@vellumai/assistant 0.8.10 → 0.8.11-staging.1

package/src/config/bundled-skills/skill-management/SKILL.md

@@ -1,10 +1,11 @@
 ---
 name: skill-management
-description: Create and delete custom managed skills
+description: Create, edit, and delete custom managed skills in the user's workspace. Use whenever the user wants to author a new skill from a description, scaffold a SKILL.md, or remove a skill they no longer need.
 metadata:
   emoji: "\U0001F9E9"
   vellum:
     display-name: "Skill Management"
+    category: "system"
     activation-hints:
       - "User wants to scaffold a new managed skill in their workspace from a description"
       - "User wants to delete or list the custom skills they have defined"
@@ -15,9 +16,108 @@ metadata:
 
 Manage the lifecycle of custom managed skills in `{workspaceDir}/skills`.
 
+## When to Use
+
+USE THIS SKILL WHEN:
+- The user says "build me a skill" or "create a skill for X"
+- The user wants to scaffold, edit, or delete a SKILL.md in their workspace
+- The user wants a repeatable workflow captured as an invocable skill
+
+Do NOT use this skill when the user just wants to run an existing skill. That is normal activation, not management.
+
 ## Capabilities
 
 - **Scaffold** a new managed skill with YAML frontmatter and markdown body
+- **Edit** an existing skill by scaffolding over it (rewrites the SKILL.md in place)
 - **Delete** an existing managed skill directory
 
 Skills created via `scaffold_managed_skill` become available for `skill_load` when a valid top-level `SKILL.md` is written under the skill directory.
+
+## Step 1 - Align with the user before building
+
+Ask before doing anything. Do not scaffold a skill until you have confirmed with the user:
+
+- What the skill should do
+- When it should activate (the trigger phrases, in their words)
+- The major steps it performs
+- Any destructive steps and the done condition
+
+> ✓ Checkpoint: Have you confirmed scope with the user? If you are guessing at any of the four points above, ask first. Do not scaffold on assumption.
+
+## Step 2 - Write a description AND activation-hints, always both
+
+The description is what makes the skill discoverable. It must cover both what the skill does and when to reach for it, phrased the way the user would say it.
+
+```yaml
+description: Build anything visual — apps, landing pages, dashboards, trackers,
+  calculators, games, tools, slide decks, or data visualizations. Use whenever
+  the user wants something built that they can see and interact with.
+```
+
+**Every skill must also ship `activation-hints` in its frontmatter. This is not optional.** Keep `activation-hints` separate from the description: the description sells the skill, the hints list the concrete trigger phrases the user confirmed in Step 1.
+
+```yaml
+metadata:
+  vellum:
+    activation-hints:
+      - "build me an app"
+      - "make a dashboard"
+      - "create a landing page"
+```
+
+> ✓ Checkpoint: Does the frontmatter have both a description and an `activation-hints` list? If hints are missing, go back and add them before writing the body.
+
+## Step 3 - Structure the body so it survives weak models
+
+Strong models tolerate loose structure. Weaker models drift. Build every body with these patterns.
+
+**Open with a `## When to Use` block.** User language, not jargon. This is what makes the model recognize when the skill applies.
+
+**Put critical warnings at the point of action.** A warning at the top of a file is forgotten by the time the model is 200 lines deep. Do not trust the top-of-file warning. Repeat the danger where the dangerous action happens.
+
+```markdown
+## Step 5 - Apply the JSON blob
+
+⚠️ CRITICAL: Use the complete blob below. Setting even one key
+wipes the entire block. Copy the whole thing or fail.
+```
+
+**Add explicit checkpoints between major steps, sparingly.** Long executions blur together. The model finishes step 3 and slides into step 4 without re-anchoring. A checkpoint forces a re-read. Use them between major sections, not on every step.
+
+**Make branching explicit with `If / →`, and always name the default.** Prose hides decisions. The model reads linearly and walks past a branch without registering it. Every `If` must cover the default case. Implicit fall-through ("otherwise figure it out") creates drift.
+
+```markdown
+If the user already has a draft → restructure it into the template.
+If not → build the steps from their description (default).
+```
+
+## Step 4 - Define done by binding tool calls to artifacts
+
+Without an explicit done condition, the model invents one. It stops too early ("the file was created, done") or overshoots ("let me add one more feature"). Both are drift.
+
+Each completion criterion must bind a **tool call** to the **user-visible artifact** it produces. Do not write criteria the model can satisfy by narration alone.
+
+```markdown
+## SKILL COMPLETE WHEN
+
+- [ ] `scaffold_managed_skill` wrote the SKILL.md and returned its path
+- [ ] User confirmed the skill loads via `skill_load`
+- [ ] User saw the trigger phrases that will activate it
+```
+
+> ✓ Checkpoint: Before scaffolding, confirm the body has a `## When to Use` block, point-of-action warnings on any dangerous step, explicit `If / →` branches with named defaults, and artifact-bound completion criteria.
+
+## Step 5 - Keep SKILL.md under 500 lines
+
+Past 500 lines the model loses things in the middle. Warnings get buried, branching loses visibility, and the file fights the task for the same context budget. If a skill is growing past 500 lines, split reference material into separate files the skill points to.
+
+## Step 6 - Test the skill before calling it done
+
+After scaffolding, load the skill and confirm it activates on the intended trigger and follows its own steps. If it does not activate or drifts, fix the body and test again.
+
+How you exercise it depends on what the skill does:
+
+- **If the skill performs side effects** (sends messages, deletes data, makes purchases, mutates external state) → do not trigger a live run during creation. Confirm it loads and activates on the intended trigger (static check), then ask the user before exercising it for real.
+- **Otherwise** (read-only or local-only skills) → run it against a realistic prompt directly. This is the default.
+
+> ⚠️ CRITICAL: Do not tell the user a skill is ready until you have confirmed it loads and activates on the intended trigger. A skill that was never loaded is a skill that was never tested. Never perform user-visible side effects just to test a skill without the user's consent.

