@vellumai/assistant 0.8.3 → 0.8.4

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

@@ -9,6 +9,14 @@
       "description": "Automatically trigger conversation analysis on the same cadence as memory extraction (batch threshold, idle debounce, end-of-conversation). The analysis agent has full tool access and writes back to memory and skills without user approval.",
       "defaultEnabled": false
     },
+    {
+      "id": "memory-retrospective-fork",
+      "scope": "assistant",
+      "key": "memory-retrospective-fork",
+      "label": "Fork-based memory retrospective",
+      "description": "Fork the source conversation through its latest message for memory retrospectives, instead of rendering the slice into a transcript and waking an empty background conversation. Lets the retrospective hit the provider prompt cache and read compaction summary + tail messages natively.",
+      "defaultEnabled": false
+    },
     {
       "id": "user-hosted-enabled",
       "scope": "client",
@@ -43,7 +51,7 @@
     },
     {
       "id": "settings-developer-nav",
-      "scope": "client",
+      "scope": "assistant",
       "key": "settings-developer-nav",
       "label": "Settings Developer Nav",
       "description": "Control Developer nav visibility in macOS settings",
@@ -89,6 +97,14 @@
       "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",
@@ -275,7 +291,7 @@
     },
     {
       "id": "account-deletion",
-      "scope": "client",
+      "scope": "assistant",
       "key": "account-deletion",
       "label": "Account Deletion",
       "description": "Surfaces the user-initiated account deletion flow in client settings.",
@@ -299,7 +315,7 @@
     },
     {
       "id": "pro-plan-adjust",
-      "scope": "assistant",
+      "scope": "client",
       "key": "pro-plan-adjust",
       "label": "Pro Plan Adjust",
       "description": "Show the rich Plan card (current plan, features, Manage/Upgrade CTA) at the top of the macOS Settings \u2192 Billing tab.",
@@ -328,6 +344,86 @@
       "label": "Velvet Theme",
       "description": "Show the Velvet theme option in the macOS appearance settings. Velvet is a dark-mode variant with red/pink accent colors.",
       "defaultEnabled": false
+    },
+    {
+      "id": "query-complexity-routing",
+      "scope": "assistant",
+      "key": "query-complexity-routing",
+      "label": "Query Complexity Routing",
+      "description": "Automatically route user messages to the most appropriate inference profile based on query complexity. Simple queries use the speed profile, complex queries escalate to the quality profile. The user is notified of each switch and can opt out by pinning a profile on the conversation.",
+      "defaultEnabled": false
+    },
+    {
+      "id": "queue-steering",
+      "scope": "assistant",
+      "key": "queue-steering",
+      "label": "Queue Steering",
+      "description": "Enable the 'Push to agent' button on queued messages, allowing users to steer the assistant to a specific queued message by aborting the current generation and promoting the message to the head of the queue.",
+      "defaultEnabled": false
+    },
+    {
+      "id": "chat-pull-to-refresh-enabled",
+      "scope": "client",
+      "key": "chat-pull-to-refresh-enabled",
+      "label": "Chat Pull to Refresh",
+      "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",
+      "key": "platform-notifications",
+      "label": "Platform Notifications",
+      "description": "Enable the Notifications tab in settings.",
+      "defaultEnabled": false
+    },
+    {
+      "id": "rollback-enabled",
+      "scope": "assistant",
+      "key": "rollback-enabled",
+      "label": "Rollback Enabled",
+      "description": "Show older versions in the version picker, allowing rollback to previous releases.",
+      "defaultEnabled": false
+    },
+    {
+      "id": "self-hosted-assistant",
+      "scope": "client",
+      "key": "self-hosted-assistant",
+      "label": "Self-Hosted Assistant",
+      "description": "Enable self-hosted assistant configuration.",
+      "defaultEnabled": false
+    },
+    {
+      "id": "settings-sleep-policy",
+      "scope": "assistant",
+      "key": "settings-sleep-policy",
+      "label": "Settings Sleep Policy",
+      "description": "Enable sleep policy settings.",
+      "defaultEnabled": false
+    },
+    {
+      "id": "velvet",
+      "scope": "client",
+      "key": "velvet",
+      "label": "Velvet",
+      "description": "Enable the Velvet design theme.",
+      "defaultEnabled": false
     }
   ]
 }

package/src/config/llm-resolver.ts

@@ -81,13 +81,27 @@ export function resolveCallSiteConfig(
 type Mergeable = Record<string, unknown>;
 
