@librechat/agents 3.4.7 → 3.5.0

package/src/__tests__/stream.streamLimits.test.ts

@@ -1,1982 +0,0 @@
-/**
- * Regression tests for the streamed tool-argument circuit breaker.
- *
- * Incident: a model streamed a single malformed tool-call argument (a SQL
- * query) to 149,923 characters over 26 minutes, never completing it, until
- * the 64k output-token ceiling ended the run. Nothing bounded a single tool
- * call's streamed argument growth mid-flight.
- *
- * The guard lives on the SDK stream dispatch path: `ChatModelStreamHandler`
- * enforces `Graph.streamLimits` on every streamed chunk event and throws
- * `StreamLimitExceededError` out of the run's `streamEvents` loop, which
- * tears down the in-flight provider request (see the mid-flight halt notes
- * in `Run.processStream`).
- */
-import { describe, it, expect, jest } from '@jest/globals';
-import type { AgentContext } from '@/agents/AgentContext';
-import type { StandardGraph } from '@/graphs';
-import type * as t from '@/types';
-import {
-  STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY,
-  STREAMED_TOOL_CALL_SEAL_METADATA_KEY,
-  OPENAI_RESPONSES_STREAMED_TOOL_CALL_ADAPTER,
-} from '@/tools/streamedToolCallSeals';
-import {
-  DEFAULT_MAX_TOOL_CALL_ARG_BYTES,
-  STREAM_LIMIT_REDISPATCH_KEY,
-  STREAM_LIMIT_EPOCH_KEY,
-  StreamLimitExceededError,
-  enforceStreamLimitsForWireChunk,
-  resolveGenerationKey,
-  resolveStreamLimits,
-} from '@/llm/streamLimits';
-import { GraphEvents, Providers, StepTypes } from '@/common';
-import { ChatModelStreamHandler } from '@/stream';
-import { HandlerRegistry } from '@/events';
-
-function createGraph(overrides: Partial<StandardGraph> = {}): StandardGraph {
-  const runSteps = new Map<string, t.RunStep>();
-  const stepIdsByKey = new Map<string, string>();
-  let stepCounter = 0;
-
-  const graph = {
-    config: {
-      configurable: { user_id: 'user_1' },
-      metadata: { run_id: 'run_1' },
-    },
-    eagerEventToolExecution: undefined,
-    eagerEventToolExecutions: new Map(),
-    eagerEventToolCallChunks: new Map(),
-    eagerEventToolSuppressions: new Set<string>(),
-    handlerRegistry: new HandlerRegistry(),
-    hookRegistry: undefined,
-    humanInTheLoop: undefined,
-    toolOutputReferences: undefined,
-    sessions: new Map(),
-    toolCallStepIds: new Map(),
-    messageIdsByStepKey: new Map(),
-    messageStepHasToolCalls: new Map(),
-    prelimMessageIdsByStepKey: new Map(),
-    getAgentContext: jest.fn(
-      (): Partial<AgentContext> => ({
-        provider: Providers.OPENAI,
-        reasoningKey: 'reasoning_content',
-        toolDefinitions: [],
-        graphTools: [],
-        agentId: 'agent_1',
-      })
-    ),
-    getStepKey: jest.fn(() => 'step-key'),
-    getStepIdByKey: jest.fn((stepKey: string) => {
-      const stepId = stepIdsByKey.get(stepKey);
-      if (stepId == null) {
-        throw new Error('no current step');
-      }
-      return stepId;
-    }),
-    getRunStep: jest.fn((stepId: string) => runSteps.get(stepId)),
-    dispatchRunStep: jest.fn(async (stepKey: string, details: unknown) => {
-      const id = `step_${++stepCounter}`;
-      stepIdsByKey.set(stepKey, id);
-      runSteps.set(id, {
-        id,
-        type: (details as { type: t.RunStep['type'] }).type,
-        stepDetails: details as t.RunStep['stepDetails'],
-      } as t.RunStep);
-      return id;
-    }),
-    dispatchRunStepDelta: jest.fn(async () => undefined),
-    ...overrides,
-  };
-
-  return graph as unknown as StandardGraph;
-}
-
-const generation = (step: number): Record<string, unknown> => ({
-  langgraph_checkpoint_ns: '',
-  langgraph_node: 'agent',
-  langgraph_step: step,
-});
-
-async function streamEvent(args: {
-  handler: ChatModelStreamHandler;
-  graph: StandardGraph;
-  chunk: Record<string, unknown>;
-  metadata?: Record<string, unknown>;
-}): Promise<void> {
-  const { handler, graph, chunk, metadata } = args;
-  await handler.handle(
-    GraphEvents.CHAT_MODEL_STREAM,
-    { chunk: chunk as unknown as t.StreamChunk },
-    metadata ?? {},
-    graph
-  );
-}
-
-async function streamToolCallChunks(args: {
-  handler: ChatModelStreamHandler;
-  graph: StandardGraph;
-  chunks: Array<Record<string, unknown>>;
-  metadata?: Record<string, unknown>;
-}): Promise<void> {
-  const { handler, graph, chunks, metadata } = args;
-  for (const toolCallChunk of chunks) {
-    await streamEvent({
-      handler,
-      graph,
-      metadata,
-      chunk: { content: '', tool_call_chunks: [toolCallChunk] },
-    });
-  }
-}
-
-describe('streamed tool-call argument circuit breaker', () => {
-  it('aborts the stream once a single call exceeds the default 64 KiB', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({ streamLimits: resolveStreamLimits() });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [
-        { id: 'call_1', name: 'query_database', args: '', index: 0 },
