@vellumai/assistant 0.10.3-dev.202606292252.adea010 → 0.10.3-dev.202606300318.cf23f52

package/openapi.yaml

@@ -8527,6 +8527,49 @@ paths:
                                     - type: "null"
                               additionalProperties: false
                             - type: "null"
+                        latency:
+                          anyOf:
+                            - type: object
+                              properties:
+                                phases:
+                                  type: array
+                                  items:
+                                    type: object
+                                    properties:
+                                      key:
+                                        type: string
+                                      label:
+                                        type: string
+                                      ms:
+                                        type: number
+                                    required:
+                                      - key
+                                      - label
+                                      - ms
+                                    additionalProperties: false
+                                ttftMs:
+                                  anyOf:
+                                    - type: number
+                                    - type: "null"
+                                totalToFirstTokenMs:
+                                  anyOf:
+                                    - type: number
+                                    - type: "null"
+                                providerDurationMs:
+                                  anyOf:
+                                    - type: number
+                                    - type: "null"
+                                firstTokenKind:
+                                  anyOf:
+                                    - type: string
+                                      enum:
+                                        - thinking
+                                        - text
+                                    - type: "null"
+                              required:
+                                - phases
+                              additionalProperties: false
+                            - type: "null"
                       required:
                         - id
                         - createdAt
@@ -15521,6 +15564,49 @@ paths:
                               - type: "null"
                         additionalProperties: false
                       - type: "null"
+                  latency:
+                    anyOf:
+                      - type: object
+                        properties:
+                          phases:
+                            type: array
+                            items:
+                              type: object
+                              properties:
+                                key:
+                                  type: string
+                                label:
+                                  type: string
+                                ms:
+                                  type: number
+                              required:
+                                - key
+                                - label
+                                - ms
+                              additionalProperties: false
+                          ttftMs:
+                            anyOf:
+                              - type: number
+                              - type: "null"
+                          totalToFirstTokenMs:
+                            anyOf:
+                              - type: number
+                              - type: "null"
+                          providerDurationMs:
+                            anyOf:
+                              - type: number
+                              - type: "null"
+                          firstTokenKind:
+                            anyOf:
+                              - type: string
+                                enum:
+                                  - thinking
+                                  - text
+                              - type: "null"
+                        required:
+                          - phases
+                        additionalProperties: false
+                      - type: "null"
                 required:
                   - id
                   - createdAt
@@ -18145,6 +18231,49 @@ paths:
                                     - type: "null"
                               additionalProperties: false
                             - type: "null"
+                        latency:
+                          anyOf:
+                            - type: object
+                              properties:
+                                phases:
+                                  type: array
+                                  items:
+                                    type: object
+                                    properties:
+                                      key:
+                                        type: string
+                                      label:
+                                        type: string
+                                      ms:
+                                        type: number
+                                    required:
+                                      - key
+                                      - label
+                                      - ms
+                                    additionalProperties: false
+                                ttftMs:
+                                  anyOf:
+                                    - type: number
+                                    - type: "null"
+                                totalToFirstTokenMs:
+                                  anyOf:
+                                    - type: number
+                                    - type: "null"
+                                providerDurationMs:
+                                  anyOf:
+                                    - type: number
+                                    - type: "null"
+                                firstTokenKind:
+                                  anyOf:
+                                    - type: string
+                                      enum:
+                                        - thinking
+                                        - text
+                                    - type: "null"
+                              required:
+                                - phases
+                              additionalProperties: false
+                            - type: "null"
                       required:
                         - id
                         - createdAt
@@ -20315,6 +20444,10 @@ paths:
                     type: boolean
                   hasWebhookSecret:
                     type: boolean
+                  clientInstallationId:
+                    anyOf:
+                      - type: string
+                      - type: "null"
                   available:
                     type: boolean
                   organizationId:
@@ -20346,6 +20479,7 @@ paths:
                   - assistantId
                   - hasAssistantApiKey
                   - hasWebhookSecret
+                  - clientInstallationId
                   - available
                   - organizationId
                   - userId

