@vellumai/assistant 0.7.3 → 0.8.1

package/src/memory/llm-request-log-source-clickhouse.ts

@@ -0,0 +1,317 @@
+/**
+ * ClickHouse-backed LLM request log read source.
+ *
+ * Reads from the ClickHouse mirror (populated out-of-band by the
+ * `mirror-llm-logs-to-clickhouse` cron). Scoped to the running
+ * assistant's own `assistant_id` — never cross-assistant. URL and
+ * password are resolved lazily from the credential store
+ * (`clickhouse:url`, `clickhouse:password`); database/table/user come
+ * from workspace config.
+ *
+ * Known limitation: the mirror is INSERT-only. A row inserted locally
+ * with `message_id = NULL` and backfilled later will appear in
+ * ClickHouse with `message_id = ''` forever. Reads via this source for
+ * the most-recent ~minute of activity therefore have lower fidelity
+ * than the local source. Acceptable for the "internal use while we
+ * finetune prompts" use case; revisit when mirror updates are added.
+ */
+import type { LlmRequestLogsClickHouseConfig } from "../config/schemas/llm-request-logs.js";
+import { credentialKey } from "../security/credential-key.js";
+import { getSecureKeyAsync } from "../security/secure-keys.js";
+import { getLogger } from "../util/logger.js";
+import {
+  getAssistantMessageIdsInTurn,
+  getMessageById,
+  messageMetadataSchema,
+} from "./conversation-crud.js";
+import type { LlmRequestLogSource } from "./llm-request-log-source.js";
+import type { LogRow } from "./llm-request-log-store.js";
+
+const log = getLogger("clickhouse-llm-request-log-source");
+
+/**
+ * Read a credential and normalize `undefined` → `null`. The credential
+ * resolver factories on this class are typed `() => Promise<string | null>`;
+ * `getSecureKeyAsync` returns `Promise<string | undefined>`. Keep the
+ * coercion in one place so TypeScript stays happy without per-call casts.
+ */
+async function readCredentialOrNull(
+  service: string,
+  field: string,
+): Promise<string | null> {
+  const value = await getSecureKeyAsync(credentialKey(service, field));
+  return value ?? null;
+}
+
+/**
+ * Wire-format row returned by ClickHouse for our query columns. Note
+ * that `created_at` arrives as a string because Int64 is emitted as a
+ * quoted string under the default `output_format_json_quote_64bit_integers=1`
+ * setting; we coerce to `number` in `toLogRow`.
+ */
+interface ClickHouseRow {
+  id: string;
+  conversation_id: string;
+  message_id: string;
+  provider: string;
+  request_payload: string;
+  response_payload: string;
+  created_at: string;
+}
+
+/** Injectable fetch override for tests. Defaults to globalThis.fetch. */
+export type ClickHouseFetch = typeof fetch;
+
+/** Minimal subset of the SQLite message row the fork-source fallback needs. */
+export interface ClickHouseMessageRow {
+  metadata: string | null;
+}
+
+export interface ClickHouseLlmRequestLogSourceDeps {
+  /** Override the credential read for `clickhouse:url`. */
+  resolveUrl?: () => Promise<string | null>;
+  /** Override the credential read for `clickhouse:password`. */
+  resolvePassword?: () => Promise<string | null>;
+  /** Override the credential read for `vellum:platform_assistant_id`. */
+  resolveAssistantId?: () => Promise<string | null>;
+  /** Override the turn-id resolver (default: `getAssistantMessageIdsInTurn`). */
+  resolveTurnMessageIds?: (messageId: string) => string[];
+  /** Override the message lookup (default: `getMessageById`). */
+  resolveMessage?: (messageId: string) => ClickHouseMessageRow | null;
+  /** Override fetch for testing. */
+  fetchImpl?: ClickHouseFetch;
+}
+
+export class ClickHouseLlmRequestLogSource implements LlmRequestLogSource {
+  private cachedUrl: string | null = null;
+  private cachedPassword: string | null = null;
+  private cachedAssistantId: string | null = null;
+
+  private readonly resolveUrl: () => Promise<string | null>;
+  private readonly resolvePassword: () => Promise<string | null>;