-        ...Array.from({ length: 8 }, () => ({
-          args: 'x'.repeat(8_192),
-          index: 0,
-        })),
-      ],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'x', index: 0 }],
-      })
-    ).rejects.toMatchObject({
-      name: 'StreamLimitExceededError',
-      kind: 'tool_call_args',
-      limit: DEFAULT_MAX_TOOL_CALL_ARG_BYTES,
-      observed: DEFAULT_MAX_TOOL_CALL_ARG_BYTES + 1,
-      toolName: 'query_database',
-    });
-
-    const dispatchDelta = graph.dispatchRunStepDelta as jest.Mock;
-    expect(dispatchDelta).toHaveBeenCalledTimes(9);
-  });
-
-  it('keeps parallel tool calls on independent budgets', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamEvent({
-      handler,
-      graph,
-      chunk: {
-        content: '',
-        tool_call_chunks: [
-          { id: 'call_1', name: 'first', args: 'a'.repeat(60), index: 0 },
-          { id: 'call_2', name: 'second', args: 'b'.repeat(60), index: 1 },
-        ],
-      },
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'b'.repeat(41), index: 1 }],
-      })
-    ).rejects.toThrow('(tool call: second)');
-  });
-
-  it('counts argument bytes, not characters', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 1_000 }),
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [
-          { id: 'call_1', name: 'writer', args: '€'.repeat(300), index: 0 },
-          { args: '€'.repeat(100), index: 0 },
-        ],
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', observed: 1_200 });
-  });
-
-  it('enforces the cap before a complete arrival-sealed call can dispatch', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-    const oversized = JSON.stringify({ payload: 'x'.repeat(200) });
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_calls: [
-            { id: 'call_1', name: 'side_effect', args: { payload: 'x'.repeat(200) } },
-          ],
-          tool_call_chunks: [
-            { id: 'call_1', name: 'side_effect', args: oversized, index: 0 },
-          ],
-        },
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', toolName: 'side_effect' });
-
-    const dispatchStep = graph.dispatchRunStep as jest.Mock;
-    expect(dispatchStep).not.toHaveBeenCalled();
-  });
-
-  it('enforces the cap for chunks that carry no numeric index', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_1', name: 'no_index_tool', args: 'x'.repeat(80) }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ id: 'call_1', args: 'x'.repeat(21) }],
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', toolName: 'no_index_tool' });
-  });
-
-  it('does not double-count OpenAI Responses seal restatements', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 50 }),
-    });
-    const fullArgs = 'x'.repeat(40);
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [
-        { args: fullArgs.slice(0, 20), index: 0 },
-        { args: fullArgs.slice(20), index: 0 },
-      ],
-    });
-    await streamEvent({
-      handler,
-      graph,
-      chunk: {
-        content: '',
-        tool_call_chunks: [{ name: 'writer', args: fullArgs, index: 0 }],
-        response_metadata: {
-          [STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY]:
-            OPENAI_RESPONSES_STREAMED_TOOL_CALL_ADAPTER,
-          [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: { kind: 'single', index: 0 },
-        },
-      },
-    });
-
-    const key = `${resolveGenerationKey({})}:0`;
-    expect(graph.streamedToolCallArgTallies.has(key)).toBe(false);
-    expect(graph.streamedToolCallArgTallies.size).toBe(0);
-  });
-
-  it('keeps parallel anonymous arrival-sealed calls on separate budgets (Google)', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 64 }),
-    });
-
-    await streamEvent({
-      handler,
-      graph,
-      chunk: {
-        content: '',
-        tool_call_chunks: [
-          { name: 'search_a', args: 'x'.repeat(40) },
-          { name: 'search_b', args: 'y'.repeat(40) },
-        ],
-        response_metadata: {
-          [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: { kind: 'all' },
-        },
-      },
-    });
-    expect(graph.streamedToolCallArgTallies.size).toBe(0);
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [{ name: 'search_c', args: 'z'.repeat(70) }],
-          response_metadata: {
-            [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: { kind: 'all' },
-          },
-        },
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', toolName: 'search_c' });
-  });
-
-  it('streams unbounded when explicitly disabled', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 0 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [
-        { id: 'call_1', name: 'query_database', args: '', index: 0 },
-        ...Array.from({ length: 4 }, () => ({
-          args: 'x'.repeat(65_536),
-          index: 0,
-        })),
-      ],
-    });
-
-    const dispatchDelta = graph.dispatchRunStepDelta as jest.Mock;