 /**
- * Returns the resolved default profile key for a call site, accounting for
- * the `custom-*` user-profile fallback when the managed profile is unavailable.
+ * Returns the effective default profile key the resolver would actually
+ * select for a call site when no per-turn `overrideProfile` is supplied.
+ *
+ * Mirrors the layering in `resolveCallSiteConfig`:
+ * - For `mainAgent`, the workspace's `activeProfile` sits ABOVE the
+ *   call-site catalog default (and above any static `llm.callSites.mainAgent`
+ *   override), so a non-disabled `activeProfile` wins.
+ * - For other call sites, the catalog default sits ABOVE `activeProfile`,
+ *   so the catalog default (with `custom-*` fallback) wins.
  */
 export function resolveDefaultProfileKey(
   callSite: LLMCallSite,
   llm: z.infer<typeof LLMSchema>,
 ): string | undefined {
+  if (callSite === "mainAgent" && llm.activeProfile != null) {
+    const active = llm.profiles?.[llm.activeProfile];
+    if (active != null && active.status !== "disabled") {
+      return llm.activeProfile;
+    }
+  }
+
   const dflt = CALL_SITE_DEFAULTS[callSite];
   if (dflt?.profile == null) return undefined;
   const target = llm.profiles?.[dflt.profile];

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

@@ -23,6 +23,7 @@ describe("MemoryV2ConfigSchema", () => {
       bm25_k1: 1.2,
       bm25_b: 0.4,
       consolidation_interval_hours: 4,
+      consolidation_max_buffer_lines: 100,
       max_page_chars: 5000,
       consolidation_prompt_path: null,
       rerank: {
@@ -36,6 +37,9 @@ describe("MemoryV2ConfigSchema", () => {
         enabled: true,
         max_page_ids: 25,
         router_prompt_path: null,
+        batch_size: null,
+        tier1_size: null,
+        tier2_size: null,
       },
     });
   });

package/src/config/schemas/call-site-catalog.ts

@@ -300,6 +300,27 @@ const CATALOG_RECORD: CatalogRecord = {
       "Builds the personalized artifact in a background conversation with tool access.",
     domain: "agentLoop",
   },
+  homeGreeting: {
+    id: "homeGreeting",
+    displayName: "Home Greeting",
+    description:
+      "Generates the personalized greeting shown on the Home page in the assistant's tone/persona.",
+    domain: "ui",
+  },
+  homeSuggestedPrompts: {
+    id: "homeSuggestedPrompts",
+    displayName: "Home Suggested Prompts",
+    description:
+      "Generates contextual conversation-starter suggestions for the Home page.",
+    domain: "ui",
+  },
+  queryComplexityRouter: {
+    id: "queryComplexityRouter",
+    displayName: "Query Complexity Router",
+    description:
+      "Classifies user message complexity to route to the appropriate inference profile.",
+    domain: "agentLoop",
+  },
 };
 
 // Source of truth for call-site display metadata. API responses and usage

package/src/config/schemas/llm.ts

@@ -76,6 +76,9 @@ export const LLMCallSiteEnum = z.enum([
   "trustRuleSuggestion",
   "proactiveArtifactDecision",
   "proactiveArtifactBuild",
+  "homeGreeting",
+  "homeSuggestedPrompts",
+  "queryComplexityRouter",
 ]);
 export type LLMCallSite = z.infer<typeof LLMCallSiteEnum>;
 