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vellumai/assistant",
-  "version": "0.10.3-dev.202606292252.adea010",
+  "version": "0.10.3-dev.202606300318.cf23f52",
   "license": "MIT",
   "type": "module",
   "exports": {

package/src/__tests__/agent-loop-compaction-events.test.ts

@@ -150,6 +150,7 @@ describe("AgentLoop compaction start/end event pair", () => {
       onEvent: (event) => {
         events.push(event);
       },
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: true, safetyMarginRatio: 0 },

package/src/__tests__/agent-loop-compaction-strip.test.ts

@@ -167,6 +167,7 @@ describe("AgentLoop compaction summarizer input", () => {
       onEvent: (event) => {
         events.push(event);
       },
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: true, safetyMarginRatio: 0 },
@@ -221,6 +222,7 @@ describe("AgentLoop compaction summarizer input", () => {
       onEvent: (event) => {
         events.push(event);
       },
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: true, safetyMarginRatio: 0 },

package/src/__tests__/agent-loop-exit-reason.test.ts

@@ -460,6 +460,7 @@ describe("AgentLoop exit-reason instrumentation", () => {
       messages: [userMessage],
       onEvent: () => {},
       trust: { sourceChannel: "vellum", trustClass: "unknown" },
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: false, safetyMarginRatio: 0 },
@@ -500,6 +501,7 @@ describe("AgentLoop exit-reason instrumentation", () => {
       onEvent: (event) => {
         events.push(event);
       },
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: true, safetyMarginRatio: 0 },

package/src/__tests__/agent-loop-regrowth-guard.test.ts

@@ -148,6 +148,7 @@ async function runOnce(args: {
     onEvent: (event) => {
       events.push(event);
     },
+    modelProfileKey: "balanced",
     resolveContextWindow: () => ({
       maxInputTokens: 10,
       overflowRecovery: { enabled: true, safetyMarginRatio: 0 },
@@ -215,6 +216,7 @@ describe("AgentLoop budget-gate regrowth guard", () => {
       onEvent: (event) => {
         events.push(event);
       },
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: true, safetyMarginRatio: 0 },
@@ -249,6 +251,7 @@ describe("AgentLoop budget-gate regrowth guard", () => {
       requestId: "req",
       messages: [userMessage],
       onEvent: () => {},
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: true, safetyMarginRatio: 0 },
@@ -310,6 +313,7 @@ describe("AgentLoop budget-gate per-turn proactive-futility suppression", () =>
       onEvent: (event) => {
         events.push(event);
       },
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: true, safetyMarginRatio: 0 },
@@ -369,6 +373,7 @@ describe("AgentLoop budget-gate per-turn proactive-futility suppression", () =>
       onEvent: (e) => {
         events.push(e);
       },
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: true, safetyMarginRatio: 0 },
@@ -411,6 +416,7 @@ describe("AgentLoop budget-gate per-turn proactive-futility suppression", () =>
       onEvent: (e) => {
         turn1.push(e);
       },
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: true, safetyMarginRatio: 0 },
@@ -428,6 +434,7 @@ describe("AgentLoop budget-gate per-turn proactive-futility suppression", () =>
       onEvent: (e) => {
         turn2.push(e);
       },
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: true, safetyMarginRatio: 0 },
@@ -483,6 +490,7 @@ describe("AgentLoop budget-gate per-turn proactive-futility suppression", () =>
       onEvent: (e) => {
         events.push(e);
       },
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens: 10,
         overflowRecovery: { enabled: true, safetyMarginRatio: 0 },

package/src/__tests__/agent-loop.test.ts

@@ -204,6 +204,7 @@ describe("AgentLoop", () => {
       trust: { sourceChannel: "vellum", trustClass: "unknown" },
       requestId: "req-1",
       callSite: "mainAgent",
+      modelProfileKey: "balanced",
       resolveContextWindow: () => ({
         maxInputTokens,
         overflowRecovery: { enabled: false, safetyMarginRatio: 0 },

package/src/__tests__/conversation-agent-loop.test.ts

@@ -32,6 +32,8 @@ const conversationDiskViewRealSnapshot = {
   ) as Record<string, unknown>),
 };
 let mockUiConfig: { userTimezone?: string; detectedTimezone?: string } = {};
+let mockLlmProfiles: Record<string, unknown> = {};
+let mockLlmActiveProfile: string | undefined;
 