-    expect(dispatchDelta).toHaveBeenCalledTimes(5);
-    expect(graph.streamedToolCallArgTallies).toBeUndefined();
-  });
-
-  it('creates the tool_calls run step before the breaker can trip', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 50 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_1', name: 'query_database', args: '{"q":', index: 0 }],
-    });
-
-    const dispatchStep = graph.dispatchRunStep as jest.Mock;
-    const dispatchedTypes = dispatchStep.mock.calls.map(
-      (call) => (call[1] as { type: string }).type
-    );
-    expect(dispatchedTypes).toContain(StepTypes.TOOL_CALLS);
-  });
-});
-
-describe('per-tool argument byte overrides', () => {
-  it('raises the cap for the named tool without loosening others', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { create_file: 1_000 },
-      }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [
-        { id: 'call_1', name: 'create_file', args: 'x'.repeat(500), index: 0 },
-      ],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [
-          { id: 'call_2', name: 'query_database', args: 'y'.repeat(101), index: 1 },
-        ],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 100,
-      toolName: 'query_database',
-    });
-  });
-
-  it('trips the named tool at its own lower cap', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 1_000,
-        maxToolCallArgBytesByTool: { chatty_tool: 50 },
-      }),
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [
-          { id: 'call_1', name: 'chatty_tool', args: 'x'.repeat(51), index: 0 },
-        ],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 50,
-      toolName: 'chatty_tool',
-    });
-  });
-
-  it('enforces a per-tool cap even when the global cap is disabled', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 0,
-        maxToolCallArgBytesByTool: { create_file: 100 },
-      }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [
-        { id: 'call_1', name: 'unbounded_tool', args: 'x'.repeat(500), index: 0 },
-      ],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [
-          { id: 'call_2', name: 'create_file', args: 'y'.repeat(101), index: 1 },
-        ],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 100,
-      toolName: 'create_file',
-    });
-  });
-
-  it('applies overrides to arrival-sealed complete calls', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 64,
-        maxToolCallArgBytesByTool: { create_file: 1_000 },
-      }),
-    });
-
-    await streamEvent({
-      handler,
-      graph,
-      chunk: {
-        content: '',
-        tool_call_chunks: [{ name: 'create_file', args: 'x'.repeat(500) }],
-        response_metadata: {
-          [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: { kind: 'all' },
-        },
-      },
-    });
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [{ name: 'search_c', args: 'z'.repeat(70) }],
-          response_metadata: {
-            [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: { kind: 'all' },
-          },
-        },
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', toolName: 'search_c' });
-  });
-
-  it('re-judges tallied bytes when the tool name arrives on a later empty chunk', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 1_000,
-        maxToolCallArgBytesByTool: { capped_tool: 50 },
-      }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ args: 'x'.repeat(60), index: 0 }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ id: 'call_1', name: 'capped_tool', args: '', index: 0 }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 50,
-      observed: 60,
-      toolName: 'capped_tool',
-    });
-  });
-
-  it('honors an override for a tool named __proto__', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: JSON.parse('{"__proto__": 1000}') as Record<
-          string,
-          number
-        >,
-      }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_1', name: '__proto__', args: 'x'.repeat(500), index: 0 }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'x'.repeat(501), index: 0 }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 1_000,
-      toolName: '__proto__',
-    });
-  });
-
-  it('keeps charging a call whose later deltas omit both id and index', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60) }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'a'.repeat(60) }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 100,
-      observed: 120,
-      toolName: 'writer',
-    });
-  });
-
-  it('does not charge argument bytes for marked OpenRouter redispatches', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60), index: 0 }],
-    });
-    await streamEvent({
-      handler,
-      graph,
-      metadata: { [STREAM_LIMIT_REDISPATCH_KEY]: true },
-      chunk: {
-        content: '',
-        tool_call_chunks: [{ args: 'a'.repeat(60), index: 0 }],
-      },
-    });
