om-memory-system 5.3.0 → 5.4.1

package/dist/config-template.d.ts

@@ -3,4 +3,4 @@
  * 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  // Skill learning: record each full turn (prompt, replies, tool calls with\n  // their outcome, and the skills used) in ~/.omms/data/skills.db, so OMMS can\n  // rank your skills and draft better versions. Off by default. A project config\n  // can turn it off but never on. No model call is made to record a turn.\n  // \"skillLearning\": false,\n  // Remove <private> text and common API key formats before a turn is stored.\n  // Global config only. Turning this off stores text as written.\n  // \"skillLearningRedact\": true,\n  // Folders OMMS reads skills from. Absolute or starting with ~/. Global config\n  // only. Unset: ~/.agents/skills and the Claude Code folder's skills folder.\n  // \"skillFolders\": [\"~/.agents/skills\"],\n\n  // Visible recall: the OpenCode tui entrypoint draws a row in the session for\n  // each memory recall. On by default. A project config can turn it off but\n  // never on.\n  // \"visibleRecall\": true,\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";
+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  // Skill learning: record each full turn (prompt, replies, tool calls with\n  // their outcome, and the skills used) in ~/.omms/data/skills.db, so OMMS can\n  // rank your skills and draft better versions. Off by default. A project config\n  // can turn it off but never on. No model call is made to record a turn.\n  // \"skillLearning\": false,\n  // Remove <private> text and common API key formats before a turn is stored.\n  // Global config only. Turning this off stores text as written.\n  // \"skillLearningRedact\": true,\n  // Folders OMMS reads skills from. Absolute or starting with ~/. Global config\n  // only. Unset: ~/.agents/skills and the Claude Code folder's skills folder.\n  // \"skillFolders\": [\"~/.agents/skills\"],\n\n  // Visible recall: the OpenCode terminal TUI and Pi's interactive terminal draw\n  // a row for each memory recall. On by default. A project config can turn it\n  // off but never on.\n  // \"visibleRecall\": true,\n\n  // Memory injection, on OpenCode V2, Pi and Claude Code. Each prompt gets the\n  // memories that match it. openingContext also searches memory with\n  // openingContextQuery and adds the matches to the\n  // first message of a new session. The query is global only, trimmed, and\n  // 500 characters at most. subagentMemory gives each subagent the memories\n  // that match its task. For both switches, a project config can only turn these off.\n  // \"openingContext\": false,\n  // \"openingContextQuery\": \"project architecture, setup, conventions and key decisions\",\n  // \"subagentMemory\": true,\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  // Restore a session's own memories after compaction and after resume, on\n  // OpenCode V2, Pi and Claude Code. memoryLimit is the most memories to\n  // restore. A project config can turn it on or off.\n  // \"compaction\": { \"enabled\": true, \"memoryLimit\": 10 },\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

@@ -254,11 +254,21 @@ export const CONFIG_TEMPLATE = `{
   // only. Unset: ~/.agents/skills and the Claude Code folder's skills folder.
   // "skillFolders": ["~/.agents/skills"],
 
-  // Visible recall: the OpenCode tui entrypoint draws a row in the session for
-  // each memory recall. On by default. A project config can turn it off but
-  // never on.
+  // Visible recall: the OpenCode terminal TUI and Pi's interactive terminal draw
+  // a row for each memory recall. On by default. A project config can turn it
+  // off but never on.
   // "visibleRecall": true,
 
+  // Memory injection, on OpenCode V2, Pi and Claude Code. Each prompt gets the
+  // memories that match it. openingContext also searches memory with
+  // openingContextQuery and adds the matches to the
+  // first message of a new session. The query is global only, trimmed, and
+  // 500 characters at most. subagentMemory gives each subagent the memories
+  // that match its task. For both switches, a project config can only turn these off.
+  // "openingContext": false,
+  // "openingContextQuery": "project architecture, setup, conventions and key decisions",
+  // "subagentMemory": true,
+
   // 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>
@@ -392,6 +402,11 @@ export const CONFIG_TEMPLATE = `{
   // restoration. Not an exact provider token limit or a spending limit.
   "retrievalMaxTokens": 2000,
 