package/src/config/bundled-skills/subagent/SKILL.md

@@ -6,6 +6,7 @@ metadata:
   emoji: "🤖"
   vellum:
     display-name: "Subagent"
+    category: "system"
     activation-hints:
       - "Spawn a background worker that runs in parallel with the main turn"
       - "Delegate a self-contained research or implementation task off the main thread"

package/src/config/bundled-skills/transcribe/SKILL.md

@@ -6,6 +6,7 @@ metadata:
   emoji: "🎙️"
   vellum:
     display-name: "Transcribe"
+    category: "voice"
     activation-hints:
       - "User has an audio or video file on disk they want converted to text"
       - "User wants speech-to-text on a recording, voice memo, podcast, or meeting capture"

package/src/config/env-registry.ts

@@ -72,6 +72,27 @@ export function isPlatformRemote(): boolean {
   return getIsPlatform();
 }
 
+/**
+ * VELLUM_DISABLE_PLATFORM — boolean, default: false
+ * When true, all outbound platform API calls in local mode are suppressed.
+ * The daemon, gateway, and web UI each no-op platform requests with a
+ * debug log. Has no effect when IS_PLATFORM is true (platform-managed
+ * instances always connect to the platform).
+ */
+export function getDisablePlatform(): boolean {
+  return flag("VELLUM_DISABLE_PLATFORM");
+}
+
+/**
+ * VELLUM_DEVICE_ID — string, default: undefined
+ * Host-stable device id injected by the CLI into containerized deployments,
+ * where the container fs cannot persist device.json across recreation.
+ * Takes precedence over device.json resolution in util/device-id.ts.
+ */
+export function getDeviceIdOverride(): string | undefined {
+  return str("VELLUM_DEVICE_ID");
+}
+
 /**
  * VELLUM_WORKSPACE_DIR — string, default: undefined
  * Overrides the default workspace directory.
@@ -191,6 +212,8 @@ const KNOWN_VELLUM_VARS = new Set([
   "VELLUM_DEBUG",
   "VELLUM_DESKTOP_APP",
   "VELLUM_DEV",
+  "VELLUM_DEVICE_ID",
+  "VELLUM_DISABLE_PLATFORM",
   "VELLUM_DOCS_BASE_URL",
   "VELLUM_ENVIRONMENT",
   "VELLUM_HATCHED_BY",

