@librechat/agents 3.1.73 → 3.1.75

package/src/agents/AgentContext.ts

@@ -20,6 +20,16 @@ import { addCacheControl } from '@/messages/cache';
 import { DEFAULT_RESERVE_RATIO } from '@/messages';
 import { toJsonSchema } from '@/utils/schema';
 
+type AgentSystemTextBlock = {
+  type: 'text';
+  text: string;
+  cache_control?: { type: 'ephemeral' };
+};
+
+type AgentSystemContentBlock =
+  | AgentSystemTextBlock
+  | { cachePoint: { type: 'default' } };
+
 /**
  * Encapsulates agent-specific state that can vary between agents in a multi-agent system
  */
@@ -249,7 +259,7 @@ export class AgentContext {
   private summaryTokenCount: number = 0;
   /**
    * Where the summary should be injected:
-   * - `'system_prompt'`: cross-run summary, included in `buildInstructionsString`
+   * - `'system_prompt'`: cross-run summary, included in the dynamic system tail
    * - `'user_message'`: mid-run compaction, injected as HumanMessage on clean slate
    * - `'none'`: no summary present
    */
@@ -417,7 +427,8 @@ export class AgentContext {
 
   /**
    * Gets the system runnable, creating it lazily if needed.
-   * Includes instructions, additional instructions, and programmatic-only tools documentation.
+   * Includes stable instructions, dynamic additional instructions, and
+   * programmatic-only tools documentation.
    * Only rebuilds when marked stale (via markToolsAsDiscovered).
    */
   get systemRunnable():
@@ -431,8 +442,10 @@ export class AgentContext {
       return this.cachedSystemRunnable;
     }
 
-    const instructionsString = this.buildInstructionsString();
-    this.cachedSystemRunnable = this.buildSystemRunnable(instructionsString);
+    this.cachedSystemRunnable = this.buildSystemRunnable({
+      stableInstructions: this.buildStableInstructionsString(),
+      dynamicInstructions: this.buildDynamicInstructionsString(),
+    });
     this.systemRunnableStale = false;
     return this.cachedSystemRunnable;
   }
@@ -443,17 +456,19 @@ export class AgentContext {
    */
   initializeSystemRunnable(): void {
     if (this.systemRunnableStale || this.cachedSystemRunnable === undefined) {
-      const instructionsString = this.buildInstructionsString();
-      this.cachedSystemRunnable = this.buildSystemRunnable(instructionsString);
+      this.cachedSystemRunnable = this.buildSystemRunnable({
+        stableInstructions: this.buildStableInstructionsString(),
+        dynamicInstructions: this.buildDynamicInstructionsString(),
+      });
       this.systemRunnableStale = false;
     }
   }
 
   /**
-   * Builds the raw instructions string (without creating SystemMessage).
+   * Builds the cacheable instructions string (without creating SystemMessage).
    * Includes agent identity preamble and handoff context when available.
    */
-  private buildInstructionsString(): string {
+  private buildStableInstructionsString(): string {
     const parts: string[] = [];
 
     const identityPreamble = this.buildIdentityPreamble();
@@ -465,6 +480,22 @@ export class AgentContext {
       parts.push(this.instructions);
     }
 
+    const programmaticToolsDoc = this.buildProgrammaticOnlyToolsInstructions();
+    if (programmaticToolsDoc) {
+      parts.push(programmaticToolsDoc);
+    }
+
+    return parts.join('\n\n');
+  }
+
+  /**
+   * Builds the dynamic system-tail string (without creating SystemMessage).
+   * Keep this out of prompt-cache-marked content so volatile context does not
+   * invalidate the stable prefix.
+   */
+  private buildDynamicInstructionsString(): string {
+    const parts: string[] = [];
+
     if (
       this.additionalInstructions != null &&
       this.additionalInstructions !== ''
@@ -472,14 +503,10 @@ export class AgentContext {
       parts.push(this.additionalInstructions);
     }
 