+  private readonly resolveAssistantId: () => Promise<string | null>;
+  private readonly resolveTurnMessageIds: (messageId: string) => string[];
+  private readonly resolveMessage: (
+    messageId: string,
+  ) => ClickHouseMessageRow | null;
+  private readonly fetchImpl: ClickHouseFetch;
+
+  constructor(
+    private readonly config: LlmRequestLogsClickHouseConfig,
+    deps: ClickHouseLlmRequestLogSourceDeps = {},
+  ) {
+    this.resolveUrl =
+      deps.resolveUrl ?? (() => readCredentialOrNull("clickhouse", "url"));
+    this.resolvePassword =
+      deps.resolvePassword ??
+      (() => readCredentialOrNull("clickhouse", "password"));
+    this.resolveAssistantId =
+      deps.resolveAssistantId ??
+      (() => readCredentialOrNull("vellum", "platform_assistant_id"));
+    this.resolveTurnMessageIds =
+      deps.resolveTurnMessageIds ?? getAssistantMessageIdsInTurn;
+    this.resolveMessage = deps.resolveMessage ?? getMessageById;
+    this.fetchImpl = deps.fetchImpl ?? globalThis.fetch.bind(globalThis);
+  }
+
+  async getRequestLogById(logId: string): Promise<LogRow | null> {
+    const aid = await this.assistantId();
+    const sql = `SELECT
+        id,
+        conversation_id,
+        message_id,
+        provider,
+        request_payload,
+        response_payload,
+        toUnixTimestamp64Milli(created_at) AS created_at
+      FROM ${this.tableRef()}
+      WHERE assistant_id = {assistant_id:String}
+        AND id = {log_id:String}
+      ORDER BY created_at DESC
+      LIMIT 1
+      FORMAT JSONEachRow`;
+    const rows = await this.exec(sql, { assistant_id: aid, log_id: logId });
+    return rows[0] ? this.toLogRow(rows[0]) : null;
+  }
+
+  async getRequestLogsByMessageId(messageId: string): Promise<LogRow[]> {
+    const turnIds = this.resolveTurnMessageIds(messageId);
+    let rows = await this.selectByMessageIds(turnIds);
+
+    if (rows.length === 0) {
+      // Fork-source fallback. Mirror behavior of the local source: when no
+      // logs match the queried message's turn, see if it was forked from
+      // another and resolve that source's turn. The fork relationship lives
+      // in local SQLite (message.metadata.forkSourceMessageId), not CH.
+      const message = this.resolveMessage(messageId);
+      if (message?.metadata) {
+        try {
+          const parsed = messageMetadataSchema.safeParse(
+            JSON.parse(message.metadata),
+          );
+          const sourceMessageId =
+            parsed.success &&
+            typeof parsed.data.forkSourceMessageId === "string"
+              ? parsed.data.forkSourceMessageId
+              : null;
+          if (sourceMessageId && sourceMessageId !== messageId) {
+            const sourceTurnIds = this.resolveTurnMessageIds(sourceMessageId);
+            rows = await this.selectByMessageIds(sourceTurnIds);
+          }
+        } catch {
+          // metadata not JSON / schema mismatch — no fork fallback, return []
+        }
+      }
+    }
+
+    return rows.sort(
+      (a, b) => a.createdAt - b.createdAt || a.id.localeCompare(b.id),
+    );
+  }
+
+  private async selectByMessageIds(ids: string[]): Promise<LogRow[]> {
+    if (ids.length === 0) return [];
+    const aid = await this.assistantId();
+    // ClickHouse Array(String) URL encoding is fiddly; the message IDs are
+    // server-generated UUIDs (or safe internal strings), so inline the
+    // literal after escaping single quotes. No SQL-injection surface — the
+    // values originate from our own SQLite messages table.
+    const idLiteral =
+      "[" +
+      ids.map((id) => `'${id.replace(/'/g, "''")}'`).join(",") +
+      "]";
+    const sql = `SELECT
+        id,
+        conversation_id,
+        message_id,
+        provider,
+        request_payload,
+        response_payload,
+        toUnixTimestamp64Milli(created_at) AS created_at
+      FROM ${this.tableRef()}
+      WHERE assistant_id = {assistant_id:String}
+        AND message_id IN ${idLiteral}
