om-memory-system 4.12.0 → 4.13.3

package/README.md

@@ -55,6 +55,12 @@ downloads a small embedding model (the part that makes memories searchable),
 so you need internet access once. The terminal import command also needs
 Node.js 22.14 or later.
 
+> [!NOTE]
+> OMMS is tested mostly on macOS and Linux. Windows runs only in the release
+> check, so it gets less testing. If something breaks on Windows,
+> [open an issue](https://github.com/cmdaltctr/omms/issues). If you fix it,
+> a pull request is welcome.
+
 ## Set up
 
 ### 1. Install

package/dist/adapters/claude-code/hook-command.d.ts

@@ -4,4 +4,6 @@ import { type ClaudeStatusOptions } from "./status.js";
  * `om-memory-system claude-hook <event>`. Always returns 0: a hook must never
  * block or break a Claude Code session, whatever goes wrong.
  */
-export declare function runClaudeHookCommand(argv: string[], io?: ClaudeHookOptions & ClaudeStatusOptions): Promise<number>;
+export declare function runClaudeHookCommand(argv: string[], io?: ClaudeHookOptions & ClaudeStatusOptions & {
+    skipSelfUpdate?: boolean;
+}): Promise<number>;

package/dist/adapters/claude-code/hook-command.js

@@ -9,6 +9,18 @@ async function printStatus(io) {
     const status = await buildClaudeStatus({ ...inputs, fetch: io.fetch ?? globalThis.fetch });
     (io.writeStdout ?? ((text) => process.stdout.write(text)))(`${JSON.stringify(status)}\n`);
 }
+/** Update the Claude plugin in the background when npm has a newer release. Never throws. */
+async function selfUpdate(io) {
+    try {
+        const { startPluginSelfUpdate, pluginSelfUpdateDeps } = await import("./plugin-self-update.js");
+        const log = io.log ??
+            (async (message, data) => (await import("../../services/logger.js")).log(message, data));
+        await startPluginSelfUpdate(pluginSelfUpdateDeps(log));
+    }
+    catch {
+        /* The memory hook already ran; a failed update check changes nothing. */
+    }
+}
 /**
  * `om-memory-system claude-hook <event>`. Always returns 0: a hook must never
  * block or break a Claude Code session, whatever goes wrong.
@@ -22,6 +34,8 @@ export async function runClaudeHookCommand(argv, io = {}) {
         }
         if (isHookEvent(event)) {
             await runClaudeHook(event, io);
+            if (event === "session-start" && !io.skipSelfUpdate)
+                await selfUpdate(io);
             return 0;
         }
         // Only a known name reaches the log; argv is not trusted text.

package/dist/adapters/claude-code/plugin-self-update.d.ts

@@ -0,0 +1,27 @@
+/** A second session does not start another update for the same release within this time. */
+export declare const SELF_UPDATE_RETRY_MS: number;
+export type PluginSelfUpdateCode = "started" | "current" | "recent" | "disabled" | "no-plugin" | "unreadable-plugin" | "unreachable" | "spawn-failed";
+export interface PluginSelfUpdateDeps {
+    /** False when `OMMS_DISABLE_UPDATE_CHECK` is `1`. */
+    enabled: boolean;
+    /** `CLAUDE_PLUGIN_ROOT`; undefined outside a Claude plugin. */
+    pluginRoot: string | undefined;
+    readPluginVersion: (root: string) => string;
+    fetch: typeof fetch;
+    now: () => number;
+    readMarker: () => {
+        version: string;
+        startedAt: number;
+    } | null;
+    writeMarker: (value: {
+        version: string;
+        startedAt: number;
+    }) => void;
+    /** Start a process that outlives this hook. */
+    spawnDetached: (command: string, args: string[]) => void;
+    log: (message: string, data: Record<string, unknown>) => void | Promise<void>;
+}
+/** Start a background plugin update when npm has a newer release. Never throws. */
+export declare function startPluginSelfUpdate(deps: PluginSelfUpdateDeps): Promise<PluginSelfUpdateCode>;
+/** Production inputs: the real plugin manifest, npm, and a marker file in ~/.omms. */
+export declare function pluginSelfUpdateDeps(log: PluginSelfUpdateDeps["log"], env?: NodeJS.ProcessEnv): PluginSelfUpdateDeps;