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ args: 'a'.repeat(40), index: 0 }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'a', index: 0 }],
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', observed: 101 });
-  });
-
-  it('judges client tool chunks arriving with a server-tool result before the early return', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-    (graph.getAgentContext as jest.Mock).mockReturnValue({
-      provider: Providers.ANTHROPIC,
-      reasoningKey: 'reasoning_content',
-      toolDefinitions: [],
-      graphTools: [],
-      agentId: 'agent_1',
-    });
-    await graph.dispatchRunStep('step-key', {
-      type: StepTypes.TOOL_CALLS,
-      tool_calls: [{ id: 'srvtoolu_1', name: 'web_search', args: {} }],
-    });
-    graph.toolCallStepIds.set('srvtoolu_1', 'step_1');
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: [
-            { type: 'web_search_tool_result', tool_use_id: 'srvtoolu_1', content: [] },
-          ],
-          tool_call_chunks: [
-            { id: 'call_big', name: 'side_effect', args: 'x'.repeat(200), index: 0 },
-          ],
-        },
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', toolName: 'side_effect' });
-  });
-
-  it('judges complete parsed tool calls that arrive without raw chunks', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { create_file: 1_000 },
-      }),
-    });
-
-    await streamEvent({
-      handler,
-      graph,
-      chunk: {
-        content: '',
-        tool_calls: [
-          { id: 'call_1', name: 'create_file', args: { content: 'x'.repeat(500) } },
-        ],
-      },
-    });
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_calls: [
-            { id: 'call_2', name: 'side_effect', args: { payload: 'x'.repeat(200) } },
-          ],
-        },
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 100,
-      toolName: 'side_effect',
-    });
-  });
-
-  it('charges a chunk once when the handler echo wins the race', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-    const wireChunk = {
-      content: '',
-      tool_call_chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60), index: 0 }],
-    };
-
-    await streamEvent({ handler, graph, chunk: wireChunk });
-    enforceStreamLimitsForWireChunk({
-      graph,
-      metadata: {},
-      chunk: wireChunk as never,
-    });
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ args: 'a'.repeat(40), index: 0 }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'a', index: 0 }],
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', observed: 101 });
-  });
-
-  it('charges wire chunks synchronously and skips the handler echo', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-    const wireChunk = {
-      content: '',
-      tool_call_chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60), index: 0 }],
-    };
-
-    enforceStreamLimitsForWireChunk({
-      graph,
-      metadata: {},
-      chunk: wireChunk as never,
-    });
-    await streamEvent({ handler, graph, chunk: wireChunk });
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ args: 'a'.repeat(40), index: 0 }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'a', index: 0 }],
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', observed: 101 });
-  });
-
-  it('keeps charging a mutable chunk object re-yielded across emissions', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-    const reused: Record<string, unknown> = {
-      content: '',
-      tool_call_chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60), index: 0 }],
-    };
-
-    await streamEvent({ handler, graph, chunk: reused });
-    reused.tool_call_chunks = [{ args: 'a'.repeat(40), index: 0 }];
-    await streamEvent({ handler, graph, chunk: reused });
-    reused.tool_call_chunks = [{ args: 'a', index: 0 }];
-
-    await expect(streamEvent({ handler, graph, chunk: reused })).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      observed: 101,
-    });
-  });
-
-  it('pairs producer and echo charges per emission for reused chunk objects', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-    const reused: Record<string, unknown> = {
-      content: '',
-      tool_call_chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60), index: 0 }],
-    };
-
-    enforceStreamLimitsForWireChunk({ graph, metadata: {}, chunk: reused as never });
-    await streamEvent({ handler, graph, chunk: reused });
-    reused.tool_call_chunks = [{ args: 'a'.repeat(40), index: 0 }];
-    enforceStreamLimitsForWireChunk({ graph, metadata: {}, chunk: reused as never });
-    await streamEvent({ handler, graph, chunk: reused });
-    reused.tool_call_chunks = [{ args: 'a', index: 0 }];
-
-    let thrown: unknown;
-    try {
-      enforceStreamLimitsForWireChunk({ graph, metadata: {}, chunk: reused as never });
-    } catch (error) {
-      thrown = error;
-    }
-    expect(thrown).toMatchObject({ kind: 'tool_call_args', observed: 101 });
-  });
-
