om-memory-system 4.23.0-next.144 → 4.24.0-next.147

package/dist/config-template.d.ts

@@ -0,0 +1,6 @@
+/**
+ * The JSONC template written to a new ~/.config/omms/omms.jsonc. Lived in
+ * src/config.ts until that file reached its size ceiling; the text is
+ * unchanged. Shared by config.ts and the global settings writer.
+ */
+export declare const CONFIG_TEMPLATE = "{\n  // ============================================\n  // omms (Opinionated Modular Memory System) Configuration\n  // ============================================\n  \n  // Storage location for the vector database. Leave unset to use the\n  // default ~/.omms/data (a legacy ~/.opencode-mem/data store is migrated\n  // there automatically on first start; see docs/omms-migration.md).\n  // \"storagePath\": \"~/.omms/data\",\n\n  \"userEmailOverride\": \"\",\n  \"userNameOverride\": \"\",\n  \n  // ============================================\n  // Embedding Model (for similarity search)\n  // ============================================\n  // Local = Hugging Face / ONNX via @huggingface/transformers (not Apple MLX).\n  // Remote = set embeddingApiUrl (OpenAI-compatible /embeddings); embeddingApiKey is optional.\n  \n  // Default: Nomic Embed v1 (768 dimensions, 8192 context, multilingual)\n  \"embeddingModel\": \"Xenova/nomic-embed-text-v1\",\n\n  // Opt-in Nomic task prefixes (search_document: / search_query:). After enabling,\n  // re-index existing memories so store and query vectors stay aligned.\n  // \"embeddingUseTaskPrefixes\": true,\n  \n  // Auto-detected dimensions (no need to set manually)\n  // \"embeddingDimensions\": 768,\n  \n  // Other recommended local models:\n  // \"embeddingModel\": \"Xenova/jina-embeddings-v2-base-en\",  // 768 dims, English-only, 8192 context\n  // \"embeddingModel\": \"Xenova/jina-embeddings-v2-small-en\", // 512 dims, faster, 8192 context\n  // \"embeddingModel\": \"Xenova/all-MiniLM-L6-v2\",            // 384 dims, very fast, 512 context\n  // \"embeddingModel\": \"Xenova/all-mpnet-base-v2\",           // 768 dims, good quality, 512 context\n  \n  // Optional: OpenAI-compatible API for embeddings (URL required; key only if the server needs one)\n  // \"embeddingApiUrl\": \"https://api.openai.com/v1\",\n  // \"embeddingApiKey\": \"env://OPENAI_API_KEY\",  // or \"sk-...\" / \"file:///path/to/key\"\n  // \"embeddingModel\": \"text-embedding-3-small\",  // 1536 dims, auto-detected\n \n  // ============================================\n  // Web Server Settings\n  // ============================================\n  \n  // Start a background import of past chats on each host's next start.\n  \"autoBackfill\": true,\n  // \"piBackfillModel\": \"inherit\", // or \"external\", or \"provider/model\"\n  // \"opencodeBackfillModel\": \"inherit\", // or \"external\", or \"provider/model\"\n  // Directory maps for history recorded in moved or deleted directories.\n  // \"importPathMaps\": [{ \"from\": \"~/code/app-feat-x\", \"to\": \"~/code/app\" }],\n  // Unresolved directories to hide on the Directory maps page. They stay unimported.\n  // \"importIgnoredDirectories\": [\"/private/tmp/scratch\"],\n  // Claude Code's folder, when it is not ~/.claude (like CLAUDE_CONFIG_DIR).\n  // \"claudeConfigDir\": \"~/.claude-work\",\n\n  // Register the web app to start when you log in.\n  \"webServerAutoStart\": true,\n  // Enable web UI for managing memories (accessible at http://localhost:4747)\n  \"webServerEnabled\": true,\n  \n  // Port for web UI server\n  \"webServerPort\": 4747,\n  \n  // Host address for web UI (use 127.0.0.1 for local only, 0.0.0.0 for network access)\n  \"webServerHost\": \"127.0.0.1\",\n\n  // Optional HTTP Basic Auth for the web UI. Accepts literal, env://, or file:// secrets.\n  // \"webServerAuthPassword\": \"\",\n  // \"webServerAuthUsername\": \"\",\n\n  // Deprecated: imported once into the API token table, then no longer read.\n  // Create API tokens on the Settings page (Keys and access) instead.