dsh-tdai-memory 0.2.9 → 0.2.11

package/index.js

@@ -1,421 +1,421 @@
-/**
- * dsh-tdai-memory — TencentDB Agent Memory ported into DeepSeek Harness.
- *
- * Reuses the host-neutral core of `tdai-memory-openclaw-plugin`
- * (TdaiCore: L0 conversation capture -> L1 structured facts -> L2 scenes ->
- * L3 persona; vector recall; memory/conversation search) through a thin
- * DshHostAdapter. The existing data directory (~/.memory-tencentdb/memory-tdai)
- * is reused as-is, so previously accumulated memories stay available.
- *
- * Wiring:
- *  - `session/event` (turn/start, turn/end) -> L0 capture + pipeline trigger
- *  - `system-prompt/assemble` waterfall -> per-turn recall injection
- *    (relevant L1 memories + L3 persona + L2 scene navigation)
- *  - tools `tdai_memory_search` / `tdai_conversation_search`
- *  - LLM calls (L1/L2/L3 extraction) go through the standalone OpenAI-
- *    compatible runner (config `llm`, falls back to TDAI_LLM_* env vars)
- *  - embeddings through the configurable OpenAI-compatible endpoint
- *    (defaults to the local Qwen3-Embedding-0.6B service on :8088)
- */
-import os from "node:os";
-import { join } from "node:path";
-import z from "@deepseek-ai/schemastery";
-import { defineTool } from "@deepseek-ai/dsh-tools";
-import { installSettingsSection, settingsNamespace } from "@deepseek-ai/dsh-settings";
-import { TdaiCore } from "./vendor/tdai/core/tdai-core.js";
-import { parseConfig } from "./vendor/tdai/config.js";
-import { StandaloneHostAdapter } from "./vendor/tdai/adapters/standalone/host-adapter.js";
-import { ensureSettingsNamespaceExposed } from "./vendor/dsh-settings-expose.js";
-
-/** Cordis plugin name. */
-const name = "tdai-memory";
-/** Services this row injects. */
-const inject = ["sessions", "systemPrompt", "tools"];
-/** Settings namespace owned by this plugin (Web UI settings section). */
-const NS = settingsNamespace("tdai-memory");
-
-/** Runtime schema for the tdai-memory row. */
-const Config = z.object({
-  /** TDAI data directory; default: ~/.memory-tencentdb/memory-tdai. */
-  dataDir: z.string().default(""),
-  /** LLM used for L1/L2/L3 extraction (OpenAI-compatible). Falls back to TDAI_LLM_* env. */
-  llm: z.object({
-    baseUrl: z.string().default(""),
-    apiKey: z.string().default("").role("secret"),
-    model: z.string().default(""),
-    maxTokens: z.number().default(4096),
-    timeoutMs: z.number().default(120000),
-  }),
-  /** Embedding endpoint (OpenAI-compatible /v1/embeddings). */
-  embedding: z.object({
-    baseUrl: z.string().default("http://127.0.0.1:8088/v1"),
-    apiKey: z.string().default("").role("secret"),
-    model: z.string().default("Qwen3-Embedding-0.6B"),
-    dimensions: z.number().default(1024),
-    sendDimensions: z.boolean().default(false),
-  }),
-  /** Capture conversations into L0. */
-  captureEnabled: z.boolean().default(true),
-  /** L1 extraction pipeline (facts from conversations). */
-  extraction: z.object({
-    enabled: z.boolean().default(true),
-    /** Conflict detection before storing; extra LLM call, disable if flaky.
-     *  Off by default: dedup LLM output parsing is flaky (README matches). */
-    enableDedup: z.boolean().default(false),
-  }),
-  /** Recall injection settings. */
-  recall: z.object({
-    enabled: z.boolean().default(true),
-    maxResults: z.number().default(5),
-    scoreThreshold: z.number().default(0.3),
-    timeoutMs: z.number().default(3000),
-  }),
-  /** Register the two search tools. */
-  toolsEnabled: z.boolean().default(true),
-});
-
-/**
- * Whether a derived message is real human input.
- *
- * A `"user"` role alone is not enough: DSH injects synthetic user-role
- * context (system reminders, runtime context, skill catalogs, goal rounds,
- * tool results, …) that must never be treated as user input. Real input is
- * distinguished by `message.source.kind === "user"` (see dsh-llm Message
- * types). Messages without a `source` (older harnesses) are treated as real
- * input to stay compatible.
