om-memory-system 5.10.1 → 6.1.0-next.237

package/dist/adapters/opencode/profile-llm-client.js

@@ -1,29 +1,20 @@
 import { CONFIG } from "../../config.js";
 import { resolveOpencodeHostModel } from "../../services/ai/live-model-choice.js";
 import { loadOpencodeProvider } from "./opencode-provider-loader.js";
-let _cachedClient = null;
-let _cachedProvider = null;
-let _cachedModel = null;
-export async function getOpenCodeClient() {
+/** The OpenCode plugin context for profile calls, once the host model's provider is signed in. */
+export async function getOpencodeProfileHost() {
     const hostModel = resolveOpencodeHostModel(CONFIG);
     if (!hostModel) {
         throw new Error("omms: live calls use the external API; no OpenCode model is configured");
     }
-    const { isProviderConnected, getV2Client, resolveOpencodeModelRef } = await loadOpencodeProvider();
+    const { isProviderConnected, getOpencodeHost, resolveOpencodeModelRef } = await loadOpencodeProvider();
     // "inherit" checks the provider of the model it resolves to.
-    const { providerID: provider, modelID: model } = resolveOpencodeModelRef(hostModel);
-    if (_cachedClient && _cachedProvider === provider && _cachedModel === model) {
-        return _cachedClient;
-    }
+    const { providerID: provider } = resolveOpencodeModelRef(hostModel);
     if (!isProviderConnected(provider)) {
         throw new Error(`opencode provider '${provider}' is not connected. Check your opencode provider configuration.`);
     }
-    const client = getV2Client();
-    if (!client) {
-        throw new Error("omms: v2 client not initialized");
-    }
-    _cachedClient = client;
-    _cachedProvider = provider;
-    _cachedModel = model;
-    return client;
+    const host = getOpencodeHost();
+    if (!host)
+        throw new Error("omms: the OpenCode plugin context is not ready");
+    return host;
 }

package/dist/adapters/opencode/profile-model.d.ts

@@ -1,23 +1,7 @@
 import type { ModelPort } from "../../core/profile-analysis.js";
 import type { OpencodeModelRef } from "./opencode-import-models.js";
-type PromptInfo = {
-    error?: {
-        name: string;
-        data?: {
-            message?: string;
-        };
-    };
-};
-/**
- * Extract assistant text from an OpenCode session.prompt result.
- * AssistantMessage has no `text` field; content lives in `parts` (#177).
- */
-export declare function extractTextFromPromptResult(promptResult: unknown): {
-    info: PromptInfo | undefined;
-    rawText: string;
-};
+import type { OpencodeHost } from "./opencode-host.js";
 /** Adapt OpenCode's host model to the shared profile code, as `adaptPiProfileModel` does for Pi. */
-export declare function adaptOpencodeProfileModel(v2Client: unknown, hostModel: OpencodeModelRef): ModelPort;
+export declare function adaptOpencodeProfileModel(host: Pick<OpencodeHost, "generate">, hostModel: OpencodeModelRef): ModelPort;
 /** Route shared profile calls in this process to OpenCode's host model while one is configured. */
 export declare function registerOpencodeProfileModel(): void;
-export {};