package/src/config/feature-flag-registry.json

@@ -59,6 +59,14 @@
       "description": "Enable the A2A (Agent-to-Agent) channel for inter-assistant communication via the open A2A protocol",
       "defaultEnabled": false
     },
+    {
+      "id": "workspace-tools-watcher",
+      "scope": "assistant",
+      "key": "workspace-tools-watcher",
+      "label": "Workspace tools file watcher",
+      "description": "Hot-reload workspace tool overrides under `<workspaceDir>/tools/` by watching the directory and re-registering tools as files change. When disabled, workspace tools still load from disk once at daemon startup, but live edits require a restart to take effect.",
+      "defaultEnabled": false
+    },
     {
       "id": "settings-developer-nav",
       "scope": "assistant",
@@ -72,7 +80,7 @@
       "scope": "assistant",
       "key": "system-schedule-toggles",
       "label": "System Schedule Toggles",
-      "description": "Show Heartbeat and Consolidation pause toggles on the schedules settings page. The toggles only pause automatic runs; manual Run now remains available.",
+      "description": "Show the Heartbeat pause toggle on the schedules settings page. The toggle only pauses automatic runs; manual Run now remains available.",
       "defaultEnabled": false
     },
     {
@@ -115,22 +123,6 @@
       "description": "Surface credential grant and audit inspection endpoints for reviewing active grants and access logs",
       "defaultEnabled": false
     },
-    {
-      "id": "chatgpt-subscription-auth",
-      "scope": "assistant",
-      "key": "chatgpt-subscription-auth",
-      "label": "ChatGPT Subscription Auth",
-      "description": "Enable ChatGPT subscription OAuth as a provider auth type for OpenAI models, using the Codex device-code flow.",
-      "defaultEnabled": false
-    },
-    {
-      "id": "deploy-to-vercel",
-      "scope": "assistant",
-      "key": "deploy-to-vercel",
-      "label": "Deploy to Vercel",
-      "description": "Enable the Deploy to Vercel / Publish option in the app workspace header share menu",
-      "defaultEnabled": false
-    },
     {
       "id": "settings-embedding-provider",
       "scope": "assistant",
@@ -163,14 +155,6 @@
       "description": "Auto-expand completed tool call step groups instead of showing them collapsed",
       "defaultEnabled": false
     },
-    {
-      "id": "sounds",
-      "scope": "assistant",
-      "key": "sounds",
-      "label": "Sounds",
-      "description": "Enable the Sounds tab in Settings and all app sound playback (event sounds, random ambient sounds)",
-      "defaultEnabled": false
-    },
     {
       "id": "message-tts",
       "scope": "assistant",
@@ -179,14 +163,6 @@
       "description": "Show a speaker button on assistant messages to generate and play the message as audio via Fish Audio TTS",
       "defaultEnabled": false
     },
-    {
-      "id": "openai-compatible-endpoints",
-      "scope": "assistant",
-      "key": "openai-compatible-endpoints",
-      "label": "OpenAI-Compatible Endpoints",
-      "description": "Enable user-configured OpenAI-compatible inference endpoints with custom base URLs and model identifiers",
-      "defaultEnabled": false
-    },
     {
       "id": "backward-releases",
       "scope": "assistant",
@@ -235,14 +211,6 @@
       "description": "Enable assistant sandboxing via Apple Containers on macOS 26+, providing a lightweight native sandbox without third-party dependencies",
       "defaultEnabled": false
     },
-    {
-      "id": "managed-gemini-embeddings-enabled",
-      "scope": "assistant",
-      "key": "managed-gemini-embeddings-enabled",
