@vellumai/assistant 0.11.1 → 0.11.2-staging.2

package/openapi.yaml

@@ -3,7 +3,7 @@
 openapi: 3.1.0
 info:
   title: Vellum Assistant API
-  version: 0.11.1
+  version: 0.11.2
   description: Auto-generated OpenAPI specification for the Vellum Assistant runtime HTTP server.
 servers:
   - url: http://127.0.0.1:7821
@@ -699,6 +699,52 @@ paths:
                 required:
                   - ok
                 additionalProperties: false
+  /v1/admin/workspace-compact-history:
+    post:
+      operationId: admin_workspacecompacthistory_post
+      summary: Compact workspace git history
+      description:
+        Squash workspace git history older than the retention window into a single base commit and prune reclaimed
+        objects. By default acts only when an oversized blob is reclaimable; with force, the squash runs
+        unconditionally.
+      tags:
+        - admin
+      requestBody:
+        required: true
+        content:
+          application/json:
+            schema:
+              type: object
+              properties:
+                force:
+                  description: Squash aged history even when no oversized blob is reclaimable
+                  type: boolean
+                retentionDays:
+                  description: Retention window override in days (default 7)
+                  type: integer
+                  exclusiveMinimum: 0
+                  maximum: 9007199254740991
+      responses:
+        "200":
+          description: Successful response
+          content:
+            application/json:
+              schema:
+                type: object
+                properties:
+                  rewrote:
+                    type: boolean
+                  squashedCommits:
+                    type: number
+                  keptCommits:
+                    type: number
+                  retryAfterMs:
+                    type: number
+                required:
+                  - rewrote
+                  - squashedCommits
+                  - keptCommits
+                additionalProperties: false
   /v1/apps:
     get:
       operationId: apps_get
@@ -10665,6 +10711,11 @@ paths:
                     type: number
                   updatedAt:
                     type: number
+                  workspacePath:
+                    anyOf:
+                      - type: string
+                      - type: "null"
+                    description: Workspace-relative file this document is bound to, or null
                 required:
                   - success
                   - surfaceId
@@ -10674,6 +10725,7 @@ paths:
                   - wordCount
                   - createdAt
                   - updatedAt
+                  - workspacePath
                 additionalProperties: false
   /v1/documents/{id}/comments:
     get:
@@ -11091,6 +11143,74 @@ paths:
               schema:
                 type: string
                 format: binary
+  /v1/documents/for-workspace-file:
+    post:
+      operationId: documents_forworkspacefile_post
+      summary: Get or create the document for a workspace markdown file
+      description:
+        "Return the document bound to a workspace markdown file, creating it on first open. The file is the source
+        of truth at open time: stored content that has drifted from the file is refreshed from disk before the document
+        is returned."
+      tags:
+        - documents
+      requestBody:
+        required: true
+        content:
+          application/json:
+            schema:
+              type: object
+              properties:
+                path:
+                  type: string
+                  description: Workspace-relative path to a .md/.markdown file
+                conversationId:
+                  type: string
+                  description: Conversation the document is opened from
+              required:
+                - path
+                - conversationId
+      responses:
+        "200":
+          description: Successful response
+          content:
+            application/json:
+              schema:
+                type: object
+                properties:
+                  success:
+                    type: boolean
+                  surfaceId:
+                    type: string
+                  conversationId:
+                    type: string
+                  title:
+                    type: string
+                  content:
+                    type: string
+                  wordCount:
+                    type: number
+                  createdAt:
+                    type: number
+                  updatedAt:
+                    type: number
+                  workspacePath:
+                    anyOf:
+                      - type: string
+                      - type: "null"
+                    description: Workspace-relative file this document is bound to, or null
+                required:
+                  - success
+                  - surfaceId
+                  - conversationId
+                  - title
+                  - content
+                  - wordCount
+                  - createdAt
+                  - updatedAt
+                  - workspacePath
+                additionalProperties: false
+        "422":
+          description: Path is not a markdown file
   /v1/domain/register:
     post:
       operationId: domain_register_post
@@ -14633,8 +14753,9 @@ paths:
       summary: Create a provider connection
       description:
         Create a new named provider connection. When auth is omitted it is derived from the provider (keyless
-        providers get none, vellum gets platform, everything else needs credential for api_key auth). Fails with 409 if
-        a connection with this name already exists.