package/dist/adapters/claude-code/plugin-self-update.js

@@ -0,0 +1,103 @@
+import { spawn } from "node:child_process";
+import { mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { homedir } from "node:os";
+import { dirname, join } from "node:path";
+import { availableUpdate, latestNpmVersion } from "../../services/update-check.js";
+// Claude Code installs the OMMS plugin from the `stable` branch, but its own
+// update pass runs at its own time and can leave a new session on the old
+// plugin. At session start OMMS compares the plugin with npm `latest` and, when
+// it is older, runs Claude Code's own update commands in the background. The
+// new copy loads in the next session or after /reload-plugins.
+/** A second session does not start another update for the same release within this time. */
+export const SELF_UPDATE_RETRY_MS = 30 * 60 * 1000;
+/** Node script that refreshes the marketplace, then updates the plugin. Run detached. */
+const UPDATE_SCRIPT = `
+const { spawnSync } = require("node:child_process");
+const shell = process.platform === "win32";
+for (const args of [["plugin", "marketplace", "update", "omms"], ["plugin", "update", "omms@omms"]]) {
+  const run = spawnSync("claude", args, { stdio: "ignore", shell, timeout: 120000 });
+  if (run.status !== 0) process.exit(1);
+}
+`;
+/** Start a background plugin update when npm has a newer release. Never throws. */
+export async function startPluginSelfUpdate(deps) {
+    const code = await decide(deps);
+    if (code.code === "started" || code.code === "spawn-failed") {
+        await deps.log("Claude plugin self-update", {
+            code: code.code,
+            plugin: code.plugin,
+            latest: code.latest,
+        });
+    }
+    return code.code;
+}
+async function decide(deps) {
+    if (!deps.enabled)
+        return { code: "disabled" };
+    if (!deps.pluginRoot)
+        return { code: "no-plugin" };
+    let plugin;
+    try {
+        plugin = deps.readPluginVersion(deps.pluginRoot);
+    }
+    catch {
+        return { code: "unreadable-plugin" };
+    }
+    const latest = await latestNpmVersion(deps.fetch);
+    if (latest === null)
+        return { code: "unreachable" };
+    const newer = availableUpdate(plugin, latest);
+    if (!newer)
+        return { code: "current" };
+    const marker = deps.readMarker();
+    const now = deps.now();
+    if (marker?.version === newer && now - marker.startedAt <= SELF_UPDATE_RETRY_MS) {
+        return { code: "recent" };
+    }
+    // Write first, so a session that starts during the update does not start another.
+    deps.writeMarker({ version: newer, startedAt: now });
+    try {
+        deps.spawnDetached(process.execPath, ["-e", UPDATE_SCRIPT]);
+    }
+    catch {
+        return { code: "spawn-failed", plugin, latest: newer };
+    }
+    return { code: "started", plugin, latest: newer };
+}
+/** Production inputs: the real plugin manifest, npm, and a marker file in ~/.omms. */
+export function pluginSelfUpdateDeps(log, env = process.env) {
+    const markerPath = join(homedir(), ".omms", "claude-plugin-update.json");
+    return {
+        enabled: env.OMMS_DISABLE_UPDATE_CHECK !== "1",
+        pluginRoot: env.CLAUDE_PLUGIN_ROOT || undefined,
+        readPluginVersion: (root) => {
+            const manifest = JSON.parse(readFileSync(join(root, ".claude-plugin", "plugin.json"), "utf8"));
+            if (typeof manifest.version !== "string")
+                throw new Error("no version");
+            return manifest.version;
+        },
+        fetch: globalThis.fetch,
+        now: Date.now,
+        readMarker: () => {
+            try {
+                const value = JSON.parse(readFileSync(markerPath, "utf8"));
+                return typeof value.version === "string" && typeof value.startedAt === "number"
+                    ? { version: value.version, startedAt: value.startedAt }
+                    : null;
+            }
+            catch {
+                return null;
+            }
+        },
+        writeMarker: (value) => {
+            mkdirSync(dirname(markerPath), { recursive: true });
+            writeFileSync(markerPath, JSON.stringify(value));
+        },
+        spawnDetached: (command, args) => {
+            const child = spawn(command, args, { detached: true, stdio: "ignore", windowsHide: true });
+            child.on("error", () => { });
+            child.unref();
+        },
+        log,
+    };
+}

