@vellumai/assistant 0.8.10-dev.202606092238.d04fd59 → 0.8.10

package/AGENTS.md

@@ -84,7 +84,7 @@ The rule exists because test machinery and production code have **inverted invar
 
 Concretely:
 
-- **Test helpers** (e.g. `src/__tests__/*-test-helpers.ts`) use only node stdlib, `bun:test`, and sibling helpers. Type-only imports (`import type`) of production types are permitted: they are erased at compile time and carry no import-time side effects, so the isolation invariant holds — prefer them over duplicating non-trivial production types structurally. If helpers need to manipulate shared state that production code also reads, both sides declare a typed slot under `globalThis.vellumAssistant.*` and read/write that slot independently. The slot shape is duplicated by design — the helper and the production module both reference the namespace, neither imports the other.
+- **Test helpers** (e.g. `src/__tests__/*-test-helpers.ts`) use only node stdlib, `bun:test`, and sibling helpers. If they need to manipulate shared state that production code also reads, both sides declare a typed slot under `globalThis.vellumAssistant.*` and read/write that slot independently. The slot shape is duplicated by design — the helper and the production module both reference the namespace, neither imports the other.
 - **The test preload** (`src/__tests__/test-preload.ts`) is the strictest: it must not import from `src/` at all. Its only static imports are node stdlib, `bun:test`, and helpers in `src/__tests__/`. Importing from a source module risks running its import-time side effects before the workspace override is set.
 - **The preload verifier** (`src/__tests__/test-preload-verifier.ts`) runs after the main preload and asserts the override took effect (`VELLUM_WORKSPACE_DIR` must resolve under `os.tmpdir()`).
 - **Destructive ops** (e.g. `rmSync(dbPath, ...)`) in tests must call `assertNotLiveDb(path)` from `src/__tests__/assert-not-live-db.js` immediately before the destructive call. The check is a per-callsite belt to the preload-verifier suspenders.

package/examples/plugins/echo/README.md

@@ -16,8 +16,8 @@ to eyeball which hooks fire during a conversation.
 
 For the full plugin authoring guide — manifest shape, every contribution
 surface, hook patterns, and conventions — see
-[`plugins/README.md`](../../../../plugins/README.md).
-[`simple-memory`](../../../../plugins/simple-memory/) is the
+[`experimental/plugins/README.md`](../../../../experimental/plugins/README.md).
+[`simple-memory`](../../../../experimental/plugins/simple-memory/) is the
 canonical reference implementation that exercises every wired surface.
 
 ## What it does
@@ -115,7 +115,7 @@ vellum restart
 
 ## Next steps
 
-- Read [`plugins/README.md`](../../../../plugins/README.md)
+- Read [`experimental/plugins/README.md`](../../../../experimental/plugins/README.md)
   for the full plugin authoring guide: manifest shape, every contribution
   surface, hook patterns (observe / transform), tool contributions, and
   conventions.

package/node_modules/@vellumai/skill-host-contracts/src/tool-types.ts

@@ -195,8 +195,6 @@ interface ToolLifecycleEventBase {
   conversationId: string;
   requestId?: string;
   executionTarget?: ExecutionTarget;
-  /** Id of the skill whose `skill_execute` dispatch triggered this tool call. Absent for direct tool calls. */
-  skillId?: string;
 }
 