+        providers get none, vellum gets platform, everything else needs credential for api_key auth). An explicit auth
+        object must agree with the provider; platform auth belongs to vellum and only to vellum. Fails with 409 if a
+        connection with this name already exists.
       tags:
         - inference
       requestBody:
@@ -14743,8 +14864,9 @@ paths:
       summary: Update a provider connection
       description:
         Update an existing connection. Cannot rename or change the provider. Omitting auth keeps the stored auth;
-        passing credential alone rotates the key via provider-derived api_key auth. For the Vellum-managed connection
-        (vellum) the auth is locked to platform; label remains editable.
+        passing credential alone rotates the key via provider-derived api_key auth. An explicit auth object must agree
+        with the connection's provider; platform auth belongs to vellum and only to vellum. For the Vellum-managed
+        connection (vellum) the auth is locked to platform; label remains editable.
       tags:
         - inference
       parameters:
@@ -19543,6 +19665,14 @@ paths:
                           additionalProperties: false
                         systemCard:
                           type: boolean
+                        providerError:
+                          type: object
+                          properties:
+                            code:
+                              type: string
+                            category:
+                              type: string
+                          additionalProperties: false
                         slackMessage:
                           type: object
                           properties:
@@ -19709,6 +19839,13 @@ paths:
                     Client OS surface ("web" | "ios" | "macos" | "android"), reported separately from `interface`. Drives the
                     per-turn `client_os` context only; does not affect transport/host-proxy capabilities.
                   type: string
+                visibleAppId:
+                  description:
+                    'Id of the app the client currently has open on screen (app viewer or the app-editing split). Drives the
+                    per-turn `visible_app:` context line so the assistant can resolve "the app" to what the user is
+                    looking at. View state only: it never affects transport, routing, or tool gating, and is omitted
+                    whenever no app is in view.'
+                  type: string
                 clientMessageId:
                   type: string
                   description:
@@ -19752,6 +19889,16 @@ paths:
                     snapshots, emits no echo, and does not supersede pending interactions. Honored on the standard send
                     path only — slash-command content bypasses it."
                   type: boolean
+                scripted:
+                  description:
+                    "When true, this turn was auto-sent on the user's behalf rather than typed by them: onboarding research
+                    prompts, the personality rewrite message, research corrections, hidden kickoff greetings, the legacy
+                    pre-chat bootstrap. Stamped onto the persisted message and forwarded to turn telemetry, where
+                    activation metrics exclude it. Send false for a genuine typed message; OMIT the field only if the
+                    client genuinely cannot tell, since absent means UNKNOWN and a wrong false is trusted downstream.
+                    Independent of `hidden`: a turn can be visible and scripted (the research prompt) or hidden and
+                    scripted (the kickoff greeting)."
+                  type: boolean
                 onboarding:
                   type: object
                   properties:
@@ -21372,6 +21519,7 @@ paths:
                   minLength: 1
                   description: Feed item id (notif:<uuid>) or bare uuid
                 title:
+                  description: New title. An empty value is ignored, never cleared.
                   type: string
                 body:
                   type: string
@@ -22765,6 +22913,145 @@ paths:
                   - disconnected
                   - previousBaseUrl
                 additionalProperties: false
+  /v1/platform/invoices:
+    get:
+      operationId: platform_invoices_get
+      summary: List one page of the organization's Stripe invoices
+      description:
+        Fetches one page of the org's Stripe invoice history (newest first) from the platform billing invoices
+        endpoint. Amounts are in the currency's minor units. When has_more is true, pass the last invoice's id as
+        starting_after to fetch the next page.
+      tags:
+        - platform
+      parameters:
+        - name: starting_after
+          in: query
+          required: false
+          schema:
+            type: string
+          description: "Cursor: return invoices older than the invoice with this id (from the previous page's last entry)."
