om-memory-system 5.2.0 → 5.3.0

package/dist/adapters/opencode/in-process-client.d.ts

@@ -0,0 +1,2 @@
+export declare function markInProcessClient<T extends object>(client: T): T;
+export declare function isInProcessClient(client: unknown): boolean;

package/dist/adapters/opencode/in-process-client.js

@@ -0,0 +1,16 @@
+/**
+ * Marks a host client that talks to OpenCode in-process. The V2 legacy client
+ * is one: it calls the plugin context directly, so it has no HTTP transport
+ * and no fetch to borrow. The transport setup skips its "no host fetch"
+ * warning for it.
+ */
+const IN_PROCESS_CLIENT = Symbol.for("omms.opencode.inProcessClient");
+export function markInProcessClient(client) {
+    Object.defineProperty(client, IN_PROCESS_CLIENT, { value: true });
+    return client;
+}
+export function isInProcessClient(client) {
+    return (typeof client === "object" &&
+        client !== null &&
+        client[IN_PROCESS_CLIENT] === true);
+}

package/dist/adapters/opencode/opencode-host-config.d.ts

@@ -2,6 +2,9 @@ export type HostClientConfig = {
     readonly baseUrl: string | undefined;
     readonly fetch: typeof fetch | undefined;
     readonly headers?: RequestInit["headers"];
+    /** Present when the host client runs in-process (OpenCode v2): no HTTP
+     * transport, so a missing fetch is expected and not worth a warning. */
+    readonly inProcess?: true;
     readonly clientKeys: readonly string[];
     readonly sdkConfigCount: number;
 };

package/dist/adapters/opencode/opencode-host-config.js

@@ -1,3 +1,4 @@
+import { isInProcessClient } from "./in-process-client.js";
 export function getHostClientConfig(ctx) {
     const client = toRecord(ctx.client);
     if (!client) {
@@ -11,6 +12,7 @@ export function getHostClientConfig(ctx) {
         baseUrl: typeof baseUrl === "string" ? baseUrl : undefined,
         fetch: isFetch(customFetch) ? customFetch : undefined,
         ...(isHeadersInit(headers) ? { headers } : {}),
+        ...(isInProcessClient(client) ? { inProcess: true } : {}),
         clientKeys: Object.keys(client),
         sdkConfigCount: configs.length,
     };

package/dist/adapters/opencode/tui-recall.d.ts

@@ -0,0 +1,68 @@
+import { type RecallSummary } from "../../core/recall-summary.js";
+/**
+ * Fold rows for OMMS recalls in the OpenCode terminal TUI (TDR-060). The TUI entry (`opencode/tui.tsx`)
+ * passes its host pieces in, so this module never imports a host SDK, storage, capture or config. Every
+ * host call sits behind a guard: a missing RPC handler, a failing storage call or a throwing render
+ * draws nothing for that part and never throws into the host. Nothing here logs: memory text must not
+ * reach a log.
+ */
+/** Most rows shown at once. Older summaries of the session are dropped from the view. */
+export declare const MAX_ROWS = 5;
+/** Persisted fold state: only open rows are stored, keyed by `rowKey`. */
+export interface FoldState {
+    open: Record<string, boolean>;
+}
+/** The slice of `ctx.client.rpc(recallRpc)` this module uses. */
+export interface RecallRpcClient {
+    list(input: {
+        sessionID: string;
+    }): Promise<{
+        enabled: boolean;
+        items: RecallSummary[];
+    }>;
+    events: {
+        on(name: string, handler: (event: any) => void): unknown;
+    };
+}
+export interface TuiRecallDeps {
+    client: {
+        rpc(definition: unknown): RecallRpcClient;
+    };
+    data: {
+        listen(handler: (event: any) => void): unknown;
+    };
+    ui: {
+        slot(spec: {
+            append: string;
+            render: (input: {
+                sessionID?: string;
+            }) => unknown;
+        }): unknown;
+    };
+    keymap: {
+        layer(factory: () => unknown): unknown;
+    };
+    storage: {
+        store(key: string, options: {
+            initial: FoldState;
+        }): [FoldState, (change: (draft: FoldState) => void) => unknown];
+    };
+    createSignal: <T>(value: T) => [() => T, (next: T | ((previous: T) => T)) => unknown];
+    jsx: (tag: string, props: Record<string, unknown>) => unknown;
+    /** Theme colour for row text; read on each paint. */
+    muted: () => unknown;
+    /** Quiet period that collapses a burst of host events into one `list` call. */
+    debounceMs?: number;
+}
+/**
+ * Stable per-summary key. A summary with a server sequence number is keyed by session id and that
+ * number, so it survives the ring dropping older entries. An older server sends none, so the key
+ * falls back to the session id, a hash of mode, query and each memory's id (or headline when it
+ * has none), and a counter for identical summaries.
+ */
+export declare function rowKeys(sessionID: string, summaries: RecallSummary[]): string[];
+/**
+ * Draw one fold row per recall summary of the session in view. Returns a stop function that removes
+ * every hook it added.
+ */
+export declare function startTuiRecall(deps: TuiRecallDeps): () => void;