 export interface AllowlistOption {

package/openapi.yaml

@@ -96,11 +96,8 @@ paths:
       operationId: acp_by_id_steer_post
       summary: Steer ACP session
       description:
-        "Send a steering instruction to an ACP session. Sessions no longer in memory (completed, or lost to a
-        daemon restart) are transparently resumed from persisted history first, when the agent supports ACP session
-        loading. Resuming a terminal session re-spawns the host agent, so it requires guardian approval: the route acks
-        immediately with approvalPending=true and performs the resume in the background once approved, streaming results
-        over SSE."
+        Send a steering instruction to an ACP session. Sessions no longer in memory (completed, or lost to a daemon
+        restart) are transparently resumed from persisted history first, when the agent supports ACP session loading.
       tags:
         - acp
       responses:
@@ -118,11 +115,6 @@ paths:
                   resumed:
                     description: True when the session was resumed from persisted history before steering.
                     type: boolean
-                  approvalPending:
-                    description:
-                      True when the steer triggered a guardian-approval-gated resume that will run asynchronously; watch SSE for
-                      the outcome.
-                    type: boolean
                 required:
                   - acpSessionId
                   - steered
@@ -15945,8 +15937,6 @@ paths:
                             - assistant
                         content:
                           type: string
-                          deprecated: true
-                          description: "Deprecated: superseded by contentBlocks. Flat plain-text body (joined text segments)."
                         timestamp:
                           type: string
                         attachments:
@@ -16251,8 +16241,6 @@ paths:
                               - name
                               - input
                             additionalProperties: false
-                          deprecated: true
-                          description: "Deprecated: superseded by contentBlocks (the tool_use variant). Flat list of tool calls."
                         surfaces:
                           type: array
                           items:
@@ -16304,28 +16292,18 @@ paths:
                               - surfaceType
                               - data
                             additionalProperties: false
-                          deprecated: true
-                          description: "Deprecated: superseded by contentBlocks (the surface variant). Flat list of surfaces."
                         textSegments:
                           type: array
                           items:
                             type: string
-                          deprecated: true
-                          description: "Deprecated: superseded by contentBlocks. Text segments split by tool-call boundaries."
                         thinkingSegments:
                           type: array
                           items:
                             type: string
-                          deprecated: true
-                          description: "Deprecated: superseded by contentBlocks. Reasoning text extracted from thinking blocks."
                         contentOrder:
                           type: array
                           items:
                             type: string
-                          deprecated: true
-                          description:
-                            'Deprecated: superseded by contentBlocks. Positional "<type>:<index>" content ordering (e.g. "text:0",
-                            "thinking:1").'
                         contentBlocks:
                           type: array
                           items:
@@ -16348,10 +16326,6 @@ paths:
                                     const: thinking
                                   thinking:
                                     type: string
-                                  startedAt:
-                                    type: number
-                                  completedAt:
-                                    type: number
                                 required:
                                   - type
                                   - thinking
@@ -19475,11 +19449,11 @@ paths:
       operationId: plugins_install_post
       summary: Install a plugin
       description:
-        "Install a plugin by name from the canonical source — a whitelisted `plugins/marketplace.json` entry, else
-        the first-party `plugins/<name>/` convention. Always resolves against the curated default git ref (no
-        caller-supplied ref): installing from an unreviewed revision would bypass the marketplace/first-party curation
-        boundary and let attacker-controlled code be loaded. Materializes the plugin under
-        `<workspaceDir>/plugins/<name>/`; the assistant must be restarted to load it. Mirrors the CLI's `assistant
+        "Install a plugin by name from the canonical source — a whitelisted `experimental/plugins/marketplace.json`
+        entry, else the first-party `experimental/plugins/<name>/` convention. Always resolves against the curated
+        default git ref (no caller-supplied ref): installing from an unreviewed revision would bypass the
+        marketplace/first-party curation boundary and let attacker-controlled code be loaded. Materializes the plugin
+        under `<workspaceDir>/plugins/<name>/`; the assistant must be restarted to load it. Mirrors the CLI's `assistant
         plugins install <name>`. An already-installed name without `force` returns 409; a name that resolves to nothing
         returns 404. Sibling to `POST /v1/skills/install`."
       tags:
@@ -19541,9 +19515,9 @@ paths:
       operationId: plugins_search_get
       summary: Search the plugin catalog
       description:
-        List installable plugins from the canonical `vellum-ai/vellum-assistant` catalog at `plugins/`. The query
-        is an ECMAScript regex matched case-insensitively against the directory name (e.g. `memory`, `^simple`). Empty
-        query returns every entry. Mirrors the CLI's `assistant plugins search`.
+        List installable plugins from the canonical `vellum-ai/vellum-assistant` catalog at
+        `experimental/plugins/`. The query is an ECMAScript regex matched case-insensitively against the directory name
+        (e.g. `memory`, `^simple`). Empty query returns every entry. Mirrors the CLI's `assistant plugins search`.
       tags:
         - plugins
       responses:
@@ -19571,7 +19545,7 @@ paths:
                         path:
                           type: string
                           description:
-                            "Human-readable origin: a repo-relative path (`plugins/<name>`) for first-party plugins or a
+                            "Human-readable origin: a repo-relative path (`experimental/plugins/<name>`) for first-party plugins or a
                             `github:owner/repo@ref` locator for external ones."
                         description:
                           description: Short description, when known (external entries only today).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vellumai/assistant",
-  "version": "0.8.10-dev.202606092238.d04fd59",
+  "version": "0.8.10",
   "license": "MIT",
   "type": "module",
   "exports": {