-      "label": "Managed Gemini Embeddings Enabled",
-      "description": "Route embedding requests through the platform runtime proxy using Vellum-managed Gemini credentials when available",
-      "defaultEnabled": false
-    },
     {
       "id": "bookmarks",
       "scope": "client",
@@ -299,14 +267,6 @@
       "description": "Expose the developer-only Compaction Playground tab in macOS Settings and enable the /playground/* HTTP endpoints for exercising compaction conditions. Dev-only; default off.",
       "defaultEnabled": false
     },
-    {
-      "id": "account-deletion",
-      "scope": "assistant",
-      "key": "account-deletion",
-      "label": "Account Deletion",
-      "description": "Surfaces the user-initiated account deletion flow in client settings.",
-      "defaultEnabled": true
-    },
     {
       "id": "app-control",
       "scope": "assistant",
@@ -331,14 +291,6 @@
       "description": "Enable the external-plugin install path: the `assistant plugins` CLI subcommand and the declarative external-plugin loader convention (`<workspaceDir>/plugins/<name>/` directories containing `package.json` plus interface dirs). Gates an unstable surface that may change shape before stabilizing.",
       "defaultEnabled": false
     },
-    {
-      "id": "provider-zai",
-      "scope": "assistant",
-      "key": "provider-zai",
-      "label": "z.ai Provider",
-      "description": "Enable the z.ai (Zhipu AI) provider and its GLM models in the provider picker and model selection UI",
-      "defaultEnabled": false
-    },
     {
       "id": "velvet-theme",
       "scope": "client",
@@ -371,22 +323,6 @@
       "description": "Enable pull-to-refresh gesture in the chat view.",
       "defaultEnabled": false
     },
-    {
-      "id": "doctor",
-      "scope": "client",
-      "key": "doctor",
-      "label": "Doctor",
-      "description": "Enable the Doctor diagnostic tab in Debug settings.",
-      "defaultEnabled": false
-    },
-    {
-      "id": "home-page",
-      "scope": "client",
-      "key": "home-page",
-      "label": "Home Page",
-      "description": "Enable the Home page as the default landing view.",
-      "defaultEnabled": false
-    },
     {
       "id": "platform-notifications",
       "scope": "client",
@@ -395,6 +331,14 @@
       "description": "Enable the Notifications tab in settings.",
       "defaultEnabled": false
     },
+    {
+      "id": "centralized-assistant-status",
+      "scope": "client",
+      "key": "centralized-assistant-status",
+      "label": "Centralized Assistant Status",
+      "description": "Enable client polling of the centralized assistant operational status endpoint.",
+      "defaultEnabled": false
+    },
     {
       "id": "preview-channel",
       "scope": "client",
@@ -443,14 +387,6 @@
       "description": "Expose the developer-only Memory Router Playground tab in macOS Settings and the /assistant/memory-router-playground web page for dry-running v4 router config overrides against the live page index. Dev-only; default off.",
       "defaultEnabled": false
     },
-    {
-      "id": "platform-features-in-local-mode",
-      "scope": "assistant",
-      "key": "platform-features-in-local-mode",
-      "label": "Platform Features in Local Mode",
-      "description": "When enabled, the assistant can call the Vellum platform API from local mode. When disabled, all platform API clients in the daemon, gateway, CES, and web UI no-op with a debug log instead of making outbound requests.",
-      "defaultEnabled": true
-    },
     {
       "id": "memory-v3-shadow",
       "scope": "assistant",