package/dist/adapters/opencode/tui-recall.js

@@ -0,0 +1,302 @@
+import { callTitle, unfoldLines } from "../../core/recall-summary.js";
+import { recallRpc } from "../../v2/recall-rpc.js";
+/**
+ * Fold rows for OMMS recalls in the OpenCode terminal TUI (TDR-060). The TUI entry (`opencode/tui.tsx`)
+ * passes its host pieces in, so this module never imports a host SDK, storage, capture or config. Every
+ * host call sits behind a guard: a missing RPC handler, a failing storage call or a throwing render
+ * draws nothing for that part and never throws into the host. Nothing here logs: memory text must not
+ * reach a log.
+ */
+/** Most rows shown at once. Older summaries of the session are dropped from the view. */
+export const MAX_ROWS = 5;
+const FOLD_STORE_KEY = "omms.recall.fold";
+const SLOT = "session.composer.top";
+const DEFAULT_DEBOUNCE_MS = 150;
+/**
+ * Host events that pull the recall list. The server records a recall when a model step starts (the
+ * `context` hook), so only the events around a step, plus session creation and view changes, can
+ * coincide with a new recall. The end events catch a missed `recalled` push right after the step.
+ * The failure events are end events too: the summary is recorded in the `context` hook after
+ * `session.step.started` fires, so if the push is absent and the step then fails, no later event
+ * in this set would pull it.
+ * Streaming, part, message, inbox and instruction events fire dozens of times per reply and never
+ * change the list, so they must stay out of this set.
+ */
+const PULL_EVENT_TYPES = new Set([
+    "session.created",
+    "session.viewed",
+    "session.execution.started",
+    "session.step.started",
+    "session.step.ended",
+    "session.execution.succeeded",
+    "session.step.failed",
+    "session.execution.failed",
+]);
+function isSummary(value) {
+    const v = value;
+    return (!!v &&
+        (v.mode === "search" || v.mode === "recent") &&
+        Array.isArray(v.items) &&
+        v.items.every((item) => !!item && typeof item.headline === "string"));
+}
+/** FNV-1a, 32 bit: short, deterministic, no crypto needed for a fold key. */
+function hash(text) {
+    let h = 0x811c9dc5;
+    for (let i = 0; i < text.length; i++) {
+        h ^= text.charCodeAt(i);
+        h = Math.imul(h, 0x01000193);
+    }
+    return (h >>> 0).toString(16);
+}
+/**
+ * Stable per-summary key. A summary with a server sequence number is keyed by session id and that
+ * number, so it survives the ring dropping older entries. An older server sends none, so the key
+ * falls back to the session id, a hash of mode, query and each memory's id (or headline when it
+ * has none), and a counter for identical summaries.
+ */
+export function rowKeys(sessionID, summaries) {
+    const seen = new Map();
+    return summaries.map((summary) => {
+        if (typeof summary.seq === "number")
+            return `${sessionID}:${summary.seq}`;
+        const body = [
+            summary.mode,
+            summary.query ?? "",
+            ...summary.items.map((item) => item.id ?? item.headline),
+        ].join("\u0000");
+        const base = `${sessionID}:${hash(body)}`;
+        const count = seen.get(base) ?? 0;
+        seen.set(base, count + 1);
+        return count === 0 ? base : `${base}#${count}`;
+    });
+}
+function guard(fn) {
+    try {
+        return fn();
+    }
+    catch {
+        return undefined;
+    }
+}
+function release(handle) {
+    if (typeof handle === "function")
+        guard(() => handle());
+}
+/**
+ * Draw one fold row per recall summary of the session in view. Returns a stop function that removes
+ * every hook it added.
+ */
+export function startTuiRecall(deps) {
+    const debounceMs = deps.debounceMs ?? DEFAULT_DEBOUNCE_MS;
+    const sessions = new Map();
+    const local = new Map();
+    let lastViewed;
+    let stopped = false;
+    let layerHandle;
+    // Two signals: the list changes when summaries arrive, the text changes when a row folds.
+    const [listRevision, setListRevision] = deps.createSignal(0);
+    const [foldRevision, setFoldRevision] = deps.createSignal(0);
+    const repaintList = () => void guard(() => setListRevision((n) => n + 1));
+    const repaintFold = () => void guard(() => setFoldRevision((n) => n + 1));
+    const rpc = guard(() => deps.client.rpc(recallRpc));
+    const fold = guard(() => deps.storage.store(FOLD_STORE_KEY, { initial: { open: {} } }));
+    const sessionState = (id) => {
+        let state = sessions.get(id);
+        if (!state) {
+            state = { enabled: false, summaries: [], inflight: false, dirty: false };
+            sessions.set(id, state);
+        }
+        return state;
+    };
+    const apply = (id, enabled, summaries) => {
+        const state = sessionState(id);
+        state.enabled = enabled;
+        state.summaries = summaries;
+        repaintList();
+    };
+    async function pull(id) {
+        const state = sessionState(id);
+        state.inflight = true;
+        try {
+            const reply = await rpc.list({ sessionID: id });
+            if (stopped)
+                return;
+            const ok = reply?.enabled === true && Array.isArray(reply.items);
+            apply(id, ok, ok ? reply.items.filter(isSummary) : []);
+        }
+        catch {
+            if (!stopped)
+                apply(id, false, []);
+        }
+        finally {
+            state.inflight = false;
+            if (state.dirty && !stopped) {
+                state.dirty = false;
+                request(id, debounceMs);
+            }
+        }
+    }
+    /** At most one `list` per session per window; a request during a call asks for one more after it. */
+    function request(id, delayMs) {
+        if (stopped || !rpc)
+            return;
+        const state = sessionState(id);
+        if (state.timer)
+            return;
+        if (state.inflight) {
+            state.dirty = true;
+            return;
+        }
+        state.timer = setTimeout(() => {
+            state.timer = undefined;
+            void pull(id);
+        }, delayMs);
+        state.timer.unref?.();
+    }
+    const isOpen = (key) => local.get(key) ?? guard(() => fold?.[0]?.open?.[key] === true) ?? false;
+    const persist = (changes) => {
+        if (!fold)
+            return;
+        guard(() => {
+            const done = fold[1]((draft) => {
+                draft.open ??= {};
+                for (const [key, open] of changes) {
+                    if (open)
+                        draft.open[key] = true;
+                    else
+                        delete draft.open[key];
+                }
+            });
+            Promise.resolve(done).catch(() => undefined);
+        });
+    };
+    const rowsFor = (id) => {
+        const state = sessions.get(id);
+        if (!state?.enabled || state.summaries.length === 0)
+            return [];
+        const keys = rowKeys(id, state.summaries);
+        return state.summaries
+            .map((summary, index) => ({ key: keys[index], summary }))
+            .slice(-MAX_ROWS);
+    };
+    const toggle = (key) => {
+        const next = !isOpen(key);
+        local.set(key, next);
+        persist([[key, next]]);
+        repaintFold();
+    };
+    const toggleAll = () => {
+        if (!lastViewed)
+            return;
+        const rows = rowsFor(lastViewed);
+        const target = rows.some((row) => !isOpen(row.key));
+        for (const row of rows)
+            local.set(row.key, target);
+        persist(rows.map((row) => [row.key, target]));
+        repaintFold();
+    };
+    // The host owns a keymap layer through the component that created it and removes it when that
+    // component is disposed. A slot remount therefore needs a fresh layer, so this runs once per render.
+    const registerCommand = () => {
+        layerHandle = guard(() => deps.keymap.layer(() => ({
+            mode: "global",
+            commands: [
+                {
+                    id: "omms.recall.fold",
+                    title: "Toggle OMMS recall rows",
+                    palette: true,
+                    slash: { name: "omms-recall" },
+                    run: () => void guard(toggleAll),
+                },
+            ],
+        })));
+    };
+    const drawRow = ({ key, summary }) => deps.jsx("box", {
+        flexDirection: "column",
+        onMouseDown: () => void guard(() => toggle(key)),
+        get children() {
+            return [
+                deps.jsx("text", {
+                    get fg() {
+                        return guard(deps.muted);
+                    },
+                    get children() {
+                        foldRevision();
+                        return guard(() => `${isOpen(key) ? "▾" : "▸"} ● ${callTitle(summary)}`) ?? "";
+                    },
+                }),
+                deps.jsx("text", {
+                    get visible() {
+                        foldRevision();
+                        return isOpen(key);
+                    },
+                    get fg() {
+                        return guard(deps.muted);
+                    },
+                    get children() {
+                        foldRevision();
+                        if (!isOpen(key))
+                            return "";
+                        return (guard(() => unfoldLines(summary)
+                            .map((line) => `  ${line}`)
+                            .join("\n")) ?? "");
+                    },
+                }),
+            ];
+        },
+    });
+    const render = (input) => {
+        try {
+            // The host passes slot input as reactive props and may keep this claim mounted while the
+            // session in view changes, so the id is read on each paint. `tracked` makes the follow-up
+            // idempotent: a repaint with the same id asks for no new pull.
+            let tracked;
+            const follow = () => {
+                const id = guard(() => input?.sessionID);
+                if (id && id !== tracked) {
+                    tracked = id;
+                    lastViewed = id;
+                    request(id, 0);
+                }
+                return id;
+            };
+            follow();
+            registerCommand();
+            return deps.jsx("box", {
+                flexDirection: "column",
+                get children() {
+                    listRevision();
+                    const id = follow();
+                    return (id ? guard(() => rowsFor(id).map(drawRow)) : undefined) ?? [];
+                },
+            });
+        }
+        catch {
+            return null;
+        }
+    };
+    const removeSlot = guard(() => deps.ui.slot({ append: SLOT, render }));
+    const offPush = guard(() => rpc?.events.on("recalled", (event) => {
+        const id = event?.data?.sessionID;
+        if (typeof id === "string" && sessions.has(id))
+            request(id, debounceMs);
+    }));
+    // Only sessions the slot has drawn are refreshed, and an event without a session id is no hint.
+    const offData = guard(() => deps.data.listen((event) => {
+        const details = event?.details;
+        if (typeof details?.type !== "string" || !PULL_EVENT_TYPES.has(details.type))
+            return;
+        const id = details.data?.sessionID ?? details.properties?.sessionID;
+        if (typeof id === "string" && sessions.has(id))
+            request(id, debounceMs);
+    }));
+    return () => {
+        stopped = true;
+        for (const state of sessions.values())
+            clearTimeout(state.timer);
+        release(offData);
+        release(offPush);
+        release(removeSlot);
+        release(layerHandle);
+    };
+}