-  it('keeps parallel index-less calls on distinct budgets when one seals', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_a', name: 'a_tool', args: 'a'.repeat(60) }],
-    });
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_b', name: 'b_tool', args: 'b'.repeat(60) }],
-    });
-    await streamEvent({
-      handler,
-      graph,
-      chunk: {
-        content: '',
-        tool_call_chunks: [{ id: 'call_a', args: '' }],
-        response_metadata: {
-          [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: { kind: 'single', id: 'call_a' },
-        },
-      },
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'b'.repeat(41) }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      observed: 101,
-      toolName: 'b_tool',
-    });
-  });
-
-  it('scopes charge credits per generation for shared chunk objects', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-    const shared: Record<string, unknown> = {
-      content: '',
-      tool_call_chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60), index: 0 }],
-    };
-
-    await streamEvent({ handler, graph, chunk: shared, metadata: generation(1) });
-    enforceStreamLimitsForWireChunk({
-      graph,
-      metadata: generation(2),
-      chunk: shared as never,
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'a'.repeat(41), index: 0 }],
-        metadata: generation(2),
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', observed: 101 });
-  });
-
-  it('judges parsed calls even when a raw chunk representation is present', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [{ id: 'call_1', name: 'side_effect', args: '', index: 0 }],
-          tool_calls: [
-            { id: 'call_1', name: 'side_effect', args: { payload: 'x'.repeat(200) } },
-          ],
-        },
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', toolName: 'side_effect' });
-  });
-
-  it('keeps one budget when a call drops its index but keeps its id', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60), index: 0 }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ id: 'call_1', args: 'a'.repeat(41) }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      observed: 101,
-      toolName: 'writer',
-    });
-  });
-
-  it('keeps one budget when a call drops both identifiers on later deltas', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60), index: 0 }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'a'.repeat(41) }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      observed: 101,
-      toolName: 'writer',
-    });
-  });
-
-  it('adopts the existing tally when a later delta adds an index', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60) }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ id: 'call_1', args: 'a'.repeat(41), index: 0 }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      observed: 101,
-      toolName: 'writer',
-    });
-  });
-
-  it('names anonymous raw chunks from parsed calls so overrides apply', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { create_file: 1_000 },
-      }),
-    });
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [{ id: 'call_1', args: 'x'.repeat(500), index: 0 }],
-          tool_calls: [
-            { id: 'call_1', name: 'create_file', args: { content: 'x'.repeat(500) } },
-          ],
-        },
-      })
-    ).resolves.toBeUndefined();
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'x'.repeat(501), index: 0 }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 1_000,
-      toolName: 'create_file',
-    });
-  });
-
-  it('stops queued events once the shared breaker has tripped', async () => {
-    const handler = new ChatModelStreamHandler();
-    const breakerAbort = new AbortController();
-    const trip = new StreamLimitExceededError({
-      kind: 'tool_call_args',
-      limit: 10,
-      observed: 11,
-      toolName: 'db_query',
-    });
-    breakerAbort.abort(trip);
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits(),
-      breakerAbort,
-    });
-
-    /** A sibling branch tripped the breaker while this event sat queued in
-     * streamEvents; the handler must rethrow instead of continuing into
-     * content handling or the eager-tool dispatch paths. */
-    await expect(
-      streamEvent({ handler, graph, chunk: { content: 'queued text' } })
-    ).rejects.toBe(trip);
-  });
-
-  it('does not charge an anonymous sparse delta to an identified call\'s override', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { create_file: 1_000 },
-      }),
-    });
-
-    /** Parallel batch: two identified calls share one event, so batch
-     * positions are NOT stable identities for later sparse events. */
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [
-            { id: 'call_a', name: 'create_file', args: 'x'.repeat(50) },
-            { id: 'call_b', name: 'db_query', args: 'x'.repeat(10) },
-          ],
-        },
-      })
-    ).resolves.toBeUndefined();
-