+      ORDER BY created_at ASC, id ASC
+      LIMIT 1 BY id
+      FORMAT JSONEachRow`;
+    const rows = await this.exec(sql, { assistant_id: aid });
+    return rows.map((r) => this.toLogRow(r));
+  }
+
+  private tableRef(): string {
+    // Database is set via the `database=` URL param in `exec`, so we only
+    // need to quote the table identifier here. Backtick-quote both to
+    // tolerate non-default names with special characters.
+    return `\`${this.config.table.replace(/`/g, "``")}\``;
+  }
+
+  private async exec(
+    sql: string,
+    params: Record<string, string>,
+  ): Promise<ClickHouseRow[]> {
+    const baseUrl = await this.url();
+    const password = await this.password();
+
+    let target: URL;
+    try {
+      target = new URL(baseUrl);
+    } catch (err) {
+      throw new Error(
+        `clickhouse:url is not a valid URL: ${err instanceof Error ? err.message : String(err)}`,
+      );
+    }
+    target.searchParams.set("database", this.config.database);
+    for (const [k, v] of Object.entries(params)) {
+      target.searchParams.set(`param_${k}`, v);
+    }
+
+    const auth =
+      "Basic " +
+      Buffer.from(`${this.config.user}:${password}`, "utf8").toString("base64");
+
+    const res = await this.fetchImpl(target.toString(), {
+      method: "POST",
+      headers: { Authorization: auth, "Content-Type": "text/plain" },
+      body: sql,
+    });
+
+    if (!res.ok) {
+      const body = await res.text().catch(() => "");
+      log.error(
+        { status: res.status, table: this.config.table, bodySnippet: body.slice(0, 200) },
+        "ClickHouse query failed",
+      );
+      throw new Error(
+        `ClickHouse query failed (HTTP ${res.status}): ${body.slice(0, 500)}`,
+      );
+    }
+
+    const text = await res.text();
+    if (text.trim().length === 0) return [];
+
+    const rows: ClickHouseRow[] = [];
+    for (const line of text.split("\n")) {
+      const trimmed = line.trim();
+      if (trimmed.length === 0) continue;
+      try {
+        rows.push(JSON.parse(trimmed) as ClickHouseRow);
+      } catch (err) {
+        throw new Error(
+          `Failed to parse ClickHouse JSONEachRow line: ${err instanceof Error ? err.message : String(err)}`,
+        );
+      }
+    }
+    return rows;
+  }
+
+  private toLogRow(row: ClickHouseRow): LogRow {
+    return {
+      id: row.id,
+      conversationId: row.conversation_id,
+      // The mirror writes empty-string for missing message_id/provider
+      // because the CH table columns have `DEFAULT ''` (Nullable adds
+      // overhead). Map empty back to null to match the local LogRow shape.
+      messageId: row.message_id === "" ? null : row.message_id,
+      provider: row.provider === "" ? null : row.provider,
+      requestPayload: row.request_payload,
+      responsePayload: row.response_payload,
+      createdAt: Number(row.created_at),
+    };
+  }
+
+  private async assistantId(): Promise<string> {
+    if (this.cachedAssistantId) return this.cachedAssistantId;
+    const val = await this.resolveAssistantId();
+    if (!val) {
+      throw new Error(
+        "vellum:platform_assistant_id credential is required when readSource=clickhouse",
+      );
+    }
+    this.cachedAssistantId = val;
+    return val;
+  }
+
+  private async url(): Promise<string> {
+    if (this.cachedUrl) return this.cachedUrl;
+    const val = await this.resolveUrl();
+    if (!val) {
+      throw new Error(
+        "clickhouse:url credential is required when readSource=clickhouse",
+      );
+    }
+    this.cachedUrl = val;
+    return val;
+  }
+
+  private async password(): Promise<string> {
+    if (this.cachedPassword) return this.cachedPassword;
+    const val = await this.resolvePassword();
+    if (!val) {
+      throw new Error(
+        "clickhouse:password credential is required when readSource=clickhouse",
+      );
+    }
+    this.cachedPassword = val;
+    return val;
+  }
+}

