@vellumai/assistant 0.8.8 → 0.8.9-staging.2

package/src/context/window-manager.ts

@@ -40,7 +40,8 @@ const INTERNAL_CONTEXT_SUMMARY_MESSAGES = new WeakSet<Message>();
 
 // ---------------------------------------------------------------------------
 // Public types — preserved for downstream consumers (agent loop, conversation,
-// plugin pipeline, applyCompactionResult, routes/playground/force-compact).
+// default compaction plugin, applyCompactionResult,
+// routes/playground/force-compact).
 // ---------------------------------------------------------------------------
 
 export interface ContextWindowResult {
@@ -198,8 +199,7 @@ export class ContextWindowManager {
    * `compactedMessages` so `compactedPersistedMessages` only reflects DB
    * rows. Decremented after a successful compaction.
    */
-  nonPersistedPrefixCount = 0;
-  summaryIsInjected = false;
+  private _nonPersistedPrefixCount = 0;
   private _resolvedSystemPrompt: string | undefined;
 
   constructor(options: ContextWindowManagerOptions) {
@@ -215,6 +215,21 @@ export class ContextWindowManager {
     this.config = config;
   }
 
+  /** Leading non-persisted inherited-context messages the compactor preserves. */
+  get nonPersistedPrefixCount(): number {
+    return this._nonPersistedPrefixCount;
+  }
+
+  /**
+   * Seed the non-persisted inherited-context prefix when a forked/sub
+   * conversation inherits its parent's in-memory context. The compactor folds
+   * this prefix into the summary and {@link consumeCompactionState} decrements
+   * it as the leading messages are compacted away.
+   */
+  seedNonPersistedPrefix(count: number): void {
+    this._nonPersistedPrefixCount = count;
+  }
+
   private get estimationProviderName(): string {
     return this.provider.tokenEstimationProvider ?? this.provider.name;
   }
@@ -352,7 +367,7 @@ export class ContextWindowManager {
       signal,
       overrideProfile: options?.overrideProfile ?? null,
       actorTrustClass: options?.actorTrustClass,
-      nonPersistedPrefixCount: this.nonPersistedPrefixCount,
+      nonPersistedPrefixCount: this._nonPersistedPrefixCount,
     };
 
     // Retry budget for the compactor itself. Lives here (not in the
@@ -390,7 +405,7 @@ export class ContextWindowManager {
         ...baseArgs,
         messages: result.messages,
         previousEstimatedInputTokens: previousEstimate,
-        nonPersistedPrefixCount: this.nonPersistedPrefixCount,
+        nonPersistedPrefixCount: this._nonPersistedPrefixCount,
       });
       if (!nextResult.compacted) break;
       const nextEstimate = this.recomputePostCompactionEstimate(
@@ -434,21 +449,20 @@ export class ContextWindowManager {
   }
 
   /**
-   * Decrement the non-persisted prefix bookkeeping and clear the
-   * injected-summary flag after a productive compaction. Called once per
-   * successful internal attempt so multi-attempt runs keep the count
-   * honest as the leading injected messages get folded into the summary.
+   * Decrement the non-persisted prefix bookkeeping after a productive
+   * compaction. Called once per successful internal attempt so multi-attempt
+   * runs keep the count honest as the leading injected messages get folded
+   * into the summary.
    */
   private consumeCompactionState(compactedMessages: number): void {
     const compactedAway = Math.min(
-      this.nonPersistedPrefixCount,
+      this._nonPersistedPrefixCount,
       compactedMessages,
     );
-    this.nonPersistedPrefixCount = Math.max(
+    this._nonPersistedPrefixCount = Math.max(
       0,
-      this.nonPersistedPrefixCount - compactedAway,
+      this._nonPersistedPrefixCount - compactedAway,
     );
-    this.summaryIsInjected = false;
   }
 }
 