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

package/dist/config-template.js

@@ -254,6 +254,11 @@ export const CONFIG_TEMPLATE = `{
   // only. Unset: ~/.agents/skills and the Claude Code folder's skills folder.
   // "skillFolders": ["~/.agents/skills"],
 
+  // Visible recall: the OpenCode tui entrypoint draws a row in the session for
+  // each memory recall. On by default. A project config can turn it off but
+  // never on.
+  // "visibleRecall": true,
+
   // Capture retry queue: when a live capture fails because the model cannot be
   // reached (network error, timeout, HTTP 408, 429 or 5xx), OMMS keeps a copy of
   // the turn in ~/.omms/data and tries it again later. The copy has <private>

package/dist/config.d.ts

@@ -188,6 +188,7 @@ declare function buildConfig(fileConfig: OmmsConfig): {
         maxAgeDays: number | undefined;
         injectOn: "first" | "always";
     };
+    visibleRecall: boolean;
     skillLearning: boolean;
     skillLearningRedact: boolean;
     skillFolders: string[] | undefined;
@@ -298,6 +299,7 @@ export declare let CONFIG: {
         maxAgeDays: number | undefined;
         injectOn: "first" | "always";
     };
+    visibleRecall: boolean;
     skillLearning: boolean;
     skillLearningRedact: boolean;
     skillFolders: string[] | undefined;

