Break Tool-Call Rendering Mid-Stream

When a model requests a tool call, the structured payload rides the same stream as prose chunks. A malformed tool-call payload is a classic production incident: the prose renders fine, then the tool-call card throws or renders raw JSON. The ai-tool-call-fails preset reproduces exactly that.

What the preset does

ai-tool-call-fails adds one fetch-stream corruption rule that matches chunk CONTENT: any chunk whose decoded text contains a tool-call wire marker ("tool_calls" for OpenAI, "tool_use" for Anthropic, "function_call" for legacy clients) is mutated with the malformed-json strategy. Prose chunks stream through untouched. The corruption event is tagged phase: 'ai:tool-call-failed' so it stands out in the chaos log and report output.

Playwright

import { test, expect } from '@playwright/test';
import { injectChaos, getChaosLog } from '@chaos-maker/playwright';

test('tool-call card falls back when the payload breaks mid-stream', async ({ page }) => {
  await injectChaos(page, { presets: ['ai-tool-call-fails'], seed: 42 });

  await page.goto('/chat');
  await page.getByRole('textbox').fill('What is the weather in Kochi?');
  await page.getByRole('button', { name: 'Send' }).click();

  // Prose still streams in.
  await expect(page.locator('[data-testid="assistant-message"]')).toContainText('weather');
  // The tool-call card must degrade, not crash the message list.
  await expect(page.locator('[data-testid="tool-call-error"]')).toBeVisible();

  const log = await getChaosLog(page);
  const toolCallFailures = log.filter(
    (e) => e.type === 'fetch-stream:chunk-corrupted' && e.detail.phase === 'ai:tool-call-failed',
  );
  expect(toolCallFailures.length).toBeGreaterThan(0);
});

Cypress

import { injectChaos, getChaosLog } from '@chaos-maker/cypress';

it('tool-call card falls back when the payload breaks mid-stream', () => {
  injectChaos({ presets: ['ai-tool-call-fails'], seed: 42 });

  cy.visit('/chat');
  cy.get('[data-testid="prompt-input"]').type('What is the weather in Kochi?');
  cy.contains('button', 'Send').click();

  cy.get('[data-testid="assistant-message"]').should('contain.text', 'weather');
  cy.get('[data-testid="tool-call-error"]').should('be.visible');

  getChaosLog().then((log) => {
    const toolCallFailures = log.filter(
      (e) => e.type === 'fetch-stream:chunk-corrupted' && e.detail.phase === 'ai:tool-call-failed',
    );
    expect(toolCallFailures.length).to.be.greaterThan(0);
  });
});

Targeting your own markers

The preset covers the common provider wire formats. If your backend uses a custom envelope, write the rule directly with chunkPattern:

await injectChaos(page, {
  seed: 42,
  fetchStream: {
    corruptions: [
      {
        urlPattern: '/api/chat',
        chunkPattern: /"kind":"tool-invocation"/,
        strategy: 'malformed-json',
        probability: 1,
        phase: 'ai:tool-call-failed',
      },
    ],
  },
});

chunkPattern accepts a string (case-sensitive substring containment) or a RegExp (g and y flags are rejected at validation time). It matches the decoded UTF-8 text of each chunk and combines with chunkIndex when both are set. Binary chunks never match; the first binary skip per connection emits a diagnostic event with applied: false and reason: 'binary-chunk'.

What to assert

• The message list survives: prose before and after the broken payload still renders.
• The tool-call UI degrades explicitly (error card, retry affordance), not silently.
• The chaos log carries at least one fetch-stream:chunk-corrupted event with phase: 'ai:tool-call-failed', proving the corruption hit the structured payload and not a prose chunk.