- */
-function isRealUserMessage(m) {
-  return m?.role === "user" && (m.source == null || m.source.kind === "user");
-}
-
-/** Extract plain text from a dsh message content (block array or string). */
-function extractText(content) {
-  if (typeof content === "string") return content.trim();
-  if (!Array.isArray(content)) return "";
-  return content
-    .filter((b) => b && b.type === "text" && typeof b.text === "string")
-    .map((b) => b.text)
-    .join("\n")
-    .trim();
-}
-
-function apply(ctx, config) {
-  // ── build-once state (rebuildable) ───────────────────────────────────────
-  // TdaiCore is built from the RESOLVED settings value (composition entry as
-  // base, `tdai-memory` settings section / settings.yaml as user layer), so
-  // keys stored in settings.yaml (llm.apiKey etc.) are honored. The settings
-  // inject callback that drives `installSettingsSection`'s setSource is
-  // scheduled by cordis as a plugin start — it ALWAYS runs after any 0ms
-  // timer — so the no-settings fallback below cannot be allowed to win the
-  // race. Instead of racing, we let the fallback build from the entry config
-  // first (guaranteeing a working core even without settings), and REBUILD
-  // from the resolved settings when the settings mount arrives. The fallback
-  // is a one-shot; the settings build always wins afterwards.
-  let core = null;
-  let tdaiConfig = null;
-  let dataDir = "";
-  let builtOnce = false;
-
-  /** turn id of the last turn/end per session (capture trigger). */
-  const lastTurnEnd = new Map();
-  /** turn ids already captured per session (dedup across flushes). */
-  const capturedTurns = new Map();
-  /** turn/start timestamp per session (L0 capture cursor floor). */
-  const turnStarts = new Map();
-  /** session keys seen this process (for final L1 flush on teardown). */
-  const seenSessions = new Set();
-  /** recall cache per session: same user text within TTL reuses the result. */
-  const recallCache = new Map();
-  const RECALL_CACHE_TTL_MS = 30_000;
-
-  // One-shot mode (dsh --profile headless): the process exits right after
-  // flush, so L1 must finish inside the flush listener. In web mode the
-  // long-lived process runs the pipeline in the background instead.
-  const isOneShot = process.argv.some((arg) => arg === "headless");
-
-  // Verbose tdai-internal logging goes to the console only when explicitly
-  // enabled; routine diagnostics use the dsh logger.
-  const verbose = !!process.env.DSH_TDAI_DEBUG;
-  const log = (level, m) => {
-    if (verbose) console.error(`[tdai-memory] ${m}`);
-    else if (level === "warn") ctx.logger.warn(`[tdai-memory] ${m}`);
-    else if (level === "error") ctx.logger.error(`[tdai-memory] ${m}`);
-  };
-  const logger = {
-    debug: (m) => log("debug", m),
-    info: (m) => log("info", m),
-    warn: (m) => log("warn", m),
-    error: (m) => log("error", m),
-  };
-
-  // ── core construction (rebuildable) ──────────────────────────────────────
-  async function build(cfg) {
-    if (cfg?.llm?.apiKey) {
-      log("debug", `build: using settings config (apiKey=${cfg.llm.apiKey.slice(0, 6)}…)`);
-    }
-    // Tear down the previous core, if any (first build has none).
-    const previous = core;
-    if (previous) {
-      await previous.destroy().catch((error) => {
-        log("warn", `previous core destroy failed: ${String(error)}`);
-      });
-    }
-    dataDir = cfg.dataDir || join(os.homedir(), ".memory-tencentdb", "memory-tdai");
-    const nextConfig = parseConfig({
-      timezone: "system",
-      storeBackend: "sqlite",
-      capture: { enabled: cfg.captureEnabled },
-      extraction: { enabled: cfg.extraction.enabled, enableDedup: cfg.extraction.enableDedup },
-      recall: { ...cfg.recall },
-      embedding: {
-        enabled: true,
-        provider: "openai",
-        baseUrl: cfg.embedding.baseUrl,
-        apiKey: cfg.embedding.apiKey,
-        model: cfg.embedding.model,
-        dimensions: cfg.embedding.dimensions,
-        sendDimensions: cfg.embedding.sendDimensions,
-        timeoutMs: 10000,
-      },
-      llm: {
-        enabled: true,
-        baseUrl: cfg.llm.baseUrl || process.env.TDAI_LLM_BASE_URL || "",
-        apiKey: cfg.llm.apiKey || process.env.TDAI_LLM_API_KEY || "",
-        model: cfg.llm.model || process.env.TDAI_LLM_MODEL || "deepseek-v4-flash",
-        maxTokens: cfg.llm.maxTokens,
-        timeoutMs: cfg.llm.timeoutMs,
-      },
-    });
-    const nextHostAdapter = new StandaloneHostAdapter({
-      dataDir,
-      llmConfig: {
-        baseUrl: nextConfig.llm.baseUrl,
-        apiKey: nextConfig.llm.apiKey,
-        model: nextConfig.llm.model,
-        maxTokens: nextConfig.llm.maxTokens,
-        timeoutMs: nextConfig.llm.timeoutMs,
-      },
-      logger,
-      defaultUserId: "dsh-user",
-      platform: "dsh",
-    });
-    const nextCore = new TdaiCore({ hostAdapter: nextHostAdapter, config: nextConfig, instanceId: "dsh" });
-    // Publish before initialize so concurrent flush listeners already use the
-    // new core; initialize() only sets up stores/pipeline.