package/dist/config.js

@@ -10,6 +10,7 @@ import { getAutoCaptureProviderStatus } from "./services/ai/live-model-choice.js
 import { parseBackfillModel } from "./importer/backfill-model.js";
 import { parseIgnoredDirectories, parseImportPathMaps } from "./importer/import-path-maps.js";
 import { normaliseSkillConfig, projectSkillOverride, } from "./services/skill-learning/skill-config.js";
+import * as visibleRecall from "./services/visible-recall-config.js";
 import { checkedMemoryLimit, MEMORY_LIMITS } from "./utils/memory-limits.js";
 import { defaultSimilarityThreshold } from "./utils/threshold-default.js";
 import { RERANKER_TIMEOUT_DEFAULT_MS } from "./services/request-timeouts.js";
@@ -371,6 +372,7 @@ function buildConfig(fileConfig) {
         captureAttemptRetentionDays: normalizeAutoCleanupRetentionDays(fileConfig.captureAttemptRetentionDays ?? DEFAULTS.captureAttemptRetentionDays),
         captureRetryRetentionHours: normalizeCaptureRetryRetentionHours(fileConfig.captureRetryRetentionHours),
         ...normaliseSkillConfig(fileConfig),
+        ...visibleRecall.visibleRecallFromFile(fileConfig),
         autoBackfill: fileConfig.autoBackfill ?? DEFAULTS.autoBackfill,
         opencodeBackfillModel: fileConfig.opencodeBackfillModel ?? DEFAULTS.opencodeBackfillModel,
         piBackfillModel: fileConfig.piBackfillModel ?? DEFAULTS.piBackfillModel,
@@ -543,6 +545,7 @@ export function initConfig(directory, options = {}) {
     delete projectOverrides.skillLearningRedact;
     delete projectOverrides.skillFolders;
     Object.assign(projectOverrides, skillRule.override);
+    visibleRecall.applyProjectVisibleRecall(projectConfig, projectOverrides, directory, log);
     delete projectOverrides.autoBackfill;
     delete projectOverrides.opencodeBackfillModel;
     delete projectOverrides.piBackfillModel;

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

@@ -0,0 +1,28 @@
+/** One memory a recall injected, reduced to what a fold row shows. */
+export type RecallItem = {
+    relevance: number;
+    id?: string;
+    headline: string;
+};
+/** One recall as OMMS made it: a prompt search, or the recent memories a session opens with. */
+export type RecallSummary = {
+    mode: "search" | "recent";
+    query: string | null;
+    items: RecallItem[];
+    /** Per-session counter the v2 reporter sets when it stores the summary; the TUI keys rows by it. */
+    seq?: number;
+};
+/** Query on one line, cut to `max` characters. */
+export declare function queryLine(query: string, max?: number): string;
+/** One line that names a memory: the first line under `## Request`, else the first line. */
+export declare function headline(body: string): string;
+/**
+ * Summarise a formatted retrieval section for a host to show. Returns null when the section holds
+ * no memory (a profile-only section included), so the host draws nothing. A `recent` recall has
+ * no query, whatever the caller passes.
+ */
+export declare function buildRecallSummary(section: string, mode: "search" | "recent", query: string | null): RecallSummary | null;
+/** `memory [mode=search, query="...", limit=4]`: the call's name and arguments, as a tool row shows them. */
+export declare function callTitle(summary: RecallSummary): string;
+/** Body of an unfolded row, one string per line, without indentation (the host indents). */
+export declare function unfoldLines(summary: RecallSummary): string[];

package/dist/core/recall-summary.js

@@ -0,0 +1,71 @@
+const QUERY_CUT = 60;
+const UNFOLDED_QUERY_CUT = 300;
+const HEADLINE_CUT = 140;
+/** Query on one line, cut to `max` characters. */
+export function queryLine(query, max = QUERY_CUT) {
+    const line = query.replace(/\s+/g, " ").trim();
+    return line.length > max ? `${line.slice(0, max - 3)}...` : line;
+}
+// Entry shape comes from `formatContextForPrompt`: `<memory id="..." relevance="N%">`, with the
+// id missing when the result had none. A body ends at its own `</memory>` when another entry, a
+// section close or the end of the text follows it. The Claude Code hook can cut a block short and
+// leave tags open, so a body also ends at a section close or at the end of the text. An opening
+// tag inside the text, even at the start of a line, does not end it: a memory can quote the tag.
+const SECTION_CLOSE = String.raw `<\/project_knowledge>|<\/memory_context>|<\/omms-retrieval>`;
+const MEMORY_ENTRY = new RegExp(String.raw `<memory(?:\s+id="([^"]*)")?\s+relevance="(\d+)%">([\s\S]*?)` +
+    String.raw `(?=<\/memory>\s*(?:<memory\s|${SECTION_CLOSE}|$)|${SECTION_CLOSE}|$)`, "g");
+/** One line that names a memory: the first line under `## Request`, else the first line. */
+export function headline(body) {
+    const lines = body
+        .replace(/<\/memory>\s*$/, "")
+        .split("\n")
+        .map((line) => line.trim())
+        .filter(Boolean);
+    const heading = lines.findIndex((line) => /^##\s*request\s*$/i.test(line));
+    const line = (heading >= 0 ? lines[heading + 1] : lines[0]) ?? "";
+    return line.length > HEADLINE_CUT ? `${line.slice(0, HEADLINE_CUT - 3)}...` : line;
+}
+/**
+ * Summarise a formatted retrieval section for a host to show. Returns null when the section holds
+ * no memory (a profile-only section included), so the host draws nothing. A `recent` recall has
+ * no query, whatever the caller passes.
+ */
+export function buildRecallSummary(section, mode, query) {
+    const items = [];
+    for (const match of section.matchAll(MEMORY_ENTRY)) {
+        const text = headline(match[3] ?? "");
+        if (!text)
+            continue;
+        const item = { relevance: Number(match[2]), headline: text };
+        if (match[1])
+            item.id = match[1];
+        items.push(item);
+    }
+    if (items.length === 0)
+        return null;
+    return { mode, query: mode === "recent" ? null : query, items };
+}
+/** Escape backslashes and double quotes so the quoted value in the title reads as one string. */
+function escapeQuoted(text) {
+    return text.replace(/[\\"]/g, "\\$&");
+}
+/** `memory [mode=search, query="...", limit=4]`: the call's name and arguments, as a tool row shows them. */
+export function callTitle(summary) {
+    const args = [`mode=${summary.mode}`];
+    if (summary.query)
+        args.push(`query="${escapeQuoted(queryLine(summary.query))}"`);
+    args.push(`limit=${summary.items.length}`);
+    return `memory [${args.join(", ")}]`;
+}
+/** Body of an unfolded row, one string per line, without indentation (the host indents). */
+export function unfoldLines(summary) {
+    const lines = [`mode: ${summary.mode}`];
+    if (summary.query)
+        lines.push(`query: ${queryLine(summary.query, UNFOLDED_QUERY_CUT)}`);
+    lines.push(`limit: ${summary.items.length}`);
+    summary.items.forEach((item, index) => {
+        const id = item.id ? ` id=${item.id}` : "";
+        lines.push(`${index + 1}. ${item.relevance}%${id} ${item.headline}`);
+    });
+    return lines;
+}