package/dist/adapters/opencode/profile-model.js

@@ -4,22 +4,7 @@ import { log } from "../../services/logger.js";
 import { OPENCODE_PROFILE_CLEANUP_TIMEOUT_MS } from "../../services/request-timeouts.js";
 import { registerHostProfileModel } from "../../services/user-profile/profile-model.js";
 import { loadOpencodeProvider } from "./opencode-provider-loader.js";
-import { getOpenCodeClient } from "./profile-llm-client.js";
-/**
- * Extract assistant text from an OpenCode session.prompt result.
- * AssistantMessage has no `text` field; content lives in `parts` (#177).
- */
-export function extractTextFromPromptResult(promptResult) {
-    const result = promptResult;
-    const info = result?.data?.info ?? result?.info;
-    const parts = result?.data?.parts ?? result?.parts ?? [];
-    const rawText = parts
-        .filter((p) => p.type === "text" && p.text)
-        .map((p) => p.text)
-        .join("\n")
-        .trim();
-    return { info, rawText };
-}
+import { getOpencodeProfileHost } from "./profile-llm-client.js";
 function raceWithTimeout(promise, ms, message) {
     let timer;
     const timeout = new Promise((_, reject) => {
@@ -30,68 +15,32 @@ function raceWithTimeout(promise, ms, message) {
             clearTimeout(timer);
     });
 }
-/**
- * Plain-text reply through a transient OpenCode session. Only AI profile cleanup
- * calls it, so the session carries the cleanup title the history reader skips.
- */
-async function completeViaSession(v2Client, hostModel, systemPrompt, userPrompt) {
+/** Plain-text reply through one sessionless OpenCode generation. Only AI profile cleanup calls it. */
+async function completeViaGeneration(host, hostModel, systemPrompt, userPrompt) {
     const t0 = Date.now();
     // Only "inherit" (or no model configured) needs resolving to the session's recent model.
     const model = hostModel.modelID === "inherit"
         ? (await loadOpencodeProvider()).resolveOpencodeModelRef(hostModel)
         : hostModel;
-    const created = (await raceWithTimeout(v2Client.session.create({
-        title: "omms profile cleanup",
-        directory: process.cwd(),
-    }), 30000, "session.create timeout"));
-    log("AI cleanup: session.create result", {
-        rawType: typeof created,
-        keys: Object.keys(created || {}),
-        hasData: !!created?.data,
-        dataId: created?.data?.id,
-    });
-    const sessionID = created?.data?.id || created?.id || created?.sessionID;
-    if (!sessionID)
-        throw new Error("session.create returned no session id");
-    log("AI cleanup: session created", { sessionID, createMs: Date.now() - t0 });
-    try {
-        const TIMEOUT_MS = OPENCODE_PROFILE_CLEANUP_TIMEOUT_MS;
-        const promptResult = await raceWithTimeout(v2Client.session.prompt({
-            sessionID,
-            model,
-            system: systemPrompt,
-            parts: [{ type: "text", text: userPrompt }],
-            // `noReply` suppresses assistant generation; cleanup needs the JSON reply (#177).
-            noReply: false,
-        }), TIMEOUT_MS, `opencodeClient prompt timeout after ${TIMEOUT_MS}ms`);
-        log("AI cleanup: session.prompt done", { promptMs: Date.now() - t0 });
-        const { info, rawText } = extractTextFromPromptResult(promptResult);
-        if (!info)
-            throw new Error("prompt response missing info");
-        if (info.error)
-            throw new Error(`opencode reported ${info.error.name}: ${info.error.data?.message ?? ""}`);
-        return rawText;
-    }
-    finally {
-        try {
-            await v2Client.session.delete({ sessionID });
-        }
-        catch {
-            // ignore cleanup failures for ephemeral sessions
-        }
-    }
+    const timeoutMs = OPENCODE_PROFILE_CLEANUP_TIMEOUT_MS;
+    const generated = await raceWithTimeout(host.generate.text({
+        prompt: [systemPrompt, userPrompt].filter(Boolean).join("\n\n"),
+        model: { providerID: model.providerID, id: model.modelID },
+    }), timeoutMs, `opencode generation timeout after ${timeoutMs}ms`);
+    log("AI cleanup: generation done", { generateMs: Date.now() - t0 });
+    return (generated?.text ?? "").trim();
 }
 /** Adapt OpenCode's host model to the shared profile code, as `adaptPiProfileModel` does for Pi. */
-export function adaptOpencodeProfileModel(v2Client, hostModel) {
+export function adaptOpencodeProfileModel(host, hostModel) {
     return {
         provider: hostModel.providerID,
         modelId: hostModel.modelID,
-        complete: (systemPrompt, userPrompt) => completeViaSession(v2Client, hostModel, systemPrompt, userPrompt),
+        complete: (systemPrompt, userPrompt) => completeViaGeneration(host, hostModel, systemPrompt, userPrompt),
         async completeStructured(systemPrompt, userPrompt, schema) {
-            const client = await getOpenCodeClient();
+            const profileHost = await getOpencodeProfileHost();
             const { generateStructuredOutput } = await loadOpencodeProvider();
             return generateStructuredOutput({
-                client,
+                host: profileHost,
                 providerID: hostModel.providerID,
                 modelID: hostModel.modelID,
                 systemPrompt,
@@ -107,11 +56,11 @@ export function registerOpencodeProfileModel() {
         const hostModel = resolveOpencodeHostModel(CONFIG);
         if (!hostModel)
             return null;
-        const v2Client = (await loadOpencodeProvider()).getV2Client();
-        if (!v2Client) {
-            log("profile model: opencode client unavailable, falling back to external API", undefined, "warn");
+        const host = (await loadOpencodeProvider()).getOpencodeHost();
+        if (!host) {
+            log("profile model: OpenCode plugin context unavailable, falling back to external API", undefined, "warn");
             return null;
         }
-        return adaptOpencodeProfileModel(v2Client, hostModel);
+        return adaptOpencodeProfileModel(host, hostModel);
     });
 }

