@vellumai/assistant 0.8.3 → 0.8.4

package/src/runtime/routes/inference-provider-connection-routes.ts

@@ -28,8 +28,14 @@ import {
   getConnection,
   listConnections,
   MANAGED_CONNECTION_NAMES,
+  PROVIDERS_REQUIRING_BASE_URL_AND_MODELS,
   updateConnection,
 } from "../../providers/inference/connections.js";
+import {
+  isPrivateOrLocalHost,
+  resolveHostAddresses,
+  resolveRequestAddress,
+} from "../../tools/network/url-safety.js";
 import { BadRequestError, ConflictError, NotFoundError } from "./errors.js";
 import type { RouteDefinition, RouteHandlerArgs } from "./types.js";
 
@@ -59,10 +65,26 @@ function rejectDisabledOpenAICompatibleProvider(provider: string): void {
 // Custom provider field parsing (openai-compatible base_url + models)
 // ---------------------------------------------------------------------------
 
-function parseCustomProviderFields(body: Record<string, unknown>): {
+/**
+ * Parse and validate `base_url` and `models` from the request body.
+ *
+ * `base_url` is only accepted for providers in
+ * `PROVIDERS_REQUIRING_BASE_URL_AND_MODELS` (currently `openai-compatible`).
+ * For all other providers, supplying `base_url` returns a 400. This prevents
+ * API-key exfiltration: an attacker cannot create an `anthropic` connection
+ * with a `base_url` pointing to their own server, which would redirect all
+ * LLM calls (and the API key) to the attacker.
+ *
+ * Even for `openai-compatible`, the `base_url` must not point to private
+ * networks or cloud metadata endpoints (SSRF protection).
+ */
+async function parseCustomProviderFields(
+  body: Record<string, unknown>,
+  provider: string,
+): Promise<{
   baseUrl?: string | null;
   models?: ConnectionModel[] | null;
-} {
+}> {
   const out: {
     baseUrl?: string | null;
     models?: ConnectionModel[] | null;
@@ -70,11 +92,24 @@ function parseCustomProviderFields(body: Record<string, unknown>): {
 
   if ("base_url" in body) {
     const raw = body.base_url;
+
+    // Gate: base_url is only valid for openai-compatible providers.
+    if (
+      raw !== null &&
+      raw !== undefined &&
+      !PROVIDERS_REQUIRING_BASE_URL_AND_MODELS.has(provider)
+    ) {
+      throw new BadRequestError(
+        `base_url is only valid for openai-compatible providers. Remove base_url or use the openai-compatible provider type.`,
+      );
+    }
+
     if (raw === null) {
       out.baseUrl = null;
     } else if (typeof raw === "string" && raw.length > 0) {
+      let parsed: URL;
       try {
-        const parsed = new URL(raw);
+        parsed = new URL(raw);
         if (parsed.protocol !== "http:" && parsed.protocol !== "https:") {
           throw new BadRequestError(`Invalid base_url: must be an http(s) URL`);
         }
@@ -84,6 +119,27 @@ function parseCustomProviderFields(body: Record<string, unknown>): {
           `Invalid base_url: must be a valid http(s) URL`,
         );
       }
+
+      // SSRF protection: reject private IPs, localhost, cloud metadata endpoints.
+      const hostname = parsed.hostname;
+      if (isPrivateOrLocalHost(hostname)) {
+        throw new BadRequestError(
+          `Invalid base_url: must not point to a private or local network address.`,
+        );
+      }
+
+      // DNS resolution check: hostname may resolve to a private IP.
+      const resolved = await resolveRequestAddress(
+        hostname,
+        resolveHostAddresses,
+        /* allowPrivateNetwork */ false,
+      );
+      if (resolved.blockedAddress) {
+        throw new BadRequestError(
+          `Invalid base_url: hostname resolves to a private network address.`,
+        );
+      }
+
       out.baseUrl = raw;
     } else {
       throw new BadRequestError(
@@ -145,7 +201,7 @@ function handleGetConnection({ pathParams = {} }: RouteHandlerArgs) {
   return conn;
 }
 