\n  // \"webServerApiToken\": \"env://OMMS_WEB_TOKEN\",\n  \n  // ============================================\n  // Database Settings\n  // ============================================\n  \n  // Maximum vectors per database shard (auto-creates new shard when limit reached)\n  \"maxVectorsPerShard\": 50000,\n  \n  // Automatically delete old memories based on retention period\n  \"autoCleanupEnabled\": true,\n  \n  // Days to keep memories before auto-cleanup (only if autoCleanupEnabled is true)\n  \"autoCleanupRetentionDays\": 30,\n  \n  // Automatically detect and remove duplicate memories\n  \"deduplicationEnabled\": true,\n  \n   // Similarity threshold (0-1) for detecting duplicates (higher = stricter)\n   \"deduplicationSimilarityThreshold\": 0.90,\n   \n  // ============================================\n  // Memory Scope Settings\n  // ============================================\n\n  // Default scope for memory list/search queries\n  // \"project\" keeps queries within the current project, \"all-projects\" searches across all project shards\n  \"memory\": {\n    \"defaultScope\": \"project\"\n  },\n\n  // ============================================\n  // Model for auto-capture and profile learning (same rule in OpenCode and Pi)\n  // ============================================\n\n   // Which model summarises your work, in this order:\n   //   1. The host model below: opencodeProvider/opencodeModel in OpenCode,\n   //      piProvider/piModel in Pi. Set the model to \"inherit\" to follow\n   //      whatever model the session is using, or to \"external\" to send\n   //      every call to the external API below.\n   //   2. If no host model is set: the external API (memoryModel/memoryApiUrl/\n   //      memoryApiKey further down).\n   //   3. If neither is set: the session's own model.\n   // If the host model fails and the external API is configured, the external\n   // API is used instead.\n   //\n   // Host models go through OpenCode's or Pi's own sign-in (OAuth like Claude\n   // Pro/Max, GitHub Copilot, ChatGPT, or your own key there), so no separate API\n   // key is needed here.\n   //\n   // Examples (OpenCode names from 'opencode providers list'; Pi names from its model list):\n   //   \"opencodeProvider\": \"anthropic\",    \"opencodeModel\": \"claude-haiku-4-5-20251001\"\n   //   \"opencodeProvider\": \"openai\",       \"opencodeModel\": \"inherit\"\n   //   \"piProvider\": \"openai-codex\",       \"piModel\": \"gpt-5.6-luna\"\n   //\n   // \"opencodeProvider\": \"anthropic\",\n   // \"opencodeModel\": \"claude-haiku-4-5-20251001\",\n   // \"piProvider\": \"openai-codex\",\n   // \"piModel\": \"gpt-5.6-luna\",\n\n   // ============================================\n   // Auto-Capture Settings\n   // ============================================\n  \n  // Auto-capture runs in the background without blocking your main session.\n  // Note: Ollama may not support tool calling. Use OpenAI, Anthropic, or Groq for best results.\n  \n  \"autoCaptureEnabled\": true,\n  \n  // Provider type: \"openai-chat\" | \"openai-responses\" | \"anthropic\" | \"minimax\" | \"orcarouter\" | \"google-gemini\"\n  // Note: \"openai-chat\" is a generic OpenAI API-compatible mode.\n  // Any service that follows the OpenAI Chat Completions API can use it via custom \"memoryApiUrl\".\n  \"memoryProvider\": \"openai-chat\",\n  \n  // External API (step 2 above, and the fallback when a host model fails).\n  // Uncomment all 3 lines and replace memoryApiKey before use:\n  // \"memoryModel\": \"gpt-4o-mini\",\n  // \"memoryApiUrl\": \"https://api.openai.com/v1\",\n  // \"memoryApiKey\": \"sk-...\",\n\n  // API Key Formats:\n  // Direct value:        \"sk-...\"\n  // From file:           \"file://~/.config/litellm-key.txt\"\n  // From env variable:   \"env://LITELLM_API_KEY\"\n  \n  // Examples for different providers:\n  // Any OpenAI-compatible endpoint can use the \"openai-chat\" provider pattern below.\n  // Common examples: DeepSeek, Qwen (via Alibaba Cloud ModelStudio),\n  // Zhipu GLM (BigModel platform), and Kimi (Moonshot AI platform).\n\n  // OpenAI Chat Completion (default, backward compatible):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"gpt-4o-mini\"\n  //   \"memoryApiUrl\": \"https://api.openai.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n\n  // DeepSeek (OpenAI-compatible example):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"deepseek-chat\"\n  //   \"memoryApiUrl\": \"https://api.deepseek.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n  \n  // OpenAI Responses API (recommended, with session support):\n  //   \"memoryProvider\": \"openai-responses\"\n  //   \"memoryModel\": \"gpt-4o\"\n  //   \"memoryApiUrl\": \"https://api.openai.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n  \n  // Anthropic (with session support):\n  //   \"memoryProvider\": \"anthropic\"\n  //   \"memoryModel\": \"claude-3-5-haiku-20241022\"\n  //   \"memoryApiUrl\": \"https://api.anthropic.com/v1\"\n  //   \"memoryApiKey\": \"sk-ant-...