-    tdaiConfig = nextConfig;
-    core = nextCore;
-    await nextCore.initialize();
-    const key = nextConfig.llm.apiKey || "";
-    log("info", `initialized (dataDir=${dataDir}, llm=${nextConfig.llm.model}, apiKey=${key ? `${key.slice(0, 6)}…` : "<empty>"}, embedding=${nextConfig.embedding.model})`);
-  }
-
-  // ── capture: triggered on session/flush (awaitable; runs before process
-  //    exit in headless, and at every request boundary in web) ─────────────
-  ctx.on("session/event", (session, event) => {
-    if (event.type === "turn/start") {
-      turnStarts.set(session.id, event.time);
-    } else if (event.type === "turn/end") {
-      lastTurnEnd.set(session.id, event.data.turn);
-    }
-  });
-
-  ctx.on("session/flush", async (session) => {
-    try {
-      const turnId = lastTurnEnd.get(session.id);
-      if (turnId === undefined || capturedTurns.get(session.id) === turnId) return;
-      capturedTurns.set(session.id, turnId);
-      seenSessions.add(session.id);
-      const startedAt = turnStarts.get(session.id) ?? Date.now();
-      await captureTurn(session, startedAt);
-      if (isOneShot) {
-        // Run pending L1 extraction to completion before the process exits.
-        await core?.handleSessionEnd(session.id).catch((error) => {
-          ctx.logger.warn(`[tdai-memory] L1 flush failed: ${String(error)}`);
-        });
-      }
-    } catch (error) {
-      ctx.logger.warn(`[tdai-memory] capture failed: ${String(error)}`);
-    }
-  });
-
-  async function captureTurn(session, startedAt) {
-    if (!tdaiConfig?.capture.enabled) return;
-    const derived = session.deriveMessages();
-    // Start from the last REAL user message. DSH injects synthetic user-role
-    // context (system reminders, runtime context, skill catalogs, goal
-    // rounds, …) whose `source.kind` is not "user"; those must not pollute
-    // L0 memory (issue #1).
-    const lastUserIdx = derived.findLastIndex(isRealUserMessage);
-    if (lastUserIdx < 0) return;
-    const messages = derived
-      .slice(lastUserIdx)
-      .map((m) => {
-        // Keep assistant replies and real user input only; drop synthetic
-        // user-role messages (plugin context, tool results, …) injected
-        // after the last real user message.
-        if (m.role !== "assistant" && !isRealUserMessage(m)) return null;
-        const text = extractText(m.content);
-        if (!text) return null;
-        return {
-          id: m.id ?? `msg_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`,
-          role: m.role === "assistant" ? "assistant" : "user",
-          content: text,
-          timestamp: Date.now(),
-        };
-      })
-      .filter(Boolean);
-    if (messages.length === 0) return;
-    const lastUser = [...messages].reverse().find((m) => m.role === "user");
-    const result = await core.handleTurnCommitted({
-      userText: lastUser?.content ?? "",
-      assistantText: "",
-      messages,
-      sessionKey: session.id,
-      startedAt,
-    });
-    ctx.logger.debug(`[tdai-memory] captured turn: l0=${result.l0RecordedCount} notified=${result.schedulerNotified} vectors=${result.l0VectorsWritten}`);
-  }
-
-  // ── recall injection ────────────────────────────────────────────────────
-  // The prompt assembly runs inside the agent's own context (scope = agent),
-  // so the waterfall listener must be registered on `agent.ctx` — root-level
-  // listeners never see agent-scoped assemblies. Agents appear right after
-  // their session is announced, so defer one tick and attach there.