-function handleCreateConnection({ body = {} }: RouteHandlerArgs) {
+async function handleCreateConnection({ body = {} }: RouteHandlerArgs) {
   const name = body.name;
   const provider = body.provider;
   const auth = body.auth;
@@ -186,7 +242,7 @@ function handleCreateConnection({ body = {} }: RouteHandlerArgs) {
     );
   }
 
-  const customFields = parseCustomProviderFields(body);
+  const customFields = await parseCustomProviderFields(body, providerResult.data);
 
   const result = createConnection(getDb(), {
     name,
@@ -224,7 +280,7 @@ function handleCreateConnection({ body = {} }: RouteHandlerArgs) {
   return result.connection;
 }
 
-function handleUpdateConnection({
+async function handleUpdateConnection({
   pathParams = {},
   body = {},
 }: RouteHandlerArgs) {
@@ -278,7 +334,7 @@ function handleUpdateConnection({
     );
   }
 
-  const customFields = parseCustomProviderFields(body);
+  const customFields = await parseCustomProviderFields(body, existing.provider);
 
   const result = updateConnection(getDb(), name, {
     auth: authResult.data,

package/src/runtime/routes/integrations/a2a.ts

@@ -7,11 +7,13 @@
  * POST   /v1/integrations/a2a/invite          — create a shareable A2A invite token
  * POST   /v1/integrations/a2a/invite/complete — sender-side invite completion
  * POST   /v1/integrations/a2a/invite/redeem   — receiver-side invite redemption
+ * POST   /v1/integrations/a2a/invite/accept   — self-hosted broker: orchestrate complete + redeem
  */
 
 import { isA2AEnabled } from "../../../a2a/feature-gate.js";
 import { getConfig } from "../../../config/loader.js";
 import {
+  acceptA2AInvite,
   clearA2AConfig,
   completeA2AInvite,
   createA2AInvite,
@@ -19,7 +21,7 @@ import {
   redeemA2AInvite,
   setA2AConfig,
 } from "../../../daemon/handlers/config-a2a.js";
-import { BadRequestError } from "../errors.js";
+import { BadGatewayError, BadRequestError } from "../errors.js";
 import type { RouteDefinition, RouteHandlerArgs } from "../types.js";
 
 // ---------------------------------------------------------------------------
@@ -164,6 +166,52 @@ function handleRedeemA2AInvite({ body = {} }: RouteHandlerArgs) {
   return result;
 }
 
+async function handleAcceptA2AInvite({ body = {} }: RouteHandlerArgs) {
+  assertA2AFlag();
+  const { senderGatewayUrl, senderAssistantId, token } = body as {
+    senderGatewayUrl?: unknown;
+    senderAssistantId?: unknown;
+    token?: unknown;
+  };
+
+  if (typeof senderGatewayUrl !== "string" || !senderGatewayUrl) {
+    throw new BadRequestError(
+      "senderGatewayUrl is required and must be a non-empty string",
+    );
+  }
+  try {
+    new URL(senderGatewayUrl);
+  } catch {
+    throw new BadRequestError("senderGatewayUrl must be a valid URL");
+  }
+  if (typeof senderAssistantId !== "string" || !senderAssistantId) {
+    throw new BadRequestError(
+      "senderAssistantId is required and must be a non-empty string",
+    );
+  }
+  if (typeof token !== "string" || !token) {
+    throw new BadRequestError(
+      "token is required and must be a non-empty string",
+    );
+  }
+
+  const result = await acceptA2AInvite({
+    senderGatewayUrl,
+    senderAssistantId,
+    token,
+  });
+  if (!result.success) {
+    const isSenderFault =
+      result.errorCode === "sender_unreachable" ||
+      result.errorCode === "complete_failed";
+    if (isSenderFault) {
+      throw new BadGatewayError(result.error ?? "Failed to accept A2A invite");
+    }
+    throw new BadRequestError(result.error ?? "Failed to accept A2A invite");
+  }
+  return result;
+}
+
 // ---------------------------------------------------------------------------
 // Route definitions
 // ---------------------------------------------------------------------------
@@ -232,4 +280,15 @@ export const ROUTES: RouteDefinition[] = [
     requirePolicyEnforcement: true,
     handler: handleRedeemA2AInvite,
   },
+  {
+    operationId: "integrations_a2a_invite_accept_post",
+    endpoint: "integrations/a2a/invite/accept",
+    method: "POST",
+    summary: "Accept A2A invite (self-hosted broker)",
+    description:
+      "Orchestrate cross-daemon invite acceptance for self-hosted deployments. Calls the sender's invite/complete, then creates a local contact via invite/redeem.",
+    tags: ["integrations"],
+    requirePolicyEnforcement: true,
+    handler: handleAcceptA2AInvite,
+  },
 ];