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

@@ -194,6 +194,19 @@ export const MemoryV2ConfigSchema = z
       .describe(
         "Hours between scheduled consolidation runs that synthesize buffered memories into concept pages",
       ),
+    consolidation_max_buffer_lines: z
+      .number({
+        error: "memory.v2.consolidation_max_buffer_lines must be a number",
+      })
+      .int("memory.v2.consolidation_max_buffer_lines must be an integer")
+      .positive(
+        "memory.v2.consolidation_max_buffer_lines must be a positive integer",
+      )
+      .nullable()
+      .default(100)
+      .describe(
+        "Size-based trigger for consolidation. When `memory/buffer.md` reaches this many non-empty lines, consolidation runs even if the time-based interval hasn't elapsed. Defaults to 100. Set to `null` to disable the size trigger and rely solely on `consolidation_interval_hours`.",
+      ),
     max_page_chars: z
       .number({ error: "memory.v2.max_page_chars must be a number" })
       .int("memory.v2.max_page_chars must be an integer")
@@ -282,8 +295,42 @@ export const MemoryV2ConfigSchema = z
           .describe(
             "Optional path to a file whose contents replace the bundled router prompt. Absolute paths are used as-is, a leading `~/` is expanded to the home directory, otherwise the path is resolved under the workspace root. The loaded contents may include `{{ASSISTANT_NAME}}`, `{{USER_NAME}}`, and `{{PAGE_INDEX}}`, which are substituted at runtime. If the file is missing, unreadable, or empty, the bundled prompt is used and a warning is logged.",
           ),
+        batch_size: z
+          .number()
+          .int()
+          .min(1)
+          .nullable()
+          .default(null)
+          .describe(
+            "Target batch size for parallel page-index routing. `null` (default) sends the entire page index in one call — identical to v3 behavior. When set, pages are split into `ceil(N / batch_size)` batches by stable FNV-1a hash on slug (so adding/removing a single page only invalidates one batch's KV cache), routed in parallel, and the selected slugs are unioned. A failure in one batch does not abort the turn as long as at least one batch succeeds.",
+          ),
+        tier1_size: z
+          .number()
+          .int()
+          .min(1)
+          .nullable()
+          .default(null)
+          .describe(
+            "Pool size for the tier-1 'recently modified' batch. `null` (default) disables tier 1 entirely — all pages flow through tier 3 batching. When set, the top-N concept pages by file mtime become their own dedicated parallel batch with mtime-desc ordering; everything else is partitioned into tier 3 batches by `batch_size`. Synthetic entries (skills, CLI commands) have mtime=0 and naturally rank below real concept pages so they don't crowd tier 1.",
+          ),
+        tier2_size: z
+          .number()
+          .int()
+          .min(1)
+          .nullable()
+          .default(null)
+          .describe(
+            "Pool size for the tier-2 'useful' batch. `null` (default) disables tier 2 — pages skip straight from tier 1 to tier 3. When set, the top-M pages by injection-frequency EMA (excluding tier 1) become their own parallel batch ordered by score desc. Pages with score 0 (never selected since EMA tracking began) are ineligible for tier 2 and stay in tier 3 regardless of `tier2_size`. Score is the time-decayed sum `Σ exp(-λ(now - tᵢ))` with 3-day half-life, computed on read from `memory_v2_injection_events`.",
+          ),
+      })
+      .default({
+        enabled: true,
+        max_page_ids: 25,
+        router_prompt_path: null,
+        batch_size: null,
+        tier1_size: null,
+        tier2_size: null,
       })
-      .default({ enabled: true, max_page_ids: 25, router_prompt_path: null })
       .describe(
         "LLM router configuration. When enabled, a single router LLM call replaces spreading activation for per-turn page selection.",
       ),

package/src/context/compactor.ts

@@ -155,6 +155,11 @@ export interface CompactionRunResult {
   thresholdTokens: number;
   compactedMessages: number;
   compactedPersistedMessages: number;
+  /**
+   * Number of recent ("tail") messages preserved verbatim alongside the
+   * summary. Omitted on no-op / skipped results — defaults to 0 at render.
+   */
+  preservedTailMessages?: number;
   summaryCalls: number;
   summaryInputTokens: number;
   summaryOutputTokens: number;
@@ -327,7 +332,7 @@ export function renderImageManifest(entries: ManifestEntry[]): string {
  * runtime emitted — typically
  * `2026-04-02 (Thursday) 01:52:33 -05:00 (America/Chicago)`).
  */
-function extractTurnContextTimestamp(message: Message): string | null {
+export function extractTurnContextTimestamp(message: Message): string | null {
   if (message.role !== "user") return null;
   for (const block of message.content) {
     if (block.type !== "text") continue;
@@ -841,6 +846,7 @@ export async function runAssistantDrivenCompaction(
     thresholdTokens,
     compactedMessages: compactableMessages.length,
     compactedPersistedMessages,
+    preservedTailMessages: args.messages.length - tailIndex,
     summaryCalls: 1,
     summaryInputTokens: response.usage.inputTokens,
     summaryOutputTokens: response.usage.outputTokens,
@@ -1090,6 +1096,7 @@ export async function runEmergencyCompaction(
     thresholdTokens,
     compactedMessages: compactedCount,
     compactedPersistedMessages: Math.max(0, compactedCount - nonPersistedAway),
+    preservedTailMessages: keptTail.length,
     summaryCalls: 1,
     summaryInputTokens: response.usage.inputTokens,
     summaryOutputTokens: response.usage.outputTokens,

package/src/context/token-estimator.ts

@@ -52,6 +52,24 @@ const IMAGE_MAX_PIXELS = 1_200_000;
 const IMAGE_TOKENS_PER_PIXEL = 1 / 750;
 const IMAGE_MAX_TOKENS = 1_600;
 