+  // Restore a session's own memories after compaction and after resume, on
+  // OpenCode V2, Pi and Claude Code. memoryLimit is the most memories to
+  // restore. A project config can turn it on or off.
+  // "compaction": { "enabled": true, "memoryLimit": 10 },
+
   // ============================================
   // Advanced Settings
   // ============================================

package/dist/config.d.ts

@@ -188,6 +188,9 @@ declare function buildConfig(fileConfig: OmmsConfig): {
         maxAgeDays: number | undefined;
         injectOn: "first" | "always";
     };
+    openingContext: boolean;
+    openingContextQuery: string;
+    subagentMemory: boolean;
     visibleRecall: boolean;
     skillLearning: boolean;
     skillLearningRedact: boolean;
@@ -299,6 +302,9 @@ export declare let CONFIG: {
         maxAgeDays: number | undefined;
         injectOn: "first" | "always";
     };
+    openingContext: boolean;
+    openingContextQuery: string;
+    subagentMemory: boolean;
     visibleRecall: boolean;
     skillLearning: boolean;
     skillLearningRedact: boolean;

package/dist/config.js

@@ -11,6 +11,7 @@ import { parseBackfillModel } from "./importer/backfill-model.js";
 import { parseIgnoredDirectories, parseImportPathMaps } from "./importer/import-path-maps.js";
 import { normaliseSkillConfig, projectSkillOverride, } from "./services/skill-learning/skill-config.js";
 import * as visibleRecall from "./services/visible-recall-config.js";
+import * as injectionOptions from "./services/injection-options-config.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";
@@ -31,12 +32,10 @@ const CONFIG_FILES = [
     join(OMMS_CONFIG_DIR, "omms.json"),
     ...LEGACY_CONFIG_FILES,
 ];
-if (!existsSync(OMMS_CONFIG_DIR)) {
+if (!existsSync(OMMS_CONFIG_DIR))
     mkdirSync(OMMS_CONFIG_DIR, { recursive: true });
-}
-if (!existsSync(DATA_DIR)) {
+if (!existsSync(DATA_DIR))
     mkdirSync(DATA_DIR, { recursive: true });
-}
 const DEFAULTS = {
     // Default store: ~/.omms/data (see resolveDefaultStoragePath for the
     // legacy-layout fallback while a legacy store exists unmigrated).
@@ -373,6 +372,7 @@ function buildConfig(fileConfig) {
         captureRetryRetentionHours: normalizeCaptureRetryRetentionHours(fileConfig.captureRetryRetentionHours),
         ...normaliseSkillConfig(fileConfig),
         ...visibleRecall.visibleRecallFromFile(fileConfig),
+        ...injectionOptions.injectionOptionsFromFile(fileConfig, log),
         autoBackfill: fileConfig.autoBackfill ?? DEFAULTS.autoBackfill,
         opencodeBackfillModel: fileConfig.opencodeBackfillModel ?? DEFAULTS.opencodeBackfillModel,
         piBackfillModel: fileConfig.piBackfillModel ?? DEFAULTS.piBackfillModel,
@@ -546,6 +546,7 @@ export function initConfig(directory, options = {}) {
     delete projectOverrides.skillFolders;
     Object.assign(projectOverrides, skillRule.override);
     visibleRecall.applyProjectVisibleRecall(projectConfig, projectOverrides, directory, log);
+    injectionOptions.applyProjectInjectionOptions(projectConfig, projectOverrides, directory, log);
     delete projectOverrides.autoBackfill;
     delete projectOverrides.opencodeBackfillModel;
     delete projectOverrides.piBackfillModel;

package/dist/core/context-budget.d.ts

@@ -59,6 +59,8 @@ export interface PackedSection {
     /** UTF-8 bytes of `text` (the host wrapper is not part of `text`). */
     bytesUsed: number;
 }
+/** The byte ceiling of a budget: `maxBytes` when set, else `maxTokens` (default 2000) in bytes. */
+export declare function resolveCeiling(options: ContextBudgetOptions): number;
 /**
  * Pack payload blocks into a section that fits one byte ceiling.
  *