\"\n\n  // MiniMax (Anthropic Messages-compatible endpoint, with session support):\n  //   \"memoryProvider\": \"minimax\"\n  //   \"memoryModel\": \"MiniMax-M3\"\n  //   \"memoryApiUrl\": \"https://api.minimax.io\"        // global endpoint\n  //   \"memoryApiKey\": \"<MiniMax API key>\"\n  //   // China endpoint: \"memoryApiUrl\": \"https://api.minimaxi.com\"\n  //   // Optional adaptive thinking for MiniMax-M3:\n  //   \"memoryExtraParams\": { \"thinking\": { \"type\": \"adaptive\" } }\n\n  // OrcaRouter (OpenAI-compatible gateway, namespaced model IDs, with session support):\n  //   \"memoryProvider\": \"orcarouter\"\n  //   \"memoryApiKey\": \"<OrcaRouter API key>\"\n  //   // memoryApiUrl and memoryModel are optional \u2014 they default to\n  //   // https://api.orcarouter.ai/v1 and \"orcarouter/auto\" (a routing alias).\n  //   // OrcaRouter rejects bare model names, so if you set memoryModel, use a\n  //   // namespaced ID such as \"openai/gpt-5.5\" or \"deepseek/deepseek-v4-flash\".\n  //   \"memoryModel\": \"openai/gpt-5.5\"\n\n  // Groq (OpenAI-compatible, use openai-chat provider):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"llama-3.3-70b-versatile\"\n  //   \"memoryApiUrl\": \"https://api.groq.com/openai/v1\"\n  //   \"memoryApiKey\": \"gsk_...\"\n  \n  // Maximum iterations for multi-turn AI analysis (for openai-responses, anthropic, and minimax)\n  \"autoCaptureMaxIterations\": 5,\n   \n  // Timeout per iteration in milliseconds (30 seconds default)\n  \"autoCaptureIterationTimeout\": 30000,\n\n  // Tries for one turn's capture, on Pi and OpenCode, 2 and 4 seconds apart. After the\n  // last one fails, a turn the model could not reach goes to the capture retry queue.\n  \"autoCaptureMaxRetries\": 3,\n\n  // Maximum UTF-8 bytes for the auto-capture markdown context sent to the summary model.\n  // Prevents HTTP 400 context overflows on models with ~131K token windows (e.g. Groq Llama).\n  // Rough guide: tokens \u2248 bytes / 4 for mixed code/prose.\n  \"autoCaptureMaxContextBytes\": 131072,\n   \n  // Days to keep AI session history before cleanup\n  \"aiSessionRetentionDays\": 7,\n\n  // Capture diagnostics: every capture attempt always writes one metadata line\n  // (model, stop reason, sizes, outcome) to ~/.omms/omms.log, with no\n  // conversation text. Set captureTrace to true to also write each attempt's\n  // prompt and raw reply to ~/.omms/traces/ for debugging. Traces can contain\n  // conversation content: <private> text and common API key formats are\n  // redacted, files are readable only by you, and they are deleted after\n  // captureTraceRetentionDays. Only this global file can turn tracing on.\n  // \"captureTrace\": false,\n  // \"captureTraceRetentionDays\": 7,\n  // \"captureAttemptRetentionDays\": 30,\n\n  // Capture retry queue: when a live capture fails because the model cannot be\n  // reached (network error, timeout, HTTP 408, 429 or 5xx), OMMS keeps a copy of\n  // the turn in ~/.omms/data and tries it again later. The copy has <private>\n  // text and common API key formats removed. Turns are deleted when the retry\n  // ends or after captureRetryRetentionHours (0 to 720). 0 turns the queue off.\n  // Only this global file can set it.\n  // \"captureRetryRetentionHours\": 72,\n\n  // Temperature for AI API requests (set to false to omit parameter for models that don't support it)\n  // Some reasoning models (like o1, o3, gpt-5) don't support temperature parameter\n  // Set to false and add \"memoryTemperature\": false in config when using such models\n  \"memoryTemperature\": 0.3,\n\n  // Extra parameters to include in API request body\n  // Useful for local inference servers (e.g. llama-server with --jinja) that support\n  // additional parameters like disabling thinking/reasoning mode\n  // Example for Qwen3 models: { \"enable_thinking\": false }\n  // \"memoryExtraParams\": {},\n\n  // Language for auto-capture summaries (default: \"auto\" for auto-detection)\n  // Options: \"auto\", \"en\", \"id\", \"zh\", \"ja\", \"es\", \"fr\", \"de\", \"ru\", \"pt\", \"ar\", \"ko\"\n  // \"autoCaptureLanguage\": \"auto\",\n\n  // ============================================\n  // Toast Notifications\n  // ============================================\n\n  // Show toast when memory is auto-captured\n  \"showAutoCaptureToasts\": true,\n\n  // Show toast when user profile is updated\n  \"showUserProfileToasts\": true,\n\n  // Show toast for error messages\n  \"showErrorToasts\": true,\n\n  // ============================================\n  // User Profile System\n  // ============================================\n\n  // Analyze user prompts every N prompts to build/update your user profile\n  // When N uncaptured prompts accumulate, AI will analyze them to identify:\n  // - User preferences (code style, communication style, tool preferences)\n  // - User patterns (recurring topics, problem domains, technical interests)\n  // - User workflows (development habits, sequences, learning style)\n  // - Skill level (overall and per-domain assessment)\n  \"userProfileAnalysisInterval\": 10,\n\n  // Days before inactive items (all types) are eligible for removal\n  \"userProfileStaleDays\": 2,\n\n  // Days before an inactive workflow is eligible for removal. A workflow seen\n  // userProfileMinEvidenceForRetention times or more is kept.