@librechat/agents 3.2.68 → 3.3.1

package/src/specs/activity-label-prompt.test.ts

@@ -0,0 +1,128 @@
+import type { ActivityLabelToolEntry } from '@/types/activityLabel';
+import { LANGFUSE_TOOL_OUTPUT_REDACTION_TEXT } from '@/langfuseToolOutputTracing';
+import { buildActivityLabelPrompt } from '@/prompts/activityLabel';
+import { resolveToolOutputTracingConfig } from '@/langfuseConfig';
+
+const entries: ActivityLabelToolEntry[] = [
+  {
+    toolName: 'web_search',
+    toolInput: { query: 'runtime versions' },
+    toolOutput: 'PUBLIC_SEARCH_RESULTS',
+    status: 'success',
+  },
+  {
+    toolName: 'db_query',
+    toolInput: { sql: 'select 1' },
+    error: 'SECRET_CONNECTION_STRING_LEAK',
+    status: 'error',
+  },
+];
+
+describe('buildActivityLabelPrompt redaction', () => {
+  it('embeds raw outputs and errors when no redaction policy resolves', () => {
+    const prompt = buildActivityLabelPrompt({ entries, charLimit: 600 });
+    expect(prompt).toContain('PUBLIC_SEARCH_RESULTS');
+    expect(prompt).toContain('SECRET_CONNECTION_STRING_LEAK');
+  });
+
+  it('redacts every outcome when tool-output tracing is globally disabled', () => {
+    const redaction = resolveToolOutputTracingConfig({
+      toolOutputTracing: { enabled: false },
+    });
+    const prompt = buildActivityLabelPrompt({
+      entries,
+      charLimit: 600,
+      redaction,
+    });
+    expect(prompt).not.toContain('PUBLIC_SEARCH_RESULTS');
+    expect(prompt).not.toContain('SECRET_CONNECTION_STRING_LEAK');
+    expect(prompt).toContain(LANGFUSE_TOOL_OUTPUT_REDACTION_TEXT);
+    /** Tool names and inputs stay — matching the span processor, which
+     *  redacts output fields only. */
+    expect(prompt).toContain('web_search');
+    expect(prompt).toContain('runtime versions');
+  });
+
+  it('drops reasoning excerpts when any batch entry is redacted', () => {
+    const redaction = resolveToolOutputTracingConfig({
+      toolOutputTracing: { redactedToolNames: ['db_query'] },
+    });
+    const prompt = buildActivityLabelPrompt({
+      entries,
+      charLimit: 600,
+      thinkingExcerpts: [
+        'The db_query returned SECRET_CONNECTION_STRING_LEAK earlier',
+      ],
+      redaction,
+    });
+    expect(prompt).not.toContain('Reasoning excerpts');
+    expect(prompt).not.toContain('SECRET_CONNECTION_STRING_LEAK');
+  });
+
+  it('drops free-form context under ANY active policy, even with no matching entry', () => {
+    /** Reasoning/intent can quote output from an EARLIER call to the
+     *  redacted tool that this batch does not contain, so an active policy
+     *  suppresses free-form prose regardless of this batch's entries. */
+    const redaction = resolveToolOutputTracingConfig({
+      toolOutputTracing: { redactedToolNames: ['unrelated_tool'] },
+    });
+    const prompt = buildActivityLabelPrompt({
+      entries,
+      charLimit: 600,
+      thinkingExcerpts: ['Comparing versions across sources'],
+      lastAssistantText: 'Checking the unrelated_tool result from before',
+      redaction,
+    });
+    expect(prompt).not.toContain('Comparing versions across sources');
+    expect(prompt).not.toContain('Intent');
+    /** Non-matching entries keep their own outcomes. */
+    expect(prompt).toContain('PUBLIC_SEARCH_RESULTS');
+  });
+
+  it('keeps free-form context when no redaction policy is configured', () => {
+    const prompt = buildActivityLabelPrompt({
+      entries,
+      charLimit: 600,
+      thinkingExcerpts: ['Comparing versions across sources'],
+      lastAssistantText: 'Verifying each runtime',
+      redaction: undefined,
+    });
+    expect(prompt).toContain('Comparing versions across sources');
+    expect(prompt).toContain('Verifying each runtime');
+  });
+
+  it('bounds serialization of oversized structured tool output', () => {
+    const huge = Array.from({ length: 50_000 }, (_, i) => ({
+      id: i,
+      blob: 'x'.repeat(200),
+    }));
+    const prompt = buildActivityLabelPrompt({
+      entries: [
+        {
+          toolName: 'db_rows',
+          toolInput: { sql: 'select *' },
+          toolOutput: huge,
+          status: 'success',
+        },
+      ],
+      charLimit: 600,
+    });
+    /** Degrades to a shape summary instead of materializing ~10MB of JSON. */
+    expect(prompt).toContain('[Array(50000)]');
+    expect(prompt.length).toBeLessThan(2000);
+  });
+
+  it('redacts only named tools, including their error text', () => {
+    const redaction = resolveToolOutputTracingConfig({
+      toolOutputTracing: { redactedToolNames: ['db_query'] },
+    });
+    const prompt = buildActivityLabelPrompt({
+      entries,
+      charLimit: 600,
+      redaction,
+    });
+    expect(prompt).toContain('PUBLIC_SEARCH_RESULTS');
+    expect(prompt).not.toContain('SECRET_CONNECTION_STRING_LEAK');
+    expect(prompt).toContain(LANGFUSE_TOOL_OUTPUT_REDACTION_TEXT);
+  });
+});

package/src/specs/activity-label-trace-seed.test.ts

@@ -0,0 +1,47 @@
+import {
+  resolveLangfuseRuntimeScope,
+  withLangfuseRuntimeScope,
+} from '@/langfuseRuntimeScope';
+import { getTraceIdSeed } from '@/langfuseRuntimeContext';
+
+/**
+ * `generateActivityLabel` passes one label seed to both the Langfuse handler
+ * and the runtime scope it invokes under. This proves the mechanism: a label
+ * scope's seed must override an inherited (parent run) seed for the duration
+ * of the label call, then restore — otherwise per-batch label generations
+ * collapse into the main run trace under deterministic tracing.
+ *
+ * Asserted through the ALS runtime-context channel (`getTraceIdSeed`), which
+ * the trace id generator consults alongside OTel context; tests run without
+ * a registered OTel context manager, so the OTel channel is a no-op here.
+ */
+describe('activity-label trace seed scoping', () => {
+  it('overrides an inherited run seed for the nested scope only', () => {
+    const runScope = resolveLangfuseRuntimeScope({ traceIdSeed: 'run-seed' });
+    withLangfuseRuntimeScope(runScope, () => {
+      expect(getTraceIdSeed()).toBe('run-seed');
+
+      const labelScope = resolveLangfuseRuntimeScope({
+        traceIdSeed: 'run-1-activity-3',
+      });
+      withLangfuseRuntimeScope(labelScope, () => {
+        expect(getTraceIdSeed()).toBe('run-1-activity-3');
+      });
+
+      expect(getTraceIdSeed()).toBe('run-seed');
+    });
+  });
+
+  it('keeps distinct seeds for successive label scopes', () => {
+    const seeds: Array<string | undefined> = [];
+    for (const seed of ['run-1-activity-0', 'run-1-activity-1']) {
+      withLangfuseRuntimeScope(
+        resolveLangfuseRuntimeScope({ traceIdSeed: seed }),
+        () => {
+          seeds.push(getTraceIdSeed());
+        }
+      );
+    }
+    expect(seeds).toEqual(['run-1-activity-0', 'run-1-activity-1']);
+  });
+});