@animalabs/agent-framework 0.8.0 → 0.10.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@animalabs/agent-framework",
-  "version": "0.8.0",
+  "version": "0.10.0",
   "description": "Multi-agent framework with pluggable modules, persistent state, and concurrent tool execution",
   "repository": {
     "type": "git",
@@ -31,7 +31,8 @@
     "clean": "rm -rf dist",
     "typecheck": "tsc --noEmit",
     "pretest": "mkdir -p dist/test/fixtures && cp test/fixtures/*.mjs dist/test/fixtures/",
-    "test": "node --test dist/test/*.test.js",
+    "test": "node --test --test-force-exit dist/test/*.test.js",
+    "version": "node scripts/release-changelog.mjs && git add CHANGELOG.md",
     "prepublishOnly": "npm run build"
   },
   "keywords": [
@@ -45,9 +46,9 @@
   "author": "Anima Research",
   "license": "MIT",
   "dependencies": {
-    "@animalabs/chronicle": "^0.2.2",
+    "@animalabs/chronicle": "^0.3.0",
     "@animalabs/context-manager": "^0.6.0",
-    "@animalabs/membrane": "^0.5.75",
+    "@animalabs/membrane": "^0.5.78",
     "chokidar": "^4.0.3",
     "discord.js": "^14.25.1",
     "ws": "^8.18.0"

package/src/agent.ts

@@ -1,6 +1,10 @@
 import type { Membrane, NormalizedMessage, NormalizedRequest, ContentBlock, YieldingStream } from '@animalabs/membrane';
 import { isAbortedResponse } from '@animalabs/membrane';
-import { toolResultDataToHistoryString } from './tool-result-history.js';
+import {
+  toolResultDataToHistoryString,
+  truncateForHistory,
+  DEFAULT_TOOL_RESULT_INLINE_MAX_CHARS,
+} from './tool-result-history.js';
 
 export interface StartStreamResult {
   stream: YieldingStream;
@@ -63,7 +67,7 @@ export class Agent {
   /** Refusal auto-rewind policy (see AgentConfig.refusalHandling). */
   readonly refusalHandling: AgentConfig['refusalHandling'];
   /** Prose delivery mode (see AgentConfig.proseRouting). Default 'locus'. */
-  readonly proseRouting: 'locus' | 'explicit';
+  readonly proseRouting: 'locus' | 'explicit' | 'hybrid';
   /** Prompt-cache TTL forwarded to the provider (see AgentConfig.cacheTtl). */
   readonly cacheTtl: NonNullable<AgentConfig['cacheTtl']>;
   /** Emit prompt-cache markers on requests (see AgentConfig.promptCaching). */
@@ -78,7 +82,17 @@ export class Agent {
   private _inferenceStartedAt = 0;
   private _streamId = 0;
   lastStreamInputTokens = 0;
+  /** Real prefix size of the last usage event: fresh + cache creation +
+   *  cache read. THE window-shaped number — `lastStreamInputTokens` alone
+   *  omits cached tokens, which are most of a warm stream's window. */
+  lastStreamRealInputTokens = 0;
+  /** Output tokens of the last usage event — the prior round's generated
+   *  thinking/text/tool_use, which becomes part of the NEXT round's input
+   *  and so belongs in the physical-window projection. */
+  lastStreamOutputTokens = 0;
   maxStreamTokens: number;
+  /** Provider hard context cap (see AgentConfig.physicalWindowTokens). */
+  readonly physicalWindowTokens?: number;
   /** Per-agent context compile budget (input tokens). When unset, the
    * ContextManager's built-in default applies. reserveForResponse uses
    * this agent's maxTokens. */
@@ -113,6 +127,7 @@ export class Agent {
     this.providerParams = config.providerParams;
     this.sameRoundThinkTextPolicy = config.sameRoundThinkTextPolicy;
     this.maxStreamTokens = config.maxStreamTokens ?? 150_000;
+    this.physicalWindowTokens = config.physicalWindowTokens;
     this.contextBudgetTokens = config.contextBudgetTokens;
     this.configuredContextBudgetTokens = config.contextBudgetTokens;
     this.contextManager = contextManager;
@@ -676,6 +691,14 @@ export class Agent {
     this._streamId++;
     this._inferenceStartedAt = Date.now();
     this.lastStreamInputTokens = 0;
+    // Reset the cache-inclusive counters too: a new stream must not inherit
+    // the prior stream's window size — on the physical-window restart path
+    // that inherited value IS the oversized number that caused the restart,
+    // and projecting on it before this stream's first usage event would
+    // restart again (loop). The `> 0` guard downstream means "no usage
+    // observed on THIS stream yet" only because of this reset.
+    this.lastStreamRealInputTokens = 0;
+    this.lastStreamOutputTokens = 0;
 
     const request = await this.buildActivationRequest(availableTools, injections, budget);
 
@@ -683,6 +706,12 @@ export class Agent {
       emitTokens: true,
       emitBlocks: false,
       emitUsage: true,
+      // Retry a content-policy refusal at the provider seam rather than at
+      // the framework level: membrane replays the SAME request immediately
+      // (cache-warm), where a framework-level requeue would recompile and
+      // land a different window. The framework's driveStream handles the
+      // resulting `retrying` event by discarding the abandoned attempt.
+      ...(this.refusalHandling?.retries ? { refusalRetries: this.refusalHandling.retries } : {}),
     });
 
     this._state = { status: 'streaming', stream };
@@ -884,12 +913,17 @@ export class Agent {
     // Tool results go as a user message with tool_result blocks. The history
     // serializer turns MCP image blocks into `[image: type, size]` so the
     // persisted transcript stays small and survives truncation.
+    //
+    // This is the direct-Agent-API path (the framework stores results itself
+    // via buildStoredToolResultContent, with workspace spill). No workspace
+    // is reachable from here, so the house default cap applies as plain
+    // truncation — bounded beats unbounded, for errors too.
     const content = results.map((r) => ({
       type: 'tool_result' as const,
       toolUseId: r.id,
       content: r.result.isError
-        ? r.result.error ?? 'Unknown error'
-        : toolResultDataToHistoryString(r.result.data),
+        ? truncateForHistory(r.result.error ?? 'Unknown error', DEFAULT_TOOL_RESULT_INLINE_MAX_CHARS)
+        : toolResultDataToHistoryString(r.result.data, DEFAULT_TOOL_RESULT_INLINE_MAX_CHARS),
       isError: r.result.isError,
     }));