-    const programmaticToolsDoc = this.buildProgrammaticOnlyToolsInstructions();
-    if (programmaticToolsDoc) {
-      parts.push(programmaticToolsDoc);
-    }
-
-    // Cross-run summary: include in system prompt so the model has context
-    // from the prior run.  Mid-run summaries are injected as a HumanMessage
-    // on the post-compaction clean slate instead (see buildSystemRunnable).
+    // Cross-run summary: include in the system tail so the model has context
+    // from the prior run without invalidating the cacheable prefix. Mid-run
+    // summaries are injected as a HumanMessage on the post-compaction clean
+    // slate instead (see buildSystemRunnable).
     if (
       this._summaryLocation === 'system_prompt' &&
       this.summaryText != null &&
@@ -523,9 +550,13 @@ export class AgentContext {
    * Build system runnable from pre-built instructions string.
    * Only called when content has actually changed.
    */
-  private buildSystemRunnable(
-    instructionsString: string
-  ):
+  private buildSystemRunnable({
+    stableInstructions,
+    dynamicInstructions,
+  }: {
+    stableInstructions: string;
+    dynamicInstructions: string;
+  }):
     | Runnable<
         BaseMessage[],
         (BaseMessage | SystemMessage)[],
@@ -537,35 +568,17 @@ export class AgentContext {
       this.summaryText != null &&
       this.summaryText !== '';
 
-    if (!instructionsString && !hasMidRunSummary) {
+    if (!stableInstructions && !dynamicInstructions && !hasMidRunSummary) {
       this.systemMessageTokens = 0;
       return undefined;
     }
 
-    let finalInstructions: string | BaseMessageFields = instructionsString;
-
-    let usePromptCache = false;
-    if (this.provider === Providers.ANTHROPIC) {
-      const anthropicOptions = this.clientOptions as
-        | t.AnthropicClientOptions
-        | undefined;
-      if (anthropicOptions?.promptCache === true) {
-        usePromptCache = true;
-        finalInstructions = {
-          content: [
-            {
-              type: 'text',
-              text: instructionsString,
-              cache_control: { type: 'ephemeral' },
-            },
-          ],
-        };
-      }
-    }
-
-    const systemMessage = instructionsString
-      ? new SystemMessage(finalInstructions)
-      : undefined;
+    const usePromptCache = this.hasAnthropicPromptCache();
+    const systemMessage = this.buildSystemMessage({
+      stableInstructions,
+      dynamicInstructions,
+      usePromptCache,
+    });
 
     if (this.tokenCounter) {
       this.systemMessageTokens = systemMessage
@@ -615,6 +628,72 @@ export class AgentContext {
     }).withConfig({ runName: 'prompt' });
   }
 