package/src/daemon/conversation-agent-loop-handlers.ts

@@ -76,10 +76,8 @@ import {
   cleanAssistantContent,
   drainDirectiveDisplayBuffer,
 } from "./assistant-attachments.js";
-import type {
-  AgentLoopConversationContext,
-  AssistantSurface,
-} from "./conversation-agent-loop.js";
+import type { Conversation } from "./conversation.js";
+import type { AssistantSurface } from "./conversation-agent-loop.js";
 import {
   buildConversationErrorMessage,
   classifyConversationError,
@@ -295,29 +293,11 @@ export interface EventHandlerState {
    * never claims content a flush has not yet written.
    */
   lastPersistedContentSeq: number | undefined;
-  /**
-   * Whether the workspace top-level block should be (re)injected on this
-   * turn. Compaction's prepare phase strips the workspace / NOW.md / PKB
-   * blocks off the tail, so it is set after any successful compaction to
-   * force the workspace overview back in. On an ordinary turn the block is
-   * already present in history, so it defaults `false` to avoid burning
-   * tokens re-injecting it redundantly.
-   */
-  shouldInjectWorkspace: boolean;
-  /**
-   * Whether the reducer has compacted `ctx.messages`, gating the Slack
-   * chronological-transcript override on re-injection. The captured
-   * transcript is the full persisted history; blindly replaying it after
-   * compaction would overwrite the reduced messages and undo compaction, so
-   * once this is `true` the override falls back to the reduced
-   * `ctx.messages`.
-   */
-  reducerCompacted: boolean;
 }
 
 /** Immutable context shared across event handlers within a single agent loop run. */
 export interface EventHandlerDeps {
-  readonly ctx: AgentLoopConversationContext;
+  readonly ctx: Conversation;
   readonly onEvent: (msg: ServerMessage) => void;
   readonly reqId: string;
   readonly isFirstMessage: boolean;
@@ -392,8 +372,6 @@ export function createEventHandlerState(): EventHandlerState {
     pendingPartialFlushPromise: undefined,
     currentMessageContent: [],
     lastPersistedContentSeq: undefined,
-    shouldInjectWorkspace: false,
-    reducerCompacted: false,
   };
 }
 
@@ -655,9 +633,7 @@ function computeToolUseStatusText(
   return `Running ${friendlyToolName(name)}`;
 }
 
-function resolveAssistantReplyTimestampTimezone(
-  ctx: AgentLoopConversationContext,
-): string {
+function resolveAssistantReplyTimestampTimezone(ctx: Conversation): string {
   const config = getConfig();
   return resolveTurnTimezoneContext({
     configuredUserTimeZone: config.ui?.userTimezone ?? null,
@@ -1659,6 +1635,8 @@ export async function handleMessageComplete(
       deps.ctx.conversationId,
       state.pendingNotifiedInferenceProfile,
     );
+    deps.ctx.lastNotifiedInferenceProfile =
+      state.pendingNotifiedInferenceProfile;
     state.pendingNotifiedInferenceProfile = null;
   }
 
@@ -2302,17 +2280,14 @@ export async function dispatchAgentEvent(
         // so re-injection re-applies onto the stripped history even when the
         // pipeline ran but did not compact. When it did compact, commit the
         // durable result (DB-record fields, Slack provenance, SSE) — which
-        // overwrites `ctx.messages` with the compacted history — and flip the
-        // per-turn re-injection guards the orchestrator reads. This runs
+        // overwrites `ctx.messages` with the compacted history. This runs
         // before the loop's `reinject` hook (the loop awaits this dispatch),
-        // so the guards are set in time. A failed durable commit re-throws
-        // below to abort the turn rather than re-injecting against
-        // half-applied state.
+        // so the committed history is in place in time. A failed durable
+        // commit re-throws below to abort the turn rather than re-injecting
+        // against half-applied state.
         deps.ctx.messages = event.basis;
         if (event.result.compacted) {
           await deps.applyCompaction(event.result, event.basis);
-          state.reducerCompacted = true;
-          state.shouldInjectWorkspace = true;
         }
         break;
       case "history_stripped":