package/src/config/loader.ts

@@ -14,7 +14,6 @@ import {
   getWorkspaceConfigPath,
   getWorkspaceDir,
 } from "../util/platform.js";
-import { isAssistantFeatureFlagEnabled } from "./assistant-feature-flags.js";
 import { AssistantConfigSchema } from "./schema.js";
 import type { AssistantConfig } from "./types.js";
 
@@ -692,16 +691,15 @@ export function loadConfig(): AssistantConfig {
 
     if (suppressConfigDiskWritesDepth === 0) {
       // Managed Gemini embedding defaults migration.
-      // When on a managed platform (IS_PLATFORM=true) with the feature flag
-      // enabled and no explicit embedding provider chosen (provider=auto),
-      // persist Gemini embedding defaults into the raw config file.
+      // When on a managed platform (IS_PLATFORM=true) and no explicit embedding
+      // provider chosen (provider=auto), persist Gemini embedding defaults into
+      // the raw config file.
       // Idempotent: once provider=gemini is written, subsequent loads skip this.
       if (config.memory.embeddings.provider === "auto") {
         try {
           if (
-            (process.env.IS_PLATFORM === "true" ||
-              process.env.IS_PLATFORM === "1") &&
-            isManagedGeminiFFEnabled(config)
+            process.env.IS_PLATFORM === "true" ||
+            process.env.IS_PLATFORM === "1"
           ) {
             setNestedValue(fileConfig, "memory.embeddings.provider", "gemini");
             setNestedValue(
@@ -797,21 +795,6 @@ export function loadConfig(): AssistantConfig {
   }
 }
 
-/**
- * Check whether the managed-gemini-embeddings-enabled feature flag is on.
- * Wrapped in a try/catch so a flag-resolver failure never breaks config loading.
- */
-function isManagedGeminiFFEnabled(config: AssistantConfig): boolean {
-  try {
-    return isAssistantFeatureFlagEnabled(
-      "managed-gemini-embeddings-enabled",
-      config,
-    );
-  } catch {
-    return false;
-  }
-}
-
 export function getConfig(): AssistantConfig {
   return loadConfig();
 }

package/src/config/schema.ts

@@ -27,6 +27,7 @@ import {
   WhatsAppConfigSchema,
 } from "./schemas/channels.js";
 import { CompactionConfigSchema } from "./schemas/compaction.js";
+import { CompactionLogsConfigSchema } from "./schemas/compaction-logs.js";
 import { ConversationsConfigSchema } from "./schemas/conversations.js";
 import { FilingConfigSchema } from "./schemas/filing.js";
 import { HeartbeatConfigSchema } from "./schemas/heartbeat.js";
@@ -107,6 +108,7 @@ export const AssistantConfigSchema = z
     compaction: CompactionConfigSchema.default(
       CompactionConfigSchema.parse({}),
     ),
+    compactionLogs: CompactionLogsConfigSchema,
     twilio: TwilioConfigSchema.default(TwilioConfigSchema.parse({})),
     calls: CallsConfigSchema.default(CallsConfigSchema.parse({})),
     whatsapp: WhatsAppConfigSchema.default(WhatsAppConfigSchema.parse({})),

package/src/config/schemas/__tests__/compaction-logs.test.ts

@@ -0,0 +1,56 @@
+import { describe, expect, test } from "bun:test";
+
+import { CompactionLogsConfigSchema } from "../compaction-logs.js";
+
+describe("CompactionLogsConfigSchema", () => {
+  test("parses undefined to the disabled default", () => {
+    expect(CompactionLogsConfigSchema.parse(undefined)).toEqual({
+      destination: "none",
+    });
+  });
+
+  test("parses an explicit none destination", () => {
+    expect(CompactionLogsConfigSchema.parse({ destination: "none" })).toEqual({
+      destination: "none",
+    });
+  });
+
+  test("parses an explicit clickhouse destination with defaulted connection fields", () => {
+    expect(
+      CompactionLogsConfigSchema.parse({ destination: "clickhouse" }),
+    ).toEqual({
+      destination: "clickhouse",
+      clickhouse: {
+        database: "default",
+        table: "compaction_logs",
+        user: "default",
+      },
+    });
+  });
+
+  test("parses explicit clickhouse connection overrides", () => {
+    expect(
+      CompactionLogsConfigSchema.parse({
+        destination: "clickhouse",
+        clickhouse: {
+          database: "analytics",
+          table: "compactions",
+          user: "writer",
+        },
+      }),
+    ).toEqual({
+      destination: "clickhouse",
+      clickhouse: {
+        database: "analytics",
+        table: "compactions",
+        user: "writer",
+      },
+    });
+  });
+
+  test("rejects an unknown destination", () => {
+    expect(() =>
+      CompactionLogsConfigSchema.parse({ destination: "sqlite" }),
+    ).toThrow();
+  });
+});