package/dist/core/context-budget.js

@@ -16,7 +16,8 @@ export function estimateTokens(text) {
 export function tokensToByteCeiling(maxTokens) {
     return maxTokens * 4;
 }
-function resolveCeiling(options) {
+/** The byte ceiling of a budget: `maxBytes` when set, else `maxTokens` (default 2000) in bytes. */
+export function resolveCeiling(options) {
     if (options.maxBytes !== undefined)
         return options.maxBytes;
     return tokensToByteCeiling(options.maxTokens ?? DEFAULT_RETRIEVAL_MAX_TOKENS);

package/dist/core/opening-context.d.ts

@@ -0,0 +1,45 @@
+import { type ContextBudgetOptions } from "./context-budget.js";
+/**
+ * The opening context (design D1c of change align-recall-host-parity): on the
+ * first prompt of a main session, one extra search with the configured query,
+ * added as a second retrieval section after the prompt's own. Shared by
+ * OpenCode V2, Pi and Claude Code. Hosts read `CONFIG` and pass the values in;
+ * this module does not import it, because many tests stub `src/config.js`.
+ */
+/**
+ * The opening search query, or null when the option is off or the query is
+ * empty. Pass `CONFIG.openingContext` and `CONFIG.openingContextQuery`.
+ */
+export declare function openingQueryFor(options: {
+    openingContext: boolean;
+    openingContextQuery: string;
+}): string | null;
+export interface FirstPromptSectionsInput {
+    prompt: string;
+    /** From `openingQueryFor`. Null runs no opening search. */
+    openingQuery: string | null;
+    directory: string;
+    sessionId: string;
+    /**
+     * The request budget, as `buildRetrievalSection` takes it. `wrapperBytes` is
+     * what the host adds around each section, the retrieval tag included; it
+     * defaults to the tag alone.
+     */
+    budget: ContextBudgetOptions;
+    /** Bytes the host adds to the opening section only, such as Claude Code's opening marker. */
+    openingReserveBytes: number;
+}
+/**
+ * The sections for the first prompt of a main session, each wrapped in the
+ * retrieval tag. The prompt section is `wrapRetrievalSection` of what
+ * `buildRetrievalSection` returns for the same budget, when the budget names
+ * its `wrapperBytes` as every host passes it. Both searches run in parallel with the same rules
+ * (`searchSectionMemories`). The opening section leaves out memories the
+ * prompt search found, carries no user profile (the prompt section holds it),
+ * and is packed with what the prompt section leaves of the budget, minus
+ * `openingReserveBytes`. Either section is null when nothing fits or matches.
+ */
+export declare function buildFirstPromptSections(input: FirstPromptSectionsInput): Promise<{
+    promptSection: string | null;
+    openingSection: string | null;
+}>;