package/src/runtime/routes/log-export-routes.ts

@@ -36,6 +36,7 @@ import {
   getWorkspaceConfigPath,
 } from "../../util/platform.js";
 import { APP_VERSION, COMMIT_SHA } from "../../version.js";
+import { assistantEventHub } from "../assistant-event-hub.js";
 import { createTarGz } from "./archive-utils.js";
 import { InternalError } from "./errors.js";
 import { collectWorkspaceData } from "./log-export/workspace-allowlist.js";
@@ -270,6 +271,44 @@ async function handleExport({
       );
     }
 
+    // --- Connected clients (`assistant clients list --json`) ---
+    try {
+      const clients = assistantEventHub.listClients();
+      const clientsList = {
+        clients: clients.map((c) => ({
+          clientId: c.clientId,
+          interfaceId: c.interfaceId,
+          capabilities: c.capabilities,
+          machineName: c.machineName,
+          connectedAt: c.connectedAt.toISOString(),
+          lastActiveAt: c.lastActiveAt.toISOString(),
+        })),
+      };
+      writeFileSync(
+        join(staging, "clients-list.json"),
+        JSON.stringify(clientsList, null, 2),
+        "utf-8",
+      );
+    } catch (err) {
+      const message = err instanceof Error ? err.message : String(err);
+      writeFileSync(
+        join(staging, "clients-list-error.json"),
+        JSON.stringify(
+          {
+            error: message,
+            collectedAt: new Date().toISOString(),
+          },
+          null,
+          2,
+        ),
+        "utf-8",
+      );
+      log.warn(
+        { err },
+        "Failed to collect connected clients list, continuing without it",
+      );
+    }
+
     // --- Export manifest ---
     const manifestType = conversationId
       ? ("conversation-export" as const)

package/src/runtime/routes/memory-v2-routes.ts

@@ -8,6 +8,8 @@ import { join } from "node:path";
 import { z } from "zod";
 
 import { loadConfig } from "../../config/loader.js";
+import type { AssistantConfig } from "../../config/types.js";
+import { getDb } from "../../memory/db-connection.js";
 import {
   enqueueMemoryJob,
   type MemoryJobType,
@@ -21,12 +23,16 @@ import {
   totalEdgeCount,
   validateEdgeTargets,
 } from "../../memory/v2/edge-index.js";
+import { computeInjectionScores } from "../../memory/v2/injection-events.js";
+import { loadNowText } from "../../memory/v2/now-text.js";
+import { getPageIndex } from "../../memory/v2/page-index.js";
 import {
   getConceptsDir,
   listPages,
   readPage,
   renderPageContent,
 } from "../../memory/v2/page-store.js";
+import { type RouterSource, runRouter } from "../../memory/v2/router.js";
 import { seedV2SkillEntries } from "../../memory/v2/skill-store.js";
 import { getLogger } from "../../util/logger.js";
 import { getWorkspaceDir } from "../../util/platform.js";
@@ -290,6 +296,195 @@ async function handleConceptFrequency({
   return getConceptFrequencySummary(workspaceDir, { conversationId, sinceMs });
 }
 