+  private hasAnthropicPromptCache(): boolean {
+    if (this.provider !== Providers.ANTHROPIC) {
+      return false;
+    }
+    const anthropicOptions = this.clientOptions as
+      | t.AnthropicClientOptions
+      | undefined;
+    return anthropicOptions?.promptCache === true;
+  }
+
+  private hasBedrockPromptCache(): boolean {
+    if (this.provider !== Providers.BEDROCK) {
+      return false;
+    }
+    const bedrockOptions = this.clientOptions as
+      | t.BedrockAnthropicClientOptions
+      | undefined;
+    return bedrockOptions?.promptCache === true;
+  }
+
+  private buildSystemMessage({
+    stableInstructions,
+    dynamicInstructions,
+    usePromptCache,
+  }: {
+    stableInstructions: string;
+    dynamicInstructions: string;
+    usePromptCache: boolean;
+  }): SystemMessage | undefined {
+    if (!stableInstructions && !dynamicInstructions) {
+      return undefined;
+    }
+
+    if (usePromptCache) {
+      const content: AgentSystemContentBlock[] = [];
+      if (stableInstructions) {
+        content.push({
+          type: 'text',
+          text: stableInstructions,
+          cache_control: { type: 'ephemeral' },
+        });
+      }
+      if (dynamicInstructions) {
+        content.push({ type: 'text', text: dynamicInstructions });
+      }
+      return new SystemMessage({ content } as BaseMessageFields);
+    }
+
+    if (this.hasBedrockPromptCache() && stableInstructions) {
+      const content: AgentSystemContentBlock[] = [
+        { type: 'text', text: stableInstructions },
+        { cachePoint: { type: 'default' } },
+      ];
+      if (dynamicInstructions) {
+        content.push({ type: 'text', text: dynamicInstructions });
+      }
+      return new SystemMessage({ content } as BaseMessageFields);
+    }
+
+    return new SystemMessage(
+      [stableInstructions, dynamicInstructions]
+        .filter((part) => part !== '')
+        .join('\n\n')
+    );
+  }
+
   /**
    * Reset context for a new run
    */
@@ -681,10 +760,47 @@ export class AgentContext {
     if (!this.toolDefinitions) {
       return [];
     }
-    return this.toolDefinitions.filter(
-      (def) =>
+    /**
+     * Mirror `getEventDrivenToolsForBinding`'s gate: a definition is only
+     * bound to the model when its `allowed_callers` include `'direct'` and
+     * (if deferred) it has been discovered. Filtering by `defer_loading`
+     * alone left programmatic-only definitions counted in
+     * `toolSchemaTokens` even though they were never bound.
+     */
+    return this.toolDefinitions.filter((def) => {
+      const allowedCallers = def.allowed_callers ?? ['direct'];
+      if (!allowedCallers.includes('direct')) {
+        return false;
+      }
+      return (
         def.defer_loading !== true || this.discoveredToolNames.has(def.name)
-    );
+      );
+    });
+  }
+
+  /**
+   * Single source of truth for "which entries of `this.tools` should be
+   * treated as actually bound". Callers:
+   *   - `getToolsForBinding` (non-event-driven branch)
+   *   - `getEventDrivenToolsForBinding` (appends instance tools alongside
+   *     schema-only definitions)
+   *   - `calculateInstructionTokens` (counts schema bytes for accounting)
+   *
+   * In event-driven mode (`toolDefinitions` present) instance tools are
+   * appended unfiltered; outside event-driven mode they pass through
+   * `filterToolsForBinding`. Centralizing the decision here prevents the
+   * accounting/binding paths from drifting apart, which was the root
+   * cause of the original miscount.
+   */
+  private getEffectiveInstanceTools(): t.GraphTools | undefined {
+    if (!this.tools) {
+      return undefined;
+    }
+    const isEventDriven = (this.toolDefinitions?.length ?? 0) > 0;
+    if (isEventDriven || !this.toolRegistry) {
+      return this.tools;
+    }
+    return this.filterToolsForBinding(this.tools);
   }
 
   /**
@@ -703,9 +819,17 @@ export class AgentContext {
      * populated after `fromConfig()` kicks off the initial calculation, so
      * callers that mutate `graphTools` must re-trigger this method to
      * refresh `toolSchemaTokens`.
+     *
+     * Use `getEffectiveInstanceTools()` so accounting reflects exactly the
+     * subset that `getToolsForBinding` would emit — preventing the
+     * worst-case-ceiling miscount that triggered spurious `empty_messages`
+     * preflight rejections at low `maxContextTokens`. Deferred and
+     * non-`'direct'` `toolDefinitions` are excluded by
+     * `getActiveToolDefinitions()` below.
      */
     const instanceTools: t.GraphTools = [
-      ...((this.tools as t.GenericTool[] | undefined) ?? []),
+      ...((this.getEffectiveInstanceTools() as t.GenericTool[] | undefined) ??
+        []),
       ...((this.graphTools as t.GenericTool[] | undefined) ?? []),
     ];
 
