@vellumai/assistant 0.8.5 → 0.8.7-dev.202606052118.34cd356

package/src/daemon/handlers/conversations.ts

@@ -3,15 +3,16 @@ import { v4 as uuid } from "uuid";
 import { clearAll, getConversation } from "../../memory/conversation-crud.js";
 import { resolveConversationId } from "../../memory/conversation-key-store.js";
 import { broadcastMessage } from "../../runtime/assistant-event-hub.js";
-import * as pendingInteractions from "../../runtime/pending-interactions.js";
 import { getSubagentManager } from "../../subagent/index.js";
 import { createAbortReason } from "../../util/abort-reasons.js";
 import { truncate } from "../../util/truncate.js";
 import { regenerate } from "../conversation-history.js";
 import {
-  clearAllActiveConversations,
   conversationEntries,
   findConversation,
+} from "../conversation-registry.js";
+import {
+  clearAllActiveConversations,
   getOrCreateConversation,
   touchConversation,
 } from "../conversation-store.js";
@@ -28,23 +29,11 @@ export function handleConfirmationResponse(msg: ConfirmationResponse): void {
   for (const [conversationId, conversation] of conversationEntries()) {
     if (conversation.hasPendingConfirmation(msg.requestId)) {
       touchConversation(conversationId);
-      conversation.handleConfirmationResponse(
-        msg.requestId,
-        decision,
-        msg.selectedPattern,
-        msg.selectedScope,
-        undefined,
-        { source: "button" },
-      );
-      // Idempotent cleanup: the prompter's `resolveConfirmation` already
-      // resolved the interaction with the correct approved/rejected state
-      // before we reach here, so the entry is typically gone. This call is
-      // a safety net for paths where the prompter no-ops; map decision to
-      // state to keep the emitted event accurate when it does fire.
-      pendingInteractions.resolve(
-        msg.requestId,
-        decision === "allow" ? "approved" : "rejected",
-      );
+      conversation.handleConfirmationResponse(msg.requestId, decision, {
+        selectedPattern: msg.selectedPattern,
+        selectedScope: msg.selectedScope,
+        emissionContext: { source: "button" },
+      });
       return;
     }
   }
@@ -225,10 +214,7 @@ export function steerToMessage(
   | { steered: true }
   | {
       steered: false;
-      reason:
-        | "conversation_not_found"
-        | "message_not_found"
-        | "not_processing";
+      reason: "conversation_not_found" | "message_not_found" | "not_processing";
     } {
   const conversation = findConversation(conversationId);
   if (!conversation) {

package/src/daemon/handlers/shared.ts

@@ -1,9 +1,16 @@
 import { v4 as uuid } from "uuid";
 
+import type {
+  ConversationContentBlock,
+  ConversationMessageAttachment,
+  ConversationMessageSurface,
+  ConversationMessageToolCall,
+} from "../../api/responses/conversation-message.js";
+import { ConfirmationDecisionSchema } from "../../api/responses/conversation-message.js";
 import { getConfig } from "../../config/loader.js";
 import type { LLMCallSite, Speed } from "../../config/schemas/llm.js";
 import type { SecretPromptResult } from "../../permissions/secret-prompter.js";
-import { isPlaceholderSentinelText } from "../../providers/anthropic/client.js";
+import { isPlaceholderSentinelText } from "../../providers/placeholder-sentinels.js";
 import { broadcastMessage } from "../../runtime/assistant-event-hub.js";
 import type { AuthContext } from "../../runtime/auth/types.js";
 import { unwrapExternalContentForDisplay } from "../../security/untrusted-content.js";
@@ -30,67 +37,18 @@ const pendingStandaloneSecrets = new Map<
   }
 >();
 