+// Gemini prices images differently: any side ≤384px counts as a single 258-token
+// tile; anything larger is resized so the longest side is ≤3072px and then
+// split into 768x768 tiles at 258 tokens each. A 4000x4000 image clamps to
+// 3072x3072 → ceil(3072/768)^2 = 16 tiles = 4,128 tokens. Without the clamp
+// we'd over-count it as 36 tiles (~9,288 tokens) and trigger spurious
+// compaction. The clamped 16-tile, 4,128-token figure is also the per-image
+// ceiling we fall back to when dimensions are unparseable (e.g. HEIC/HEIF
+// from iOS attachments) — the generic 1,600 cap can under-count Gemini
+// images by ~2.5x.
+// See: https://ai.google.dev/gemini-api/docs/tokens#multimodal-tokens
+const GEMINI_IMAGE_SMALL_THRESHOLD = 384;
+const GEMINI_IMAGE_TILE_SIZE = 768;
+const GEMINI_IMAGE_TOKENS_PER_TILE = 258;
+const GEMINI_IMAGE_MAX_DIMENSION = 3072;
+const GEMINI_IMAGE_MAX_TOKENS =
+  Math.ceil(GEMINI_IMAGE_MAX_DIMENSION / GEMINI_IMAGE_TILE_SIZE) ** 2 *
+  GEMINI_IMAGE_TOKENS_PER_TILE;
+
 // Anthropic renders each PDF page as an image (~1,568 tokens at standard
 // resolution) plus any extracted text. Typical PDF pages are 50-150 KB.
 // Using ~100 KB/page and ~1,600 tokens/page gives ~0.016 tokens/byte.
@@ -129,16 +147,41 @@ function estimateImageTokensByDimensions(
   return Math.ceil(scaledWidth * scaledHeight * IMAGE_TOKENS_PER_PIXEL);
 }
 
+function estimateGeminiImageTokens(width: number, height: number): number {
+  if (
+    width <= GEMINI_IMAGE_SMALL_THRESHOLD &&
+    height <= GEMINI_IMAGE_SMALL_THRESHOLD
+  ) {
+    return GEMINI_IMAGE_TOKENS_PER_TILE;
+  }
+  // Gemini resizes images so the longest side is ≤3072px before tiling.
+  const clampedWidth = Math.min(width, GEMINI_IMAGE_MAX_DIMENSION);
+  const clampedHeight = Math.min(height, GEMINI_IMAGE_MAX_DIMENSION);
+  const tilesWide = Math.ceil(clampedWidth / GEMINI_IMAGE_TILE_SIZE);
+  const tilesHigh = Math.ceil(clampedHeight / GEMINI_IMAGE_TILE_SIZE);
+  return tilesWide * tilesHigh * GEMINI_IMAGE_TOKENS_PER_TILE;
+}
+
 function estimateImageTokens(
   block: Extract<ContentBlock, { type: "image" }>,
+  options?: TokenEstimatorOptions,
 ): number {
   const dims = parseImageDimensions(block.source.data, block.source.media_type);
   if (dims) {
+    if (options?.providerName === "gemini") {
+      return estimateGeminiImageTokens(dims.width, dims.height);
+    }
     return estimateImageTokensByDimensions(dims.width, dims.height);
   }
-  // Dimensions unparseable (corrupt header, exotic format): use the per-image
-  // cap rather than the raw base64 length, which over-counts by 30-100x for
-  // non-Anthropic providers and trips spurious compaction.
+  // Dimensions unparseable (corrupt header, or formats parseImageDimensions
+  // doesn't recognize like HEIC/HEIF coming from iOS attachments). Fall back
+  // to the per-provider per-image ceiling rather than the raw base64 length,
+  // which over-counts by 30-100x. Gemini's tile pricing tops out well above
+  // the universal 1,600-token cap, so use its max-tile budget instead to
+  // avoid under-counting large iPhone screenshots.
+  if (options?.providerName === "gemini") {
+    return GEMINI_IMAGE_MAX_TOKENS;
+  }
   return IMAGE_MAX_TOKENS;
 }
 