package/dist/index.js

@@ -161,7 +161,7 @@ export async function configureOpencodeHostTransport(ctx) {
     if (hostConfig.fetch) {
         setHostFetch(hostConfig.fetch);
     }
-    else {
+    else if (!hostConfig.inProcess) {
         log("OpenCode host fetch unavailable; falling back to global fetch", {
             clientKeys: hostConfig.clientKeys,
             sdkConfigCount: hostConfig.sdkConfigCount,

package/dist/services/single-flight.d.ts

@@ -0,0 +1,2 @@
+/** Like setInterval, but skips a tick while the previous run of the task is still pending. */
+export declare function setSingleFlightInterval(task: () => Promise<void>, ms: number): NodeJS.Timeout;

package/dist/services/single-flight.js

@@ -0,0 +1,12 @@
+/** Like setInterval, but skips a tick while the previous run of the task is still pending. */
+export function setSingleFlightInterval(task, ms) {
+    let running = false;
+    return setInterval(() => {
+        if (running)
+            return;
+        running = true;
+        void task().finally(() => {
+            running = false;
+        });
+    }, ms);
+}

package/dist/services/visible-recall-config.d.ts

@@ -0,0 +1,34 @@
+/**
+ * The `visibleRecall` setting (change add-opencode-v2-visible-recall), kept out
+ * of `src/config.ts` to save its lines. Pure on purpose: callers pass the file
+ * config and a log function, and the loaded `CONFIG` is never imported here,
+ * because many tests stub `src/config.js`.
+ *
+ * Read the effective value from `CONFIG.visibleRecall`. It is on by default,
+ * and only an explicit `false` turns it off.
+ */
+/**
+ * The key as it appears in a config file. `OmmsConfig` in `src/config.ts` has
+ * no field for it, to save lines, so the functions take `FileConfigWithKey`,
+ * which any config object satisfies. A hand-edited file can hold any value,
+ * which is why they compare with `=== false`.
+ */
+export interface VisibleRecallFileKey {
+    /** Draw a row in the OpenCode TUI for each memory recall. Default true. A project can only turn it off. */
+    visibleRecall?: boolean;
+}
+type FileConfigWithKey = VisibleRecallFileKey & object;
+/** Read the setting from a merged file config. Unset means on. */
+export declare function visibleRecallFromFile(fileConfig: FileConfigWithKey): {
+    visibleRecall: boolean;
+};
+/**
+ * Apply the project rule to `overrides`, the copy of the project config that
+ * is merged over the global one. A project may turn the rows off, so a value
+ * of `false` stays. Any other project value is dropped, so the global value
+ * decides. A project value of `true` is also reported through `log`.
+ */
+export declare function applyProjectVisibleRecall(project: FileConfigWithKey, overrides: FileConfigWithKey, directory: string, log: (message: string, data: {
+    directory: string;
+}) => void): void;
+export {};

package/dist/services/visible-recall-config.js

@@ -0,0 +1,27 @@
+/**
+ * The `visibleRecall` setting (change add-opencode-v2-visible-recall), kept out
+ * of `src/config.ts` to save its lines. Pure on purpose: callers pass the file
+ * config and a log function, and the loaded `CONFIG` is never imported here,
+ * because many tests stub `src/config.js`.
+ *
+ * Read the effective value from `CONFIG.visibleRecall`. It is on by default,
+ * and only an explicit `false` turns it off.
+ */
+/** Read the setting from a merged file config. Unset means on. */
+export function visibleRecallFromFile(fileConfig) {
+    return { visibleRecall: fileConfig.visibleRecall !== false };
+}
+/**
+ * Apply the project rule to `overrides`, the copy of the project config that
+ * is merged over the global one. A project may turn the rows off, so a value
+ * of `false` stays. Any other project value is dropped, so the global value
+ * decides. A project value of `true` is also reported through `log`.
+ */
+export function applyProjectVisibleRecall(project, overrides, directory, log) {
+    if (project.visibleRecall === false)
+        return;
+    delete overrides.visibleRecall;
+    if (project.visibleRecall === true) {
+        log("Project config cannot turn on visibleRecall; the value was ignored", { directory });
+    }
+}

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

@@ -48,6 +48,7 @@ export declare class WebServer {
     private holdOffTimer;
     private portsExhaustedNotified;
     private takeoverFailures;
+    private lastHealthFailure;
     private readonly maxFallbackPort;
     private settingsImportJobs?;
     private backfillControlsInstance?;
@@ -99,6 +100,8 @@ export declare class WebServer {
     isRunning(): boolean;
     isServerOwner(): boolean;
     getUrl(): string;
+    /** Why the last `checkServerAvailable` call said no, or null when it said yes. */
+    getLastHealthFailure(): string | null;
     checkServerAvailable(): Promise<boolean>;
     private handleRequest;
     private contentTypeFor;