+      responses:
+        "200":
+          description: Successful response
+          content:
+            application/json:
+              schema:
+                type: object
+                properties:
+                  invoices:
+                    type: array
+                    items:
+                      type: object
+                      properties:
+                        id:
+                          type: string
+                        number:
+                          anyOf:
+                            - type: string
+                            - type: "null"
+                        status:
+                          anyOf:
+                            - type: string
+                            - type: "null"
+                        currency:
+                          type: string
+                        amount_due:
+                          type: number
+                        amount_paid:
+                          type: number
+                        amount_remaining:
+                          type: number
+                        created:
+                          type: number
+                        hosted_invoice_url:
+                          anyOf:
+                            - type: string
+                            - type: "null"
+                        invoice_pdf:
+                          anyOf:
+                            - type: string
+                            - type: "null"
+                      required:
+                        - id
+                        - number
+                        - status
+                        - currency
+                        - amount_due
+                        - amount_paid
+                        - amount_remaining
+                        - created
+                        - hosted_invoice_url
+                        - invoice_pdf
+                      additionalProperties: false
+                  has_more:
+                    type: boolean
+                required:
+                  - invoices
+                  - has_more
+                additionalProperties: false
+  /v1/platform/invoices/{id}:
+    get:
+      operationId: platform_invoices_by_id_get
+      summary: Get a single Stripe invoice by ID
+      description:
+        Pages through the org's invoice list from the platform and returns the invoice matching the given Stripe
+        invoice ID (e.g. in_xxx). 404 if no such invoice.
+      tags:
+        - platform
+      parameters:
+        - name: id
+          in: path
+          required: true
+          schema:
+            type: string
+      responses:
+        "200":
+          description: Successful response
+          content:
+            application/json:
+              schema:
+                type: object
+                properties:
+                  id:
+                    type: string
+                  number:
+                    anyOf:
+                      - type: string
+                      - type: "null"
+                  status:
+                    anyOf:
+                      - type: string
+                      - type: "null"
+                  currency:
+                    type: string
+                  amount_due:
+                    type: number
+                  amount_paid:
+                    type: number
+                  amount_remaining:
+                    type: number
+                  created:
+                    type: number
+                  hosted_invoice_url:
+                    anyOf:
+                      - type: string
+                      - type: "null"
+                  invoice_pdf:
+                    anyOf:
+                      - type: string
+                      - type: "null"
+                required:
+                  - id
+                  - number
+                  - status
+                  - currency
+                  - amount_due
+                  - amount_paid
+                  - amount_remaining
+                  - created
+                  - hosted_invoice_url
+                  - invoice_pdf
+                additionalProperties: false
   /v1/platform/plans:
     get:
       operationId: platform_plans_get

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vellumai/assistant",
-  "version": "0.11.1",
+  "version": "0.11.2-staging.2",
   "license": "MIT",
   "type": "module",
   "exports": {

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

@@ -0,0 +1,223 @@
+/**
+ * Cover for the loop's interrupted-call recovery.
+ *
+ * The reported failure: the model runs several tools, the next call streams for
+ * a while and then dies (the upstream killed a generation that blew its decode
+ * deadline), and the loop surfaces the error and ends the turn — leaving the
+ * task half-done until the user types "continue". These tests assert the loop
+ * now carries on by itself, and that it still stops when carrying on would be
+ * pointless.
+ *
+ * The recovery is built into {@link AgentLoop}, not contributed by a plugin, so
+ * these run it as the loop's own behavior. The default plugin stack is
+ * registered because the loop fires hook chains during a turn regardless.