-    /** A sparse continuation event with a single anonymous delta lands on
-     * batch position 0 — inheriting call_a's tally would judge these bytes
-     * under create_file's 1000-byte override and bypass the global cap. */
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'x'.repeat(200) }],
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', limit: 100 });
-  });
-
-  it('judges ambiguous sparse deltas under each candidate\'s own limit', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 1_000,
-        maxToolCallArgBytesByTool: { tight_tool: 50 },
-      }),
-    });
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [
-            { id: 'call_a', name: 'tight_tool', args: 'x'.repeat(10) },
-            { id: 'call_b', name: 'db_query', args: 'x'.repeat(10) },
-          ],
-        },
-      })
-    ).resolves.toBeUndefined();
-
-    /** The anonymous continuation might belong to tight_tool, whose 50-byte
-     * override is LOWER than the raised global cap — judging candidates
-     * only against the global limit would let it stream past indefinitely. */
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'x'.repeat(60) }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 50,
-      toolName: 'tight_tool',
-    });
-  });
-
-  it('keeps the sole live call\'s budget continuous across anonymous continuations', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_a', name: 'writer', args: 'x'.repeat(60) }],
-    });
-    /** A second id-only call takes over the #0 position alias, then seals —
-     * releasing it removes the alias while call_a stays live under its id. */
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_b', name: 'other', args: 'x'.repeat(10) }],
-    });
-    await streamEvent({
-      handler,
-      graph,
-      chunk: {
-        content: '',
-        tool_call_chunks: [{ id: 'call_b', args: '' }],
-        response_metadata: {
-          [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: {
-            kind: 'single',
-            id: 'call_b',
-          },
-        },
-      },
-    });
-
-    /** The anonymous continuation must charge call_a's existing 60-byte
-     * tally — a fresh #0 tally would reset the sole remaining call's
-     * budget and let combined args exceed the cap. */
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'x'.repeat(60) }],
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', limit: 100 });
-  });
-
-  it('still accounts when per-tool overrides exist with the global cap off', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 0,
-        maxToolCallArgBytesByTool: { create_file: 500 },
-      }),
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [
-          { id: 'call_1', name: 'create_file', args: 'x'.repeat(501), index: 0 },
-        ],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 500,
-      toolName: 'create_file',
-    });
-  });
-
-  it('allocates nothing when overrides contain only zero-valued disables', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 0,
-        maxToolCallArgBytesByTool: { writer: 0 },
-      }),
-    });
-
-    /** `{ writer: 0 }` is a per-tool disable, not a limit — with the global
-     * cap off too, no argument limit can fire and accounting must not
-     * allocate. */
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [
-          { id: 'call_1', name: 'writer', args: 'x'.repeat(100_000), index: 0 },
-        ],
-      })
-    ).resolves.toBeUndefined();
-    expect(graph.streamLimitChargeCredits).toBeUndefined();
-    expect(graph.streamedToolCallArgTallies).toBeUndefined();
-  });
-
-  it('allocates no charge bookkeeping when every guard is disabled', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 0 }),
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [
-          { id: 'call_1', name: 'writer', args: 'x'.repeat(100_000), index: 0 },
-        ],
-      })
-    ).resolves.toBeUndefined();
-    expect(graph.streamLimitChargeCredits).toBeUndefined();
-    expect(graph.streamedToolCallArgTallies).toBeUndefined();
-  });
-
-  it('binds consumer trips to the event epoch, sparing a newer run', async () => {
-    const handler = new ChatModelStreamHandler();
-    const newRunBreaker = new AbortController();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-      breakerAbort: newRunBreaker,
-      breakerEpoch: 5,
-    });
-
-    /** A straggling oversized chunk from a failed run (epoch 4) handled
-     * after resetValues installed epoch 5 is dropped outright — acting on
-     * it (content handling, eager dispatch) would compose the new run's
-     * live controller — and the new run's controller stays live. */
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [
-          { id: 'call_1', name: 'writer', args: 'x'.repeat(101), index: 0 },
-        ],
-        metadata: { [STREAM_LIMIT_EPOCH_KEY]: 4 },
-      })
-    ).resolves.toBeUndefined();
-    expect(newRunBreaker.signal.aborted).toBe(false);
-