package/dist/adapters/opencode/prompt-model.d.ts

@@ -0,0 +1,9 @@
+/**
+ * Remember a model request's model as the session model, and store it on its recorded
+ * prompt row so capture can follow it. The row is written only while live calls follow
+ * the session model ("inherit", or nothing configured).
+ */
+export declare function recordOpencodePromptModel(messageID: string, model: {
+    readonly providerID: string;
+    readonly id: string;
+}): Promise<void>;

package/dist/adapters/opencode/prompt-model.js

@@ -0,0 +1,21 @@
+import { CONFIG, isConfigured } from "../../config.js";
+import { resolveOpencodeHostModel } from "../../services/ai/live-model-choice.js";
+import { log } from "../../services/logger.js";
+import { userPromptManager } from "../../services/user-prompt/user-prompt-manager.js";
+import { rememberOpencodeSessionModel } from "./opencode-host.js";
+/**
+ * Remember a model request's model as the session model, and store it on its recorded
+ * prompt row so capture can follow it. The row is written only while live calls follow
+ * the session model ("inherit", or nothing configured).
+ */
+export async function recordOpencodePromptModel(messageID, model) {
+    rememberOpencodeSessionModel(model);
+    if (!isConfigured() || resolveOpencodeHostModel(CONFIG)?.modelID !== "inherit")
+        return;
+    try {
+        await userPromptManager.setPromptModel(messageID, model.providerID, model.id);
+    }
+    catch (error) {
+        log("v2 context: failed to record the prompt model", { error: String(error) }, "error");
+    }
+}

package/dist/adapters/opencode/skills-tool.d.ts

@@ -1,3 +1,3 @@
-import { type ToolDefinition } from "@opencode-ai/plugin";
-/** The OpenCode `skills` tool. V2 reaches it through the V1 plugin's `tool.skills`. */
-export declare function opencodeSkillsTool(directory: string): ToolDefinition;
+import { type SkillsToolArgs } from "../../core/skills-tool.js";
+/** Run the OpenCode `skills` tool. The V2 plugin registers it in `src/v2/tools.ts`. */
+export declare function runOpencodeSkillsTool(args: SkillsToolArgs, directory: string): Promise<string>;

package/dist/adapters/opencode/skills-tool.js

@@ -1,31 +1,5 @@
-import { tool } from "@opencode-ai/plugin";
-import { SKILLS_TOOL_DESCRIPTION, SKILLS_TOOL_MODES, runSkillsTool, } from "../../core/skills-tool.js";
-/** The OpenCode `skills` tool. V2 reaches it through the V1 plugin's `tool.skills`. */
-export function opencodeSkillsTool(directory) {
-    return tool({
-        description: SKILLS_TOOL_DESCRIPTION,
-        args: {
-            mode: tool.schema.enum(SKILLS_TOOL_MODES).optional(),
-            skill: tool.schema.string().optional(),
-            id: tool.schema.string().optional(),
-            state: tool.schema.string().optional(),
-            order: tool.schema.string().optional(),
-            folder: tool.schema.string().optional(),
-            idea: tool.schema
-                .object({
-                name: tool.schema.string().optional(),
-                description: tool.schema.string().optional(),
-                reason: tool.schema.string().optional(),
-                evidenceTurnIds: tool.schema.array(tool.schema.number()).optional(),
-            })
-                .optional(),
-            limit: tool.schema.number().optional(),
-            content: tool.schema.string().optional(),
-            summary: tool.schema.string().optional(),
-            evidence: tool.schema.array(tool.schema.number()).optional(),
-        },
-        async execute(args) {
-            return runSkillsTool(args, directory);
-        },
-    });
+import { runSkillsTool } from "../../core/skills-tool.js";
+/** Run the OpenCode `skills` tool. The V2 plugin registers it in `src/v2/tools.ts`. */
+export function runOpencodeSkillsTool(args, directory) {
+    return runSkillsTool(args, directory);
 }