+// ── EMA scores ──────────────────────────────────────────────────────────
+
+const MemoryV2EmaScoresParams = z.object({}).strict();
+
+export interface MemoryV2EmaScoresEntry {
+  slug: string;
+  /** Time-decayed injection frequency; 0 when no events in the read window. */
+  score: number;
+  /** File mtime in epoch ms; 0 for synthetic entries (skills, CLI commands). */
+  modifiedAt: number;
+}
+
+export interface MemoryV2EmaScoresResult {
+  /** Every page index entry, sorted by score descending then slug ASCII. */
+  entries: MemoryV2EmaScoresEntry[];
+}
+
+async function handleEmaScores({
+  body = {},
+}: RouteHandlerArgs): Promise<MemoryV2EmaScoresResult> {
+  // Intentionally NOT gated on `memory.v2.enabled` — operators inspecting
+  // EMA data before flipping tier-2 routing on is a legitimate dry-run
+  // use case, mirroring `memory_v2_validate`.
+  MemoryV2EmaScoresParams.parse(body);
+
+  const pageIndex = await getPageIndex(getWorkspaceDir());
+  const slugs = pageIndex.entries.map((e) => e.slug);
+  const scores = computeInjectionScores(getDb(), slugs, Date.now());
+
+  const entries: MemoryV2EmaScoresEntry[] = pageIndex.entries.map((entry) => ({
+    slug: entry.slug,
+    score: scores.get(entry.slug) ?? 0,
+    modifiedAt: entry.modifiedAt,
+  }));
+  entries.sort((a, b) => {
+    if (a.score !== b.score) return b.score - a.score;
+    return a.slug < b.slug ? -1 : a.slug > b.slug ? 1 : 0;
+  });
+  return { entries };
+}
+
+// ── Simulate router (dry-run playground) ────────────────────────────────
+
+const SimulateRouterOverridesSchema = z
+  .object({
+    tier1_size: z.number().int().min(1).nullable().optional(),
+    tier2_size: z.number().int().min(1).nullable().optional(),
+    batch_size: z.number().int().min(1).nullable().optional(),
+  })
+  .strict();
+
+const MemoryV2SimulateRouterParams = z
+  .object({
+    query: z.string().min(1, "query must be non-empty"),
+    configOverrides: SimulateRouterOverridesSchema.optional(),
+  })
+  .strict();
+
+export interface MemoryV2SimulateRouterEffectiveConfig {
+  tier1_size: number | null;
+  tier2_size: number | null;
+  batch_size: number | null;
+  max_page_ids: number;
+}
+
+export interface MemoryV2SimulateRouterResult {
+  /** Slugs the router would select, in model-returned order. */
+  selectedSlugs: string[];
+  /** Per-slug provenance: `"tier1"`, `"tier2"`, or `"tier3:<bucket>"`. */
+  sourceBySlug: Record<string, RouterSource>;
+  /** EMA scores for the selected slugs (0 when the slug has no events). */
+  scores: Record<string, number>;
+  /** `null` on success; otherwise one of the router failure reasons. */
+  failureReason: string | null;
+  /** The router config that actually ran (live merged with overrides). */
+  effectiveConfig: MemoryV2SimulateRouterEffectiveConfig;
+  /** The overrides the caller submitted, for display. */
+  overrides: {
+    tier1_size?: number | null;
+    tier2_size?: number | null;
+    batch_size?: number | null;
+  };
+  /** Page index size the router was given (post-tier-carve, all batches). */
+  totalCandidatePages: number;
+}
+
+/**
+ * Build the config the router will see by overlaying override values on top
+ * of the live workspace config. Only the three new tier knobs are exposed —
+ * everything else (provider, prompts, weights) stays exactly as it would on
+ * a real turn. `undefined` means "inherit live"; `null` is a valid override
+ * value (meaning "disable this tier").
+ */
+function applySimulateOverrides(
+  live: AssistantConfig,
+  overrides: z.infer<typeof SimulateRouterOverridesSchema> | undefined,
+): AssistantConfig {
+  if (!overrides) return live;
+  const liveRouter = live.memory.v2.router;
+  const mergedRouter = {
+    ...liveRouter,
+    ...("tier1_size" in overrides && overrides.tier1_size !== undefined
+      ? { tier1_size: overrides.tier1_size }
+      : {}),
+    ...("tier2_size" in overrides && overrides.tier2_size !== undefined
+      ? { tier2_size: overrides.tier2_size }
+      : {}),
+    ...("batch_size" in overrides && overrides.batch_size !== undefined
+      ? { batch_size: overrides.batch_size }
+      : {}),
+  };
+  return {
+    ...live,
+    memory: {
+      ...live.memory,
+      v2: {
+        ...live.memory.v2,
+        router: mergedRouter,
+      },
+    },
+  };
+}
+
+export async function handleSimulateRouter({
+  body = {},
+}: RouteHandlerArgs): Promise<MemoryV2SimulateRouterResult> {
+  requireMemoryV2Enabled();
+  const { query, configOverrides } = MemoryV2SimulateRouterParams.parse(body);
+
+  const liveConfig = loadConfig();
+  const mergedConfig = applySimulateOverrides(liveConfig, configOverrides);
+  const effectiveRouter = mergedConfig.memory.v2.router;
+
+  const workspaceDir = getWorkspaceDir();
+  const nowText = await loadNowText(workspaceDir);
+
+  const routerResult = await runRouter({
+    workspaceDir,
+    userMessage: query,
+    assistantMessage: "",
+    nowText,
+    priorEverInjected: [],
+    config: mergedConfig,
+    database: getDb(),
+  });
+
+  const pageIndex = await getPageIndex(workspaceDir);
+  const scores = computeInjectionScores(
+    getDb(),
+    routerResult.selectedSlugs,
+    Date.now(),
+  );
+
+  const sourceBySlug: Record<string, RouterSource> = {};
+  for (const [slug, source] of routerResult.sourceBySlug.entries()) {
+    sourceBySlug[slug] = source;
+  }
+
+  const scoresOut: Record<string, number> = {};
+  for (const slug of routerResult.selectedSlugs) {
+    scoresOut[slug] = scores.get(slug) ?? 0;
+  }
+
+  return {
+    selectedSlugs: routerResult.selectedSlugs,
+    sourceBySlug,
+    scores: scoresOut,
+    failureReason: routerResult.failureReason,
+    effectiveConfig: {
+      tier1_size: effectiveRouter.tier1_size,
+      tier2_size: effectiveRouter.tier2_size,
+      batch_size: effectiveRouter.batch_size,
+      max_page_ids: effectiveRouter.max_page_ids,
+    },
+    overrides: {
+      ...(configOverrides?.tier1_size !== undefined
+        ? { tier1_size: configOverrides.tier1_size }
+        : {}),
+      ...(configOverrides?.tier2_size !== undefined
+        ? { tier2_size: configOverrides.tier2_size }
+        : {}),
+      ...(configOverrides?.batch_size !== undefined
+        ? { batch_size: configOverrides.batch_size }
+        : {}),
+    },
+    totalCandidatePages: pageIndex.entries.length,
+  };
+}
+
 // ── Route definitions ───────────────────────────────────────────────────
 