-export interface HistoryToolCall {
-  name: string;
-  input: Record<string, unknown>;
-  result?: string;
-  isError?: boolean;
-  /** Base64-encoded image data from tool contentBlocks (e.g. browser_screenshot). @deprecated Use imageDataList. */
-  imageData?: string;
-  /** Base64-encoded image data from tool contentBlocks (e.g. browser_screenshot, image generation). */
-  imageDataList?: string[];
-  /** Unix ms when the tool started executing. */
-  startedAt?: number;
-  /** Unix ms when the tool completed. */
-  completedAt?: number;
-  /** Confirmation decision for this tool call: "approved" | "denied" | "timed_out". */
-  confirmationDecision?: string;
-  /** Friendly label for the confirmation (e.g. "Edit File", "Run Command"). */
-  confirmationLabel?: string;
-  /** Risk level classification at invocation time ("low" | "medium" | "high" | "unknown"). */
-  riskLevel?: string;
-  /** Human-readable reason for the risk classification. */
-  riskReason?: string;
-  /** ID of the trust rule that matched this invocation (if any). */
-  matchedTrustRuleId?: string;
-  /**
-   * @deprecated Use `approvalMode` and `approvalReason` instead.
-   * Kept for backward compatibility during the migration window.
-   */
-  autoApproved?: boolean;
-  /** How the approval decision was reached: prompted, auto, blocked, or unknown (legacy). */
-  approvalMode?: string;
-  /** Why the approval decision was reached (stable enum for client display). */
-  approvalReason?: string;
-  /** Snapshot of the auto-approve threshold at execution time. */
-  riskThreshold?: string;
-  /**
-   * Display-only regex ladder for the rule editor (narrowest → broadest).
-   * Persisted on tool_use blocks by `annotatePersistedAssistantMessage` so
-   * historical chips render the same ladder as live tool_result events.
-   */
-  riskScopeOptions?: Array<{ pattern: string; label: string }>;
-  /** Minimatch save patterns for the rule editor (narrowest → broadest). */
-  riskAllowlistOptions?: Array<{
-    label: string;
-    description: string;
-    pattern: string;
-  }>;
-  /** Directory scope ladder for the rule editor. */
-  riskDirectoryScopeOptions?: Array<{ scope: string; label: string }>;
-}
+/**
+ * A single tool call rendered into a history row. Alias of the canonical
+ * wire-contract type so `renderHistoryContent` (the producer) cannot drift
+ * from what the messages endpoint serializes.
+ */
+export type HistoryToolCall = ConversationMessageToolCall;
 
-export interface HistorySurface {
-  surfaceId: string;
-  surfaceType: string;
-  title?: string;
-  data: Record<string, unknown>;
-  actions?: Array<{ id: string; label: string; style?: string }>;
-  display?: string;
-  persistent?: boolean;
-  completed?: boolean;
-  completionSummary?: string;
-}
+/**
+ * A UI surface (widget) embedded in a history row. Alias of the canonical
+ * wire-contract type so the producer matches the serialized shape.
+ */
+export type HistorySurface = ConversationMessageSurface;
 
 /**
  * Positional reference to a file attachment captured while walking the
@@ -124,6 +82,16 @@ export interface RenderedHistoryContent {
    * attachment metadata to this ordering for inline placement.
    */
   attachments: HistoryAttachmentRef[];
+  /**
+   * Unified ordered content blocks built directly from the model-native
+   * content during the single walk — the wire `contentBlocks` projection.
+   * `attachment` blocks are inlined for file blocks whose DB-hydrated metadata
+   * the caller supplies via the `attachmentBlocks` argument (matched by
+   * attachment-ref order); a file block with no supplied metadata produces no
+   * block. Every other block type is always complete, so the serializer ships
+   * this array as-is with no post-processing.
+   */
+  contentBlocks: ConversationContentBlock[];
 }
 