\n  \"userProfileWorkflowStaleDays\": 30,\n\n  // Number of preferences shown in UI\n  \"userProfileDisplayPreferences\": 20,\n  \n  // Number of patterns shown in UI\n  \"userProfileDisplayPatterns\": 15,\n  \n  // Number of workflows shown in UI\n  \"userProfileDisplayWorkflows\": 10,\n  \n  // Number of preferences injected into LLM conversation context\n  // Keep this small \u2014 the strongest signals are enough; more dilute LLM attention\n  \"userProfileInjectPreferences\": 5,\n  \n  // Number of patterns injected into LLM conversation context\n  \"userProfileInjectPatterns\": 5,\n  \n  // Number of workflows injected into LLM conversation context\n  \"userProfileInjectWorkflows\": 3,\n  \n  // Days before preference confidence starts to decay (if not reinforced)\n  // Preferences that aren't seen again will gradually lose confidence and be removed\n  \"userProfileConfidenceDecayDays\": 30,\n  \n  // Number of profile versions to keep in changelog (for rollback/debugging)\n  // Older versions are automatically cleaned up\n  \"userProfileChangelogRetentionCount\": 5,\n\n  // Minimum evidence count for a preference/pattern to survive confidence decay\n  // Items confirmed fewer times are more likely to be pruned when confidence decays\n  \"userProfileMinEvidenceForRetention\": 3,\n\n  // Merge duplicate or vague profile items every N analysed prompts (about 1 model call per 100 items)\n  \"userProfileAutoCleanupEnabled\": true,\n  \"userProfileAutoCleanupInterval\": 500,\n\n  // Enable LLM validation of existing preferences against recent behavior.\n  // When enabled, each analysis round checks if top-5 preferences still match recent prompts.\n  // Experimental \u2014 disabled by default.\n  \"userProfileValidationEnabled\": false,\n\n  // ============================================\n  // Search Settings\n  // ============================================\n   \n  // Minimum similarity score (0-1) for memory search results. Leave unset to\n  // follow the default for your embedding model: 0.5 for qwen3-embedding:0.6b\n  // (the only model measured), 0.6 for every other embedder.\n  // \"similarityThreshold\": 0.5,\n\n  // Maximum number of memories to return in search results.\n  // Manual searches can ask for fewer; prompt retrieval uses this ceiling.\n  \"maxMemories\": 10,\n\n  // ============================================\n  // Memory Context Controls\n  // ============================================\n\n  // Recent memories added at fresh session start (OpenCode V1 and Claude Code).\n  // Pi and OpenCode V2 search the store with each prompt instead.\n  \"chatMessage\": {\n    \"enabled\": true,\n    // Recent memories injected at session start.\n    \"maxMemories\": 3,\n    \"excludeCurrentSession\": true,\n    // When to inject: \"first\" message of a session or \"always\".\n    // \"injectOn\": \"first\",\n  },\n\n  // OpenCode profile-learning input limit, in UTF-8 bytes. Smaller values can\n  // omit prompts from that input. Does not limit Pi or Claude Code profile input.\n  \"userProfileMaxContextBytes\": 32768,\n\n  // Shared budget for automatic memory context (profile text, memories, and\n  // formatting) across all hosts, in approximate tokens: ceil(UTF-8 bytes / 4).\n  // Covers prompt retrieval, recent-memory injection, and session-memory\n  // restoration. Not an exact provider token limit or a spending limit.\n  \"retrievalMaxTokens\": 2000,\n\n  // ============================================\n  // Advanced Settings\n  // ============================================\n  \n  // Inject user profile into AI context (preferences, patterns, workflows)\n  \"injectProfile\": true\n}\n";

package/dist/config-template.js

@@ -0,0 +1,385 @@
+/**
+ * The JSONC template written to a new ~/.config/omms/omms.jsonc. Lived in
+ * src/config.ts until that file reached its size ceiling; the text is
+ * unchanged. Shared by config.ts and the global settings writer.
+ */
+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. Leave unset to
+  // follow the default for your embedding model: 0.5 for qwen3-embedding:0.6b
+  // (the only model measured), 0.6 for every other embedder.