 export const ROUTES: RouteDefinition[] = [
@@ -359,4 +554,26 @@ export const ROUTES: RouteDefinition[] = [
     tags: ["memory"],
     requestBody: MemoryV2ConceptFrequencyParams,
   },
+  {
+    operationId: "memory_v2_ema_scores",
+    method: "POST",
+    endpoint: "memory/v2/ema-scores",
+    handler: handleEmaScores,
+    summary: "List every concept page with its injection-frequency EMA score",
+    description:
+      "Computes the time-decayed injection frequency (3-day half-life) for every entry in the current page index by reading memory_v2_injection_events. Returns entries sorted by score descending then slug ASCII, including zero-score pages so callers can decide whether to filter. Read-only; tier 2 of the v4 router uses the same computation to pick its top-M.",
+    tags: ["memory"],
+    requestBody: MemoryV2EmaScoresParams,
+  },
+  {
+    operationId: "memory_v2_simulate_router",
+    method: "POST",
+    endpoint: "memory/v2/simulate-router",
+    handler: handleSimulateRouter,
+    summary: "Dry-run the v4 router with config overrides (read-only)",
+    description:
+      "Runs the memory router against the live page index + EMA scores with optional tier_size / batch_size overrides, without recording an injection event or writing an activation log. Returns the slugs that would have been selected, per-slug tier provenance, EMA scores, and the effective router config so operators can validate knob changes before flipping them in workspace config.",
+    tags: ["memory"],
+    requestBody: MemoryV2SimulateRouterParams,
+  },
 ];