+ */
+
+import { beforeEach, describe, expect, test } from "bun:test";
+
+import type { AgentEvent } from "../agent/loop.js";
+import { AgentLoop } from "../agent/loop.js";
+import type { LLMCallSite } from "../config/schemas/llm.js";
+import { resetPluginRegistryAndRegisterDefaults } from "../plugins/defaults/index.js";
+import type {
+  Message,
+  Provider,
+  ProviderResponse,
+  SendMessageOptions,
+  ToolDefinition,
+} from "../providers/types.js";
+import { ProviderError } from "../util/errors.js";
+
+const tools: ToolDefinition[] = [
+  {
+    name: "read_file",
+    description: "Read a file",
+    input_schema: { type: "object", properties: { path: { type: "string" } } },
+  },
+];
+
+const userMessage: Message = {
+  role: "user",
+  content: [{ type: "text", text: "finish the course page" }],
+};
+
+/** The rejection from the reported turn, thrown from inside a live stream. */
+function midStreamRejection(): ProviderError {
+  return new ProviderError(
+    "OpenAI-compatible API error (unknown status): litellm.MidStreamFallbackError: " +
+      "litellm.APIConnectionError: OpenAIException - Decode wall clock timeout after 600s",
+    "openai-compatible",
+  );
+}
+
+/** A scripted turn: return a response, or stream `thinking` and then throw. */
+type Turn =
+  | { kind: "reply"; response: ProviderResponse }
+  | { kind: "interrupt"; thinking: string }
+  | { kind: "refuse" };
+
+function scriptedProvider(turns: Turn[]): {
+  provider: Provider;
+  callCount: () => number;
+} {
+  let index = 0;
+  const provider: Provider = {
+    name: "openai-compatible",
+    async sendMessage(
+      _messages: Message[],
+      options?: SendMessageOptions,
+    ): Promise<ProviderResponse> {
+      const turn = turns[Math.min(index, turns.length - 1)]!;
+      index++;
+      if (turn.kind === "refuse") {
+        // Rejected before generating: nothing ever streams.
+        throw midStreamRejection();
+      }
+      if (turn.kind === "interrupt") {
+        options?.onEvent?.({ type: "thinking_delta", thinking: turn.thinking });
+        throw midStreamRejection();
+      }
+      for (const block of turn.response.content) {
+        if (block.type === "text") {
+          options?.onEvent?.({ type: "text_delta", text: block.text });
+        }
+      }
+      return turn.response;
+    },
+  };
+  return { provider, callCount: () => index };
+}
+
+function reply(text: string): Turn {
+  return {
+    kind: "reply",
+    response: {
+      content: [{ type: "text", text }],
+      model: "glm-5.2",
+      usage: { inputTokens: 10, outputTokens: 5 },
+      stopReason: "end_turn",
+    },
+  };
+}
+
+function callsTool(id: string): Turn {
+  return {
+    kind: "reply",
+    response: {
+      content: [
+        { type: "tool_use", id, name: "read_file", input: { path: "app.js" } },
+      ],
+      model: "glm-5.2",
+      usage: { inputTokens: 10, outputTokens: 5 },
+      stopReason: "tool_use",
+    },
+  };
+}
+
+function runLoop(provider: Provider, callSite: LLMCallSite = "mainAgent") {
+  const events: AgentEvent[] = [];
+  const loop = new AgentLoop({
+    provider,
+    systemPrompt: "system",
+    conversationId: "conv-resume-e2e",
+    tools,
+    toolExecutor: async () => ({ content: "file data", isError: false }),
+  });
+  return loop
+    .run({
+      requestId: "req-resume-e2e",
+      messages: [userMessage],
+      callSite,
+      onEvent: (event) => {
+        events.push(event);
+      },
+      trust: { sourceChannel: "vellum", trustClass: "unknown" },
+    })
+    .then((result) => ({ ...result, events }));
+}
+
+function assistantTextOf(history: Message[]): string {
+  return history
+    .filter((message) => message.role === "assistant")
+    .flatMap((message) => message.content)
+    .map((block) => (block.type === "text" ? block.text : ""))
+    .join("");
+}
+
+describe("AgentLoop — a call interrupted mid-generation", () => {
+  beforeEach(() => {
+    resetPluginRegistryAndRegisterDefaults();
+  });
+
+  test("carries on instead of ending the turn", async () => {
+    const { provider, callCount } = scriptedProvider([
+      callsTool("call-1"),
+      { kind: "interrupt", thinking: "Now I need to understand the layout" },
+      reply("Done — the page is wired up."),
+    ]);
+
+    const { history, events } = await runLoop(provider);
+
+    // The turn finished on its own: tool call, interrupted call, resumed call.