-    /** Same breach with the matching epoch trips the controller. */
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [
-          { id: 'call_2', name: 'writer', args: 'x'.repeat(101), index: 0 },
-        ],
-        metadata: { [STREAM_LIMIT_EPOCH_KEY]: 5 },
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args' });
-    expect(newRunBreaker.signal.aborted).toBe(true);
-  });
-
-  it('lets stale-epoch events pass the queued-event breaker check', async () => {
-    const handler = new ChatModelStreamHandler();
-    const breakerAbort = new AbortController();
-    breakerAbort.abort(
-      new StreamLimitExceededError({
-        kind: 'tool_call_args',
-        limit: 10,
-        observed: 11,
-        toolName: 'db_query',
-      })
-    );
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits(),
-      breakerAbort,
-      breakerEpoch: 5,
-    });
-
-    /** The tripped controller belongs to the CURRENT run (epoch 5); an
-     * old-run chunk (epoch 4) must not be failed against it. The same
-     * event WITH the matching epoch throws, proving only the epoch gate
-     * differs. */
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: { content: '' },
-        metadata: { [STREAM_LIMIT_EPOCH_KEY]: 4 },
-      })
-    ).resolves.toBeUndefined();
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: { content: '' },
-        metadata: { [STREAM_LIMIT_EPOCH_KEY]: 5 },
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args' });
-  });
-
-  it('treats empty tool-call ids as anonymous, keeping parallel calls apart', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    /** OpenAI-compatible adapters can emit index-less parallel calls whose
-     * ids are empty placeholders; a shared ':c:' identity would merge the
-     * two 60-byte calls into a phantom 120-byte one and falsely trip. */
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [
-            { id: '', args: 'x'.repeat(60) },
-            { id: '', args: 'x'.repeat(60) },
-          ],
-        },
-      })
-    ).resolves.toBeUndefined();
-  });
-
-  it('does not hand a lone parsed call\'s override to multiple id-less chunks', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { create_file: 1_000 },
-      }),
-    });
-
-    /** Mixed event: two id-less raw chunks while only one call is parseable.
-     * Positional correlation is ambiguous here — naming both chunks
-     * 'create_file' would judge the unrelated 500-byte call against the
-     * higher override and bypass the global cap. */
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [
-            { args: 'x'.repeat(50), index: 0 },
-            { args: 'x'.repeat(500), index: 1 },
-          ],
-          tool_calls: [
-            { id: 'call_1', name: 'create_file', args: { content: 'x' } },
-          ],
-        },
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args', limit: 100 });
-  });
-
-  it('prefers the id-correlated complete name over a raw name fragment', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { create_file: 1_000 },
-      }),
-    });
-
-    /** A custom adapter can emit a raw chunk whose name is only a FRAGMENT
-     * ("create_") alongside the complete parsed call for the same id. The
-     * complete call's name must select the per-tool override — judging the
-     * bytes under the fragment would trip the global cap instead. */
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [
-            { id: 'call_1', name: 'create_', args: 'x'.repeat(500), index: 0 },
-          ],
-          tool_calls: [
-            {
-              id: 'call_1',
-              name: 'create_file',
-              args: { content: 'x'.repeat(500) },
-            },
-          ],
-        },
-      })
-    ).resolves.toBeUndefined();
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'x'.repeat(501), index: 0 }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 1_000,
-      toolName: 'create_file',
-    });
-  });
-
-  it('keeps an index 0 call and an id "0" call on separate budgets', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ name: 'a_tool', args: 'a'.repeat(60), index: 0 }],
-    });
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: '0', name: 'b_tool', args: 'b'.repeat(60) }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ id: '0', args: 'b'.repeat(41) }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      observed: 101,
-      toolName: 'b_tool',
-    });
-  });
-
-  it('adopts an anonymous tally when a later delta adds only an index', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ name: 'writer', args: 'a'.repeat(60) }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'a'.repeat(41), index: 0 }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      observed: 101,
-      toolName: 'writer',
-    });
-  });
-
-  it('adopts a sparse indexed continuation by its stable call index', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { wide_tool: 1_000 },