-  ctx.on("session/created", (session) => {
-    setTimeout(() => {
-      try {
-        const agents = ctx.get("agents");
-        const agent = agents?.get?.(session.id);
-        if (!agent?.ctx) return;
-        agent.ctx.on("system-prompt/assemble", async (assembly, context) => {
-          try {
-            if (!tdaiConfig?.recall.enabled) return assembly;
-            const lastUser = [...session.deriveMessages()].reverse().find(isRealUserMessage);
-            if (!lastUser) return assembly;
-            const text = extractText(lastUser.content);
-            if (!text) return assembly;
-            const now = Date.now();
-            const cached = recallCache.get(session.id);
-            if (cached && cached.text === text && now - cached.ts < RECALL_CACHE_TTL_MS) {
-              return injectRecall(assembly, cached.result);
-            }
-            const result = await Promise.race([
-              core.handleBeforeRecall(text, session.id),
-              new Promise((resolve) => setTimeout(() => resolve(null), tdaiConfig.recall.timeoutMs + 1000)),
-            ]).catch(() => null);
-            recallCache.set(session.id, { text, result: result ?? {}, ts: now });
-            return injectRecall(assembly, result ?? {});
-          } catch (error) {
-            ctx.logger.warn(`[tdai-memory] recall failed: ${String(error)}`);
-            return assembly;
-          }
-        });
-      } catch (error) {
-        ctx.logger.warn(`[tdai-memory] recall attach failed: ${String(error)}`);
-      }
-    }, 0);
-  });
-
-  function injectRecall(assembly, result) {
-    const extras = [];
-    if (result.prependContext) extras.push({ name: "tdai:recall", order: 1000, text: result.prependContext });
-    if (result.appendSystemContext) extras.push({ name: "tdai:profile", order: 1001, text: result.appendSystemContext });
-    if (extras.length === 0) return assembly;
-    return { ...assembly, contexts: [...assembly.contexts, ...extras] };
-  }
-
-  // ── tools ────────────────────────────────────────────────────────────────
-  if (config.toolsEnabled) {
-    ctx.tools.register(defineTool({
-      name: "tdai_memory_search",
-      description:
-        "Search structured long-term memories (facts, preferences, and experiences extracted from past conversations). " +
-        "Use this to recall what the user told you or what happened in earlier sessions.",
-      parameters: {
-        query: { type: "string", required: true, description: "The memory to search for." },
-        limit: { type: "integer", description: "Maximum results (default 5)." },
-      },
-      output: {
-        schema: { type: "string" },
-        render: (_args, value) => [{ type: "text", text: value }],
-      },
-      async execute(args) {
-        const result = await core.searchMemories({ query: args.query, limit: args.limit ?? 5 });
-        return result.text;
-      },
-    }));
-    ctx.tools.register(defineTool({
-      name: "tdai_conversation_search",
-      description:
-        "Search raw past conversation logs (L0) by keyword. Use this to look up exact quotes or details " +
-        "that structured memory search does not cover.",
-      parameters: {
-        query: { type: "string", required: true, description: "Text to find in past conversations." },
-        limit: { type: "integer", description: "Maximum results (default 5)." },
-      },
-      output: {
-        schema: { type: "string" },
-        render: (_args, value) => [{ type: "text", text: value }],
-      },
-      async execute(args) {
-        const result = await core.searchConversations({ query: args.query, limit: args.limit ?? 5 });
-        return result.text;
-      },
-    }));
-  }
-
-  // ── settings-backed configuration ─────────────────────────────────────────
-  // The composition entry stays the base layer; a registered `tdai-memory`
-  // settings section overlays it (settings.yaml). `installSettingsSection`
-  // hands `setSource` a GETTER (`() => scope.get()`), not the config object —
-  // we must call it to obtain the resolved settings. The settings mount
-  // (cordis inject) always arrives after our fallback timer, so the fallback
-  // builds the core from the entry config first (one-shot) and the settings
-  // build REBUILDS the core with the resolved values (apiKey included).
-  installSettingsSection(ctx, NS, Config, config, {
-    setSource: (getter) => {
-      builtOnce = true;
-      void build(getter());
-    },
-    onChange: () => {
-      ctx.logger.warn("[tdai-memory] settings updated; restart to apply (TdaiCore is built at startup)");
-    },
-  });
-
-  // dsh-host-apiproxy hard-codes which settings namespaces the Web client may
-  // see; without this, the settings section answers `settings-not-exposed`
-  // on any stock install. Patch the allowlist idempotently (self-heals after
-  // dsh updates overwrite the file).
-  ensureSettingsNamespaceExposed(ctx, "tdai-memory", ctx.logger);
-
-  // Fallback when no settings service ever mounts: build from the entry
-  // config. One-shot — if the settings mount arrives later, its build
-  // supersedes this one. Delayed past the 0ms window so headless/web boots
-  // are never blocked on this timer.
-  setTimeout(() => {
-    if (!builtOnce) {
-      builtOnce = true;
-      void build(config);
-    }
-  }, 500);
-
-  // ── lifecycle ────────────────────────────────────────────────────────────
-  ctx.effect(() => {
-    return async () => {
-      // Final flush: run pending L1 extraction for every session seen this
-      // process and wait for it to finish (headless exits right after the
-      // tree disposes, so this is what makes extraction complete in one-shot
-      // runs). Unknown keys are tolerated as no-ops.
-      for (const sessionKey of seenSessions) {
-        await core?.handleSessionEnd(sessionKey).catch((error) => {
-          ctx.logger.warn(`[tdai-memory] final L1 flush failed for ${sessionKey}: ${String(error)}`);
-        });
-      }
-      await core?.destroy().catch((error) => {
-        ctx.logger.warn(`[tdai-memory] destroy failed: ${String(error)}`);
-      });
-    };
-  }, "tdai-memory.cleanup()");
-}
-
-export { Config, apply, inject, name };
+/**
+ * dsh-tdai-memory — TencentDB Agent Memory ported into DeepSeek Harness.
