = [];
+    const text = answer
+        .replace(MARKER, (match, label) => {
+        if (!sent.has(label))
+            return "";
+        if (!cited.includes(label))
+            cited.push(label);
+        return match;
+    })
+        .trim();
+    return { text, cited };
+}
+/** The checked answer, or null when the reply holds no answer text. */
+export function finaliseAnswer(raw, sources) {
+    if (!raw || typeof raw !== "object")
+        return null;
+    const reply = raw;
+    if (typeof reply.answer !== "string" || !reply.answer.trim())
+        return null;
+    const insufficient = reply.insufficient === true;
+    const { text, cited } = validateCitations(reply.answer, sources.map((source) => source.label));
+    return { text, cited, insufficient, uncited: !insufficient && cited.length === 0 };
+}

package/dist/services/chat/chat-plan.d.ts

@@ -0,0 +1,60 @@
+import type { ChatPlanRecord, ChatRole } from "./chat-types.js";
+export declare const MAX_PLAN_TAGS = 200;
+export declare const MAX_PLAN_TURNS = 6;
+export declare const MAX_TURN_CHARS = 500;
+export declare const MAX_PLAN_QUERIES = 3;
+export interface PlanInput {
+    question: string;
+    turns: Array<{
+        role: ChatRole;
+        content: string;
+    }>;
+    tags: Array<{
+        keyword: string;
+        count: number;
+    }>;
+    categories: string[];
+}
+export declare const PLAN_TOOL: {
+    readonly type: "function";
+    readonly function: {
+        readonly name: "plan_search";
+        readonly description: "Choose how to search the user's stored memories and profile";
+        readonly parameters: {
+            readonly type: "object";
+            readonly properties: {
+                readonly queries: {
+                    readonly type: "array";
+                    readonly items: {
+                        readonly type: "string";
+                    };
+                    readonly description: "1 to 3 short search phrases";
+                };
+                readonly tags: {
+                    readonly type: "array";
+                    readonly items: {
+                        readonly type: "string";
+                    };
+                    readonly description: "Tags from the given list that match the question";
+                };
+                readonly profileCategories: {
+                    readonly type: "array";
+                    readonly items: {
+                        readonly type: "string";
+                    };
+                    readonly description: "Profile categories from the given list that match the question";
+                };
+            };
+            readonly required: readonly ["queries", "tags", "profileCategories"];
+        };
+    };
+};
+export declare function buildPlanPrompt(input: PlanInput): {
+    system: string;
+    user: string;
+};
+/**
+ * The model's plan with unknown tags and categories removed. A missing or
+ * broken plan searches with the question alone and is marked skipped.
+ */
+export declare function validatePlan(raw: unknown, input: PlanInput): ChatPlanRecord;