package/dist/core/opening-context.js

@@ -0,0 +1,60 @@
+import { formatContextForPrompt } from "../services/context.js";
+import { getTags } from "../services/tags.js";
+import { utf8ByteLength } from "../utils/context-limit.js";
+import { resolveCeiling } from "./context-budget.js";
+import { retrievalWrapperBytes, searchablePrompt, searchSectionMemories, wrapRetrievalSection, } from "./retrieval.js";
+/**
+ * The opening context (design D1c of change align-recall-host-parity): on the
+ * first prompt of a main session, one extra search with the configured query,
+ * added as a second retrieval section after the prompt's own. Shared by
+ * OpenCode V2, Pi and Claude Code. Hosts read `CONFIG` and pass the values in;
+ * this module does not import it, because many tests stub `src/config.js`.
+ */
+/**
+ * The opening search query, or null when the option is off or the query is
+ * empty. Pass `CONFIG.openingContext` and `CONFIG.openingContextQuery`.
+ */
+export function openingQueryFor(options) {
+    if (!options.openingContext)
+        return null;
+    return options.openingContextQuery.trim() || null;
+}
+/**
+ * The sections for the first prompt of a main session, each wrapped in the
+ * retrieval tag. The prompt section is `wrapRetrievalSection` of what
+ * `buildRetrievalSection` returns for the same budget, when the budget names
+ * its `wrapperBytes` as every host passes it. Both searches run in parallel with the same rules
+ * (`searchSectionMemories`). The opening section leaves out memories the
+ * prompt search found, carries no user profile (the prompt section holds it),
+ * and is packed with what the prompt section leaves of the budget, minus
+ * `openingReserveBytes`. Either section is null when nothing fits or matches.
+ */
+export async function buildFirstPromptSections(input) {
+    // A fully private prompt gets no section at all, the opening section included.
+    const promptQuery = searchablePrompt(input.prompt);
+    if (!promptQuery)
+        return { promptSection: null, openingSection: null };
+    const tags = getTags(input.directory);
+    const [promptSearch, openingSearch] = await Promise.all([
+        searchSectionMemories(promptQuery, tags.project.tag, input.sessionId),
+        input.openingQuery
+            ? searchSectionMemories(input.openingQuery, tags.project.tag, input.sessionId)
+            : null,
+    ]);
+    const wrapperBytes = input.budget.wrapperBytes ?? retrievalWrapperBytes();
+    const promptResults = promptSearch?.results ?? [];
+    const promptText = promptResults.length > 0
+        ? await formatContextForPrompt(tags.user.userEmail || null, { results: promptResults }, { ...input.budget, wrapperBytes }, promptSearch?.queryVector)
+        : "";
+    const promptIds = new Set(promptResults.map((result) => result.id));
+    const openingResults = (openingSearch?.results ?? []).filter((result) => !promptIds.has(result.id));
+    const promptBytes = promptText ? utf8ByteLength(promptText) + wrapperBytes : 0;
+    const remaining = resolveCeiling(input.budget) - promptBytes - input.openingReserveBytes;
+    const openingText = openingResults.length > 0 && remaining > 0
+        ? await formatContextForPrompt(null, { results: openingResults }, { maxBytes: remaining, wrapperBytes })
+        : "";
+    return {
+        promptSection: promptText ? wrapRetrievalSection(promptText) : null,
+        openingSection: openingText ? wrapRetrievalSection(openingText) : null,
+    };
+}

package/dist/core/recall-summary.d.ts

@@ -1,12 +1,15 @@
-/** One memory a recall injected, reduced to what a fold row shows. */
+/** One memory a recall injected, reduced to what a fold row shows. A restored memory has no relevance. */
 export type RecallItem = {
-    relevance: number;
+    relevance: number | null;
     id?: string;
     headline: string;
 };
-/** One recall as OMMS made it: a prompt search, or the recent memories a session opens with. */
+/**
+ * One recall as OMMS made it: a prompt search, the opening search with the configured query, a
+ * restored session after compaction or resume, or the recent memories a session opens with.
+ */
 export type RecallSummary = {
-    mode: "search" | "recent";
+    mode: "search" | "recent" | "opening" | "restore";
     query: string | null;
     items: RecallItem[];
     /** Per-session counter the v2 reporter sets when it stores the summary; the TUI keys rows by it. */
@@ -18,10 +21,11 @@ export declare function queryLine(query: string, max?: number): string;
 export declare function headline(body: string): string;
 /**
  * Summarise a formatted retrieval section for a host to show. Returns null when the section holds
- * no memory (a profile-only section included), so the host draws nothing. A `recent` recall has
- * no query, whatever the caller passes.
+ * no memory (a profile-only section included), so the host draws nothing. A `restore` summary
+ * reads the restored-session entries, with no relevance. A `restore` or `recent` recall has no
+ * query, whatever the caller passes.
  */
-export declare function buildRecallSummary(section: string, mode: "search" | "recent", query: string | null): RecallSummary | null;
+export declare function buildRecallSummary(section: string, mode: RecallSummary["mode"], query: string | null): RecallSummary | null;
 /** `memory [mode=search, query="...", limit=4]`: the call's name and arguments, as a tool row shows them. */
 export declare function callTitle(summary: RecallSummary): string;
 /** Body of an unfolded row, one string per line, without indentation (the host indents). */