@@ -900,8 +1024,16 @@ export class AgentContext {
    */
   getTokenBudgetBreakdown(messages?: BaseMessage[]): t.TokenBudgetBreakdown {
     const maxContextTokens = this.maxContextTokens ?? 0;
-    const toolCount =
-      (this.tools?.length ?? 0) + this.getActiveToolDefinitions().length;
+    /**
+     * Derive `toolCount` from `getToolsForBinding()` so the diagnostic stays
+     * aligned with what is actually bound to the model — and with what
+     * `calculateInstructionTokens` counts into `toolSchemaTokens`. Using raw
+     * `this.tools.length` would inflate the count whenever the registry
+     * marks instance tools as deferred-undiscovered or non-`'direct'`,
+     * producing the same misleading "N tools" diagnostic this fix is meant
+     * to eliminate.
+     */
+    const toolCount = this.getToolsForBinding()?.length ?? 0;
     const messageCount = messages?.length ?? 0;
 
     let messageTokens = 0;
@@ -1014,10 +1146,7 @@ export class AgentContext {
       return this.getEventDrivenToolsForBinding();
     }
 
-    const filtered =
-      !this.tools || !this.toolRegistry
-        ? this.tools
-        : this.filterToolsForBinding(this.tools);
+    const filtered = this.getEffectiveInstanceTools();
 
     if (this.graphTools && this.graphTools.length > 0) {
       return [...(filtered ?? []), ...this.graphTools];
@@ -1032,21 +1161,9 @@ export class AgentContext {
       return this.graphTools ?? [];
     }
 
-    const defsToInclude = this.toolDefinitions.filter((def) => {
-      const allowedCallers = def.allowed_callers ?? ['direct'];
-      if (!allowedCallers.includes('direct')) {
-        return false;
-      }
-      if (
-        def.defer_loading === true &&
-        !this.discoveredToolNames.has(def.name)
-      ) {
-        return false;
-      }
-      return true;
-    });
-
-    const schemaTools = createSchemaOnlyTools(defsToInclude) as t.GraphTools;
+    const schemaTools = createSchemaOnlyTools(
+      this.getActiveToolDefinitions()
+    ) as t.GraphTools;
 
     const allTools = [...schemaTools];
 
@@ -1054,8 +1171,9 @@ export class AgentContext {
       allTools.push(...this.graphTools);
     }
 
-    if (this.tools && this.tools.length > 0) {
-      allTools.push(...this.tools);
+    const instanceTools = this.getEffectiveInstanceTools();
+    if (instanceTools && instanceTools.length > 0) {
+      allTools.push(...instanceTools);
     }
 
     return allTools;