package/src/config/schemas/__tests__/memory-v2.test.ts

@@ -23,7 +23,6 @@ describe("MemoryV2ConfigSchema", () => {
       bm25_k1: 1.2,
       bm25_b: 0.4,
       consolidation_interval_hours: 4,
-      consolidation_enabled: true,
       consolidation_max_buffer_lines: 100,
       max_page_chars: 5000,
       consolidation_prompt_path: null,

package/src/config/schemas/__tests__/memory-v3.test.ts

@@ -6,13 +6,73 @@ describe("MemoryV3ConfigSchema", () => {
   test("parses an empty object to documented defaults", () => {
     const parsed = MemoryV3ConfigSchema.parse({});
     expect(parsed).toEqual({
-      workingSet: { maxPages: 150, evictWindow: 5 },
+      prune: { maxResidentBytes: 393216, targetResidentBytes: 262144 },
+      hotSet: { k: 40, halfLifeDays: 14 },
+      spotlight: { n: 6, windowTurns: 2 },
       needleK: 100,
       denseK: 100,
       edge: { hubDegree: 30, seedCount: 18, perSeed: 6, cap: 45 },
     });
   });
 
+  test("accepts and ignores the retired workingSet key from legacy configs", () => {
+    // Existing user config files may still carry the removed `workingSet`
+    // sub-config; it must parse (unknown-key strip), not throw.
+    const parsed = MemoryV3ConfigSchema.parse({
+      workingSet: { maxPages: 150, evictWindow: 5 },
+      needleK: 50,
+    });
+    expect(parsed.needleK).toBe(50);
+    expect("workingSet" in parsed).toBe(false);
+  });
+
+  test("prune: accepts partial overrides and enforces target < max", () => {
+    const parsed = MemoryV3ConfigSchema.parse({
+      prune: { maxResidentBytes: 500000 },
+    });
+    expect(parsed.prune).toEqual({
+      maxResidentBytes: 500000,
+      targetResidentBytes: 262144,
+    });
+    // target must be strictly below max (defaults included in the check).
+    expect(() =>
+      MemoryV3ConfigSchema.parse({
+        prune: { maxResidentBytes: 100000 },
+      }),
+    ).toThrow();
+    expect(() =>
+      MemoryV3ConfigSchema.parse({
+        prune: { maxResidentBytes: 1000, targetResidentBytes: 1000 },
+      }),
+    ).toThrow();
+    expect(() =>
+      MemoryV3ConfigSchema.parse({
+        prune: { maxResidentBytes: 0 },
+      }),
+    ).toThrow();
+    expect(() =>
+      MemoryV3ConfigSchema.parse({
+        prune: { targetResidentBytes: 1.5 },
+      }),
+    ).toThrow();
+  });
+
+  test("accepts a partial spotlight override and rejects invalid knobs", () => {
+    const parsed = MemoryV3ConfigSchema.parse({ spotlight: { n: 3 } });
+    expect(parsed.spotlight).toEqual({ n: 3, windowTurns: 2 });
+    // windowTurns: 0 is valid — current turn only, no carried window.
+    expect(
+      MemoryV3ConfigSchema.parse({ spotlight: { windowTurns: 0 } }).spotlight,
+    ).toEqual({ n: 6, windowTurns: 0 });
+    expect(() => MemoryV3ConfigSchema.parse({ spotlight: { n: 0 } })).toThrow();
+    expect(() =>
+      MemoryV3ConfigSchema.parse({ spotlight: { windowTurns: -1 } }),
+    ).toThrow();
+    expect(() =>
+      MemoryV3ConfigSchema.parse({ spotlight: { n: 1.5 } }),
+    ).toThrow();
+  });
+
   test("accepts explicit lane-K overrides", () => {
     const parsed = MemoryV3ConfigSchema.parse({ needleK: 50, denseK: 75 });
     expect(parsed.needleK).toBe(50);