package/dist/core/recall-summary.js

@@ -14,6 +14,9 @@ export function queryLine(query, max = QUERY_CUT) {
 const SECTION_CLOSE = String.raw `<\/project_knowledge>|<\/memory_context>|<\/omms-retrieval>`;
 const MEMORY_ENTRY = new RegExp(String.raw `<memory(?:\s+id="([^"]*)")?\s+relevance="(\d+)%">([\s\S]*?)` +
     String.raw `(?=<\/memory>\s*(?:<memory\s|${SECTION_CLOSE}|$)|${SECTION_CLOSE}|$)`, "g");
+// A restored-session section from `formatMemoriesForCompaction`: `### Memory N (id)`, no relevance.
+// Same rule as the Claude Code parser (`hooks/omms-recall-parse.js`).
+const RESTORED_ENTRY = /### Memory \d+(?: \(([^)\n]+)\))?\n([\s\S]*?)(?=### Memory \d+|<\/omms-retrieval>|$)/g;
 /** One line that names a memory: the first line under `## Request`, else the first line. */
 export function headline(body) {
     const lines = body
@@ -27,23 +30,25 @@ export function headline(body) {
 }
 /**
  * Summarise a formatted retrieval section for a host to show. Returns null when the section holds
- * no memory (a profile-only section included), so the host draws nothing. A `recent` recall has
- * no query, whatever the caller passes.
+ * no memory (a profile-only section included), so the host draws nothing. A `restore` summary
+ * reads the restored-session entries, with no relevance. A `restore` or `recent` recall has no
+ * query, whatever the caller passes.
  */
 export function buildRecallSummary(section, mode, query) {
     const items = [];
-    for (const match of section.matchAll(MEMORY_ENTRY)) {
-        const text = headline(match[3] ?? "");
+    const restore = mode === "restore";
+    for (const match of section.matchAll(restore ? RESTORED_ENTRY : MEMORY_ENTRY)) {
+        const text = headline((restore ? match[2] : match[3]) ?? "");
         if (!text)
             continue;
-        const item = { relevance: Number(match[2]), headline: text };
+        const item = { relevance: restore ? null : Number(match[2]), headline: text };
         if (match[1])
             item.id = match[1];
         items.push(item);
     }
     if (items.length === 0)
         return null;
-    return { mode, query: mode === "recent" ? null : query, items };
+    return { mode, query: restore || mode === "recent" ? null : query, items };
 }
 /** Escape backslashes and double quotes so the quoted value in the title reads as one string. */
 function escapeQuoted(text) {
@@ -64,8 +69,9 @@ export function unfoldLines(summary) {
         lines.push(`query: ${queryLine(summary.query, UNFOLDED_QUERY_CUT)}`);
     lines.push(`limit: ${summary.items.length}`);
     summary.items.forEach((item, index) => {
-        const id = item.id ? ` id=${item.id}` : "";
-        lines.push(`${index + 1}. ${item.relevance}%${id} ${item.headline}`);
+        const relevance = item.relevance === null ? "" : `${item.relevance}%`;
+        const id = item.id ? `id=${item.id}` : "";
+        lines.push(`${index + 1}. ${[relevance, id, item.headline].filter(Boolean).join(" ")}`);
     });
     return lines;
 }

package/dist/core/retrieval.d.ts

@@ -1,5 +1,21 @@
 import { type ContextBudgetOptions } from "./context-budget.js";
 export declare const RETRIEVAL_SECTION_TAG = "omms-retrieval";
+/**
+ * The search step of every query-driven section (`buildRetrievalSection` and
+ * the opening search in `opening-context.ts`), so they share one rule set: the
+ * store applies the result limit and the relevance threshold, then memories
+ * from the current session are dropped when `chatMessage.excludeCurrentSession`
+ * is on. Returns null when the search failed.
+ */
+export declare function searchSectionMemories(query: string, projectTag: string, sessionId: string): Promise<{
+    results: any[];
+    queryVector?: ArrayLike<number>;
+} | null>;
+/**
+ * The prompt text that may reach a search: `<private>` regions removed. Null when nothing is
+ * left, so a fully private prompt runs no search and nothing of it reaches an embedding provider.
+ */
+export declare function searchablePrompt(prompt: string): string | null;
 /**
  * Prompt-aware semantic retrieval shared by every host (Pi `before_agent_start`,
  * OpenCode v2 `prompt`/`context` hooks): search the shared project memory with

