@vellumai/assistant 0.8.6 → 0.8.7

package/src/providers/__tests__/retry-callsite.test.ts

@@ -74,7 +74,7 @@ function makeProvider(
 ): Provider {
   return {
     name,
-    async sendMessage(_messages, _tools, _systemPrompt, options) {
+    async sendMessage(_messages: Message[], options?: SendMessageOptions) {
       onCall(options);
       const config = options?.config as Record<string, unknown> | undefined;
       return makeResponse(
@@ -116,7 +116,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "memoryRetrieval" },
     });
 
@@ -155,7 +155,7 @@ describe("RetryProvider — callSite resolution", () => {
       { forwardUsageAttributionHeaders: true },
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: {
         callSite: "memoryRetrieval",
         overrideProfile: "conversation-profile",
@@ -197,7 +197,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "memoryRetrieval" },
     });
 
@@ -227,7 +227,7 @@ describe("RetryProvider — callSite resolution", () => {
       { forwardUsageAttributionHeaders: true },
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "memoryRetrieval" },
     });
 
@@ -256,7 +256,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "memoryRetrieval" },
     });
 
@@ -288,7 +288,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "heartbeatAgent" },
     });
 
@@ -327,7 +327,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent" },
     });
 
@@ -358,7 +358,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent" },
     });
 
@@ -383,7 +383,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent" },
     });
 
@@ -417,7 +417,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent" },
     });
 
@@ -446,7 +446,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "memoryRetrieval" },
     });
 
@@ -474,7 +474,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent" },
     });
 
@@ -503,7 +503,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent" },
     });
 
@@ -528,7 +528,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent" },
     });
 
@@ -550,7 +550,7 @@ describe("RetryProvider — callSite resolution", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent", model: "explicit-override" },
     });
 
@@ -587,7 +587,7 @@ describe("RetryProvider — thinking/temperature conflict guard", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       // Hardcoded per-call temperature — the pattern that caused the PR
       // #29560 bug class. Without the guard this would forward to Anthropic
       // and 400.
@@ -619,7 +619,7 @@ describe("RetryProvider — thinking/temperature conflict guard", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "recall", temperature: 0 },
     });
 
@@ -645,7 +645,7 @@ describe("RetryProvider — thinking/temperature conflict guard", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent", temperature: 1 },
     });
 
@@ -674,7 +674,7 @@ describe("RetryProvider — thinking/temperature conflict guard", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent", temperature: 0.7 },
     });
 
@@ -700,7 +700,7 @@ describe("RetryProvider — thinking/temperature conflict guard", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent", temperature: 0.7 },
     });
 
@@ -730,7 +730,7 @@ describe("RetryProvider — thinking/temperature conflict guard", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { callSite: "mainAgent", temperature: 0.7 },
     });
 
@@ -759,7 +759,7 @@ describe("RetryProvider — thinking/temperature conflict guard", () => {
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: {
         callSite: "trustRuleSuggestion",
         temperature: 0.7,
@@ -794,7 +794,7 @@ describe("RetryProvider — no callSite (pre-resolved config passes through)", (
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: { model: "explicit-model", max_tokens: 1234 },
     });
 
@@ -817,7 +817,7 @@ describe("RetryProvider — no callSite (pre-resolved config passes through)", (
       }),
     );
 
-    await wrapped.sendMessage(DUMMY_MESSAGES, undefined, undefined, {
+    await wrapped.sendMessage(DUMMY_MESSAGES, {
       config: {
         model: "explicit-model",
         usageAttributionHeaders: {

package/src/providers/__tests__/satellite-connection-routing.test.ts

@@ -216,9 +216,7 @@ describe("CallSiteRoutingProvider honors provider_connection (satellite gate)",
 
     const response = await wrapped.sendMessage(
       [{ role: "user", content: [{ type: "text", text: "hello" }] }],
-      [],
-      undefined,
-      { config: { callSite: "replySuggestion" } },
+      { tools: [], config: { callSite: "replySuggestion" } },
     );
 
     // Hard gate #1: connection-resolution hook fired with the right name.
@@ -261,9 +259,7 @@ describe("CallSiteRoutingProvider honors provider_connection (satellite gate)",
 
     await wrapped.sendMessage(
       [{ role: "user", content: [{ type: "text", text: "hello" }] }],
-      [],
-      undefined,
-      { config: { callSite: "memoryRetrieval" } },
+      { tools: [], config: { callSite: "memoryRetrieval" } },
     );
 
     // No connection lookup attempted at all.
@@ -303,9 +299,7 @@ describe("CallSiteRoutingProvider honors provider_connection (satellite gate)",
     try {
       await wrapped.sendMessage(
         [{ role: "user", content: [{ type: "text", text: "hello" }] }],
-        [],
-        undefined,
-        { config: { callSite: "memoryRetrieval" } },
+        { tools: [], config: { callSite: "memoryRetrieval" } },
       );
     } catch (err) {
       caught = err;
@@ -346,9 +340,7 @@ describe("CallSiteRoutingProvider honors provider_connection (satellite gate)",
     try {
       await wrapped.sendMessage(
         [{ role: "user", content: [{ type: "text", text: "hello" }] }],
-        [],
-        undefined,
-        { config: { callSite: "conversationTitle" } },
+        { tools: [], config: { callSite: "conversationTitle" } },
       );
     } catch (err) {
       caught = err;
@@ -406,9 +398,7 @@ describe("CallSiteRoutingProvider honors provider_connection (satellite gate)",
     try {
       await wrapped.sendMessage(
         [{ role: "user", content: [{ type: "text", text: "hello" }] }],
-        [],
-        undefined,
-        { config: { callSite: "replySuggestion" } },
+        { tools: [], config: { callSite: "replySuggestion" } },
       );
     } catch (err) {
       caught = err;
@@ -467,9 +457,7 @@ describe("CallSiteRoutingProvider honors provider_connection (satellite gate)",
     // This MUST NOT throw — the resolve failure is contained.
     await wrapped.sendMessage(
       [{ role: "user", content: [{ type: "text", text: "hello" }] }],
-      [],
-      undefined,
-      { config: { callSite: "replySuggestion" } },
+      { tools: [], config: { callSite: "replySuggestion" } },
     );
 