+  // "similarityThreshold": 0.5,
+
+  // 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
+}
+`;

package/dist/config.d.ts

@@ -14,6 +14,12 @@ interface OmmsConfig {
     embeddingUseTaskPrefixes?: boolean;
     similarityThreshold?: number;
     maxMemories?: number;
+    /** Opt-in reranker (global only): reorder search candidates with an external server. */
+    rerankerEnabled?: boolean;
+    rerankerApiUrl?: string;
+    rerankerModel?: string;
+    rerankerApiKey?: string;
+    rerankerTimeoutMs?: number;
     /** Shared budget for automatic memory context in approximate tokens (ceil(UTF-8 bytes / 4)). */
     retrievalMaxTokens?: number;
     maxProfileItems?: number;
@@ -107,7 +113,8 @@ interface OmmsConfig {
         injectOn?: "first" | "always";
     };
 }
-export declare const CONFIG_TEMPLATE = "{\n  // ============================================\n  // omms (Opinionated Modular Memory System) Configuration\n  // ============================================\n  \n  // Storage location for the vector database. Leave unset to use the\n  // default ~/.omms/data (a legacy ~/.opencode-mem/data store is migrated\n  // there automatically on first start; see docs/omms-migration.md).\n  // \"storagePath\": \"~/.omms/data\",\n\n  \"userEmailOverride\": \"\",\n  \"userNameOverride\": \"\",\n  \n  // ============================================\n  // Embedding Model (for similarity search)\n  // ============================================\n  // Local = Hugging Face / ONNX via @huggingface/transformers (not Apple MLX).\n  // Remote = set embeddingApiUrl (OpenAI-compatible /embeddings); embeddingApiKey is optional.\n  \n  // Default: Nomic Embed v1 (768 dimensions, 8192 context, multilingual)\n  \"embeddingModel\": \"Xenova/nomic-embed-text-v1\",\n\n  // Opt-in Nomic task prefixes (search_document: / search_query:). After enabling,\n  // re-index existing memories so store and query vectors stay aligned.\n  // \"embeddingUseTaskPrefixes\": true,\n  \n  // Auto-detected dimensions (no need to set manually)\n  // \"embeddingDimensions\": 768,\n  \n  // Other recommended local models:\n  // \"embeddingModel\": \"Xenova/jina-embeddings-v2-base-en\",  // 768 dims, English-only, 8192 context\n  // \"embeddingModel\": \"Xenova/jina-embeddings-v2-small-en\", // 512 dims, faster, 8192 context\n  // \"embeddingModel\": \"Xenova/all-MiniLM-L6-v2\",            // 384 dims, very fast, 512 context\n  // \"embeddingModel\": \"Xenova/all-mpnet-base-v2\",           // 768 dims, good quality, 512 context\n  \n  // Optional: OpenAI-compatible API for embeddings (URL required; key only if the server needs one)\n  // \"embeddingApiUrl\": \"https://api.openai.com/v1\",\n  // \"embeddingApiKey\": \"env://OPENAI_API_KEY\",  // or \"sk-...\" / \"file:///path/to/key\"\n  // \"embeddingModel\": \"text-embedding-3-small\",  // 1536 dims, auto-detected\n \n  // ============================================\n  // Web Server Settings\n  // ============================================\n  \n  // Start a background import of past chats on each host's next start.\n  \"autoBackfill\": true,\n  // \"piBackfillModel\": \"inherit\", // or \"external\", or \"provider/model\"\n  // \"opencodeBackfillModel\": \"inherit\", // or \"external\", or \"provider/model\"\n  // Directory maps for history recorded in moved or deleted directories.\n  // \"importPathMaps\": [{ \"from\": \"~/code/app-feat-x\", \"to\": \"~/code/app\" }],\n  // Unresolved directories to hide on the Directory maps page. They stay unimported.\n  // \"importIgnoredDirectories\": [\"/private/tmp/scratch\"],\n  // Claude Code's folder, when it is not ~/.claude (like CLAUDE_CONFIG_DIR).\n  // \"claudeConfigDir\": \"~/.claude-work\",\n\n  // Register the web app to start when you log in.\n  \"webServerAutoStart\": true,\n  // Enable web UI for managing memories (accessible at http://localhost:4747)\n  \"webServerEnabled\": true,\n  \n  // Port for web UI server\n  \"webServerPort\": 4747,\n  \n  // Host address for web UI (use 127.0.0.1 for local only, 0.0.0.0 for network access)\n  \"webServerHost\": \"127.0.0.1\",\n\n  // Optional HTTP Basic Auth for the web UI. Accepts literal, env://, or file:// secrets.\n  // \"webServerAuthPassword\": \"\",\n  // \"webServerAuthUsername\": \"\",\n\n  // Deprecated: imported once into the API token table, then no longer read.\n  // Create API tokens on the Settings page (Keys and access) instead.