package/src/agents/__tests__/AgentContext.anthropic.live.test.ts

@@ -0,0 +1,116 @@
+// src/agents/__tests__/AgentContext.anthropic.live.test.ts
+/**
+ * Live Anthropic prompt-cache verification.
+ *
+ * Run with:
+ * RUN_ANTHROPIC_PROMPT_CACHE_LIVE_TESTS=1 ANTHROPIC_API_KEY=... npm test -- AgentContext.anthropic.live.test.ts --runInBand
+ */
+import { config as dotenvConfig } from 'dotenv';
+dotenvConfig();
+
+import { describe, expect, it } from '@jest/globals';
+import type * as t from '@/types';
+import {
+  runLiveTurn,
+  assertSystemPayloadShape,
+  buildDynamicInstructions,
+  buildStableInstructions,
+  waitForCachePropagation,
+} from './promptCacheLiveHelpers';
+import { Providers } from '@/common';
+
+const shouldRunLive =
+  process.env.RUN_ANTHROPIC_PROMPT_CACHE_LIVE_TESTS === '1' &&
+  process.env.ANTHROPIC_API_KEY != null &&
+  process.env.ANTHROPIC_API_KEY !== '';
+
+const describeIfLive = shouldRunLive ? describe : describe.skip;
+
+const modelName =
+  process.env.ANTHROPIC_PROMPT_CACHE_MODEL ?? 'claude-sonnet-4-5';
+const providerLabel = 'Anthropic';
+
+function createClientOptions(): t.AnthropicClientOptions {
+  return {
+    modelName,
+    temperature: 0,
+    maxTokens: 8,
+    streaming: true,
+    streamUsage: true,
+    promptCache: true,
+    clientOptions: {
+      defaultHeaders: {
+        'anthropic-beta': 'prompt-caching-2024-07-31',
+      },
+    },
+  };
+}
+
+describeIfLive('AgentContext Anthropic prompt cache live API', () => {
+  it('caches only the stable system prefix while dynamic tail changes', async () => {
+    const nonce = `agent-cache-live-${Date.now()}`;
+    const clientOptions = createClientOptions();
+    const stableInstructions = buildStableInstructions({
+      nonce,
+      providerLabel,
+    });
+    const firstDynamicInstructions = buildDynamicInstructions({
+      marker: 'alpha',
+      tailDescription:
+        'The Dynamic Marker line is runtime context and must remain outside the cached prefix.',
+    });
+    const secondDynamicInstructions = buildDynamicInstructions({
+      marker: 'bravo',
+      tailDescription:
+        'The Dynamic Marker line is runtime context and must remain outside the cached prefix.',
+    });
+
+    await assertSystemPayloadShape({
+      agentId: 'live-cache-shape-check',
+      provider: Providers.ANTHROPIC,
+      clientOptions,
+      stableInstructions,
+      dynamicInstructions: firstDynamicInstructions,
+      expectedContent: [
+        {
+          type: 'text',
+          text: stableInstructions,
+          cache_control: { type: 'ephemeral' },
+        },
+        {
+          type: 'text',
+          text: firstDynamicInstructions,
+        },
+      ],
+    });
+
+    const first = await runLiveTurn({
+      provider: Providers.ANTHROPIC,
+      providerLabel,
+      clientOptions,
+      runId: `${nonce}-first`,
+      threadId: `${nonce}-thread`,
+      stableInstructions,
+      dynamicInstructions: firstDynamicInstructions,
+    });
+
+    expect(first.text.toLowerCase()).toContain('alpha');
+    expect(first.usage.input_token_details?.cache_creation).toBeGreaterThan(0);
+    expect(first.usage.input_token_details?.cache_read ?? 0).toBe(0);
+
+    await waitForCachePropagation();
+
+    const second = await runLiveTurn({
+      provider: Providers.ANTHROPIC,
+      providerLabel,
+      clientOptions,
+      runId: `${nonce}-second`,
+      threadId: `${nonce}-thread`,
+      stableInstructions,
+      dynamicInstructions: secondDynamicInstructions,
+    });
+
+    expect(second.text.toLowerCase()).toContain('bravo');
+    expect(second.usage.input_token_details?.cache_read).toBeGreaterThan(0);
+  }, 120_000);
+});