package/dist/config.d.ts

@@ -76,6 +76,7 @@ interface OmmsConfig {
     userProfileDisplayPatterns?: number;
     userProfileDisplayWorkflows?: number;
     userProfileStaleDays?: number;
+    userProfileWorkflowStaleDays?: number;
     userProfileInjectPreferences?: number;
     userProfileInjectPatterns?: number;
     userProfileInjectWorkflows?: number;
@@ -106,7 +107,7 @@ 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  // 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  // Periodically merge duplicate or irrelevant profile items with the configured AI provider\n  \"userProfileAutoCleanupEnabled\": true,\n  // Number of analyzed user prompts between automatic AI cleanup runs\n  \"userProfileAutoCleanupInterval\": 100,\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";
+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  // Periodically merge duplicate or irrelevant profile items with the configured AI provider\n  \"userProfileAutoCleanupEnabled\": true,\n  // Number of analyzed user prompts between automatic AI cleanup runs\n  \"userProfileAutoCleanupInterval\": 100,\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";
 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;
@@ -187,6 +188,7 @@ declare function buildConfig(fileConfig: OmmsConfig): {
     userProfileAutoCleanupInterval: number;
     userProfileValidationEnabled: boolean;
     userProfileStaleDays: number;
+    userProfileWorkflowStaleDays: number;
     showAutoCaptureToasts: boolean;
     showUserProfileToasts: boolean;
     showErrorToasts: boolean;
@@ -284,6 +286,7 @@ export declare let CONFIG: {
     userProfileAutoCleanupInterval: number;
     userProfileValidationEnabled: boolean;
     userProfileStaleDays: number;
+    userProfileWorkflowStaleDays: number;
     showAutoCaptureToasts: boolean;
     showUserProfileToasts: boolean;
     showErrorToasts: boolean;

package/dist/config.js

@@ -78,6 +78,7 @@ const DEFAULTS = {
     userProfileDisplayPatterns: 15,
     userProfileDisplayWorkflows: 10,
     userProfileStaleDays: 2,
+    userProfileWorkflowStaleDays: 30,
     userProfileInjectPreferences: 5,
     userProfileInjectPatterns: 5,
     userProfileInjectWorkflows: 3,
@@ -450,6 +451,10 @@ export const CONFIG_TEMPLATE = `{
   // 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,
   
@@ -762,6 +767,7 @@ function buildConfig(fileConfig) {
         userProfileAutoCleanupInterval,
         userProfileValidationEnabled: fileConfig.userProfileValidationEnabled ?? DEFAULTS.userProfileValidationEnabled,
         userProfileStaleDays: fileConfig.userProfileStaleDays ?? DEFAULTS.userProfileStaleDays,
+        userProfileWorkflowStaleDays: fileConfig.userProfileWorkflowStaleDays ?? DEFAULTS.userProfileWorkflowStaleDays,
         showAutoCaptureToasts: fileConfig.showAutoCaptureToasts ?? DEFAULTS.showAutoCaptureToasts,
         showUserProfileToasts: fileConfig.showUserProfileToasts ?? DEFAULTS.showUserProfileToasts,
         showErrorToasts: fileConfig.showErrorToasts ?? DEFAULTS.showErrorToasts,

package/dist/importer/claude-import.js

@@ -54,6 +54,7 @@ export async function importClaudeHistory(deps, filters) {
             host: "claude-code",
             signal: deps.signal,
             dryRun: Boolean(filters.dryRun),
+            force: Boolean(filters.force),
             model: deps.profile.model,
             batchSize: deps.profile.batchSize,
             ...(deps.profile.onProgress ? { onProgress: deps.profile.onProgress } : {}),

package/dist/importer/import-args.js

@@ -247,7 +247,7 @@ ${source}
   --skip-memories           Record profile prompts only
   --skip-profile            Import memories only
   --profile-batch <n>       Prompts per profile analysis batch (default: 50)
-  --force                   Reprocess already-handled memory work units
+  --force                   Reprocess already-handled memory units and profile prompts
   --help                    Show this help
 