package/dist/adapters/opencode/toast.d.ts

@@ -0,0 +1,11 @@
+export type OpencodeNotice = {
+    readonly title?: string;
+    readonly message: string;
+    readonly variant?: "info" | "success" | "warning" | "error";
+    readonly duration?: number;
+};
+/**
+ * Show a notice from the OpenCode server plugin. The V2 server plugin context has no
+ * toast API, so the notice goes to the console, which OpenCode writes to its log.
+ */
+export declare function showOpencodeNotice(notice: OpencodeNotice): void;

package/dist/adapters/opencode/toast.js

@@ -0,0 +1,11 @@
+/**
+ * Show a notice from the OpenCode server plugin. The V2 server plugin context has no
+ * toast API, so the notice goes to the console, which OpenCode writes to its log.
+ */
+export function showOpencodeNotice(notice) {
+    const text = [notice.title, notice.message].filter(Boolean).join(": ");
+    if (!text)
+        return;
+    const method = notice.variant === "error" ? "error" : notice.variant === "warning" ? "warn" : "info";
+    console[method](`[omms] ${text}`);
+}

package/dist/adapters/opencode/user-prompt.d.ts

@@ -1,10 +1,10 @@
 import { isStructuredSummaryPromptMessage } from "../../core/internal-prompt.js";
 export { isStructuredSummaryPromptMessage };
 /** True when a prompt is omms's own internal traffic and must not be recorded or searched. */
-export declare function isInternalPrompt(sessionID: string, userMessage: string): boolean;
+export declare function isInternalPrompt(userMessage: string): boolean;
 /**
  * Record a real user prompt so idle auto-capture can pair it with the session's
- * response. Shared by the V1 `chat.message` hook and the v2 `prompt` hook.
+ * response. The V2 `context` hook records each admitted prompt.
  * Returns false when the prompt is empty or internal and was not recorded.
  */
 export declare function recordUserPrompt(sessionID: string, messageID: string, directory: string, userMessage: string): Promise<boolean>;

package/dist/adapters/opencode/user-prompt.js

@@ -1,19 +1,18 @@
 import { isStructuredSummaryPromptMessage } from "../../core/internal-prompt.js";
-import { isInternalStructuredSession } from "./opencode-provider.js";
 import { userPromptManager } from "../../services/user-prompt/user-prompt-manager.js";
 import { isFullyPrivate, stripPrivateContent } from "../../services/privacy.js";
 export { isStructuredSummaryPromptMessage };
 /** True when a prompt is omms's own internal traffic and must not be recorded or searched. */
-export function isInternalPrompt(sessionID, userMessage) {
-    return isStructuredSummaryPromptMessage(userMessage) || isInternalStructuredSession(sessionID);
+export function isInternalPrompt(userMessage) {
+    return isStructuredSummaryPromptMessage(userMessage);
 }
 /**
  * Record a real user prompt so idle auto-capture can pair it with the session's
- * response. Shared by the V1 `chat.message` hook and the v2 `prompt` hook.
+ * response. The V2 `context` hook records each admitted prompt.
  * Returns false when the prompt is empty or internal and was not recorded.
  */
 export async function recordUserPrompt(sessionID, messageID, directory, userMessage) {
-    if (!userMessage.trim() || isInternalPrompt(sessionID, userMessage))
+    if (!userMessage.trim() || isInternalPrompt(userMessage))
         return false;
     // Profile learning quotes saved prompts as evidence, so <private> text must
     // never reach the store (the Claude Code hook strips it the same way).

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 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  // Memorisation settings (global config only).\n  // Repeat for a session snapshot: \"add\" or \"replace\" (default \"add\").\n  // \"memorisationRepeat\": \"add\",\n  // Maximum stored text: 100 to 20,000 characters or 25 to 5,000 tokens.\n  // Tokens are estimated with the same rule as retrievalMaxTokens.\n  // \"memorisationMaxSize\": 2000,\n  // \"memorisationSizeUnit\": \"characters\", // or \"tokens\"\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";
+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  // Per-prompt memory retrieval on every host. \"enabled\": false turns it off;\n  // \"excludeCurrentSession\" leaves out memories captured in the current session.\n  \"chatMessage\": {\n    \"enabled\": true,\n    \"excludeCurrentSession\": true,\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  // Memorisation settings (global config only).\n  // Repeat for a session snapshot: \"add\" or \"replace\" (default \"add\").\n  // \"memorisationRepeat\": \"add\",\n  // Maximum stored text: 100 to 20,000 characters or 25 to 5,000 tokens.\n  // Tokens are estimated with the same rule as retrievalMaxTokens.\n  // \"memorisationMaxSize\": 2000,\n  // \"memorisationSizeUnit\": \"characters\", // or \"tokens\"\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