package/dist/services/web-server.js

@@ -13,10 +13,13 @@ import { AUTH_HEADER, getOrCreateAuthToken, isAuthorizedApiRequest } from "./aut
 import { WebAuth } from "./web-auth.js";
 import { NODE_HTTP_IDLE_TIMEOUT_MS } from "./request-timeouts.js";
 import { testReranker } from "./reranker-test.js";
+import { setSingleFlightInterval } from "./single-flight.js";
 import { handleListTags, handleListKeywords, handleListMemories, handleAddMemory, handleDeleteMemory, handleBulkDelete, handleUpdateMemory, handleSearch, handleStats, handlePinMemory, handleUnpinMemory, handleRunCleanup, handleRunDeduplication, handleDetectMigration, handleRunMigration, handleDetectTagMigration, handleRunTagMigrationBatch, handleGetTagMigrationProgress, handleDeletePrompt, handleBulkDeletePrompts, handleGetUserProfile, handleGetProfileChangelog, handleGetProfileSnapshot, handleRefreshProfile, handleUpdateProfileItem, } from "./api-handlers.js";
 import { routeModules } from "./web-route-modules.js";
 const isBun = typeof globalThis.Bun !== "undefined";
 const MIN_FAILED_TAKEOVERS = 3;
+/** A busy owner can answer late. Two seconds was short enough to call a live owner dead. */
+const HEALTH_CHECK_TIMEOUT_MS = 8_000;
 export function attachNodeDisconnectHandlers(req, res, onDisconnect) {
     let disconnected = false;
     const disconnectOnce = () => {
@@ -175,6 +178,7 @@ export class WebServer {
     holdOffTimer = null;
     portsExhaustedNotified = false;
     takeoverFailures = 0;
+    lastHealthFailure = null;
     maxFallbackPort;
     settingsImportJobs;
     backfillControlsInstance;
@@ -584,7 +588,7 @@ export class WebServer {
         if (this.healthCheckInterval) {
             return;
         }
-        this.healthCheckInterval = setInterval(() => void this.waiterTick(), 5000);
+        this.healthCheckInterval = setSingleFlightInterval(() => this.waiterTick(), 5000);
     }
     /** One check of a web app that waits for the port. */
     async waiterTick() {
@@ -667,6 +671,8 @@ export class WebServer {
             log("Web server port held by a non-responsive process; falling back to next port", {
                 previousPort: this.config.port,
                 newPort: nextPort,
+                lastFailure: this.lastHealthFailure,
+                healthTimeoutMs: HEALTH_CHECK_TIMEOUT_MS,
             }, "warn");
             this.config.port = nextPort;
         }
@@ -743,6 +749,10 @@ export class WebServer {
     getUrl() {
         return webServerUrl(this.config.host, this.config.port);
     }
+    /** Why the last `checkServerAvailable` call said no, or null when it said yes. */
+    getLastHealthFailure() {
+        return this.lastHealthFailure;
+    }
     async checkServerAvailable() {
         try {
             // Only a web app on this machine answers here, so the local token file is enough.
@@ -751,20 +761,28 @@ export class WebServer {
             const response = await fetch(`${this.getUrl()}${endpoint}`, {
                 method: "GET",
                 headers,
-                signal: AbortSignal.timeout(2000),
+                signal: AbortSignal.timeout(HEALTH_CHECK_TIMEOUT_MS),
             });
-            if (!response.ok)
+            if (!response.ok) {
+                this.lastHealthFailure = `http ${response.status}`;
                 return false;
+            }
             // Fallback spans 10 neighbor ports; any 2xx from an unrelated local
             // service must not be mistaken for an omms owner. Require the
             // response to carry our API envelope.
             const body = (await response.json());
-            if (endpoint === "/api/health") {
-                return body.success === true && body.status === "ok";
-            }
-            return body.success === true;
+            const ok = body.success === true && (endpoint !== "/api/health" || body.status === "ok");
+            this.lastHealthFailure = ok ? null : "unexpected response";
+            return ok;
         }
-        catch {
+        catch (error) {
+            const name = error instanceof Error ? error.name : "";
+            this.lastHealthFailure =
+                name === "TimeoutError"
+                    ? "timeout"
+                    : error instanceof Error
+                        ? error.message
+                        : String(error);
             return false;
         }
     }

package/dist/v2/adapter.js

@@ -4,6 +4,7 @@ import { SKILLS_TOOL_INPUT_SCHEMA } from "../core/skills-tool.js";
 import { log } from "../services/logger.js";
 import { utf8ByteLength } from "../utils/context-limit.js";
 import { eventBelongsToLocation, legacyToolResult, toLegacyEvent } from "./legacy-client.js";
+import { createRecallReporter } from "./recall-reporter.js";
 const memoryInput = {
     type: "object",
     properties: {
@@ -75,7 +76,9 @@ export async function registerV2Adapter(ctx, legacy, memory) {
     // until their message appears in a model request's history (the context
     // hook), and are only then recorded for auto-capture and profile learning.
     const pendingPrompts = new Map();
+    const recall = await createRecallReporter(ctx, memory, setBounded);
     const forgetSession = (sessionID) => {
+        recall.forget(sessionID);
         retrievals.delete(sessionID);
         restored.delete(sessionID);
         messageIDs.delete(sessionID);
@@ -186,7 +189,7 @@ export async function registerV2Adapter(ctx, legacy, memory) {
             log("v2 prompt: memory retrieval failed", { sessionID, error: String(error) }, "error");
             return null;
         });
-        setBounded(retrievals, sessionID, { promise, budgetBytes });
+        setBounded(retrievals, sessionID, { promise, budgetBytes, query: text });
     });
     await ctx.session.hook("context", async (event) => {
         const sessionID = event.sessionID;
@@ -201,6 +204,7 @@ export async function registerV2Adapter(ctx, legacy, memory) {
             const section = await settledRetrieval(retrieval);
             if (section) {
                 event.system.push({ type: "text", text: section });
+                recall.record(sessionID, retrieval, retrieval.query, section);
                 remainingBytes = retrieval.budgetBytes - utf8ByteLength(section);
             }
             else {
@@ -266,6 +270,7 @@ export async function registerV2Adapter(ctx, legacy, memory) {
         restored.clear();
         messageIDs.clear();
         pendingPrompts.clear();
+        recall.clear();
         await legacy.dispose?.();
     };
 }

package/dist/v2/legacy-client.js

@@ -1,5 +1,6 @@
 import { randomUUID } from "node:crypto";
 import { resolve } from "node:path";
+import { markInProcessClient } from "../adapters/opencode/in-process-client.js";
 function textFromParts(parts = []) {
     return parts
         .filter((part) => part.type === "text")
@@ -216,7 +217,7 @@ export function createLegacyClient(ctx) {
             },
         },
     };
-    return client;
+    return markInProcessClient(client);
 }
 export function legacyToolResult(value) {
     if (typeof value === "string")