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

@@ -13,7 +13,7 @@
  * Sites not exercised here (`aborted_via_error`) require deeper provider
  * fakery and are best covered by integration tests.
  */
-import { afterEach, describe, expect, test } from "bun:test";
+import { afterEach, describe, expect, mock, test } from "bun:test";
 
 import type {
   AgentEvent,
@@ -23,14 +23,12 @@ import type {
 import { AgentLoop, isMaxTokensStopReason } from "../agent/loop.js";
 import type { ContextWindowConfig } from "../config/types.js";
 import type { TrustContext } from "../daemon/trust-context.js";
-import { HOOKS } from "../plugin-api/constants.js";
 import {
   createContextWindowManager,
   disposeContextWindowManager,
   getContextWindowManager,
 } from "../plugins/defaults/compaction/manager-store.js";
 import type { PostCompactContext } from "../plugins/defaults/memory-retrieval/hooks/post-compact.js";
-import { registerPlugin } from "../plugins/registry.js";
 import type {
   Message,
   Provider,
@@ -39,25 +37,22 @@ import type {
   ToolDefinition,
 } from "../providers/types.js";
 
-// The agent loop runs the default post-compaction re-injection through the
-// `post-compact` hook chain when it compacts in place. Register a test hook on
-// that chain so these unit tests can drive the re-injection result without the
-// daemon-level injector. The hook writes the re-injected history back onto the
-// context; tests assign `postCompactImpl` to observe the call or force a
-// failure; when unset the hook leaves the history it was handed untouched.
+// The agent loop invokes the default post-compaction re-injection hook directly
+// when it compacts in place. Stub it so these unit tests can drive the
+// re-injection result without the daemon-level injector chain. Tests assign
+// `postCompactImpl` to observe the call or force a failure; when unset
+// the hook is a no-op that returns the history it was handed.
 let postCompactImpl:
   | ((input: PostCompactContext) => Promise<Message[]>)
   | null = null;
-registerPlugin({
-  manifest: { name: "test-post-compact", version: "0.0.0" },
-  hooks: {
-    [HOOKS.POST_COMPACT]: async (input: PostCompactContext): Promise<void> => {
-      input.history = postCompactImpl
-        ? await postCompactImpl(input)
-        : input.history;
-    },
-  },
-});
+mock.module(
+  "../plugins/defaults/memory-retrieval/hooks/post-compact.js",
+  () => ({
+    default: async (input: PostCompactContext) => ({
+      messages: postCompactImpl ? await postCompactImpl(input) : input.history,
+    }),
+  }),
+);
 
 // ---------------------------------------------------------------------------
 // Helpers (mirrored from agent-loop.test.ts so this file is self-contained)

package/src/__tests__/build-persisted-content.test.ts

@@ -54,10 +54,7 @@ mock.module("../memory/llm-request-log-store.js", () => ({
 }));
 
 // ── Imports (after mocks) ─────────────────────────────────────────────────────
-import {
-  buildPersistedAssistantContent,
-  stampThinkingTiming,
-} from "../daemon/conversation-agent-loop-handlers.js";
+import { buildPersistedAssistantContent } from "../daemon/conversation-agent-loop-handlers.js";
 import type { ToolActivityMetadata } from "../daemon/message-types/web-activity.js";
 import type { ContentBlock } from "../providers/types.js";
 
