@vellumai/assistant 0.8.8 → 0.8.9-dev.202606091853.fbaa2ae

package/src/plugins/defaults/memory-retrieval/hooks/user-prompt-submit.ts

@@ -0,0 +1,285 @@
+/**
+ * Default `user-prompt-submit` hook: prepares a turn's prompt before the first
+ * provider call by running memory-graph retrieval and then applying the
+ * runtime injections that assemble the turn.
+ *
+ * **Memory graph** via {@link ConversationGraphMemory.prepareMemory} —
+ * dispatches to context-load or per-turn retrieval depending on initialization
+ * state; gated on the actor being trusted (guardian). The hook owns the
+ * retrieval's side effects — persisting the injected block onto the user
+ * message's metadata, writing the recall log, and emitting the
+ * `memory_recalled` event. The PKB query-vector pair is recorded on the
+ * conversation's graph handle for the PKB-reminder injector to read back. PKB
+ * context and NOW.md are sourced directly by their injectors (gated on block
+ * presence), not produced here.
+ *
+ * **Runtime injection** via {@link applyRuntimeInjections} — runs for every
+ * actor (not just trusted ones) on the retrieved history, assembling the Slack
+ * chronological transcript, the unified `<turn_context>` block,
+ * channel/voice/transport hints, and the NOW.md / PKB / memory-v2 / workspace
+ * blocks, then persisting the assembled blocks onto the user message's metadata
+ * so they survive reloads. This is the first-call counterpart of the
+ * post-compaction hook, which re-applies the same injections after a mid-loop
+ * compaction.
+ *
+ * Registered first in the `user-prompt-submit` chain, so it heads the chain
+ * ahead of history repair and title generation: those later hooks see the
+ * fully assembled, memory-injected history.
+ */
+
+import type { PluginHookFn, UserPromptSubmitContext } from "@vellumai/plugin-api";
+
+import { getConfig } from "../../../../config/loader.js";
+import { findConversationOrSubagent } from "../../../../daemon/conversation-registry.js";
+import {
+  applyRuntimeInjections,
+  resolveTurnInboundActorContext,
+  resolveTurnModelProfileLabel,
+  type RuntimeInjectionResult,
+} from "../../../../daemon/conversation-runtime-assembly.js";
+import type { MemoryRecalled } from "../../../../daemon/message-types/memory.js";
+import { resolveTrustClass } from "../../../../daemon/trust-context.js";
+import { updateMessageMetadata } from "../../../../memory/conversation-crud.js";
+import { recordMemoryRecallLog } from "../../../../memory/memory-recall-log-store.js";
+import { broadcastMessage } from "../../../../runtime/assistant-event-hub.js";
+import type { GraphMemoryResult } from "../../../types.js";
+
+/**
+ * Persist and broadcast the retrieval's side effects: the injected block on
+ * the user message's metadata (so it survives reloads), a recall-log row, and
+ * the `memory_recalled` debug event. All three are best-effort — a failure to
+ * persist must not abort the turn.
+ */
+function recordRecallSideEffects(
+  graphResult: GraphMemoryResult,
+  ctx: UserPromptSubmitContext,
+): void {
+  // Persist the injected block text in message metadata so it survives
+  // conversation reloads (eviction, restart, fork). loadFromDb re-injects
+  // from metadata.
+  if (graphResult.injectedBlockText) {
+    try {
+      updateMessageMetadata(ctx.userMessageId, {
+        memoryInjectedBlock: graphResult.injectedBlockText,
+      });
+    } catch (err) {
+      ctx.logger.warn(
+        { err },
+        "Failed to persist memory injection to metadata (non-fatal)",
+      );
+    }
+  }
+
+  const m = graphResult.metrics;
+
+  try {
+    recordMemoryRecallLog({
+      conversationId: ctx.conversationId,
+      enabled: true,
+      degraded: false,
+      provider: m?.embeddingProvider ?? undefined,
+      model: m?.embeddingModel ?? undefined,
+      semanticHits: m?.semanticHits ?? 0,
+      mergedCount: m?.mergedCount ?? 0,
+      selectedCount: m?.selectedCount ?? 0,
+      tier1Count: m?.tier1Count ?? 0,
+      tier2Count: m?.tier2Count ?? 0,
+      hybridSearchLatencyMs: m?.hybridSearchLatencyMs ?? 0,
+      sparseVectorUsed: m?.sparseVectorUsed ?? false,
+      injectedTokens: graphResult.injectedTokens,