package/dist/v2/memory-bridge.d.ts

@@ -8,6 +8,8 @@ export interface V2MemoryBridge {
     isConfigured(): boolean;
     isInjectionEnabled(): boolean;
     isCompactionEnabled(): boolean;
+    /** Effective `visibleRecall` setting for this project, from a fresh config read. */
+    isVisibleRecallEnabled(): boolean;
     isInternalPrompt(sessionID: string, text: string): boolean;
     recordPrompt(sessionID: string, messageID: string, text: string): Promise<void>;
     /**

package/dist/v2/memory-bridge.js

@@ -11,6 +11,10 @@ export function createV2MemoryBridge(directory) {
         isConfigured,
         isInjectionEnabled: () => isConfigured() && Boolean(CONFIG.chatMessage.enabled),
         isCompactionEnabled: () => isConfigured() && Boolean(CONFIG.compaction.enabled),
+        isVisibleRecallEnabled() {
+            refreshConfigIfChanged(directory);
+            return CONFIG.visibleRecall !== false;
+        },
         isInternalPrompt,
         async recordPrompt(sessionID, messageID, text) {
             await recordUserPrompt(sessionID, messageID, directory, text);

package/dist/v2/recall-reporter.d.ts

@@ -0,0 +1,24 @@
+import type { Context } from "@opencode/plugin/promise/plugin";
+import type { V2MemoryBridge } from "./memory-bridge.js";
+/** Latest summaries kept per session. The TUI shows recent rows, so older ones are dropped. */
+export declare const MAX_RECALL_SUMMARIES = 20;
+/** Longest query kept per summary. The TUI shows at most 300 characters of it. */
+export declare const MAX_RECALL_QUERY_CHARS = 300;
+type Bound = <V>(map: Map<string, V>, key: string, value: V) => void;
+export interface RecallReporter {
+    /**
+     * Record the summary of one prompt retrieval. `token` identifies the retrieval: model steps
+     * that reuse it record once. A section with no memory entries records nothing, and neither
+     * does a subagent session (one with a parent). The call does not wait for the session lookup.
+     */
+    record(sessionID: string, token: object, query: string, section: string | null): void;
+    forget(sessionID: string): void;
+    clear(): void;
+}
+/**
+ * Keeps a bounded per-session ring of recall summaries and serves it over plugin RPC to the TUI
+ * entrypoint. It never throws and never touches what is injected. Logs carry the mode, count and
+ * session id only, never memory text or the query.
+ */
+export declare function createRecallReporter(ctx: Context, memory: V2MemoryBridge, bound: Bound): Promise<RecallReporter>;
+export {};

package/dist/v2/recall-reporter.js

@@ -0,0 +1,128 @@
+import { buildRecallSummary } from "../core/recall-summary.js";
+import { log } from "../services/logger.js";
+import { recallRpc } from "./recall-rpc.js";
+/** Latest summaries kept per session. The TUI shows recent rows, so older ones are dropped. */
+export const MAX_RECALL_SUMMARIES = 20;
+/** Longest query kept per summary. The TUI shows at most 300 characters of it. */
+export const MAX_RECALL_QUERY_CHARS = 300;
+/**
+ * Keeps a bounded per-session ring of recall summaries and serves it over plugin RPC to the TUI
+ * entrypoint. It never throws and never touches what is injected. Logs carry the mode, count and
+ * session id only, never memory text or the query.
+ */
+export async function createRecallReporter(ctx, memory, bound) {
+    const rings = new Map();
+    // Last number given per session. It keeps rising when the ring drops old entries.
+    const sequences = new Map();
+    const recorded = new WeakSet();
+    // One lookup per session. The promise is cached so concurrent records keep their order.
+    const subagents = new Map();
+    // Records waiting on a session lookup. `forget` and `clear` cancel them, so a late lookup cannot
+    // store a summary for a session whose state is gone. Entries leave when their lookup settles.
+    const pending = new Map();
+    let emit;
+    try {
+        const registration = await ctx.rpc.register(recallRpc, {
+            // The contract is plain JSON schema, so the host types the input as unknown.
+            list: async (input) => ({
+                enabled: memory.isVisibleRecallEnabled(),
+                items: [...(rings.get(input.sessionID) ?? [])],
+            }),
+        });
+        emit = (sessionID) => registration.events.emit("recalled", { sessionID });
+    }
+    catch (error) {
+        log("v2 recall: rpc registration failed", { error: String(error) }, "error");
+    }
+    const notify = (sessionID) => {
+        try {
+            emit?.(sessionID).catch((error) => log("v2 recall: push failed", { sessionID, error: String(error) }, "error"));
+        }
+        catch (error) {
+            log("v2 recall: push failed", { sessionID, error: String(error) }, "error");
+        }
+    };
+    // A session with a parent is a subagent session. A failed lookup counts as a root session, so
+    // the summary still shows.
+    const isSubagent = (sessionID) => {
+        const known = subagents.get(sessionID);
+        if (known)
+            return known;
+        const lookup = (async () => {
+            try {
+                const info = await ctx.session.get({ sessionID });
+                return Boolean(info?.parentID);
+            }
+            catch (error) {
+                log("v2 recall: session lookup failed", { sessionID, error: String(error) }, "error");
+                return false;
+            }
+        })();
+        bound(subagents, sessionID, lookup);
+        return lookup;
+    };
+    const store = (sessionID, built) => {
+        const seq = (sequences.get(sessionID) ?? 0) + 1;
+        bound(sequences, sessionID, seq);
+        const summary = { ...built, seq };
+        const ring = [...(rings.get(sessionID) ?? []), summary].slice(-MAX_RECALL_SUMMARIES);
+        bound(rings, sessionID, ring);
+        log("v2 recall: summary recorded", {
+            sessionID,
+            mode: summary.mode,
+            count: summary.items.length,
+        });
+        notify(sessionID);
+    };
+    return {
+        record(sessionID, token, query, section) {
+            if (recorded.has(token))
+                return;
+            recorded.add(token);
+            try {
+                const summary = section
+                    ? buildRecallSummary(section, "search", query.slice(0, MAX_RECALL_QUERY_CHARS))
+                    : null;
+                if (!summary)
+                    return;
+                const ticket = { cancelled: false };
+                const tickets = pending.get(sessionID) ?? new Set();
+                tickets.add(ticket);
+                pending.set(sessionID, tickets);
+                void isSubagent(sessionID).then((subagent) => {
+                    tickets.delete(ticket);
+                    if (tickets.size === 0 && pending.get(sessionID) === tickets)
+                        pending.delete(sessionID);
+                    if (subagent || ticket.cancelled)
+                        return;
+                    try {
+                        store(sessionID, summary);
+                    }
+                    catch (error) {
+                        log("v2 recall: could not record summary", { sessionID, error: String(error) }, "error");
+                    }
+                });
+            }
+            catch (error) {
+                log("v2 recall: could not record summary", { sessionID, error: String(error) }, "error");
+            }
+        },
+        forget: (sessionID) => {
+            for (const ticket of pending.get(sessionID) ?? [])
+                ticket.cancelled = true;
+            pending.delete(sessionID);
+            rings.delete(sessionID);
+            sequences.delete(sessionID);
+            subagents.delete(sessionID);
+        },
+        clear: () => {
+            for (const tickets of pending.values())
+                for (const ticket of tickets)
+                    ticket.cancelled = true;
+            pending.clear();
+            rings.clear();
+            sequences.clear();
+            subagents.clear();
+        },
+    };
+}