 Values may follow the flag or use --flag=value.`;

package/dist/importer/importer.js

@@ -388,6 +388,7 @@ export async function importPiHistory(deps, filters) {
             host: "pi",
             signal: deps.signal,
             dryRun,
+            force: Boolean(filters.force),
             model: deps.profile.model,
             batchSize: deps.profile.batchSize,
             ...(deps.profile.onProgress ? { onProgress: deps.profile.onProgress } : {}),

package/dist/importer/opencode-import.js

@@ -139,6 +139,7 @@ export async function importOpencodeHistory(options) {
                 host: "opencode",
                 signal: options.signal,
                 dryRun: options.dryRun,
+                force: options.force,
                 batchSize: options.profileBatch,
                 model: options.profileModel,
                 ...(options.onProfileProgress ? { onProgress: options.onProfileProgress } : {}),

package/dist/importer/profile-import.d.ts

@@ -12,12 +12,14 @@ export interface ProfileImportReport {
     remaining: number;
     error?: string;
 }
-type PromptStore = Pick<typeof userPromptManager, "savePrompt" | "markAsCaptured" | "countUnanalyzedForUserLearning" | "getPromptsForUserLearning" | "markMultipleAsUserLearningCaptured">;
+type PromptStore = Pick<typeof userPromptManager, "savePrompt" | "markAsCaptured" | "countUnanalyzedForUserLearning" | "getPromptsForUserLearning" | "markMultipleAsUserLearningCaptured" | "markForUserLearning">;
 type ProfileStore = Pick<typeof userProfileManager, "getActiveProfile" | "createProfile" | "updateProfile">;
 type Ledger = Pick<ImportLedger, "get" | "begin" | "complete"> & Partial<Pick<ImportLedger, "peek">>;
 export interface ProfileImportOptions {
     host: MemoryHost;
     dryRun?: boolean;
+    /** Re-analyse prompts the ledger already records as done. */
+    force?: boolean;
     signal?: AbortSignal;
     batchSize?: number;
     model?: ModelPort;

package/dist/importer/profile-import.js

@@ -46,12 +46,18 @@ export async function importProfileFromHistory(source, options) {
                 continue;
             directory ??= session.directory;
             const profileKey = `${key}#profile`;
-            const existing = !ledgerReadable
+            // A forced run re-analyses each prompt once; this key records that it did.
+            const rebuildKey = `${key}#profile-rebuild`;
+            const read = async (ledgerKey) => !ledgerReadable
                 ? null
                 : options.dryRun && ledger.peek
-                    ? await ledger.peek(profileKey)
-                    : await ledger.get(profileKey);
-            if (existing?.status === "imported") {
+                    ? await ledger.peek(ledgerKey)
+                    : await ledger.get(ledgerKey);
+            const existing = await read(profileKey);
+            const reanalyse = existing?.status === "imported" &&
+                Boolean(options.force) &&
+                (await read(rebuildKey))?.status !== "imported";
+            if (existing?.status === "imported" && !reanalyse) {
                 report.promptsAlreadyHandled++;
                 continue;
             }
@@ -60,7 +66,7 @@ export async function importProfileFromHistory(source, options) {
                 continue;
             }
             await ledger.begin({
-                key: profileKey,
+                key: reanalyse ? rebuildKey : profileKey,
                 sessionId: session.sessionId,
                 sourceFile: session.sourceFile,
                 projectHash: "profile",
@@ -68,7 +74,10 @@ export async function importProfileFromHistory(source, options) {
             // savePrompt also deduplicates a retry after a crash before ledger completion.
             const promptId = await prompts.savePrompt(session.sessionId, unit.userEntryId, session.directory, prompt);
             await prompts.markAsCaptured(promptId);
-            await ledger.complete(profileKey, promptId);
+            // A forced unit may find its prompt still stored and already learned.
+            if (reanalyse)
+                await prompts.markForUserLearning(promptId);
+            await ledger.complete(reanalyse ? rebuildKey : profileKey, promptId);
             report.promptsRecorded++;
         }
     }