@@ -186,7 +229,7 @@ export function estimateContentBlockTokens(
       return (
         IMAGE_BLOCK_OVERHEAD_TOKENS +
         estimateTextTokens(block.source.media_type) +
-        estimateImageTokens(block)
+        estimateImageTokens(block, options)
       );
     case "file":
       return (

package/src/context/window-manager.ts

@@ -51,6 +51,11 @@ export interface ContextWindowResult {
   thresholdTokens: number;
   compactedMessages: number;
   compactedPersistedMessages: number;
+  /**
+   * Number of recent ("tail") messages preserved verbatim alongside the
+   * summary. Omitted on no-op / skipped results — defaults to 0 at render.
+   */
+  preservedTailMessages?: number;
   summaryCalls: number;
   summaryInputTokens: number;
   summaryOutputTokens: number;
@@ -216,6 +221,26 @@ export class ContextWindowManager {
     return getConfig().compaction;
   }
 
+  get maxInputTokens(): number {
+    return this.config.maxInputTokens;
+  }
+
+  /**
+   * Estimate the prompt-token cost of `messages` using the same path as the
+   * auto-compaction pre-check. Clears the system-prompt cache so the next
+   * turn re-resolves it (the system prompt is lazy and may have changed).
+   */
+  estimateInputTokens(messages: Message[]): number {
+    try {
+      return estimatePromptTokens(messages, this.systemPrompt, {
+        providerName: this.estimationProviderName,
+        toolTokenBudget: this.toolTokenBudget,
+      });
+    } finally {
+      this.clearSystemPromptCache();
+    }
+  }
+
   /**
    * Cheap pre-check — estimate the current token count and compare against
    * `compaction.autoThreshold`. Callers pass the estimate back through

package/src/credential-health/credential-health-service.ts

@@ -315,9 +315,11 @@ async function checkManagedProvider(
     const client = await VellumPlatformClient.create();
     if (!client?.platformAssistantId) return results;
 
+    // Query without a status filter so we can distinguish "never
+    // connected" (empty result) from "previously connected but now
+    // inactive" (non-empty result with no ACTIVE entries).
     const params = new URLSearchParams();
     params.set("provider", providerRow.provider);
-    params.set("status", "ACTIVE");
 
     const path = `/v1/assistants/${encodeURIComponent(client.platformAssistantId)}/oauth/connections/?${params.toString()}`;
     const response = await client.fetch(path);
@@ -331,19 +333,30 @@ async function checkManagedProvider(
     }
 
     const body = (await response.json()) as unknown;
-    const connections = (
+    const allConnections = (
       Array.isArray(body)
         ? body
         : ((body as Record<string, unknown>).results ?? [])
-    ) as Array<{ id: string; account_label?: string }>;
+    ) as Array<{ id: string; account_label?: string; status?: string }>;
+
+    if (allConnections.length === 0) {
+      // No connections of any status — the user has never connected this
+      // provider. The suggested-prompts system handles prompting them to
+      // connect; this is not a health issue.
+      return results;
+    }
+
+    const connections = allConnections.filter(
+      (c) => (c.status ?? "ACTIVE").toUpperCase() === "ACTIVE",
+    );
 
     if (connections.length === 0) {
-      // No active managed connections — report as missing so the
-      // heartbeat can notify the user.
+      // Connections exist but none are active — the user previously
+      // connected and the connection was revoked/deactivated.
       results.push({
         connectionId: `managed:${providerRow.provider}`,
         provider: providerRow.provider,
-        accountInfo: null,
+        accountInfo: allConnections[0]?.account_label ?? null,
         status: "missing_token",
         details: `No active managed connection for ${providerRow.provider}. Reconnect on the Vellum platform.`,
         missingScopes: [],
@@ -550,9 +563,20 @@ export async function checkAllCredentials(): Promise<CredentialHealthReport> {
   for (const providerRow of providers) {
     if (!(await isManagedProvider(providerRow))) continue;
 