\n  // \"webServerApiToken\": \"env://OMMS_WEB_TOKEN\",\n  \n  // ============================================\n  // Database Settings\n  // ============================================\n  \n  // Maximum vectors per database shard (auto-creates new shard when limit reached)\n  \"maxVectorsPerShard\": 50000,\n  \n  // Automatically delete old memories based on retention period\n  \"autoCleanupEnabled\": true,\n  \n  // Days to keep memories before auto-cleanup (only if autoCleanupEnabled is true)\n  \"autoCleanupRetentionDays\": 30,\n  \n  // Automatically detect and remove duplicate memories\n  \"deduplicationEnabled\": true,\n  \n   // Similarity threshold (0-1) for detecting duplicates (higher = stricter)\n   \"deduplicationSimilarityThreshold\": 0.90,\n   \n  // ============================================\n  // Memory Scope Settings\n  // ============================================\n\n  // Default scope for memory list/search queries\n  // \"project\" keeps queries within the current project, \"all-projects\" searches across all project shards\n  \"memory\": {\n    \"defaultScope\": \"project\"\n  },\n\n  // ============================================\n  // Model for auto-capture and profile learning (same rule in OpenCode and Pi)\n  // ============================================\n\n   // Which model summarises your work, in this order:\n   //   1. The host model below: opencodeProvider/opencodeModel in OpenCode,\n   //      piProvider/piModel in Pi. Set the model to \"inherit\" to follow\n   //      whatever model the session is using, or to \"external\" to send\n   //      every call to the external API below.\n   //   2. If no host model is set: the external API (memoryModel/memoryApiUrl/\n   //      memoryApiKey further down).\n   //   3. If neither is set: the session's own model.\n   // If the host model fails and the external API is configured, the external\n   // API is used instead.\n   //\n   // Host models go through OpenCode's or Pi's own sign-in (OAuth like Claude\n   // Pro/Max, GitHub Copilot, ChatGPT, or your own key there), so no separate API\n   // key is needed here.\n   //\n   // Examples (OpenCode names from 'opencode providers list'; Pi names from its model list):\n   //   \"opencodeProvider\": \"anthropic\",    \"opencodeModel\": \"claude-haiku-4-5-20251001\"\n   //   \"opencodeProvider\": \"openai\",       \"opencodeModel\": \"inherit\"\n   //   \"piProvider\": \"openai-codex\",       \"piModel\": \"gpt-5.6-luna\"\n   //\n   // \"opencodeProvider\": \"anthropic\",\n   // \"opencodeModel\": \"claude-haiku-4-5-20251001\",\n   // \"piProvider\": \"openai-codex\",\n   // \"piModel\": \"gpt-5.6-luna\",\n\n   // ============================================\n   // Auto-Capture Settings\n   // ============================================\n  \n  // Auto-capture runs in the background without blocking your main session.\n  // Note: Ollama may not support tool calling. Use OpenAI, Anthropic, or Groq for best results.\n  \n  \"autoCaptureEnabled\": true,\n  \n  // Provider type: \"openai-chat\" | \"openai-responses\" | \"anthropic\" | \"minimax\" | \"orcarouter\" | \"google-gemini\"\n  // Note: \"openai-chat\" is a generic OpenAI API-compatible mode.\n  // Any service that follows the OpenAI Chat Completions API can use it via custom \"memoryApiUrl\".\n  \"memoryProvider\": \"openai-chat\",\n  \n  // External API (step 2 above, and the fallback when a host model fails).\n  // Uncomment all 3 lines and replace memoryApiKey before use:\n  // \"memoryModel\": \"gpt-4o-mini\",\n  // \"memoryApiUrl\": \"https://api.openai.com/v1\",\n  // \"memoryApiKey\": \"sk-...\",\n\n  // API Key Formats:\n  // Direct value:        \"sk-...\"\n  // From file:           \"file://~/.config/litellm-key.txt\"\n  // From env variable:   \"env://LITELLM_API_KEY\"\n  \n  // Examples for different providers:\n  // Any OpenAI-compatible endpoint can use the \"openai-chat\" provider pattern below.\n  // Common examples: DeepSeek, Qwen (via Alibaba Cloud ModelStudio),\n  // Zhipu GLM (BigModel platform), and Kimi (Moonshot AI platform).\n\n  // OpenAI Chat Completion (default, backward compatible):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"gpt-4o-mini\"\n  //   \"memoryApiUrl\": \"https://api.openai.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n\n  // DeepSeek (OpenAI-compatible example):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"deepseek-chat\"\n  //   \"memoryApiUrl\": \"https://api.deepseek.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n  \n  // OpenAI Responses API (recommended, with session support):\n  //   \"memoryProvider\": \"openai-responses\"\n  //   \"memoryModel\": \"gpt-4o\"\n  //   \"memoryApiUrl\": \"https://api.openai.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n  \n  // Anthropic (with session support):\n  //   \"memoryProvider\": \"anthropic\"\n  //   \"memoryModel\": \"claude-3-5-haiku-20241022\"\n  //   \"memoryApiUrl\": \"https://api.anthropic.com/v1\"\n  //   \"memoryApiKey\": \"sk-ant-...