package/dist/core/retrieval.js

@@ -2,31 +2,53 @@ import { CONFIG } from "../config.js";
 import { formatContextForPrompt } from "../services/context.js";
 import { memoryClient } from "../services/client.js";
 import { getTags } from "../services/tags.js";
+import { isFullyPrivate, stripPrivateContent } from "../services/privacy.js";
 import { utf8ByteLength } from "../utils/context-limit.js";
 import { packContextSection } from "./context-budget.js";
 export const RETRIEVAL_SECTION_TAG = "omms-retrieval";
 /**
- * Prompt-aware semantic retrieval shared by every host (Pi `before_agent_start`,
- * OpenCode v2 `prompt`/`context` hooks): search the shared project memory with
- * the incoming prompt and format a bounded context section (memory + user
- * profile). Returns null when nothing relevant is found.
+ * The search step of every query-driven section (`buildRetrievalSection` and
+ * the opening search in `opening-context.ts`), so they share one rule set: the
+ * store applies the result limit and the relevance threshold, then memories
+ * from the current session are dropped when `chatMessage.excludeCurrentSession`
+ * is on. Returns null when the search failed.
  */
-export async function buildRetrievalSection(prompt, directory, sessionId, budget) {
-    if (!prompt.trim())
-        return null;
+export async function searchSectionMemories(query, projectTag, sessionId) {
     // Snapshot before the async search: an excludeCurrentSession edit during
     // the search must not change this operation's filtering.
     const excludeCurrentSession = CONFIG.chatMessage.excludeCurrentSession;
-    const tags = getTags(directory);
-    const search = await memoryClient.searchMemories(prompt, tags.project.tag);
+    const search = await memoryClient.searchMemories(query, projectTag);
     if (!search.success)
         return null;
     const results = excludeCurrentSession
         ? search.results.filter((result) => result.metadata?.sessionID !== sessionId)
         : search.results;
-    if (results.length === 0)
+    return { results, queryVector: search.queryVector };
+}
+/**
+ * The prompt text that may reach a search: `<private>` regions removed. Null when nothing is
+ * left, so a fully private prompt runs no search and nothing of it reaches an embedding provider.
+ */
+export function searchablePrompt(prompt) {
+    if (isFullyPrivate(prompt))
+        return null;
+    return stripPrivateContent(prompt).trim() || null;
+}
+/**
+ * Prompt-aware semantic retrieval shared by every host (Pi `before_agent_start`,
+ * OpenCode v2 `prompt`/`context` hooks): search the shared project memory with
+ * the incoming prompt and format a bounded context section (memory + user
+ * profile). Returns null when nothing relevant is found.
+ */
+export async function buildRetrievalSection(prompt, directory, sessionId, budget) {
+    const query = searchablePrompt(prompt);
+    if (!query)
+        return null;
+    const tags = getTags(directory);
+    const search = await searchSectionMemories(query, tags.project.tag, sessionId);
+    if (!search || search.results.length === 0)
         return null;
-    const memoryContext = await formatContextForPrompt(tags.user.userEmail || null, { results }, budget, search.queryVector);
+    const memoryContext = await formatContextForPrompt(tags.user.userEmail || null, { results: search.results }, budget, search.queryVector);
     return memoryContext || null;
 }
 /**

package/dist/importer/claude-hook-api.d.ts

@@ -15,10 +15,11 @@ export declare class ClaudeHookRequestError extends Error {
     constructor(message: string);
 }
 /**
- * `POST /api/claude/retrieve`: recent project memories for a fresh session,
- * the session's own memories after compaction or resume, or the retrieval
- * section for a prompt. Every text is wrapped in the retrieval tag; an empty
- * string means nothing to add.
+ * `POST /api/claude/retrieve`: the retrieval section for a prompt (and the
+ * opening section on a session's first prompt, when on), the memories for a
+ * starting subagent's task, or the session's own memories after compaction or
+ * resume. A fresh or cleared session adds nothing. Every text is wrapped in the
+ * retrieval tag; an empty string means nothing to add.
  */
 export declare function handleClaudeRetrieve(body: unknown, deps?: ClaudeHookDeps): Promise<{
     additionalContext: string;