 /**
@@ -249,6 +217,44 @@ function extractFileBlockMetadata(
   };
 }
 
+/**
+ * Build the positional attachment reference for a `file` content block:
+ * filename/mime/size from the block's source plus the persisted
+ * `_attachmentId` when present (user-uploaded files).
+ */
+function fileBlockToAttachmentRef(
+  block: Record<string, unknown>,
+  meta: FileBlockMetadata,
+): HistoryAttachmentRef {
+  const ref: HistoryAttachmentRef = {
+    filename: meta.filename,
+    mimeType: meta.mediaType,
+    sizeBytes: meta.sizeBytes,
+  };
+  if (typeof block._attachmentId === "string" && block._attachmentId) {
+    ref.attachmentId = block._attachmentId;
+  }
+  return ref;
+}
+
+/**
+ * Collect file-block attachment references in content-walk order without
+ * building the full history projection. The serializer aligns its DB-hydrated
+ * attachment rows against this ordering, then feeds the resolved metadata back
+ * into `renderHistoryContent` so it inlines `attachment` blocks during the walk.
+ */
+export function collectAttachmentRefs(
+  content: unknown,
+): HistoryAttachmentRef[] {
+  if (!Array.isArray(content)) return [];
+  const refs: HistoryAttachmentRef[] = [];
+  for (const block of content) {
+    if (!isRecord(block) || block.type !== "file") continue;
+    refs.push(fileBlockToAttachmentRef(block, extractFileBlockMetadata(block)));
+  }
+  return refs;
+}
+
 function renderFileBlockForHistory(
   block: Record<string, unknown>,
   meta: FileBlockMetadata,
@@ -270,7 +276,13 @@ function renderFileBlockForHistory(
   )}`;
 }
 
-export function renderHistoryContent(content: unknown): RenderedHistoryContent {
+export function renderHistoryContent(
+  content: unknown,
+  attachmentBlocks?: ReadonlyArray<
+    ConversationMessageAttachment | null | undefined
+  >,
+  messageId?: string,
+): RenderedHistoryContent {
   if (!Array.isArray(content)) {
     let text: string;
     if (content == null) {
@@ -289,6 +301,7 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
       surfaces: [],
       thinkingSegments: [],
       attachments: [],
+      contentBlocks: text ? [{ type: "text", text }] : [],
     };
   }
 
@@ -309,6 +322,13 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
   let currentSegmentParts: string[] = [];
   let hasOpenSegment = false;
 
+  // Unified content blocks built in lockstep with the positional arrays as we
+  // walk the model-native content. `attachment` blocks are inlined here when
+  // the caller supplied DB-hydrated metadata in `attachmentBlocks`, matched by
+  // attachment-ref order; otherwise the file block contributes no block.
+  const contentBlocks: ConversationContentBlock[] = [];
+  let currentTextBlock: { type: "text"; text: string } | null = null;
+
   function joinWithSpacing(parts: string[]): string {
     let result = parts[0] ?? "";
     for (let i = 1; i < parts.length; i++) {
@@ -334,18 +354,42 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
 
   function finalizeSegment(): void {
     if (hasOpenSegment) {
-      textSegments[textSegments.length - 1] =
-        joinWithSpacing(currentSegmentParts);
+      const joined = joinWithSpacing(currentSegmentParts);
+      textSegments[textSegments.length - 1] = joined;
+      if (currentTextBlock) {
+        currentTextBlock.text = joined;
+        currentTextBlock = null;
+      }
       currentSegmentParts = [];
       hasOpenSegment = false;
     }
   }
 
-  function ensureSegment(): void {
+  // Flush the open text segment into its tracked block and stop tracking it,
+  // without closing the segment. Used before folding the synthetic attachment
+  // description into the trailing segment: it stays in the legacy
+  // `textSegments`/`text` body but must not pollute the clean contentBlocks,
+  // since `attachment` blocks already carry that metadata.
+  function detachTextBlock(): void {
+    if (currentTextBlock) {
+      currentTextBlock.text = joinWithSpacing(currentSegmentParts);
+      currentTextBlock = null;
+    }
+  }
+
+  // `trackBlock` mirrors the segment into `contentBlocks`. The trailing
+  // attachment-description segment (legacy `message.text` for clients without
+  // attachment UI) sets it false so it isn't duplicated as a text block —
+  // attachments surface as `attachment` blocks instead.
+  function ensureSegment(trackBlock = true): void {
     if (!hasOpenSegment) {
       textSegments.push("");
       contentOrder.push(`text:${textSegments.length - 1}`);
       hasOpenSegment = true;
+      if (trackBlock) {
+        currentTextBlock = { type: "text", text: "" };
+        contentBlocks.push(currentTextBlock);
+      }
     }
   }
 