@@ -185,74 +182,3 @@ describe("buildPersistedAssistantContent — native activityMetadata", () => {
     expect(block._activityMetadata).toBeUndefined();
   });
 });
-
-describe("stampThinkingTiming", () => {
-  test("stamps internal timing onto thinking blocks by position", () => {
-    // GIVEN a turn that interleaves text and two thinking blocks AND the
-    // per-block timing captured while streaming (one entry per thinking block,
-    // in stream order)
-    const content = [
-      { type: "thinking", thinking: "first", signature: "s1" },
-      { type: "text", text: "answer" },
-      { type: "thinking", thinking: "second", signature: "s2" },
-    ] as unknown as ContentBlock[];
-    const timings = [
-      { startedAt: 100, completedAt: 250 },
-      { startedAt: 400, completedAt: 480 },
-    ];
-
-    // WHEN the content is stamped before persistence
-    const stamped = stampThinkingTiming(content, timings) as unknown as Array<
-      Record<string, unknown>
-    >;
-
-    // THEN each thinking block carries the `_`-prefixed timing for its position
-    expect(stamped[0]).toMatchObject({
-      type: "thinking",
-      _startedAt: 100,
-      _completedAt: 250,
-    });
-    expect(stamped[2]).toMatchObject({
-      type: "thinking",
-      _startedAt: 400,
-      _completedAt: 480,
-    });
-    // AND the interleaved text block is left untouched
-    expect(stamped[1]).toEqual({ type: "text", text: "answer" });
-  });
-
-  test("leaves thinking blocks unstamped when no timing was captured", () => {
-    // GIVEN thinking content but an empty timing list (thinking streaming was
-    // disabled, so no per-block timing was recorded this turn)
-    const content = [
-      { type: "thinking", thinking: "first", signature: "s1" },
-    ] as unknown as ContentBlock[];
-
-    // WHEN the content is stamped with no timing
-    const stamped = stampThinkingTiming(content, []);
-
-    // THEN the original content is returned unchanged so the UI hides duration,
-    // exactly as a tool call with no timing
-    expect(stamped).toBe(content);
-    expect(stamped[0]).not.toHaveProperty("_startedAt");
-  });
-
-  test("stamps only the thinking blocks that have a matching timing entry", () => {
-    // GIVEN two thinking blocks but only one captured timing entry (e.g. the
-    // second block opened after the timing array was already finalized)
-    const content = [
-      { type: "thinking", thinking: "first", signature: "s1" },
-      { type: "thinking", thinking: "second", signature: "s2" },
-    ] as unknown as ContentBlock[];
-    const timings = [{ startedAt: 100, completedAt: 250 }];
-
-    // WHEN the content is stamped
-    const stamped = stampThinkingTiming(content, timings) as unknown as Array<
-      Record<string, unknown>
-    >;
-
-    // THEN the first block is stamped and the unmatched second block is left as-is
-    expect(stamped[0]).toMatchObject({ _startedAt: 100, _completedAt: 250 });
-    expect(stamped[1]).not.toHaveProperty("_startedAt");
-  });
-});

package/src/__tests__/config-managed-gemini-defaults.test.ts

@@ -57,8 +57,17 @@ mock.module("../util/logger.js", () => ({
   getLogger: () => makeLoggerStub(),
 }));
 
+// ---------------------------------------------------------------------------
+// Feature flag mock — controls whether managed-gemini-embeddings-enabled is on
+// ---------------------------------------------------------------------------
+
+let featureFlagEnabled = false;
+
 mock.module("../config/assistant-feature-flags.js", () => ({
-  isAssistantFeatureFlagEnabled: () => true,
+  isAssistantFeatureFlagEnabled: (key: string) => {
+    if (key === "managed-gemini-embeddings-enabled") return featureFlagEnabled;
+    return true;
+  },
   clearFeatureFlagOverridesCache: () => {},
   initFeatureFlagOverrides: async () => {},
   getAssistantFeatureFlagDefaults: () => ({}),
@@ -110,6 +119,8 @@ describe("managed Gemini embedding defaults (via loadConfig)", () => {
     setStorePathForTesting(join(WORKSPACE_DIR, "keys.enc"));
     invalidateConfigCache();
 