package/src/memory/llm-request-log-source-local.ts

@@ -0,0 +1,26 @@
+/**
+ * Default LLM request log read source: thin async wrapper around the
+ * existing synchronous `llm-request-log-store.ts` functions.
+ *
+ * Keeps the local read path identical to its pre-pluggable behavior —
+ * including turn resolution, orphan/unlinked recovery, opportunistic
+ * backfill, and fork-source fallback. The wrapper exists only to satisfy
+ * the `LlmRequestLogSource` interface (which is `Promise`-returning so
+ * implementations with real I/O can fit).
+ */
+import type { LlmRequestLogSource } from "./llm-request-log-source.js";
+import {
+  getRequestLogById,
+  getRequestLogsByMessageId,
+  type LogRow,
+} from "./llm-request-log-store.js";
+
+export class LocalLlmRequestLogSource implements LlmRequestLogSource {
+  async getRequestLogById(logId: string): Promise<LogRow | null> {
+    return getRequestLogById(logId);
+  }
+
+  async getRequestLogsByMessageId(messageId: string): Promise<LogRow[]> {
+    return getRequestLogsByMessageId(messageId);
+  }
+}

package/src/memory/llm-request-log-source.ts

@@ -0,0 +1,97 @@
+/**
+ * Pluggable read source for LLM request logs.
+ *
+ * The Inspector view at `GET /v1/messages/:id/llm-context` (and the
+ * single-log payload route) historically read directly from the local
+ * SQLite `llm_request_logs` table via `llm-request-log-store.ts`. The
+ * source-of-truth remains local, but the *read path* is now configurable
+ * via `llmRequestLogs.readSource` in workspace config.
+ *
+ * - `local` (default): wraps the existing store functions verbatim.
+ * - `clickhouse`: queries the ClickHouse mirror (longer retention, but
+ *   only sees rows the mirror cron has flushed). See
+ *   `llm-request-log-source-clickhouse.ts`.
+ *
+ * The active source is cached at module level and invalidated on config
+ * change (see `daemon/config-watcher.ts`) so a config edit takes effect
+ * without restarting the daemon.
+ */
+import { getConfig } from "../config/loader.js";
+import { getLogger } from "../util/logger.js";
+import type { LogRow } from "./llm-request-log-store.js";
+
+const log = getLogger("llm-request-log-source");
+
+export interface LlmRequestLogSource {
+  /** Fetch a single log row by its primary key. Returns null if not found. */
+  getRequestLogById(logId: string): Promise<LogRow | null>;
+
+  /**
+   * Fetch every LLM request log associated with the given message,
+   * including all assistant messages in the same agent turn. Implementations
+   * MAY additionally apply orphan/unlinked/fork-source recovery — the
+   * local implementation does, the ClickHouse mirror does not (it is
+   * INSERT-only against the source-of-truth).
+   */
+  getRequestLogsByMessageId(messageId: string): Promise<LogRow[]>;
+}
+
+let cached: LlmRequestLogSource | null = null;
+let cachedKind: "local" | "clickhouse" | null = null;
+
+/**
+ * Return the currently configured LLM request log source.
+ *
+ * The result is cached for the lifetime of the process. Callers should
+ * never hang on to the instance across config reloads — always re-resolve
+ * through this function. The factory is async because BOTH implementations
+ * are loaded via dynamic `import()` on first use. This is deliberate: it
+ * keeps the static module graph for `llm-request-log-source.ts` (and for
+ * everything that transitively imports it, including `config-watcher`)
+ * free of `llm-request-log-store → conversation-crud → indexer → embedding-backend`,
+ * which would otherwise force test files that stub `embedding-backend.js`
+ * to also stub every export `indexer.ts` reaches for. Callers MUST `await`
+ * the source methods because the active source may swap to one with real
+ * I/O at any time.
+ */
+export async function getLlmRequestLogSource(): Promise<LlmRequestLogSource> {
+  if (cached) return cached;
+
+  const config = getConfig();
+  const kind = config.llmRequestLogs?.readSource ?? "local";
+
+  if (kind === "clickhouse") {
+    const { ClickHouseLlmRequestLogSource } = await import(
+      "./llm-request-log-source-clickhouse.js"
+    );
+    cached = new ClickHouseLlmRequestLogSource(config.llmRequestLogs.clickhouse);
+    cachedKind = "clickhouse";
+    log.info(
+      { table: config.llmRequestLogs.clickhouse.table },
+      "Using ClickHouse for LLM request log reads",
+    );
+  } else {
+    const { LocalLlmRequestLogSource } = await import(
+      "./llm-request-log-source-local.js"
+    );
+    cached = new LocalLlmRequestLogSource();
+    cachedKind = "local";
+  }
+
+  return cached;
+}
+
+/**
+ * Drop the cached source so the next `getLlmRequestLogSource()` call
+ * resolves fresh from config. Called on workspace config reload.
+ */
+export function invalidateLlmRequestLogSourceCache(): void {
+  if (cached !== null) {
+    log.debug(
+      { previousKind: cachedKind },
+      "Invalidating LLM request log source cache",
+    );
+  }
+  cached = null;
+  cachedKind = null;
+}

package/src/memory/llm-request-log-store.ts

@@ -10,7 +10,7 @@ import {
 import { getDb } from "./db-connection.js";
 import { llmRequestLogs, messages } from "./schema.js";
 