+      latencyMs: graphResult.latencyMs,
+      topCandidatesJson: (m?.topCandidates ?? []).map((c) => ({
+        key: c.nodeId,
+        type: c.type,
+        kind: "graph",
+        finalScore: c.score,
+        semantic: c.semanticSimilarity,
+        recency: c.recencyBoost,
+      })),
+      injectedText: graphResult.injectedBlockText ?? undefined,
+      reason: `graph:${graphResult.mode}`,
+      queryContext: m?.queryContext ?? undefined,
+    });
+  } catch (err) {
+    ctx.logger.warn({ err }, "Failed to persist memory recall log (non-fatal)");
+  }
+
+  if (m) {
+    const memoryRecalledEvent: MemoryRecalled = {
+      type: "memory_recalled",
+      provider: m.embeddingProvider ?? "unknown",
+      model: m.embeddingModel ?? "unknown",
+      semanticHits: m.semanticHits,
+      mergedCount: m.mergedCount,
+      selectedCount: m.selectedCount,
+      tier1Count: m.tier1Count,
+      tier2Count: m.tier2Count,
+      hybridSearchLatencyMs: m.hybridSearchLatencyMs,
+      sparseVectorUsed: m.sparseVectorUsed,
+      injectedTokens: graphResult.injectedTokens,
+      latencyMs: graphResult.latencyMs,
+      topCandidates: m.topCandidates.map((c) => ({
+        key: c.nodeId,
+        type: c.type,
+        kind: "graph",
+        finalScore: c.score,
+        semantic: c.semanticSimilarity,
+        recency: c.recencyBoost,
+      })),
+    };
+    broadcastMessage(memoryRecalledEvent);
+  }
+}
+
+/**
+ * Persist the assembled runtime-injection blocks onto the user message's
+ * metadata so they survive conversation reloads (eviction, restart, fork) —
+ * loadFromDb re-injects from metadata. Only this first-call site persists: the
+ * mid-loop re-entry sites send identical bytes and their tail row may not
+ * correspond to `userMessageId`. All present blocks are written in a single
+ * update to avoid doubling SQLite SELECT+UPDATE work each turn. Best-effort — a
+ * persistence failure must not abort the turn.
+ */
+function persistInjectionBlocks(
+  blocks: RuntimeInjectionResult["blocks"],
+  ctx: UserPromptSubmitContext,
+): void {
+  if (
+    !blocks.unifiedTurnContext &&
+    !blocks.pkbSystemReminder &&
+    !blocks.workspaceBlock &&
+    !blocks.nowScratchpadBlock &&
+    !blocks.pkbContextBlock &&
+    !blocks.memoryV2StaticBlock
+  ) {
+    return;
+  }
+  try {
+    const metadataUpdates: Record<string, unknown> = {};
+    if (blocks.unifiedTurnContext) {
+      metadataUpdates.turnContextBlock = blocks.unifiedTurnContext;
+    }
+    if (blocks.pkbSystemReminder) {
+      metadataUpdates.pkbSystemReminderBlock = blocks.pkbSystemReminder;
+    }
+    if (blocks.workspaceBlock) {
+      metadataUpdates.workspaceBlock = blocks.workspaceBlock;
+    }
+    if (blocks.nowScratchpadBlock) {
+      metadataUpdates.nowScratchpadBlock = blocks.nowScratchpadBlock;
+    }
+    if (blocks.pkbContextBlock) {
+      metadataUpdates.pkbContextBlock = blocks.pkbContextBlock;
+    }
+    if (blocks.memoryV2StaticBlock) {
+      metadataUpdates.memoryV2StaticBlock = blocks.memoryV2StaticBlock;
+    }
+    updateMessageMetadata(ctx.userMessageId, metadataUpdates);
+  } catch (err) {
+    ctx.logger.warn(
+      { err },
+      "Failed to persist injection metadata (non-fatal)",
+    );
+  }
+}
+
+/**
+ * Run the default retrieval and assemble the turn's prompt. Memory-graph
+ * retrieval is gated on the actor being trusted (guardian); untrusted actors
+ * skip it, leaving `latestMessages` as the seeded input. Runtime injection then
+ * runs for every actor, writing the fully injected result back onto
+ * `latestMessages` and persisting the assembled blocks.
+ *
+ * Memory retrieval blocks the turn — there is no soft timeout here. Memory is
+ * critical context, and silently dropping it produces a worse outcome than a
+ * slower turn. Cancellation still works via `ctx.signal`, which is threaded
+ * into `prepareMemory`.
+ */
+const userPromptSubmitMemoryRetrieval: PluginHookFn<
+  UserPromptSubmitContext
+> = async (ctx) => {
+  // The conversation-scoped retrieval state — graph handle, abort signal, and
+  // trust class — is resolved from the live conversation by id rather than
+  // threaded in, mirroring how `applyRuntimeInjections` self-resolves its