package/dist/v2/recall-rpc.d.ts

@@ -0,0 +1,51 @@
+/**
+ * Plugin RPC contract between the OpenCode v2 server entrypoint (records recall summaries) and the
+ * terminal TUI entrypoint (draws them). Plain data on purpose: no host SDK import, so both sides and
+ * the tests can load it. The TUI pulls `list` and treats `recalled` as a refresh hint only; a missing
+ * push never breaks the rows (TDR-060).
+ */
+export declare const RECALL_RPC_ID = "omms-recall";
+export declare const recallRpc: {
+    readonly id: "omms-recall";
+    readonly methods: {
+        readonly list: {
+            readonly input: {
+                readonly type: "object";
+                readonly properties: {
+                    readonly sessionID: {
+                        readonly type: "string";
+                    };
+                };
+                readonly required: readonly ["sessionID"];
+            };
+            readonly output: {
+                readonly type: "object";
+                readonly properties: {
+                    readonly enabled: {
+                        readonly type: "boolean";
+                    };
+                    readonly items: {
+                        readonly type: "array";
+                        readonly items: {
+                            readonly type: "object";
+                        };
+                    };
+                };
+                readonly required: readonly ["enabled", "items"];
+            };
+        };
+    };
+    readonly events: {
+        readonly recalled: {
+            readonly schema: {
+                readonly type: "object";
+                readonly properties: {
+                    readonly sessionID: {
+                        readonly type: "string";
+                    };
+                };
+                readonly required: readonly ["sessionID"];
+            };
+        };
+    };
+};

package/dist/v2/recall-rpc.js

@@ -0,0 +1,38 @@
+/**
+ * Plugin RPC contract between the OpenCode v2 server entrypoint (records recall summaries) and the
+ * terminal TUI entrypoint (draws them). Plain data on purpose: no host SDK import, so both sides and
+ * the tests can load it. The TUI pulls `list` and treats `recalled` as a refresh hint only; a missing
+ * push never breaks the rows (TDR-060).
+ */
+export const RECALL_RPC_ID = "omms-recall";
+export const recallRpc = {
+    id: RECALL_RPC_ID,
+    methods: {
+        // `enabled` is the effective `visibleRecall` setting for the server's project, so the TUI never
+        // reads config. Summaries are recorded and served even when it is false.
+        list: {
+            input: {
+                type: "object",
+                properties: { sessionID: { type: "string" } },
+                required: ["sessionID"],
+            },
+            output: {
+                type: "object",
+                properties: {
+                    enabled: { type: "boolean" },
+                    items: { type: "array", items: { type: "object" } },
+                },
+                required: ["enabled", "items"],
+            },
+        },
+    },
+    events: {
+        recalled: {
+            schema: {
+                type: "object",
+                properties: { sessionID: { type: "string" } },
+                required: ["sessionID"],
+            },
+        },
+    },
+};

package/opencode/tui.tsx

@@ -1,7 +1,8 @@
 /**
- * OpenCode TUI entry: shows "omms: ✔" in the prompt footer, and a toast
- * when a newer om-memory-system is on npm. OpenCode compiles this file itself;
- * the logic lives in dist/adapters/opencode/tui-status.js.
+ * OpenCode TUI entry: shows "omms: ✔" in the prompt footer, a toast when a
+ * newer om-memory-system is on npm, and one fold row per OMMS recall above the
+ * prompt. OpenCode compiles this file itself; the logic lives in
+ * dist/adapters/opencode/tui-status.js and tui-recall.js.
  *
  * Installed from npm, this file sits in node_modules, where OpenCode does not
  * map bare `solid-js` or JSX runtime imports to its own copies. The
@@ -41,7 +42,25 @@ export default {
     } catch {
       setText("omms: ✘");
     }
+    // Recall rows are independent of the footer: a failure here leaves the footer as it is.
+    let stopRecall: (() => void) | undefined;
+    try {
+      const { startTuiRecall } = await import("../dist/adapters/opencode/tui-recall.js");
+      stopRecall = startTuiRecall({
+        client: context.client,
+        data: context.data,
+        ui: context.ui,
+        keymap: context.keymap,
+        storage: context.storage,
+        createSignal,
+        jsx,
+        muted: () => context.theme?.text?.muted,
+      });
+    } catch {
+      // No rows; the footer keeps working.
+    }
     return () => {
+      stopRecall?.();
       stop?.();
       removeSlot();
     };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "om-memory-system",
-  "version": "5.2.0",
+  "version": "5.3.0",
   "description": "Opinionated modular memory system for coding agents (OpenCode + Pi)",
   "type": "module",
   "main": "dist/plugin.js",