@@ -381,15 +381,11 @@ export const CONFIG_TEMPLATE = `{
   // Memory Context Controls
   // ============================================
 
-  // Recent memories added at fresh session start (OpenCode V1 and Claude Code).
-  // Pi and OpenCode V2 search the store with each prompt instead.
+  // Per-prompt memory retrieval on every host. "enabled": false turns it off;
+  // "excludeCurrentSession" leaves out memories captured in the current session.
   "chatMessage": {
     "enabled": true,
-    // Recent memories injected at session start.
-    "maxMemories": 3,
     "excludeCurrentSession": true,
-    // When to inject: "first" message of a session or "always".
-    // "injectOn": "first",
   },
 
   // OpenCode profile-learning input limit, in UTF-8 bytes. Smaller values can

package/dist/config.d.ts

@@ -124,10 +124,7 @@ interface OmmsConfig {
     };
     chatMessage?: {
         enabled?: boolean;
-        maxMemories?: number;
         excludeCurrentSession?: boolean;
-        maxAgeDays?: number;
-        injectOn?: "first" | "always";
     };
 }
 import { CONFIG_TEMPLATE } from "./config-template.js";
@@ -189,10 +186,7 @@ declare function buildConfig(fileConfig: OmmsConfig): {
     };
     chatMessage: {
         enabled: boolean | undefined;
-        maxMemories: number;
         excludeCurrentSession: boolean | undefined;
-        maxAgeDays: number | undefined;
-        injectOn: "first" | "always";
     };
     openingContext: boolean;
     openingContextQuery: string;
@@ -306,10 +300,7 @@ export declare let CONFIG: {
     };
     chatMessage: {
         enabled: boolean | undefined;
-        maxMemories: number;
         excludeCurrentSession: boolean | undefined;
-        maxAgeDays: number | undefined;
-        injectOn: "first" | "always";
     };
     openingContext: boolean;
     openingContextQuery: string;

package/dist/config.js

@@ -108,10 +108,7 @@ const DEFAULTS = {
     },
     chatMessage: {
         enabled: true,
-        maxMemories: 3,
         excludeCurrentSession: true,
-        maxAgeDays: undefined,
-        injectOn: "first",
     },
 };
 /**
@@ -266,9 +263,6 @@ function buildConfig(fileConfig) {
     const maxMemories = checkedOwnMemoryLimit(MEMORY_LIMITS.maxMemories, fileConfig, "maxMemories", DEFAULTS.maxMemories);
     const retrievalMaxTokens = checkedOwnMemoryLimit(MEMORY_LIMITS.retrievalMaxTokens, fileConfig, "retrievalMaxTokens", DEFAULTS.retrievalMaxTokens);
     const userProfileMaxContextBytes = checkedOwnMemoryLimit(MEMORY_LIMITS.userProfileMaxContextBytes, fileConfig, "userProfileMaxContextBytes", DEFAULTS.userProfileMaxContextBytes);
-    const chatMessageMaxMemories = checkedMemoryLimit(MEMORY_LIMITS["chatMessage.maxMemories"], fileConfig.chatMessage && Object.hasOwn(fileConfig.chatMessage, "maxMemories")
-        ? fileConfig.chatMessage.maxMemories
-        : DEFAULTS.chatMessage.maxMemories);
     const userProfileAutoCleanupInterval = fileConfig.userProfileAutoCleanupInterval ?? DEFAULTS.userProfileAutoCleanupInterval;
     parseBackfillModel(fileConfig, "pi");
     parseBackfillModel(fileConfig, "opencode");
@@ -410,10 +404,7 @@ function buildConfig(fileConfig) {
         },
         chatMessage: {
             enabled: fileConfig.chatMessage?.enabled ?? DEFAULTS.chatMessage.enabled,
-            maxMemories: chatMessageMaxMemories,
             excludeCurrentSession: fileConfig.chatMessage?.excludeCurrentSession ?? DEFAULTS.chatMessage.excludeCurrentSession,
-            maxAgeDays: fileConfig.chatMessage?.maxAgeDays,
-            injectOn: (fileConfig.chatMessage?.injectOn ?? DEFAULTS.chatMessage.injectOn),
         },
     };
 }