+ *
+ * Reuses the host-neutral core of `tdai-memory-openclaw-plugin`
+ * (TdaiCore: L0 conversation capture -> L1 structured facts -> L2 scenes ->
+ * L3 persona; vector recall; memory/conversation search) through a thin
+ * DshHostAdapter. The existing data directory (~/.memory-tencentdb/memory-tdai)
+ * is reused as-is, so previously accumulated memories stay available.
+ *
+ * Wiring:
+ *  - `session/event` (turn/start, turn/end) -> L0 capture + pipeline trigger
+ *  - `system-prompt/assemble` waterfall -> per-turn recall injection
+ *    (relevant L1 memories + L3 persona + L2 scene navigation)
+ *  - tools `tdai_memory_search` / `tdai_conversation_search`
+ *  - LLM calls (L1/L2/L3 extraction) go through the standalone OpenAI-
+ *    compatible runner (config `llm`, falls back to TDAI_LLM_* env vars)
+ *  - embeddings through the configurable OpenAI-compatible endpoint
+ *    (defaults to the local Qwen3-Embedding-0.6B service on :8088)
+ */
+import os from "node:os";
+import { join } from "node:path";
+import z from "@deepseek-ai/schemastery";
+import { defineTool } from "@deepseek-ai/dsh-tools";
+import { installSettingsSection, settingsNamespace } from "@deepseek-ai/dsh-settings";
+import { TdaiCore } from "./vendor/tdai/core/tdai-core.js";
+import { parseConfig } from "./vendor/tdai/config.js";
+import { StandaloneHostAdapter } from "./vendor/tdai/adapters/standalone/host-adapter.js";
+import { ensureSettingsNamespaceExposed } from "./vendor/dsh-settings-expose.js";
+
+/** Cordis plugin name. */
+const name = "tdai-memory";
+/** Services this row injects. */
+const inject = ["sessions", "systemPrompt", "tools"];
+/** Settings namespace owned by this plugin (Web UI settings section). */
+const NS = settingsNamespace("tdai-memory");
+
+/** Runtime schema for the tdai-memory row. */
+const Config = z.object({
+  /** TDAI data directory; default: ~/.memory-tencentdb/memory-tdai. */
+  dataDir: z.string().default(""),
+  /** LLM used for L1/L2/L3 extraction (OpenAI-compatible). Falls back to TDAI_LLM_* env. */
+  llm: z.object({
+    baseUrl: z.string().default(""),
+    apiKey: z.string().default("").role("secret"),
+    model: z.string().default(""),
+    maxTokens: z.number().default(4096),
+    timeoutMs: z.number().default(120000),
+  }),
+  /** Embedding endpoint (OpenAI-compatible /v1/embeddings). */
+  embedding: z.object({
+    baseUrl: z.string().default("http://127.0.0.1:8088/v1"),
+    apiKey: z.string().default("").role("secret"),
+    model: z.string().default("Qwen3-Embedding-0.6B"),
+    dimensions: z.number().default(1024),
+    sendDimensions: z.boolean().default(false),
+  }),
+  /** Capture conversations into L0. */
+  captureEnabled: z.boolean().default(true),
+  /** L1 extraction pipeline (facts from conversations). */
+  extraction: z.object({
+    enabled: z.boolean().default(true),
+    /** Conflict detection before storing; extra LLM call, disable if flaky.
+     *  Off by default: dedup LLM output parsing is flaky (README matches). */
+    enableDedup: z.boolean().default(false),
+  }),
+  /** Recall injection settings. */
+  recall: z.object({
+    enabled: z.boolean().default(true),
+    maxResults: z.number().default(5),
+    scoreThreshold: z.number().default(0.3),
+    timeoutMs: z.number().default(3000),
+  }),
+  /** Register the two search tools. */
+  toolsEnabled: z.boolean().default(true),
+});
+
+/**
+ * Whether a derived message is real human input.
+ *
+ * A `"user"` role alone is not enough: DSH injects synthetic user-role
+ * context (system reminders, runtime context, skill catalogs, goal rounds,
+ * tool results, …) that must never be treated as user input. Real input is
+ * distinguished by `message.source.kind === "user"` (see dsh-llm Message
+ * types). Messages without a `source` (older harnesses) are treated as real
+ * input to stay compatible.