package/dist/services/turso/cross-process-write-lock.js

@@ -1,4 +1,5 @@
-import { existsSync, mkdirSync, readFileSync, unlinkSync, writeFileSync } from "node:fs";
+import { randomUUID } from "node:crypto";
+import { existsSync, linkSync, mkdirSync, readFileSync, unlinkSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 import { CONFIG } from "../../config.js";
 /**
@@ -51,6 +52,26 @@ function readLiveLock(path) {
 function lockFilePath(scope, hash) {
     return join(CONFIG.storagePath, WRITE_LOCK_DIR, `${scope}_${hash}.lock`);
 }
+async function publishLock(candidate, path) {
+    // Match the bounded Windows hard-link retries used by the web start lock (TDR-034).
+    const waits = [1, 2, 5, 10, 20, 50, 100, 200];
+    for (let attempt = 0;; attempt++) {
+        try {
+            linkSync(candidate, path);
+            return;
+        }
+        catch (error) {
+            const code = error.code;
+            if (process.platform !== "win32" ||
+                !code ||
+                !["EPERM", "EACCES", "EBUSY"].includes(code) ||
+                waits[attempt] === undefined) {
+                throw error;
+            }
+            await new Promise((resolve) => setTimeout(resolve, waits[attempt]));
+        }
+    }
+}
 export async function withCrossProcessWriteLock(scope, hash, fn) {
     const dir = join(CONFIG.storagePath, WRITE_LOCK_DIR);
     if (!existsSync(dir)) {
@@ -59,31 +80,47 @@ export async function withCrossProcessWriteLock(scope, hash, fn) {
     const path = lockFilePath(scope, hash);
     const state = { pid: process.pid, timestamp: new Date().toISOString() };
     const deadline = Date.now() + LOCK_TIMEOUT_MS;
-    for (;;) {
-        try {
-            writeFileSync(path, JSON.stringify(state), { flag: "wx" });
-            break;
-        }
-        catch {
-            const holder = readLiveLock(path);
-            if (!holder)
-                continue;
-            if (Date.now() > deadline) {
-                throw new Error(`Timed out acquiring the cross-process write lock for ${scope}/${hash}: ` +
-                    `held by pid ${holder.pid} since ${holder.timestamp}`);
+    const candidate = `${path}.${process.pid}.${randomUUID()}.tmp`;
+    let acquired = false;
+    try {
+        // Exclusive open exposes an empty file before its PID is written. Publish
+        // the complete payload with a hard link, which cannot replace a live lock.
+        writeFileSync(candidate, JSON.stringify(state), { flag: "wx" });
+        for (;;) {
+            try {
+                await publishLock(candidate, path);
+                acquired = true;
+                break;
+            }
+            catch (error) {
+                if (error.code !== "EEXIST")
+                    throw error;
+                const holder = readLiveLock(path);
+                if (!holder)
+                    continue;
+                if (Date.now() > deadline) {
+                    throw new Error(`Timed out acquiring the cross-process write lock for ${scope}/${hash}: ` +
+                        `held by pid ${holder.pid} since ${holder.timestamp}`, { cause: error });
+                }
+                await new Promise((resolve) => setTimeout(resolve, LOCK_POLL_MS + Math.random() * 10));
             }
-            await new Promise((resolve) => setTimeout(resolve, LOCK_POLL_MS + Math.random() * 10));
         }
-    }
-    try {
         return await fn();
     }
     finally {
+        if (acquired) {
+            try {
+                unlinkSync(path);
+            }
+            catch {
+                // Best effort: a stale file is cleaned up by the next acquirer.
+            }
+        }
         try {
-            unlinkSync(path);
+            unlinkSync(candidate);
         }
         catch {
-            // Best effort: a stale file is cleaned up by the next acquirer.
+            // A failed publication must not leave its temporary payload behind.
         }
     }
 }