package/src/runtime/routes/notification-routes.ts

@@ -7,6 +7,7 @@ import { z } from "zod";
 
 import { getDb } from "../../memory/db-connection.js";
 import { notificationDeliveries } from "../../memory/schema.js";
+import { bufferIfDeferred } from "../../notifications/deferred-emit.js";
 import { emitNotificationSignal } from "../../notifications/emit-signal.js";
 import { listEvents } from "../../notifications/events-store.js";
 import type { AttentionHints } from "../../notifications/signal.js";
@@ -78,6 +79,9 @@ const EmitSignalParams = z.object({
   conversationAffinityHint: z.record(z.string(), z.string()).optional(),
   dedupeKey: z.string().optional(),
   throwOnError: z.boolean().optional(),
+  // Conversation that originated this signal — used by `deferred-emit` to
+  // buffer notifications during in-band background-job tool calls.
+  originatingConversationId: z.string().optional(),
 });
 
 const ListNotificationEventsParams = z.object({
@@ -89,7 +93,7 @@ const ListNotificationEventsParams = z.object({
 
 async function handleEmitSignal({ body = {} }: RouteHandlerArgs) {
   const validated = EmitSignalParams.parse(body);
-  const result = await emitNotificationSignal({
+  const params = {
     sourceEventName: validated.sourceEventName,
     sourceChannel: validated.sourceChannel,
     sourceContextId: validated.sourceContextId,
@@ -99,7 +103,20 @@ async function handleEmitSignal({ body = {} }: RouteHandlerArgs) {
     conversationAffinityHint: validated.conversationAffinityHint,
     dedupeKey: validated.dedupeKey,
     throwOnError: validated.throwOnError,
-  });
+  };
+  const buffered = bufferIfDeferred(
+    validated.originatingConversationId,
+    params,
+  );
+  if (buffered) {
+    return {
+      signalId: buffered.signalId,
+      dispatched: buffered.dispatched,
+      deduplicated: buffered.deduplicated,
+      reason: buffered.reason,
+    };
+  }
+  const result = await emitNotificationSignal(params);
   return {
     signalId: result.signalId,
     dispatched: result.dispatched,

package/src/runtime/routes/question-routes.ts

@@ -148,7 +148,10 @@ function handleQuestionResponse({ body }: RouteHandlerArgs) {
 
   // Validation passed — deregister now to clear the prompter timer, then
   // hand the result to the prompter's caller via rpcResolve.
-  pendingInteractions.resolve(requestId);
+  pendingInteractions.resolve(
+    requestId,
+    response.kind === "close" ? "cancelled" : "answered",
+  );
 
   log.info(
     {