@@ -371,9 +415,12 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
           typeof block.completionSummary === "string"
             ? block.completionSummary
             : undefined,
+        toolCallId:
+          typeof block.toolCallId === "string" ? block.toolCallId : undefined,
       };
       surfaces.push(surface);
       contentOrder.push(`surface:${surfaces.length - 1}`);
+      contentBlocks.push({ type: "surface", surface });
       continue;
     }
 
@@ -381,6 +428,7 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
       finalizeSegment();
       thinkingSegments.push(block.thinking);
       contentOrder.push(`thinking:${thinkingSegments.length - 1}`);
+      contentBlocks.push({ type: "thinking", thinking: block.thinking });
       continue;
     }
 
@@ -407,16 +455,13 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
       const meta = extractFileBlockMetadata(block);
       attachmentParts.push(renderFileBlockForHistory(block, meta));
       finalizeSegment();
-      const ref: HistoryAttachmentRef = {
-        filename: meta.filename,
-        mimeType: meta.mediaType,
-        sizeBytes: meta.sizeBytes,
-      };
-      if (typeof block._attachmentId === "string" && block._attachmentId) {
-        ref.attachmentId = block._attachmentId;
+      attachments.push(fileBlockToAttachmentRef(block, meta));
+      const refIndex = attachments.length - 1;
+      contentOrder.push(`attachment:${refIndex}`);
+      const hydrated = attachmentBlocks?.[refIndex];
+      if (hydrated) {
+        contentBlocks.push({ type: "attachment", attachment: hydrated });
       }
-      attachments.push(ref);
-      contentOrder.push(`attachment:${attachments.length - 1}`);
       continue;
     }
     if (block.type === "image") {
@@ -433,13 +478,18 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
         : {};
       const id = typeof block.id === "string" ? block.id : "";
       const entry: HistoryToolCall = { name, input };
+      if (id) entry.id = id;
       // Extract persisted timing/confirmation metadata
       if (typeof block._startedAt === "number")
         entry.startedAt = block._startedAt;
       if (typeof block._completedAt === "number")
         entry.completedAt = block._completedAt;
-      if (typeof block._confirmationDecision === "string")
-        entry.confirmationDecision = block._confirmationDecision;
+      const confirmationDecision = ConfirmationDecisionSchema.safeParse(
+        block._confirmationDecision,
+      );
+      if (confirmationDecision.success) {
+        entry.confirmationDecision = confirmationDecision.data;
+      }
       if (typeof block._confirmationLabel === "string")
         entry.confirmationLabel = block._confirmationLabel;
       if (typeof block._riskLevel === "string")
@@ -468,9 +518,18 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
       if (Array.isArray(block._riskDirectoryScopeOptions))
         entry.riskDirectoryScopeOptions =
           block._riskDirectoryScopeOptions as HistoryToolCall["riskDirectoryScopeOptions"];
+      // Read back tool activity (web_search / web_fetch) persisted by
+      // `annotatePersistedAssistantMessage` so the activity card survives a
+      // history reopen instead of degrading to the plain result text.
+      if (isRecord(block._activityMetadata))
+        entry.activityMetadata =
+          block._activityMetadata as HistoryToolCall["activityMetadata"];
       toolCalls.push(entry);
       if (id) pendingToolUses.set(id, entry);
       contentOrder.push(`tool:${toolCalls.length - 1}`);
+      // Same `entry` reference the block carries: a later tool_result pairs its
+      // output onto `entry`, so the content block reflects it automatically.
+      contentBlocks.push({ type: "tool_use", toolCall: entry });
       if (!seenToolUse) {
         seenToolUse = true;
         if (!seenText) toolCallsBeforeText = true;
@@ -485,9 +544,16 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
         : {};
       const id = typeof block.id === "string" ? block.id : "";
       const entry: HistoryToolCall = { name, input };
+      if (id) entry.id = id;
+      // Native server tools (Anthropic web_search) persist their activity on
+      // the server_tool_use block, so read it back here too.
+      if (isRecord(block._activityMetadata))
+        entry.activityMetadata =
+          block._activityMetadata as HistoryToolCall["activityMetadata"];
       toolCalls.push(entry);
       if (id) pendingToolUses.set(id, entry);
       contentOrder.push(`tool:${toolCalls.length - 1}`);
+      contentBlocks.push({ type: "tool_use", toolCall: entry });
       if (!seenToolUse) {
         seenToolUse = true;
         if (!seenText) toolCallsBeforeText = true;
@@ -577,14 +643,28 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
   // The macOS client handles this by selecting the *first* non-empty text
   // segment in interleaved content, so trailing attachment segments are safe.
   if (attachmentParts.length > 0) {
+    detachTextBlock();
     const attachmentText = attachmentParts.join("\n");
     const prefix = textParts.length > 0 ? "\n" : "";
-    ensureSegment();
+    ensureSegment(false);
     currentSegmentParts.push(prefix + attachmentText);
   }
 