+ */
+function isRealUserMessage(m) {
+  return m?.role === "user" && (m.source == null || m.source.kind === "user");
+}
+
+/** Extract plain text from a dsh message content (block array or string). */
+function extractText(content) {
+  if (typeof content === "string") return content.trim();
+  if (!Array.isArray(content)) return "";
+  return content
+    .filter((b) => b && b.type === "text" && typeof b.text === "string")
+    .map((b) => b.text)
+    .join("\n")
+    .trim();
+}
+
+function apply(ctx, config) {
+  // ── build-once state (rebuildable) ───────────────────────────────────────
+  // TdaiCore is built from the RESOLVED settings value (composition entry as
+  // base, `tdai-memory` settings section / settings.yaml as user layer), so
+  // keys stored in settings.yaml (llm.apiKey etc.) are honored. The settings
+  // inject callback that drives `installSettingsSection`'s setSource is
+  // scheduled by cordis as a plugin start — it ALWAYS runs after any 0ms
+  // timer — so the no-settings fallback below cannot be allowed to win the
+  // race. Instead of racing, we let the fallback build from the entry config
+  // first (guaranteeing a working core even without settings), and REBUILD
+  // from the resolved settings when the settings mount arrives. The fallback
+  // is a one-shot; the settings build always wins afterwards.
+  let core = null;
+  let tdaiConfig = null;
+  let dataDir = "";
+  let builtOnce = false;
+
+  /** turn id of the last turn/end per session (capture trigger). */
+  const lastTurnEnd = new Map();
+  /** turn ids already captured per session (dedup across flushes). */
+  const capturedTurns = new Map();
+  /** turn/start timestamp per session (L0 capture cursor floor). */
+  const turnStarts = new Map();
+  /** session keys seen this process (for final L1 flush on teardown). */
+  const seenSessions = new Set();
+  /** recall cache per session: same user text within TTL reuses the result. */
+  const recallCache = new Map();
+  const RECALL_CACHE_TTL_MS = 30_000;
+
+  // One-shot mode (dsh --profile headless): the process exits right after
+  // flush, so L1 must finish inside the flush listener. In web mode the
+  // long-lived process runs the pipeline in the background instead.
+  const isOneShot = process.argv.some((arg) => arg === "headless");
+
+  // Verbose tdai-internal logging goes to the console only when explicitly
+  // enabled; routine diagnostics use the dsh logger.
+  const verbose = !!process.env.DSH_TDAI_DEBUG;
+  const log = (level, m) => {
+    if (verbose) console.error(`[tdai-memory] ${m}`);
+    else if (level === "warn") ctx.logger.warn(`[tdai-memory] ${m}`);
+    else if (level === "error") ctx.logger.error(`[tdai-memory] ${m}`);
+  };
+  const logger = {
+    debug: (m) => log("debug", m),
+    info: (m) => log("info", m),
+    warn: (m) => log("warn", m),
+    error: (m) => log("error", m),
+  };
+
+  // ── core construction (rebuildable) ──────────────────────────────────────
+  async function build(cfg) {
+    if (cfg?.llm?.apiKey) {
+      log("debug", `build: using settings config (apiKey=${cfg.llm.apiKey.slice(0, 6)}…)`);
+    }
+    // Tear down the previous core, if any (first build has none).
+    const previous = core;
+    if (previous) {
+      await previous.destroy().catch((error) => {
+        log("warn", `previous core destroy failed: ${String(error)}`);
+      });
+    }
+    dataDir = cfg.dataDir || join(os.homedir(), ".memory-tencentdb", "memory-tdai");
+    const nextConfig = parseConfig({
+      timezone: "system",
+      storeBackend: "sqlite",
+      capture: { enabled: cfg.captureEnabled },
+      extraction: { enabled: cfg.extraction.enabled, enableDedup: cfg.extraction.enableDedup },
+      recall: { ...cfg.recall },
+      embedding: {
+        enabled: true,
+        provider: "openai",
+        baseUrl: cfg.embedding.baseUrl,
+        apiKey: cfg.embedding.apiKey,
+        model: cfg.embedding.model,
+        dimensions: cfg.embedding.dimensions,
+        sendDimensions: cfg.embedding.sendDimensions,
+        timeoutMs: 10000,
+      },
+      llm: {
+        enabled: true,
+        baseUrl: cfg.llm.baseUrl || process.env.TDAI_LLM_BASE_URL || "",
+        apiKey: cfg.llm.apiKey || process.env.TDAI_LLM_API_KEY || "",
+        model: cfg.llm.model || process.env.TDAI_LLM_MODEL || "deepseek-v4-flash",
+        maxTokens: cfg.llm.maxTokens,
+        timeoutMs: cfg.llm.timeoutMs,
+      },
+    });
+    const nextHostAdapter = new StandaloneHostAdapter({
+      dataDir,
+      llmConfig: {
+        baseUrl: nextConfig.llm.baseUrl,
+        apiKey: nextConfig.llm.apiKey,
+        model: nextConfig.llm.model,
+        maxTokens: nextConfig.llm.maxTokens,
+        timeoutMs: nextConfig.llm.timeoutMs,
+      },
+      logger,
+      defaultUserId: "dsh-user",
+      platform: "dsh",
+    });
+    const nextCore = new TdaiCore({ hostAdapter: nextHostAdapter, config: nextConfig, instanceId: "dsh" });
+    // Publish before initialize so concurrent flush listeners already use the
+    // new core; initialize() only sets up stores/pipeline.