-type LogRow = {
+export type LogRow = {
   id: string;
   conversationId: string;
   messageId: string | null;

package/src/memory/memory-retrospective-constants.ts

@@ -0,0 +1,13 @@
+/**
+ * Sentinel value for the `source` column of memory-retrospective background
+ * conversations. Used both when creating them and when filtering them out of
+ * recursion / orphan-cleanup queries.
+ */
+export const MEMORY_RETROSPECTIVE_SOURCE = "memory-retrospective";
+
+/**
+ * Dedicated `group_id` value for memory-retrospective background
+ * conversations. Placed under `system:background` alongside auto-analysis,
+ * heartbeat, and filing conversations.
+ */
+export const MEMORY_RETROSPECTIVE_GROUP_ID = "system:background";

package/src/memory/memory-retrospective-enqueue.ts

@@ -0,0 +1,114 @@
+// ---------------------------------------------------------------------------
+// Memory retrospective — enqueue helper.
+// ---------------------------------------------------------------------------
+//
+// Conditionally enqueue a `memory_retrospective` job for the given
+// conversation. Gates on:
+//   - `memory-retrospective` feature flag enabled.
+//   - Source conversation isn't a memory-retrospective conversation itself
+//     (recursion guard — we never run a retrospective over reflective
+//     musings from the retrospective agent's own writes).
+//
+// All four trigger types funnel through `upsertMemoryRetrospectiveJob` which
+// coalesces rapid enqueues into a single pending row per conversation.
+// `lifecycle` and `compaction` triggers get a small debounce so the job runs
+// after the corresponding signal settles; `interval` and `message_count`
+// fire immediately.
+
+import { isAssistantFeatureFlagEnabled } from "../config/assistant-feature-flags.js";
+import { getConfig } from "../config/loader.js";
+import {
+  isUntrustedTrustClass,
+  type TrustClass,
+} from "../runtime/actor-trust-resolver.js";
+import { getLogger } from "../util/logger.js";
+import { getConversationSource } from "./conversation-crud.js";
+import { upsertMemoryRetrospectiveJob } from "./jobs-store.js";
+import { MEMORY_RETROSPECTIVE_SOURCE } from "./memory-retrospective-constants.js";
+
+const log = getLogger("memory-retrospective-enqueue");
+
+export type MemoryRetrospectiveTrigger =
+  | "interval"
+  | "message_count"
+  | "compaction"
+  | "lifecycle";
+
+const COMPACTION_DEBOUNCE_MS = 500;
+
+export function enqueueMemoryRetrospectiveIfEnabled(args: {
+  conversationId: string;
+  trigger: MemoryRetrospectiveTrigger;
+}): void {
+  const { conversationId, trigger } = args;
+
+  let config;
+  try {
+    config = getConfig();
+  } catch (err) {
+    log.warn(
+      { err, conversationId, trigger },
+      "Skipping memory-retrospective enqueue: failed to load config",
+    );
+    return;
+  }
+
+  if (!isAssistantFeatureFlagEnabled("memory-retrospective", config)) {
+    return;
+  }
+
+  if (isMemoryRetrospectiveConversation(conversationId)) {
+    log.debug(
+      { conversationId, trigger },
+      "Skipping memory-retrospective enqueue: source is a memory-retrospective conversation",
+    );
+    return;
+  }
+
+  const runAfter =
+    trigger === "compaction" ? Date.now() + COMPACTION_DEBOUNCE_MS : Date.now();
+
+  try {
+    upsertMemoryRetrospectiveJob({ conversationId }, runAfter);
+  } catch (err) {
+    log.warn(
+      { err, conversationId, trigger },
+      "Failed to upsert memory-retrospective job",
+    );
+  }
+}
+
+/**
+ * Recursion guard. The retrospective bootstraps its own background
+ * conversation; without this check, that conversation's lifecycle would
+ * enqueue another retrospective on top of it, recursing.
+ */
+export function isMemoryRetrospectiveConversation(
+  conversationId: string,
+): boolean {
+  return getConversationSource(conversationId) === MEMORY_RETROSPECTIVE_SOURCE;
+}
+
+/**
+ * Fire a memory-retrospective enqueue from the compaction site. Mirrors
+ * `enqueueAutoAnalysisOnCompaction` — same trust-class gate (don't run a
+ * guardian-trust background loop over untrusted-actor conversations) and
+ * same best-effort error swallowing (never block compaction on enqueue
+ * failures).
+ */
+export function enqueueMemoryRetrospectiveOnCompaction(
+  conversationId: string,
+  trustClass: TrustClass | undefined,
+): void {
+  if (isUntrustedTrustClass(trustClass)) {
+    return;
+  }
+  try {
+    enqueueMemoryRetrospectiveIfEnabled({
+      conversationId,
+      trigger: "compaction",
+    });
+  } catch {
+    // Best-effort — never block compaction on enqueue failures.
+  }
+}