   finalizeSegment();
 
+  // Default any tool call the provider left without an `id` to the same
+  // positional id the web client historically synthesized, so every wire tool
+  // call is self-identifying and snapshot/stream ids line up. `idx` indexes the
+  // final `toolCalls` array (the client keys off the same positions); the
+  // shared `entry` references mean `contentBlocks` reflect this for free.
+  if (messageId !== undefined) {
+    toolCalls.forEach((toolCall, idx) => {
+      if (toolCall.id === undefined) {
+        toolCall.id = `tool-history-${messageId}-${idx}`;
+      }
+    });
+  }
+
   const text = joinWithSpacing(textParts);
   let rendered: string;
   if (attachmentParts.length === 0) {
@@ -604,6 +684,7 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
     surfaces,
     thinkingSegments,
     attachments,
+    contentBlocks,
   };
 }
 

package/src/daemon/handlers/skills.ts

@@ -31,7 +31,10 @@ import {
   isTextMimeType as isTextMime,
   MAX_INLINE_TEXT_SIZE,
 } from "../../runtime/routes/workspace-utils.js";
-import { getCatalog } from "../../skills/catalog-cache.js";
+import {
+  getCachedCatalogSync,
+  getCatalog,
+} from "../../skills/catalog-cache.js";
 import type { SkillFileEntry } from "../../skills/catalog-files.js";
 import {
   catalogSkillToSlim,
@@ -48,7 +51,6 @@ import {
   installSkillLocally,
 } from "../../skills/catalog-install.js";
 import { filterByQuery } from "../../skills/catalog-search.js";
-import { inferCategory } from "../../skills/category-inference.js";
 import type { ClawhubInspectResult } from "../../skills/clawhub.js";
 import {
   clawhubCheckUpdates,
@@ -285,6 +287,25 @@ function postInstallSkill(skillId: string): void {
   refreshSkillCapabilityMemories(getConfig());
 }
 
+// ─── Catalog category lookup ────────────────────────────────────────────────
+
+let _catalogCategoryMap: Map<string, string> | null = null;
+let _catalogCategoryRef: readonly CatalogSkill[] | null = null;
+
+function getCatalogCategoryMap(): Map<string, string> {
+  const catalog = getCachedCatalogSync();
+  if (_catalogCategoryMap && _catalogCategoryRef === catalog) {
+    return _catalogCategoryMap;
+  }
+  _catalogCategoryMap = new Map();
+  for (const s of catalog) {
+    const cat = s.metadata?.vellum?.category;
+    if (cat) _catalogCategoryMap.set(s.id, cat);
+  }
+  _catalogCategoryRef = catalog;
+  return _catalogCategoryMap;
+}
+
 // ─── Kind / origin / status derivation ───────────────────────────────────────
 
 /** Map the old `source` field to the new `kind` axis. */
@@ -328,13 +349,16 @@ function toSlimSkillResponse(
   const origin = deriveOrigin(kind, summary.directoryPath, installMeta);
   const status: SlimSkillResponse["status"] = state;
 