 // ── Module mocks (must precede imports of the module under test) ─────
 
@@ -80,8 +82,9 @@ mock.module("../config/loader.js", () => ({
           },
         },
       },
-      profiles: {},
+      profiles: mockLlmProfiles,
       callSites: {},
+      activeProfile: mockLlmActiveProfile,
       pricingOverrides: [],
     },
     rateLimit: { maxRequestsPerMinute: 0 },
@@ -835,6 +838,8 @@ function makeCompactionResult(
 
 beforeEach(() => {
   mockUiConfig = {};
+  mockLlmProfiles = {};
+  mockLlmActiveProfile = undefined;
   mockEstimateTokens = 1000;
   mockReducerStepFn = null;
   mockOverflowAction = "fail_gracefully";
@@ -903,6 +908,39 @@ beforeEach(() => {
 
 describe("session-agent-loop", () => {
   describe("user-prompt-submit hook failures", () => {
+    test("passes the effective profile to hooks even when it was already announced", async () => {
+      mockLlmProfiles = {
+        balanced: {
+          label: "Balanced",
+          model: "accounts/fireworks/models/glm-5p2",
+        },
+        quality: { label: "Quality", model: "claude-opus-4-8" },
+      };
+      mockLlmActiveProfile = "quality";
+      const observedProfileKeys: string[] = [];
+      registerPlugin({
+        manifest: {
+          name: "test-observe-model-profile",
+          version: "1.0.0",
+        },
+        hooks: {
+          "user-prompt-submit": async (ctx: UserPromptSubmitContext) => {
+            observedProfileKeys.push(ctx.modelProfileKey);
+          },
+        },
+      });
+
+      const ctx = makeCtx({
+        inferenceProfile: "balanced",
+        lastNotifiedInferenceProfile: "balanced",
+        providerResponses: [textResponse("ok")],
+      });
+
+      await runAgentLoopImpl(ctx, "hello", "msg-1", () => {});
+
+      expect(observedProfileKeys).toEqual(["balanced"]);
+    });
+
     test("logs and continues with prior hook mutations", async () => {
       registerPlugin({
         manifest: {

package/src/__tests__/conversation-runtime-assembly.test.ts

@@ -1104,7 +1104,7 @@ describe("applyRuntimeInjections — injection mode", () => {
       requestId: "reinject-req",
       conversationId: "injection-mode-conv",
       isNonInteractive: false,
-      modelProfileKey: null,
+      modelProfileKey: "balanced",
     };
     await postCompact(postCompactCtx);
     const result = postCompactCtx.history;
@@ -1628,16 +1628,16 @@ describe("resolveTurnInboundActorContext", () => {
 
 describe("resolveTurnModelProfileLabel", () => {
   /**
-   * A null key means the active profile is unchanged since the last notified
-   * one, so there is no `model_profile` line to render this turn.
+   * A null notice key means there is no `model_profile` line to render this
+   * turn.
    */
-  test("returns null when the profile key is null", () => {
-    // GIVEN a turn whose profile is unchanged since the last notification
+  test("returns null when the profile notice key is null", () => {
+    // GIVEN a turn without a profile notice to render
     const llm = LLMSchema.parse({
       default: { provider: "anthropic", model: "claude-sonnet-4-7" },
     });
 
-    // WHEN the label is resolved for a null key
+    // WHEN the label is resolved for a null notice key
     const label = resolveTurnModelProfileLabel(null, "mainAgent", llm);
 
     // THEN there is no profile line to inject

package/src/__tests__/device-id.test.ts

@@ -19,7 +19,11 @@ mock.module("../util/logger.js", () => ({
     }),
 }));
 