     // The resolver DID fire (we got past the connection lookup + validation).
@@ -498,9 +486,7 @@ describe("CallSiteRoutingProvider honors provider_connection (satellite gate)",
 
     await wrapped.sendMessage(
       [{ role: "user", content: [{ type: "text", text: "hello" }] }],
-      [],
-      undefined,
-      {},
+      { tools: [] },
     );
 
     expect(resolveProviderCalls.length).toBe(0);

package/src/providers/anthropic/client.ts

@@ -12,7 +12,6 @@ import type {
   Provider,
   ProviderResponse,
   SendMessageOptions,
-  ToolDefinition,
 } from "../types.js";
 import {
   ContextOverflowError,
@@ -809,11 +808,9 @@ export class AnthropicProvider implements Provider {
 
   async sendMessage(
     messages: Message[],
-    tools?: ToolDefinition[],
-    systemPrompt?: string,
     options?: SendMessageOptions,
   ): Promise<ProviderResponse> {
-    const { config, onEvent, signal } = options ?? {};
+    const { tools, systemPrompt, config, onEvent, signal } = options ?? {};
     const cacheTtl: "5m" | "1h" =
       ((config as Record<string, unknown> | undefined)?.cacheTtl as
         | "5m"
@@ -827,6 +824,13 @@ export class AnthropicProvider implements Provider {
     const disableTurnStartCache =
       (config as Record<string, unknown> | undefined)?.disableTurnStartCache ===
       true;
+    // When true, the latest user message carries per-turn-volatile content
+    // (e.g. an injected memory block), so the long-TTL anchor must skip it and
+    // land on the most recent stable user message instead. See the breakpoint
+    // block below.
+    const mutableLatestUserMessage =
+      (config as Record<string, unknown> | undefined)
+        ?.mutableLatestUserMessage === true;
     let sentMessages: Anthropic.MessageParam[] | undefined;
     const startedAt = Date.now();
     // Hoisted so the catch block can distinguish our inner stream timeout
@@ -1015,6 +1019,7 @@ export class AnthropicProvider implements Provider {
         output_config,
         cacheTtl: _cacheTtl,
         disableTurnStartCache: _disableTurnStartCache,
+        mutableLatestUserMessage: _mutableLatestUserMessage,
         max_tokens: callerMaxTokens,
         usageAttributionHeaders,
         ...restConfig
@@ -1145,7 +1150,21 @@ export class AnthropicProvider implements Provider {
         }
       };
       const turnStartIdx = findUserTextMsgIdx(msgs.length - 1);
-      if (turnStartIdx >= 0 && !disableTurnStartCache) {
+      // When the latest user message is volatile (`mutableLatestUserMessage`)
+      // and it is itself the turn-start (first-of-turn, no tool-use loop yet),
+      // skip the long-TTL anchor here — it would land on per-turn-varying
+      // content and never hit again. The previous-turn anchor below becomes the
+      // primary stable breakpoint. During tool-use loops (`turnStartIdx <
+      // msgs.length - 1`) the turn-start block is fixed within the turn, so
+      // behavior is unchanged. Independent from `disableTurnStartCache`, which
+      // expresses a different intent (one-shot callers with no future hit).
+      const skipVolatileTurnStartAnchor =
+        mutableLatestUserMessage && turnStartIdx === msgs.length - 1;
+      if (
+        turnStartIdx >= 0 &&
+        !disableTurnStartCache &&
+        !skipVolatileTurnStartAnchor
+      ) {
         applyCacheControlToLastBlock(turnStartIdx);
       }
 

package/src/providers/call-site-routing.ts

@@ -39,7 +39,6 @@ import type {
   Provider,
   ProviderResponse,
   SendMessageOptions,
-  ToolDefinition,
 } from "./types.js";
 
 export class CallSiteRoutingProvider implements Provider {
@@ -93,20 +92,13 @@ export class CallSiteRoutingProvider implements Provider {
 
   async sendMessage(
     messages: Message[],
-    tools?: ToolDefinition[],
-    systemPrompt?: string,
     options?: SendMessageOptions,
   ): Promise<ProviderResponse> {
     const target = await this.selectProvider(options);
     const isRouted = target !== this.defaultProvider;
 
     const doSend = async (): Promise<ProviderResponse> => {
-      const response = await target.sendMessage(
-        messages,
-        tools,
-        systemPrompt,
-        options,
-      );
+      const response = await target.sendMessage(messages, options);
       // Also stamp actualProvider on the response so that handleUsage /
       // llm_call_finished (which read event.actualProvider, not provider.name)
       // attribute the call to the right provider.