+  const category = getCatalogCategoryMap().get(summary.id) ?? "system";
   const base = {
     id: summary.id,
     name: summary.displayName,
     description: summary.description,
+    icon: summary.icon,
     emoji: summary.emoji,
     kind,
     status,
+    category,
   } as const;
 
   switch (origin) {
@@ -393,14 +417,20 @@ export function listSkills(): SlimSkillResponse[] {
  * Installed skills take precedence when deduplicating by ID.
  */
 async function listSkillsWithCatalog(): Promise<SlimSkillResponse[]> {
-  const installed = listSkills();
-  const installedIds = new Set(installed.map((s) => s.id));
-
+  // Warm the catalog cache before converting installed skills so
+  // getCatalogCategoryMap() in toSlimSkillResponse() sees real categories
+  // instead of falling back to "system" on a cold cache.
   let catalogSkills: CatalogSkill[];
   try {
     catalogSkills = await getCatalog();
   } catch {
-    // If catalog fetch fails, return installed-only
+    catalogSkills = [];
+  }
+
+  const installed = listSkills();
+  const installedIds = new Set(installed.map((s) => s.id));
+
+  if (catalogSkills.length === 0) {
     return installed;
   }
 
@@ -511,16 +541,13 @@ export async function listSkillsFiltered(filter: SkillListFilter): Promise<{
   // Compute category counts BEFORE applying the category filter
   const categoryCounts: Record<string, number> = {};
   for (const s of skills) {
-    const cat = inferCategory(s.name, s.description);
-    categoryCounts[cat] = (categoryCounts[cat] ?? 0) + 1;
+    categoryCounts[s.category] = (categoryCounts[s.category] ?? 0) + 1;
   }
   const totalCount = skills.length;
 
   // Apply category filter
   if (filter.category) {
-    skills = skills.filter(
-      (s) => inferCategory(s.name, s.description) === filter.category,
-    );
+    skills = skills.filter((s) => s.category === filter.category);
   }
 
   // Sort: installed first, community origins before core within installed,
@@ -606,10 +633,12 @@ export async function getSkill(
       id: slim.id,
       name: slim.name,
       description: slim.description,
+      icon: slim.icon,
       emoji: slim.emoji,
       kind: slim.kind,
       origin: slim.origin,
       status: slim.status,
+      category: slim.category,
       slug: slim.slug,
       author: slim.author,
       stars: slim.stars,
@@ -645,10 +674,12 @@ export async function getSkill(
       id: slim.id,
       name: slim.name,
       description: slim.description,
+      icon: slim.icon,
       emoji: slim.emoji,
       kind: slim.kind,
       origin: slim.origin,
       status: slim.status,
+      category: slim.category,
       slug: slim.slug,
       sourceRepo: slim.sourceRepo,
       installs: slim.installs,
@@ -675,10 +706,12 @@ export async function getSkill(
     id: slim.id,
     name: slim.name,
     description: slim.description,
+    icon: slim.icon,
     emoji: slim.emoji,
     kind: slim.kind,
     origin: slim.origin,
     status: slim.status,
+    category: slim.category,
   };
   return { skill: detail };
 }
@@ -1363,10 +1396,12 @@ export async function searchSkills(
         id: s.id,
         name: s.displayName,
         description: s.description,
+        icon: s.icon,
         emoji: s.emoji,
         kind: "catalog" as const,
         origin: "vellum" as const,
         status: "available" as const,
+        category: getCatalogCategoryMap().get(s.id) ?? "system",
       };
     });
 
@@ -1385,6 +1420,7 @@ export async function searchSkills(
         kind: "catalog" as const,
         origin: "clawhub" as const,
         status: "available" as const,
+        category: "integrations",
         slug: s.slug,
         author: s.author,
         stars: s.stars,
@@ -1411,6 +1447,7 @@ export async function searchSkills(
         kind: "catalog" as const,
         origin: "skillssh" as const,
         status: "available" as const,
+        category: "integrations",
         slug: r.id,
         sourceRepo: r.source,
         installs: r.installs,
@@ -1562,15 +1599,11 @@ export async function draftSkill(params: {
               body.slice(0, 2000),
             ].join("\n");
 
-            const response = await provider.sendMessage(
-              [userMessage(prompt)],
-              [],
-              undefined,
-              {
-                config: { callSite: "skillCategoryInference", max_tokens: 256 },
-                signal,
-              },
-            );
+            const response = await provider.sendMessage([userMessage(prompt)], {
+              tools: [],
+              config: { callSite: "skillCategoryInference", max_tokens: 256 },
+              signal,
+            });
             cleanup();
 
             const responseText = extractText(response);