package/dist/core/retrieval.d.ts

@@ -24,27 +24,6 @@ export declare function searchablePrompt(prompt: string): string | null;
  * profile). Returns null when nothing relevant is found.
  */
 export declare function buildRetrievalSection(prompt: string, directory: string, sessionId: string, budget?: ContextBudgetOptions): Promise<string | null>;
-export interface RecentMemoriesOptions {
-    /** Project container tag, from `getTags(directory).project.tag`. */
-    projectTag: string;
-    /** Profile owner, from `getTags(directory).user.userEmail`. */
-    userEmail: string | null | undefined;
-    sessionId: string;
-    /** `CONFIG.chatMessage.maxMemories` */
-    maxMemories?: number;
-    /** `CONFIG.chatMessage.excludeCurrentSession` */
-    excludeCurrentSession?: boolean;
-    /** `CONFIG.chatMessage.maxAgeDays` */
-    maxAgeDays?: number;
-    /** Shared packing budget; hosts pass the snapshotted `retrievalMaxTokens`. */
-    budget?: ContextBudgetOptions;
-}
-/**
- * The project's most recent memories as a context section, shared by every host
- * that injects them at session start (OpenCode's first `chat.message`, Claude
- * Code `SessionStart`). Returns null when no memory qualifies.
- */
-export declare function buildRecentMemoriesSection(options: RecentMemoriesOptions): Promise<string | null>;
 /** Delimit a retrieval section so it is recognisable as injected memory, not user text. */
 export declare function wrapRetrievalSection(section: string): string;
 /** Byte overhead the `omms-retrieval` wrapper adds around one section. */

package/dist/core/retrieval.js

@@ -61,35 +61,6 @@ export async function buildRetrievalSection(prompt, directory, sessionId, budget
     const memoryContext = await formatContextForPrompt(tags.user.userEmail || null, { results: search.results }, budget, search.queryVector);
     return memoryContext || null;
 }
-/**
- * The project's most recent memories as a context section, shared by every host
- * that injects them at session start (OpenCode's first `chat.message`, Claude
- * Code `SessionStart`). Returns null when no memory qualifies.
- */
-export async function buildRecentMemoriesSection(options) {
-    const listResult = await memoryClient.listMemories(options.projectTag, options.maxMemories, "project", { includeLinked: true });
-    let memories = listResult.success ? listResult.memories : [];
-    if (options.excludeCurrentSession) {
-        memories = memories.filter((m) => m.metadata?.sessionID !== options.sessionId);
-    }
-    if (options.maxAgeDays) {
-        const cutoffDate = Date.now() - options.maxAgeDays * 86400000;
-        memories = memories.filter((m) => new Date(m.createdAt).getTime() > cutoffDate);
-    }
-    if (memories.length === 0)
-        return null;
-    const projectMemories = {
-        results: memories.map((m) => ({
-            id: m.id,
-            similarity: 1.0,
-            memory: m.summary,
-        })),
-        total: memories.length,
-        timing: 0,
-    };
-    const memoryContext = await formatContextForPrompt(options.userEmail || null, projectMemories, options.budget);
-    return memoryContext || null;
-}
 /** Delimit a retrieval section so it is recognisable as injected memory, not user text. */
 export function wrapRetrievalSection(section) {
     return `<${RETRIEVAL_SECTION_TAG}>\n${section}\n</${RETRIEVAL_SECTION_TAG}>`;

package/dist/importer/handled-lookup.d.ts

@@ -0,0 +1,17 @@
+import type { MemoryHost } from "../types/index.js";
+import type { ImportWindow } from "./importer.js";
+import type { ImportLedgerRow, PiImportLedger } from "./ledger.js";
+/**
+ * How a host finds a ledger row that already handled a unit. The default is
+ * the exact key. OpenCode matches a session and user message with any last
+ * answer id: the answer ids of old imports no longer exist in the database,
+ * and the 2.x tables can give a unit another last answer.
+ */
+export type HandledLookup = (ledger: PiImportLedger, query: {
+    key: string;
+    sessionId: string;
+    window: ImportWindow;
+    readOnly: boolean;
+}) => Promise<ImportLedgerRow | null>;
+/** The lookup for a host, chosen once per import run. */
+export declare function handledLookupFor(host: MemoryHost): HandledLookup;