-    // If the provider is in managed mode and also has BYO connections,
-    // remove the stale BYO results — managed mode takes priority.
-    if (byoProviders.has(providerRow.provider)) {
+    let managedResults: CredentialHealthResult[] = [];
+    try {
+      managedResults = await checkManagedProvider(providerRow);
+    } catch (err) {
+      log.warn(
+        { err, provider: providerRow.provider },
+        "Failed to check managed provider health",
+      );
+    }
+
+    // Only replace BYO results with managed results when the managed
+    // check returned something. If managed returned empty (user never
+    // connected via managed mode), keep any existing BYO results.
+    if (managedResults.length > 0 && byoProviders.has(providerRow.provider)) {
       const beforeLen = results.length;
       const filtered = results.filter(
         (r) => r.provider !== providerRow.provider,
@@ -562,16 +586,7 @@ export async function checkAllCredentials(): Promise<CredentialHealthReport> {
         results.push(...filtered);
       }
     }
-
-    try {
-      const managedResults = await checkManagedProvider(providerRow);
-      results.push(...managedResults);
-    } catch (err) {
-      log.warn(
-        { err, provider: providerRow.provider },
-        "Failed to check managed provider health",
-      );
-    }
+    results.push(...managedResults);
   }
 
   const unhealthy = results.filter((r) => r.status !== "healthy");

package/src/daemon/__tests__/conversation-tool-setup.test.ts

@@ -49,11 +49,8 @@ mock.module("../../runtime/assistant-event-hub.js", () => ({
 
 // Dynamic imports after mock.module calls so the stubs take effect
 // before the modules under test are loaded.
-const {
-  HOST_TOOL_NAMES,
-  HOST_TOOL_TO_CAPABILITY,
-  isToolActiveForContext,
-} = await import("../conversation-tool-setup.js");
+const { HOST_TOOL_NAMES, HOST_TOOL_TO_CAPABILITY, isToolActiveForContext } =
+  await import("../conversation-tool-setup.js");
 type SkillProjectionContext =
   import("../conversation-tool-setup.js").SkillProjectionContext;
 type SkillProjectionCache =
@@ -75,6 +72,66 @@ beforeEach(() => {
   mockClientCountByCapability.clear();
 });
 
+describe("isToolActiveForContext — Slack task_progress UI exception", () => {
+  test("ui_show and ui_update are active for Slack task_progress turns", () => {
+    const ctx = makeCtx({
+      hasNoClient: false,
+      channelCapabilities: {
+        channel: "slack",
+        supportsDynamicUi: false,
+      },
+    });
+
+    expect(isToolActiveForContext("ui_show", ctx)).toBe(true);
+    expect(isToolActiveForContext("ui_update", ctx)).toBe(true);
+  });
+
+  test("ui_dismiss remains hidden for Slack without dynamic UI support", () => {
+    expect(
+      isToolActiveForContext(
+        "ui_dismiss",
+        makeCtx({
+          hasNoClient: false,
+          channelCapabilities: {
+            channel: "slack",
+            supportsDynamicUi: false,
+          },
+        }),
+      ),
+    ).toBe(false);
+  });
+
+  test("Slack task_progress UI tools still require a connected client path", () => {
+    expect(
+      isToolActiveForContext(
+        "ui_show",
+        makeCtx({
+          hasNoClient: true,
+          channelCapabilities: {
+            channel: "slack",
+            supportsDynamicUi: false,
+          },
+        }),
+      ),
+    ).toBe(false);
+  });
+
+  test("other non-dynamic channels still hide UI surface tools", () => {
+    expect(
+      isToolActiveForContext(
+        "ui_show",
+        makeCtx({
+          hasNoClient: false,
+          channelCapabilities: {
+            channel: "telegram",
+            supportsDynamicUi: false,
+          },
+        }),
+      ),
+    ).toBe(false);
+  });
+});
+
 describe("isToolActiveForContext — host tool capability gating", () => {
   // macOS transport: SSE-based interactive approval required.
   test("host_bash is active for macOS with a connected client", () => {
@@ -331,7 +388,10 @@ describe("isToolActiveForContext — cross-client exposure for host_file_*", ()
       expect(
         isToolActiveForContext(
           tool,
-          makeCtx({ hasNoClient: true, transportInterface: "chrome-extension" }),
+          makeCtx({
+            hasNoClient: true,
+            transportInterface: "chrome-extension",
+          }),
         ),
       ).toBe(false);
     });