+    tdaiConfig = nextConfig;
+    core = nextCore;
+    await nextCore.initialize();
+    const key = nextConfig.llm.apiKey || "";
+    log("info", `initialized (dataDir=${dataDir}, llm=${nextConfig.llm.model}, apiKey=${key ? `${key.slice(0, 6)}…` : "<empty>"}, embedding=${nextConfig.embedding.model})`);
+  }
+
+  // ── capture: triggered on session/flush (awaitable; runs before process
+  //    exit in headless, and at every request boundary in web) ─────────────
+  ctx.on("session/event", (session, event) => {
+    if (event.type === "turn/start") {
+      turnStarts.set(session.id, event.time);
+    } else if (event.type === "turn/end") {
+      lastTurnEnd.set(session.id, event.data.turn);
+    }
+  });
+
+  ctx.on("session/flush", async (session) => {
+    try {
+      const turnId = lastTurnEnd.get(session.id);
+      if (turnId === undefined || capturedTurns.get(session.id) === turnId) return;
+      capturedTurns.set(session.id, turnId);
+      seenSessions.add(session.id);
+      const startedAt = turnStarts.get(session.id) ?? Date.now();
+      await captureTurn(session, startedAt);
+      if (isOneShot) {
+        // Run pending L1 extraction to completion before the process exits.
+        await core?.handleSessionEnd(session.id).catch((error) => {
+          ctx.logger.warn(`[tdai-memory] L1 flush failed: ${String(error)}`);
+        });
+      }
+    } catch (error) {
+      ctx.logger.warn(`[tdai-memory] capture failed: ${String(error)}`);
+    }
+  });
+
+  async function captureTurn(session, startedAt) {
+    if (!tdaiConfig?.capture.enabled) return;
+    const derived = session.deriveMessages();
+    // Start from the last REAL user message. DSH injects synthetic user-role
+    // context (system reminders, runtime context, skill catalogs, goal
+    // rounds, …) whose `source.kind` is not "user"; those must not pollute
+    // L0 memory (issue #1).
+    const lastUserIdx = derived.findLastIndex(isRealUserMessage);
+    if (lastUserIdx < 0) return;
+    const messages = derived
+      .slice(lastUserIdx)
+      .map((m) => {
+        // Keep assistant replies and real user input only; drop synthetic
+        // user-role messages (plugin context, tool results, …) injected
+        // after the last real user message.
+        if (m.role !== "assistant" && !isRealUserMessage(m)) return null;
+        const text = extractText(m.content);
+        if (!text) return null;
+        return {
+          id: m.id ?? `msg_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`,
+          role: m.role === "assistant" ? "assistant" : "user",
+          content: text,
+          timestamp: Date.now(),
+        };
+      })
+      .filter(Boolean);
+    if (messages.length === 0) return;
+    const lastUser = [...messages].reverse().find((m) => m.role === "user");
+    const result = await core.handleTurnCommitted({
+      userText: lastUser?.content ?? "",
+      assistantText: "",
+      messages,
+      sessionKey: session.id,
+      startedAt,
+    });
+    ctx.logger.debug(`[tdai-memory] captured turn: l0=${result.l0RecordedCount} notified=${result.schedulerNotified} vectors=${result.l0VectorsWritten}`);
+  }
+
+  // ── recall injection ────────────────────────────────────────────────────
+  // The prompt assembly runs inside the agent's own context (scope = agent),
+  // so the waterfall listener must be registered on `agent.ctx` — root-level
+  // listeners never see agent-scoped assemblies. Agents appear right after
+  // their session is announced, so defer one tick and attach there.