package/dist/services/user-profile/user-profile-manager.js

@@ -348,9 +348,9 @@ export class UserProfileManager {
     }
     decayInMemory(data) {
         const now = Date.now();
-        const prefResult = this.decayItems(data.preferences, now);
-        const patResult = this.decayItems(data.patterns, now);
-        const wfResult = this.decayItems(data.workflows, now);
+        const prefResult = this.decayItems(data.preferences, now, "preference");
+        const patResult = this.decayItems(data.patterns, now, "pattern");
+        const wfResult = this.decayItems(data.workflows, now, "workflow");
         return {
             data: {
                 ...data,
@@ -361,7 +361,7 @@ export class UserProfileManager {
             hasChanges: prefResult.hasChanges || patResult.hasChanges || wfResult.hasChanges,
         };
     }
-    decayItems(items, now) {
+    decayItems(items, now, itemType) {
         const before = items.length;
         let hasChanges = false;
         const filtered = items.filter((item) => {
@@ -376,16 +376,25 @@ export class UserProfileManager {
                 hasChanges = true;
             const age = now - (item.lastSeen || now);
             const ageDays = age / (24 * 60 * 60 * 1000);
-            const staleDays = CONFIG.userProfileStaleDays ?? 2;
+            const isWorkflow = itemType === "workflow";
+            const itemStaleDays = CONFIG.userProfileStaleDays ?? 2;
+            // A workflow never decays sooner than other items, even with a long userProfileStaleDays.
+            const staleDays = isWorkflow
+                ? Math.max(CONFIG.userProfileWorkflowStaleDays ?? 30, itemStaleDays)
+                : itemStaleDays;
             const minEvidence = CONFIG.userProfileMinEvidenceForRetention ?? 3;
             const evidenceCount = Array.isArray(item.evidence)
                 ? item.evidence.length
                 : 0;
+            // The analysis schema gives workflows no evidence, so repeat sightings count instead.
+            const support = isWorkflow
+                ? Math.max(evidenceCount, item.frequency ?? 1)
+                : evidenceCount;
             // Remove inactive, low-evidence items using configured retention thresholds.
-            if (ageDays > staleDays && evidenceCount < minEvidence) {
+            if (ageDays > staleDays && support < minEvidence) {
                 log("profile decay: removed stale", {
                     cat: item.category || item.description?.substring(0, 30),
-                    evidenceCount,
+                    evidenceCount: support,
                     minEvidence,
                     ageDays: Math.round(ageDays),
                     staleDays,

package/dist/services/user-prompt/user-prompt-manager.d.ts

@@ -49,6 +49,8 @@ export declare class UserPromptManager {
     countTrivialPromptsForLearning(): Promise<number>;
     private waitingTrivialPromptIds;
     markAsUserLearningCaptured(promptId: string): Promise<void>;
+    /** Put a learned prompt back in the profile learning queue. */
+    markForUserLearning(promptId: string): Promise<void>;
     markMultipleAsUserLearningCaptured(promptIds: string[]): Promise<void>;
     deleteOldPrompts(cutoffTime: number): Promise<{
         deleted: number;

package/dist/services/user-prompt/user-prompt-manager.js

@@ -246,6 +246,11 @@ export class UserPromptManager {
         const db = await this.ready();
         await db.run(`UPDATE user_prompts SET user_learning_captured = 1 WHERE id = ?`, [promptId]);
     }
+    /** Put a learned prompt back in the profile learning queue. */
+    async markForUserLearning(promptId) {
+        const db = await this.ready();
+        await db.run(`UPDATE user_prompts SET user_learning_captured = 0 WHERE id = ?`, [promptId]);
+    }
     async markMultipleAsUserLearningCaptured(promptIds) {
         if (promptIds.length === 0)
             return;