\"\n\n  // MiniMax (Anthropic Messages-compatible endpoint, with session support):\n  //   \"memoryProvider\": \"minimax\"\n  //   \"memoryModel\": \"MiniMax-M3\"\n  //   \"memoryApiUrl\": \"https://api.minimax.io\"        // global endpoint\n  //   \"memoryApiKey\": \"<MiniMax API key>\"\n  //   // China endpoint: \"memoryApiUrl\": \"https://api.minimaxi.com\"\n  //   // Optional adaptive thinking for MiniMax-M3:\n  //   \"memoryExtraParams\": { \"thinking\": { \"type\": \"adaptive\" } }\n\n  // OrcaRouter (OpenAI-compatible gateway, namespaced model IDs, with session support):\n  //   \"memoryProvider\": \"orcarouter\"\n  //   \"memoryApiKey\": \"<OrcaRouter API key>\"\n  //   // memoryApiUrl and memoryModel are optional \u2014 they default to\n  //   // https://api.orcarouter.ai/v1 and \"orcarouter/auto\" (a routing alias).\n  //   // OrcaRouter rejects bare model names, so if you set memoryModel, use a\n  //   // namespaced ID such as \"openai/gpt-5.5\" or \"deepseek/deepseek-v4-flash\".\n  //   \"memoryModel\": \"openai/gpt-5.5\"\n\n  // Groq (OpenAI-compatible, use openai-chat provider):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"llama-3.3-70b-versatile\"\n  //   \"memoryApiUrl\": \"https://api.groq.com/openai/v1\"\n  //   \"memoryApiKey\": \"gsk_...\"\n  \n  // Maximum iterations for multi-turn AI analysis (for openai-responses, anthropic, and minimax)\n  \"autoCaptureMaxIterations\": 5,\n   \n  // Timeout per iteration in milliseconds (30 seconds default)\n  \"autoCaptureIterationTimeout\": 30000,\n\n  // Tries for one turn's capture, on Pi and OpenCode, 2 and 4 seconds apart. After the\n  // last one fails, a turn the model could not reach goes to the capture retry queue.\n  \"autoCaptureMaxRetries\": 3,\n\n  // Maximum UTF-8 bytes for the auto-capture markdown context sent to the summary model.\n  // Prevents HTTP 400 context overflows on models with ~131K token windows (e.g. Groq Llama).\n  // Rough guide: tokens \u2248 bytes / 4 for mixed code/prose.\n  \"autoCaptureMaxContextBytes\": 131072,\n   \n  // Days to keep AI session history before cleanup\n  \"aiSessionRetentionDays\": 7,\n\n  // Capture diagnostics: every capture attempt always writes one metadata line\n  // (model, stop reason, sizes, outcome) to ~/.omms/omms.log, with no\n  // conversation text. Set captureTrace to true to also write each attempt's\n  // prompt and raw reply to ~/.omms/traces/ for debugging. Traces can contain\n  // conversation content: <private> text and common API key formats are\n  // redacted, files are readable only by you, and they are deleted after\n  // captureTraceRetentionDays. Only this global file can turn tracing on.\n  // \"captureTrace\": false,\n  // \"captureTraceRetentionDays\": 7,\n  // \"captureAttemptRetentionDays\": 30,\n\n  // Capture retry queue: when a live capture fails because the model cannot be\n  // reached (network error, timeout, HTTP 408, 429 or 5xx), OMMS keeps a copy of\n  // the turn in ~/.omms/data and tries it again later. The copy has <private>\n  // text and common API key formats removed. Turns are deleted when the retry\n  // ends or after captureRetryRetentionHours (0 to 720). 0 turns the queue off.\n  // Only this global file can set it.\n  // \"captureRetryRetentionHours\": 72,\n\n  // Temperature for AI API requests (set to false to omit parameter for models that don't support it)\n  // Some reasoning models (like o1, o3, gpt-5) don't support temperature parameter\n  // Set to false and add \"memoryTemperature\": false in config when using such models\n  \"memoryTemperature\": 0.3,\n\n  // Extra parameters to include in API request body\n  // Useful for local inference servers (e.g. llama-server with --jinja) that support\n  // additional parameters like disabling thinking/reasoning mode\n  // Example for Qwen3 models: { \"enable_thinking\": false }\n  // \"memoryExtraParams\": {},\n\n  // Language for auto-capture summaries (default: \"auto\" for auto-detection)\n  // Options: \"auto\", \"en\", \"id\", \"zh\", \"ja\", \"es\", \"fr\", \"de\", \"ru\", \"pt\", \"ar\", \"ko\"\n  // \"autoCaptureLanguage\": \"auto\",\n\n  // ============================================\n  // Toast Notifications\n  // ============================================\n\n  // Show toast when memory is auto-captured\n  \"showAutoCaptureToasts\": true,\n\n  // Show toast when user profile is updated\n  \"showUserProfileToasts\": true,\n\n  // Show toast for error messages\n  \"showErrorToasts\": true,\n\n  // ============================================\n  // User Profile System\n  // ============================================\n\n  // Analyze user prompts every N prompts to build/update your user profile\n  // When N uncaptured prompts accumulate, AI will analyze them to identify:\n  // - User preferences (code style, communication style, tool preferences)\n  // - User patterns (recurring topics, problem domains, technical interests)\n  // - User workflows (development habits, sequences, learning style)\n  // - Skill level (overall and per-domain assessment)\n  \"userProfileAnalysisInterval\": 10,\n\n  // Days before inactive items (all types) are eligible for removal\n  \"userProfileStaleDays\": 2,\n\n  // Days before an inactive workflow is eligible for removal. A workflow seen\n  // userProfileMinEvidenceForRetention times or more is kept.