+  // per-turn inputs.
+  const conversation = findConversationOrSubagent(ctx.conversationId);
+  const config = getConfig();
+  const abortSignal = conversation?.abortController?.signal;
+  const isTrustedActor =
+    resolveTrustClass(conversation?.trustContext) === "guardian";
+  const actorContext = resolveTurnInboundActorContext(
+    conversation?.trustContext,
+    conversation?.assistantId,
+  );
+
+  if (conversation && isTrustedActor && abortSignal) {
+    // Retrieval progress (`memory_status`) and the `memory_recalled` summary
+    // publish to the shared `broadcastMessage` hub — the sink every turn
+    // publisher converges to — rather than a threaded event callback. This
+    // keeps any raw client-emit capability off the hook contract.
+    const graphResult = await conversation.graphMemory.prepareMemory(
+      ctx.latestMessages,
+      config,
+      abortSignal,
+      broadcastMessage,
+    );
+
+    recordRecallSideEffects(graphResult, ctx);
+
+    ctx.latestMessages = graphResult.runMessages;
+    // Select dense+sparse as a matched pair so RRF fusion combines two signals
+    // aligned to the same query text:
+    //   1. Context-load with a user query: user-query dense + user-query sparse
+    //      — the cleanest pairing.
+    //   2. Otherwise (context-load without a user query, or per-turn): whatever
+    //      `queryVector` / `sparseVector` the retriever produced, which are
+    //      themselves co-aligned (both summary-derived in context-load, both
+    //      user-last-message-derived in per-turn).
+    // Never pair a user-query dense with a summary-aligned sparse.
+    // The PKB-reminder injector reads this pair back off the same graph handle
+    // (looked up by conversation id) rather than receiving it threaded through
+    // the agent loop.
+    if (graphResult.userQueryVector) {
+      conversation.graphMemory.recordPkbQueryVectors(
+        graphResult.userQueryVector,
+        graphResult.userQuerySparseVector,
+      );
+    } else {
+      conversation.graphMemory.recordPkbQueryVectors(
+        graphResult.queryVector,
+        graphResult.sparseVector,
+      );
+    }
+  }
+
+  // Runtime injection assembles the turn's prompt onto the retrieved history.
+  // It runs for every actor (untrusted actors skip only the memory-graph step
+  // above). `applyRuntimeInjections` self-resolves every per-turn injection
+  // input — the Slack chronological transcript, the unified `<turn_context>`
+  // block, channel/voice/transport hints, and the turn's trust/index/call-site
+  // — from the live conversation, so we hand in only the request id and
+  // conversation id plus the field resolved once at turn start
+  // (`isNonInteractive`). The `model_profile` label is rendered here from the
+  // turn's resolved profile key, using the call site self-resolved from the
+  // live conversation. The unified `<turn_context>` actor input is
+  // self-resolved from the live conversation's trust context. This first-call
+  // assembly always runs at `"full"` volume; overflow reduction only downgrades
+  // the mode on later re-injection.
+  const modelProfile = resolveTurnModelProfileLabel(
+    ctx.modelProfileKey,
+    conversation?.currentCallSite ?? "mainAgent",
+    config.llm,
+  );
+  const injection = await applyRuntimeInjections(ctx.latestMessages, {
+    isNonInteractive: ctx.isNonInteractive,
+    modelProfile,
+    actorContext,
+    mode: "full",
+    requestId: ctx.requestId,
+    conversationId: ctx.conversationId,
+  });
+  ctx.latestMessages = injection.messages;
+  persistInjectionBlocks(injection.blocks, ctx);
+};
+
+export default userPromptSubmitMemoryRetrieval;

package/src/plugins/defaults/memory-retrieval/injector-chain.ts

@@ -18,9 +18,9 @@
  * ordering assumptions about when `defaultInjectors` finishes initializing.
  */
 
-import { memoryV3Injector } from "../../../memory/v3/shadow-plugin.js";
 import type { Injector } from "../../types.js";
-import { defaultInjectors } from "../injectors/register.js";
+import { memoryV3Injector } from "../memory-v3-shadow/injector.js";
+import { defaultInjectors } from "./injectors.js";
 
 let cachedChain: Injector[] | null = null;