package/src/agents/__tests__/AgentContext.bedrock.live.test.ts

@@ -0,0 +1,149 @@
+// src/agents/__tests__/AgentContext.bedrock.live.test.ts
+/**
+ * Live Bedrock prompt-cache verification.
+ *
+ * Run with:
+ * RUN_BEDROCK_PROMPT_CACHE_LIVE_TESTS=1 BEDROCK_AWS_REGION=... BEDROCK_AWS_ACCESS_KEY_ID=... BEDROCK_AWS_SECRET_ACCESS_KEY=... npm test -- AgentContext.bedrock.live.test.ts --runInBand
+ *
+ * Standard AWS credential env vars or AWS_PROFILE can also be used.
+ */
+import { config as dotenvConfig } from 'dotenv';
+dotenvConfig();
+
+import { describe, expect, it } from '@jest/globals';
+import type * as t from '@/types';
+import {
+  runLiveTurn,
+  assertSystemPayloadShape,
+  buildDynamicInstructions,
+  buildStableInstructions,
+  waitForCachePropagation,
+} from './promptCacheLiveHelpers';
+import { Providers } from '@/common';
+
+const accessKeyId =
+  process.env.BEDROCK_AWS_ACCESS_KEY_ID ?? process.env.AWS_ACCESS_KEY_ID;
+const secretAccessKey =
+  process.env.BEDROCK_AWS_SECRET_ACCESS_KEY ??
+  process.env.AWS_SECRET_ACCESS_KEY;
+const sessionToken =
+  process.env.BEDROCK_AWS_SESSION_TOKEN ?? process.env.AWS_SESSION_TOKEN;
+const hasCredentialPair =
+  accessKeyId != null &&
+  accessKeyId !== '' &&
+  secretAccessKey != null &&
+  secretAccessKey !== '';
+const hasAmbientCredentials =
+  process.env.AWS_PROFILE != null ||
+  process.env.AWS_WEB_IDENTITY_TOKEN_FILE != null;
+
+const shouldRunLive =
+  process.env.RUN_BEDROCK_PROMPT_CACHE_LIVE_TESTS === '1' &&
+  (hasCredentialPair || hasAmbientCredentials);
+
+const describeIfLive = shouldRunLive ? describe : describe.skip;
+
+const model =
+  process.env.BEDROCK_PROMPT_CACHE_MODEL ??
+  'us.anthropic.claude-sonnet-4-5-20250929-v1:0';
+const region =
+  process.env.BEDROCK_AWS_REGION ?? process.env.AWS_REGION ?? 'us-east-1';
+const providerLabel = 'Bedrock';
+
+function getCredentials():
+  | t.BedrockAnthropicClientOptions['credentials']
+  | undefined {
+  if (!hasCredentialPair) {
+    return undefined;
+  }
+
+  return {
+    accessKeyId,
+    secretAccessKey,
+    ...(sessionToken != null && sessionToken !== '' ? { sessionToken } : {}),
+  };
+}
+
+function createClientOptions(): t.BedrockAnthropicClientOptions {
+  const credentials = getCredentials();
+  return {
+    model,
+    region,
+    maxTokens: 8,
+    streaming: true,
+    streamUsage: true,
+    promptCache: true,
+    ...(credentials != null ? { credentials } : {}),
+  };
+}
+
+describeIfLive('AgentContext Bedrock prompt cache live API', () => {
+  it('caches only the stable system prefix while dynamic tail changes', async () => {
+    const nonce = `agent-bedrock-cache-live-${Date.now()}`;
+    const clientOptions = createClientOptions();
+    const stableInstructions = buildStableInstructions({
+      nonce,
+      providerLabel,
+    });
+    const firstDynamicInstructions = buildDynamicInstructions({
+      marker: 'alpha',
+      tailDescription:
+        'The Dynamic Marker line is runtime context and must remain after the Bedrock cache point.',
+    });
+    const secondDynamicInstructions = buildDynamicInstructions({
+      marker: 'bravo',
+      tailDescription:
+        'The Dynamic Marker line is runtime context and must remain after the Bedrock cache point.',
+    });
+
+    await assertSystemPayloadShape({
+      agentId: 'live-bedrock-cache-shape-check',
+      provider: Providers.BEDROCK,
+      clientOptions,
+      stableInstructions,
+      dynamicInstructions: firstDynamicInstructions,
+      expectedContent: [
+        {
+          type: 'text',
+          text: stableInstructions,
+        },
+        {
+          cachePoint: { type: 'default' },
+        },
+        {
+          type: 'text',
+          text: firstDynamicInstructions,
+        },
+      ],
+    });
+
+    const first = await runLiveTurn({
+      provider: Providers.BEDROCK,
+      providerLabel,
+      clientOptions,
+      runId: `${nonce}-first`,
+      threadId: `${nonce}-thread`,
+      stableInstructions,
+      dynamicInstructions: firstDynamicInstructions,
+    });
+
+    expect(first.text.toLowerCase()).toContain('alpha');
+    expect(first.usage.input_token_details?.cache_creation).toBeGreaterThan(0);
+    expect(first.usage.input_token_details?.cache_read ?? 0).toBe(0);
+
+    await waitForCachePropagation();
+
+    const second = await runLiveTurn({
+      provider: Providers.BEDROCK,
+      providerLabel,
+      clientOptions,
+      runId: `${nonce}-second`,
+      threadId: `${nonce}-thread`,
+      stableInstructions,
+      dynamicInstructions: secondDynamicInstructions,
+    });
+
+    expect(second.text.toLowerCase()).toContain('bravo');
+    expect(second.usage.input_token_details?.cache_read).toBeGreaterThan(0);
+  }, 180_000);
+});