\n  \"userProfileWorkflowStaleDays\": 30,\n\n  // Number of preferences shown in UI\n  \"userProfileDisplayPreferences\": 20,\n  \n  // Number of patterns shown in UI\n  \"userProfileDisplayPatterns\": 15,\n  \n  // Number of workflows shown in UI\n  \"userProfileDisplayWorkflows\": 10,\n  \n  // Number of preferences injected into LLM conversation context\n  // Keep this small \u2014 the strongest signals are enough; more dilute LLM attention\n  \"userProfileInjectPreferences\": 5,\n  \n  // Number of patterns injected into LLM conversation context\n  \"userProfileInjectPatterns\": 5,\n  \n  // Number of workflows injected into LLM conversation context\n  \"userProfileInjectWorkflows\": 3,\n  \n  // Days before preference confidence starts to decay (if not reinforced)\n  // Preferences that aren't seen again will gradually lose confidence and be removed\n  \"userProfileConfidenceDecayDays\": 30,\n  \n  // Number of profile versions to keep in changelog (for rollback/debugging)\n  // Older versions are automatically cleaned up\n  \"userProfileChangelogRetentionCount\": 5,\n\n  // Minimum evidence count for a preference/pattern to survive confidence decay\n  // Items confirmed fewer times are more likely to be pruned when confidence decays\n  \"userProfileMinEvidenceForRetention\": 3,\n\n  // Merge duplicate or vague profile items every N analysed prompts (about 1 model call per 100 items)\n  \"userProfileAutoCleanupEnabled\": true,\n  \"userProfileAutoCleanupInterval\": 500,\n\n  // Enable LLM validation of existing preferences against recent behavior.\n  // When enabled, each analysis round checks if top-5 preferences still match recent prompts.\n  // Experimental \u2014 disabled by default.\n  \"userProfileValidationEnabled\": false,\n\n  // ============================================\n  // Search Settings\n  // ============================================\n  \n  // Minimum similarity score (0-1) for memory search results\n  \"similarityThreshold\": 0.6,\n\n  // Maximum number of memories to return in search results.\n  // Manual searches can ask for fewer; prompt retrieval uses this ceiling.\n  \"maxMemories\": 10,\n\n  // ============================================\n  // Memory Context Controls\n  // ============================================\n\n  // Recent memories added at fresh session start (OpenCode V1 and Claude Code).\n  // Pi and OpenCode V2 search the store with each prompt instead.\n  \"chatMessage\": {\n    \"enabled\": true,\n    // Recent memories injected at session start.\n    \"maxMemories\": 3,\n    \"excludeCurrentSession\": true,\n    // When to inject: \"first\" message of a session or \"always\".\n    // \"injectOn\": \"first\",\n  },\n\n  // OpenCode profile-learning input limit, in UTF-8 bytes. Smaller values can\n  // omit prompts from that input. Does not limit Pi or Claude Code profile input.\n  \"userProfileMaxContextBytes\": 32768,\n\n  // Shared budget for automatic memory context (profile text, memories, and\n  // formatting) across all hosts, in approximate tokens: ceil(UTF-8 bytes / 4).\n  // Covers prompt retrieval, recent-memory injection, and session-memory\n  // restoration. Not an exact provider token limit or a spending limit.\n  \"retrievalMaxTokens\": 2000,\n\n  // ============================================\n  // Advanced Settings\n  // ============================================\n  \n  // Inject user profile into AI context (preferences, patterns, workflows)\n  \"injectProfile\": true\n}\n";
+import { CONFIG_TEMPLATE } from "./config-template.js";
+export { CONFIG_TEMPLATE };
 export declare function normalizeAutoCaptureMaxContextBytes(value: number): number;
 /** Whole hours from 0 to 720. A value that is not a number gives the default. */
 export declare function normalizeCaptureRetryRetentionHours(value: unknown): number;
@@ -122,6 +129,11 @@ declare function buildConfig(fileConfig: OmmsConfig): {
     embeddingApiUrl: string | undefined;
     embeddingApiKey: string | undefined;
     similarityThreshold: number;
+    rerankerEnabled: boolean;
+    rerankerApiUrl: string | undefined;
+    rerankerModel: string | undefined;
+    rerankerApiKey: string | undefined;
+    rerankerTimeoutMs: number;
     maxMemories: number;
     retrievalMaxTokens: number;
     maxProfileItems: number;
@@ -220,6 +232,11 @@ export declare let CONFIG: {
     embeddingApiUrl: string | undefined;
     embeddingApiKey: string | undefined;
     similarityThreshold: number;
+    rerankerEnabled: boolean;
+    rerankerApiUrl: string | undefined;
+    rerankerModel: string | undefined;
+    rerankerApiKey: string | undefined;
+    rerankerTimeoutMs: number;
     maxMemories: number;
     retrievalMaxTokens: number;
     maxProfileItems: number;