-import { getDeviceId, resetDeviceIdCache } from "../util/device-id.js";
+import {
+  getDeviceId,
+  getExistingDeviceId,
+  resetDeviceIdCache,
+} from "../util/device-id.js";
 
 const originalVellumEnvironment = process.env.VELLUM_ENVIRONMENT;
 const originalXdgConfigHome = process.env.XDG_CONFIG_HOME;
@@ -179,3 +183,36 @@ describe("getDeviceId VELLUM_DEVICE_ID precedence", () => {
     expect(getDeviceId()).not.toBe("env-id");
   });
 });
+
+describe("getExistingDeviceId", () => {
+  beforeEach(() => {
+    delete process.env.IS_CONTAINERIZED;
+    process.env.VELLUM_ENVIRONMENT = "dev";
+    process.env.XDG_CONFIG_HOME = tempDir;
+  });
+
+  test("returns null without creating device.json when no id exists", () => {
+    const expectedPath = join(tempDir, "vellum-dev", "device.json");
+
+    expect(getExistingDeviceId()).toBeNull();
+    expect(existsSync(expectedPath)).toBe(false);
+  });
+
+  test("reads an existing device.json id", () => {
+    const dir = join(tempDir, "vellum-dev");
+    mkdirSync(dir, { recursive: true });
+    writeFileSync(
+      join(dir, "device.json"),
+      JSON.stringify({ deviceId: "file-id" }),
+    );
+
+    expect(getExistingDeviceId()).toBe("file-id");
+  });
+
+  test("returns the env override without writing device.json", () => {
+    process.env.VELLUM_DEVICE_ID = "env-id";
+
+    expect(getExistingDeviceId()).toBe("env-id");
+    expect(readdirSync(tempDir)).toEqual([]);
+  });
+});

package/src/__tests__/history-repair-hook.test.ts

@@ -72,7 +72,7 @@ function makeCtx(messages: Message[]): UserPromptSubmitContext {
     conversationId: "conv-test",
     userMessageId: "msg-test",
     requestId: "req-test",
-    modelProfileKey: null,
+    modelProfileKey: "balanced",
     isNonInteractive: false,
     prompt: "",
     originalMessages: messages,

package/src/__tests__/llm-request-log-source-clickhouse.test.ts

@@ -106,6 +106,7 @@ describe("ClickHouseLlmRequestLogSource", () => {
       createdAt: 1778465138786,
       agentLoopExitReason: "no_tool_calls",
       callSite: "mainAgent",
+      latencyBreakdown: null,
     });
   });
 

package/src/__tests__/memory-retrieval-hook.test.ts

@@ -158,7 +158,7 @@ function makeHookCtx(
     latestMessages: [],
     requestId: "req-test",
     isNonInteractive: false,
-    modelProfileKey: null,
+    modelProfileKey: "balanced",
     prompt: "",
     originalMessages: [],
     ...overrides,

package/src/__tests__/plugin-import-boundary-guard.test.ts

@@ -130,6 +130,8 @@ const BASELINE: Record<string, readonly string[]> = {
     "../../../daemon/pkb-context-tracker.js",
     "../../../daemon/pkb-reminder-builder.js",
     "../../../daemon/trust-context.js",
+    "../../../memory/context-search/agent-runner.js",
+    "../../../memory/context-search/types.js",
     "../../../memory/graph/bootstrap.js",
     "../../../memory/graph/consolidation.js",
     "../../../memory/graph/conversation-graph-memory.js",
@@ -138,6 +140,8 @@ const BASELINE: Record<string, readonly string[]> = {
     "../../../memory/graph/graph-search.js",
     "../../../memory/graph/narrative.js",
     "../../../memory/graph/pattern-scan.js",
+    "../../../memory/graph/tool-handlers.js",
+    "../../../memory/graph/tools.js",
     "../../../memory/job-handlers/backfill.js",
     "../../../memory/job-handlers/embedding.js",
     "../../../memory/job-handlers/index-maintenance.js",
@@ -152,7 +156,10 @@ const BASELINE: Record<string, readonly string[]> = {
     "../../../memory/v2/consolidation-job.js",
     "../../../memory/v2/static-context.js",
     "../../../memory/v2/sweep-job.js",
+    "../../../permissions/types.js",
     "../../../persistence/jobs-store.js",
+    "../../../runtime/capabilities.js",
+    "../../../tools/types.js",
     "../../../types.js",
     "../../../util/logger.js",
     "../../../util/platform.js",

package/src/__tests__/title-generate-hook.test.ts

@@ -109,7 +109,7 @@ function makeCtx(
     conversationId: "conv-1",
     userMessageId: "msg-1",
     requestId: "req-1",
-    modelProfileKey: null,
+    modelProfileKey: "balanced",
     isNonInteractive: false,
     prompt: "first message",
     originalMessages: messages,