package/dist/services/web-update.d.ts

@@ -1,4 +1,6 @@
 export declare const UPDATE_CHECK_INTERVAL_MS: number;
+/** An open page triggers a new check once the last one is this old. */
+export declare const STATUS_RECHECK_AFTER_MS: number;
 /** An install that takes longer is stopped, and the web app keeps serving. */
 export declare const INSTALL_TIMEOUT_MS: number;
 export type UpdateState = "idle" | "installing" | "restarting" | "failed";
@@ -60,11 +62,14 @@ export declare class WebUpdate {
     private state;
     private code;
     private started;
+    private lastCheck;
+    private checking;
     constructor(deps: WebUpdateDeps);
-    /** Check now, then every 6 hours. A second call does nothing. */
+    /** Check now, then every 10 minutes. A second call does nothing. */
     start(): void;
-    /** Read npm `latest`. A failed read keeps the last result. */
+    /** Read npm `latest`. A failed read keeps the last result. Calls during a read share it. */
     check(): Promise<void>;
+    private read;
     status(): UpdateStatus;
     private canInstall;
     /** Start the install and the restart. A request while one runs joins it. */

package/dist/services/web-update.js

@@ -1,9 +1,12 @@
 import { posix, win32 } from "node:path";
 import { availableUpdate, latestNpmVersion } from "./update-check.js";
-// The web app's npm release check. It reads npm `latest` at start and then every
-// 6 hours, and keeps the result in memory for `GET /api/web/status`. It imports
+// The web app's npm release check. It reads npm `latest` at start, every 10
+// minutes, and when an open page reads the status and the last check is over a
+// minute old. It keeps the result in memory for `GET /api/web/status`. It imports
 // no host adapter and no config, so tests pass every input.
-export const UPDATE_CHECK_INTERVAL_MS = 6 * 60 * 60 * 1000;
+export const UPDATE_CHECK_INTERVAL_MS = 10 * 60 * 1000;
+/** An open page triggers a new check once the last one is this old. */
+export const STATUS_RECHECK_AFTER_MS = 60 * 1000;
 /** An install that takes longer is stopped, and the web app keeps serving. */
 export const INSTALL_TIMEOUT_MS = 5 * 60 * 1000;
 /** Only a plain release reaches the npm arguments; the value comes from the registry. */
@@ -24,10 +27,12 @@ export class WebUpdate {
     state = "idle";
     code = null;
     started = false;
+    lastCheck = -Infinity;
+    checking = null;
     constructor(deps) {
         this.deps = deps;
     }
-    /** Check now, then every 6 hours. A second call does nothing. */
+    /** Check now, then every 10 minutes. A second call does nothing. */
     start() {
         if (this.started || !this.deps.enabled)
             return;
@@ -35,10 +40,15 @@ export class WebUpdate {
         void this.check();
         this.deps.setInterval(() => void this.check(), UPDATE_CHECK_INTERVAL_MS).unref?.();
     }
-    /** Read npm `latest`. A failed read keeps the last result. */
-    async check() {
+    /** Read npm `latest`. A failed read keeps the last result. Calls during a read share it. */
+    check() {
         if (!this.deps.enabled)
-            return;
+            return Promise.resolve();
+        this.checking ??= this.read().finally(() => (this.checking = null));
+        return this.checking;
+    }
+    async read() {
+        this.lastCheck = this.deps.now();
         const latest = await latestNpmVersion(this.deps.fetch);
         if (latest === null) {
             this.deps.log("Web app update check failed", { code: "unreachable" });
@@ -47,6 +57,9 @@ export class WebUpdate {
         this.available = availableUpdate(this.deps.version, latest);
     }
     status() {
+        // An open page polls the status, so a release shows within about a minute.
+        if (this.deps.now() - this.lastCheck > STATUS_RECHECK_AFTER_MS)
+            void this.check();
         return {
             available: this.available,
             state: this.state,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "om-memory-system",
-  "version": "4.12.0",
+  "version": "4.13.3",
   "description": "Opinionated modular memory system for coding agents (OpenCode + Pi)",
   "type": "module",
   "main": "dist/plugin.js",
@@ -36,7 +36,7 @@
     "web:dev": "cd web && bun run dev",
     "web:build": "cd web && bun run build",
     "dev": "tsc --watch",
-    "test": "bun test",
+    "test": "bash scripts/run-tests-isolated.sh",
     "typecheck": "tsc --noEmit",
     "lint": "eslint . --max-warnings=0",
     "format": "prettier --write .",