+    expect(callCount()).toBe(3);
+    expect(assistantTextOf(history)).toContain("Done — the page is wired up.");
+    expect(events.filter((event) => event.type === "error")).toHaveLength(0);
+  });
+
+  test("keeps the tool work the interrupted turn had already done", async () => {
+    const { provider } = scriptedProvider([
+      callsTool("call-1"),
+      { kind: "interrupt", thinking: "reading the file" },
+      reply("Summary of the file."),
+    ]);
+
+    const { history } = await runLoop(provider);
+
+    // The resumed call re-sends the same history, so the completed tool call
+    // and its result are still there rather than being replayed or dropped.
+    const toolUses = history
+      .flatMap((message) => message.content)
+      .filter((block) => block.type === "tool_use");
+    const toolResults = history
+      .flatMap((message) => message.content)
+      .filter((block) => block.type === "tool_result");
+    expect(toolUses).toHaveLength(1);
+    expect(toolResults).toHaveLength(1);
+  });
+
+  test("surfaces the error when the resumed call is interrupted too", async () => {
+    const { provider, callCount } = scriptedProvider([
+      { kind: "interrupt", thinking: "starting" },
+    ]);
+
+    const { events } = await runLoop(provider);
+
+    // One resume, then the error stands rather than looping.
+    expect(callCount()).toBe(2);
+    expect(events.filter((event) => event.type === "error")).toHaveLength(1);
+  });
+
+  test("does not resume a background call site", async () => {
+    // A compaction, subagent, or other background call answers to a caller
+    // that owns its own failure handling; nobody is sitting there to type
+    // "continue" at it.
+    const { provider, callCount } = scriptedProvider([
+      { kind: "interrupt", thinking: "summarizing" },
+    ]);
+
+    await runLoop(provider, "compactionAgent");
+
+    expect(callCount()).toBe(1);
+  });
+
+  test("does not resume a request the provider refused outright", async () => {
+    // Nothing streamed, so re-sending the identical request would fail the same
+    // way; the error surfaces on the first rejection.
+    const { provider, callCount } = scriptedProvider([{ kind: "refuse" }]);
+
+    const { events } = await runLoop(provider);
+
+    expect(callCount()).toBe(1);
+    expect(events.filter((event) => event.type === "error")).toHaveLength(1);
+  });
+});

package/src/__tests__/byok-default-profile-ensure.test.ts

@@ -1092,4 +1092,47 @@ describe("ensureByokDefaultProfiles", () => {
 
     expect(readFileSync(configPath(), "utf-8")).toBe(before);
   });
+
+  // The hatch wrote each copy's model by resolving the cost profile's intent
+  // at hatch time. These are the ids that resolved per provider, pinned as
+  // literals rather than derived from the template: a test that materializes
+  // the template agrees with itself no matter which intent the template names,
+  // and so cannot catch a template whose model comparison has drifted off what
+  // is on disk.
+  test.each([
+    ["openrouter", "anthropic/claude-haiku-4.5"],
+    ["gemini", "gemini-3.1-flash-lite"],
+    ["openai", "gpt-5.6-luna"],
+  ] as const)(
+    "an unedited %s cost copy still converts",
+    (provider, hatchModel) => {
+      writeConfig({
+        llm: {
+          defaultProvider: { provider },
+          activeProfile: "custom-cost-optimized",
+          profiles: {
+            "custom-cost-optimized": {
+              ...(hatchBody("cost-optimized", provider) as Record<
+                string,
+                unknown
+              >),
+              model: hatchModel,
+            },
+            "custom-balanced": hatchBody("balanced", provider),
+            "custom-quality-optimized": hatchBody(
+              "quality-optimized",
+              provider,
+            ),
+          },
+        },
+      });
+
+      ensureByokDefaultProfiles(workspaceDir);
+
+      expect(profiles()["custom-cost-optimized"]).toBeUndefined();
+      expect(profiles()["cost-optimized"]).toBeUndefined();
+      expect(llm().activeProfile).toBe("cost-optimized");
+      expectSecondRunNoop();
+    },
+  );
 });