+    // Reset mock state
+    featureFlagEnabled = false;
     originalIsPlatform = process.env.IS_PLATFORM;
     delete process.env.IS_PLATFORM;
   });
@@ -126,9 +137,10 @@ describe("managed Gemini embedding defaults (via loadConfig)", () => {
     }
   });
 
-  test("applies managed Gemini defaults when IS_PLATFORM + provider auto", () => {
+  test("applies managed Gemini defaults when FF on + IS_PLATFORM + provider auto", () => {
     writeConfig({});
 
+    featureFlagEnabled = true;
     process.env.IS_PLATFORM = "true";
 
     const config = loadConfig();
@@ -150,9 +162,22 @@ describe("managed Gemini embedding defaults (via loadConfig)", () => {
     expect(qdrantRaw.vectorSize).toBe(3072);
   });
 
+  test("does NOT apply when feature flag is OFF", () => {
+    writeConfig({});
+
+    featureFlagEnabled = false;
+    process.env.IS_PLATFORM = "true";
+
+    const config = loadConfig();
+
+    expect(config.memory.embeddings.provider).toBe("auto");
+    expect(config.memory.qdrant.vectorSize).toBe(384);
+  });
+
   test("does NOT apply when IS_PLATFORM is not set", () => {
     writeConfig({});
 
+    featureFlagEnabled = true;
     delete process.env.IS_PLATFORM;
 
     const config = loadConfig();
@@ -166,6 +191,7 @@ describe("managed Gemini embedding defaults (via loadConfig)", () => {
       memory: { embeddings: { provider: "local" } },
     });
 
+    featureFlagEnabled = true;
     process.env.IS_PLATFORM = "true";
 
     const config = loadConfig();
@@ -178,6 +204,7 @@ describe("managed Gemini embedding defaults (via loadConfig)", () => {
       memory: { embeddings: { provider: "openai" } },
     });
 
+    featureFlagEnabled = true;
     process.env.IS_PLATFORM = "true";
 
     const config = loadConfig();
@@ -192,6 +219,7 @@ describe("managed Gemini embedding defaults (via loadConfig)", () => {
       },
     });
 
+    featureFlagEnabled = true;
     process.env.IS_PLATFORM = "true";
 
     const config = loadConfig();
@@ -207,6 +235,7 @@ describe("managed Gemini embedding defaults (via loadConfig)", () => {
       memory: { embeddings: { provider: "ollama" } },
     });
 
+    featureFlagEnabled = true;
     process.env.IS_PLATFORM = "true";
 
     const config = loadConfig();
@@ -216,6 +245,7 @@ describe("managed Gemini embedding defaults (via loadConfig)", () => {
   test("is idempotent — second loadConfig is a no-op after migration", () => {
     writeConfig({});
 
+    featureFlagEnabled = true;
     process.env.IS_PLATFORM = "true";
 
     const config = loadConfig();
@@ -244,6 +274,7 @@ describe("managed Gemini embedding defaults (via loadConfig)", () => {
       },
     });
 
+    featureFlagEnabled = true;
     process.env.IS_PLATFORM = "true";
 
     const config = loadConfig();
@@ -264,9 +295,22 @@ describe("managed Gemini embedding defaults (via loadConfig)", () => {
     expect(qdrantRaw.onDisk).toBe(false);
   });
 
+  test("does NOT apply when both FF off and IS_PLATFORM not set", () => {
+    writeConfig({});
+
+    featureFlagEnabled = false;
+    delete process.env.IS_PLATFORM;
+
+    const config = loadConfig();
+
+    expect(config.memory.embeddings.provider).toBe("auto");
+    expect(config.memory.qdrant.vectorSize).toBe(384);
+  });
+
   test("applies when IS_PLATFORM is '1'", () => {
     writeConfig({});
 
+    featureFlagEnabled = true;
     process.env.IS_PLATFORM = "1";
 
     const config = loadConfig();