package/src/config/schemas/compaction-logs.ts

@@ -0,0 +1,79 @@
+/**
+ * Configuration for the compaction log destination.
+ *
+ * Compaction logging is **off by default** (`destination: "none"`). An
+ * assistant can opt into a ClickHouse-only destination; there is no SQLite
+ * destination — compaction events are derived from the agent loop's
+ * compaction start/end events and written straight to ClickHouse.
+ *
+ * When `destination === "clickhouse"` the URL and password are resolved
+ * from the credential store (`clickhouse:url`, `clickhouse:password`),
+ * matching the `llmRequestLogs.clickhouse` convention. The connection
+ * options below describe everything else (database/table/user).
+ *
+ * The shape is a discriminated union on `destination` so the `clickhouse`
+ * block only exists on the ClickHouse branch — there's no stray defaults
+ * sitting around when logging is disabled.
+ */
+import { z } from "zod";
+
+export const CompactionLogsClickHouseConfigSchema = z
+  .object({
+    database: z
+      .string({ error: "compactionLogs.clickhouse.database must be a string" })
+      .min(1, "compactionLogs.clickhouse.database cannot be empty")
+      .default("default")
+      .describe("ClickHouse database containing the compaction logs table"),
+    table: z
+      .string({ error: "compactionLogs.clickhouse.table must be a string" })
+      .min(1, "compactionLogs.clickhouse.table cannot be empty")
+      .default("compaction_logs")
+      .describe("ClickHouse table name to write compaction logs to"),
+    user: z
+      .string({ error: "compactionLogs.clickhouse.user must be a string" })
+      .min(1, "compactionLogs.clickhouse.user cannot be empty")
+      .default("default")
+      .describe("ClickHouse user (password is read from credential store)"),
+  })
+  .describe(
+    "ClickHouse connection settings used when `destination` is `clickhouse`",
+  );
+
+const NoneCompactionLogsConfigSchema = z
+  .object({
+    destination: z.literal("none"),
+  })
+  .describe("Compaction logging disabled (default).");
+
+const ClickHouseCompactionLogsConfigSchema = z
+  .object({
+    destination: z.literal("clickhouse"),
+    clickhouse: CompactionLogsClickHouseConfigSchema.default(
+      CompactionLogsClickHouseConfigSchema.parse({}),
+    ),
+  })
+  .describe(
+    "Write compaction logs to ClickHouse. Requires the `clickhouse:url` and `clickhouse:password` credentials to be set.",
+  );
+
+// The default is baked into the export so the schema matches the sibling
+// pattern across `assistant/src/config/schemas/*` — `Schema.parse(undefined)`
+// returns documented defaults. The discriminated union has no inherent
+// default (no shared discriminator value), so we explicitly select the
+// `none` branch.
+//
+// Note: `CompactionLogsConfigSchema.parse({})` still throws — a discriminated
+// union cannot pick a branch without a discriminator. Use `parse(undefined)`
+// or omit the field entirely to get the default.
+export const CompactionLogsConfigSchema = z
+  .discriminatedUnion("destination", [
+    NoneCompactionLogsConfigSchema,
+    ClickHouseCompactionLogsConfigSchema,
+  ])
+  .default({ destination: "none" })
+  .describe("Compaction log destination configuration");
+
+export type CompactionLogsConfig = z.infer<typeof CompactionLogsConfigSchema>;
+export type CompactionLogsClickHouseConfig = z.infer<
+  typeof CompactionLogsClickHouseConfigSchema
+>;

package/src/config/schemas/memory-v2.ts

@@ -194,14 +194,6 @@ export const MemoryV2ConfigSchema = z
       .describe(
         "Hours between scheduled consolidation runs that synthesize buffered memories into concept pages",
       ),
-    consolidation_enabled: z
-      .boolean({
-        error: "memory.v2.consolidation_enabled must be a boolean",
-      })
-      .default(true)
-      .describe(
-        "Whether automatic scheduled consolidation runs are enabled. Manual run-now requests remain available while memory.v2.enabled is true.",
-      ),
     consolidation_max_buffer_lines: z
       .number({
         error: "memory.v2.consolidation_max_buffer_lines must be a number",