+  ctx.on("session/created", (session) => {
+    setTimeout(() => {
+      try {
+        const agents = ctx.get("agents");
+        const agent = agents?.get?.(session.id);
+        if (!agent?.ctx) return;
+        agent.ctx.on("system-prompt/assemble", async (assembly, context) => {
+          try {
+            if (!tdaiConfig?.recall.enabled) return assembly;
+            const lastUser = [...session.deriveMessages()].reverse().find(isRealUserMessage);
+            if (!lastUser) return assembly;
+            const text = extractText(lastUser.content);
+            if (!text) return assembly;
+            const now = Date.now();
+            const cached = recallCache.get(session.id);
+            if (cached && cached.text === text && now - cached.ts < RECALL_CACHE_TTL_MS) {
+              return injectRecall(assembly, cached.result);
+            }
+            const result = await Promise.race([
+              core.handleBeforeRecall(text, session.id),
+              new Promise((resolve) => setTimeout(() => resolve(null), tdaiConfig.recall.timeoutMs + 1000)),
+            ]).catch(() => null);
+            recallCache.set(session.id, { text, result: result ?? {}, ts: now });
+            return injectRecall(assembly, result ?? {});
+          } catch (error) {
+            ctx.logger.warn(`[tdai-memory] recall failed: ${String(error)}`);
+            return assembly;
+          }
+        });
+      } catch (error) {
+        ctx.logger.warn(`[tdai-memory] recall attach failed: ${String(error)}`);
+      }
+    }, 0);
+  });
+
+  function injectRecall(assembly, result) {
+    const extras = [];
+    if (result.prependContext) extras.push({ name: "tdai:recall", order: 1000, text: result.prependContext });
+    if (result.appendSystemContext) extras.push({ name: "tdai:profile", order: 1001, text: result.appendSystemContext });
+    if (extras.length === 0) return assembly;
+    return { ...assembly, contexts: [...assembly.contexts, ...extras] };
+  }
+
+  // ── tools ────────────────────────────────────────────────────────────────
+  if (config.toolsEnabled) {
+    ctx.tools.register(defineTool({
+      name: "tdai_memory_search",
+      description:
+        "Search structured long-term memories (facts, preferences, and experiences extracted from past conversations). " +
+        "Use this to recall what the user told you or what happened in earlier sessions.",
+      parameters: {
+        query: { type: "string", required: true, description: "The memory to search for." },
+        limit: { type: "integer", description: "Maximum results (default 5)." },
+      },
+      output: {
+        schema: { type: "string" },
+        render: (_args, value) => [{ type: "text", text: value }],
+      },
+      async execute(args) {
+        const result = await core.searchMemories({ query: args.query, limit: args.limit ?? 5 });
+        return result.text;
+      },
+    }));
+    ctx.tools.register(defineTool({
+      name: "tdai_conversation_search",
+      description:
+        "Search raw past conversation logs (L0) by keyword. Use this to look up exact quotes or details " +
+        "that structured memory search does not cover.",
+      parameters: {
+        query: { type: "string", required: true, description: "Text to find in past conversations." },
+        limit: { type: "integer", description: "Maximum results (default 5)." },
+      },
+      output: {
+        schema: { type: "string" },
+        render: (_args, value) => [{ type: "text", text: value }],
+      },
+      async execute(args) {
+        const result = await core.searchConversations({ query: args.query, limit: args.limit ?? 5 });
+        return result.text;
+      },
+    }));
+  }
+
+  // ── settings-backed configuration ─────────────────────────────────────────
+  // The composition entry stays the base layer; a registered `tdai-memory`
+  // settings section overlays it (settings.yaml). `installSettingsSection`
+  // hands `setSource` a GETTER (`() => scope.get()`), not the config object —
+  // we must call it to obtain the resolved settings. The settings mount
+  // (cordis inject) always arrives after our fallback timer, so the fallback
+  // builds the core from the entry config first (one-shot) and the settings
+  // build REBUILDS the core with the resolved values (apiKey included).
+  installSettingsSection(ctx, NS, Config, config, {
+    setSource: (getter) => {
+      builtOnce = true;
+      void build(getter());
+    },
+    onChange: () => {
+      ctx.logger.warn("[tdai-memory] settings updated; restart to apply (TdaiCore is built at startup)");
+    },
+  });
+
+  // dsh-host-apiproxy hard-codes which settings namespaces the Web client may
+  // see; without this, the settings section answers `settings-not-exposed`
+  // on any stock install. Patch the allowlist idempotently (self-heals after
+  // dsh updates overwrite the file).
+  ensureSettingsNamespaceExposed(ctx, "tdai-memory", ctx.logger);
+
+  // Fallback when no settings service ever mounts: build from the entry
+  // config. One-shot — if the settings mount arrives later, its build
+  // supersedes this one. Delayed past the 0ms window so headless/web boots
+  // are never blocked on this timer.
+  setTimeout(() => {
+    if (!builtOnce) {
+      builtOnce = true;
+      void build(config);
+    }
+  }, 500);
+
+  // ── lifecycle ────────────────────────────────────────────────────────────
+  ctx.effect(() => {
+    return async () => {
+      // Final flush: run pending L1 extraction for every session seen this
+      // process and wait for it to finish (headless exits right after the
+      // tree disposes, so this is what makes extraction complete in one-shot
+      // runs). Unknown keys are tolerated as no-ops.
+      for (const sessionKey of seenSessions) {
+        await core?.handleSessionEnd(sessionKey).catch((error) => {
+          ctx.logger.warn(`[tdai-memory] final L1 flush failed for ${sessionKey}: ${String(error)}`);
+        });
+      }
+      await core?.destroy().catch((error) => {
+        ctx.logger.warn(`[tdai-memory] destroy failed: ${String(error)}`);
+      });
+    };
+  }, "tdai-memory.cleanup()");
+}
+
+export { Config, apply, inject, name };