-      }),
-    });
-
-    await streamEvent({
-      handler,
-      graph,
-      chunk: {
-        content: '',
-        tool_call_chunks: [
-          { name: 'wide_tool', args: 'x'.repeat(60) },
-          { name: 'tight_tool', args: 'x'.repeat(40) },
-        ],
-      },
-    });
-
-    /** The event position is 0, but index 1 identifies the second call from
-     * the original parallel batch. Its existing 40 bytes must stay under
-     * the global cap rather than adopting wide_tool's raised allowance. */
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'x'.repeat(61), index: 1 }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 100,
-      observed: 101,
-      toolName: 'tight_tool',
-    });
-  });
-
-  it('adopts the sole anonymous tally when a later nonzero index appears', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ name: 'writer', args: 'x'.repeat(60) }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'x'.repeat(41), index: 1 }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      observed: 101,
-      toolName: 'writer',
-    });
-  });
-
-  it('does not use the sole anonymous fallback while another call is live', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { wide_tool: 1_000 },
-      }),
-    });
-
-    await streamEvent({
-      handler,
-      graph,
-      chunk: {
-        content: '',
-        tool_call_chunks: [
-          { id: 'call_a', name: 'tight_tool', args: 'x'.repeat(10) },
-          { name: 'wide_tool', args: 'x'.repeat(10) },
-        ],
-      },
-    });
-
-    /** Index 0 could identify call_a; the one anonymous tally belongs to
-     * batch position 1 and must not lend wide_tool's raised override. */
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'x'.repeat(101), index: 0 }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 100,
-    });
-  });
-
-  it('applies a late anonymous name before charging its lower override', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 1_000,
-        maxToolCallArgBytesByTool: { tight_tool: 50 },
-      }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ args: 'x'.repeat(40) }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ name: 'tight_tool', args: 'x'.repeat(11) }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 50,
-      observed: 51,
-      toolName: 'tight_tool',
-    });
-  });
-
-  it('assembles a late anonymous name before applying its higher override', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { create_file: 1_000 },
-      }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ name: 'create_', args: 'x'.repeat(60) }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ name: 'file', args: 'x'.repeat(60) }],
-      })
-    ).resolves.toBeUndefined();
-  });
-
-  it('names id-less raw chunks from a single parsed call so overrides apply', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { create_file: 1_000 },
-      }),
-    });
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [{ args: 'x'.repeat(500), index: 0 }],
-          tool_calls: [
-            { id: 'call_1', name: 'create_file', args: { content: 'x'.repeat(500) } },
-          ],
-        },
-      })
-    ).resolves.toBeUndefined();
-  });
-
-  it('matches a dual-identifier seal against a chunk carrying only its id', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60), index: 0 }],
-    });
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [{ id: 'call_1', args: 'a'.repeat(60) }],
-          response_metadata: {
-            [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: {
-              kind: 'single',
-              index: 0,
-              id: 'call_1',
-            },
-          },
-        },
-      })
-    ).resolves.toBeUndefined();
-    expect(graph.streamedToolCallArgTallies.size).toBe(0);
-  });
-
-  it('adopts an anonymous tally when a later delta adds only an id', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ name: 'writer', args: 'a'.repeat(60) }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ id: 'call_1', args: 'a'.repeat(41) }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      observed: 101,
-      toolName: 'writer',
-    });
-  });
-
-  it('does not adopt a live id-bearing call through its position alias', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_a', name: 'a_tool', args: 'a'.repeat(60) }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ id: 'call_b', name: 'b_tool', args: 'b'.repeat(60) }],
-      })
-    ).resolves.toBeUndefined();
-  });
-
-  it('judges invalid parsed tool calls that arrive without raw chunks', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          invalid_tool_calls: [
-            {
-              id: 'call_1',
-              name: 'side_effect',
-              args: 'x'.repeat(200),
-              error: 'Unparseable arguments',
-            },
-          ],
-        },
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 100,
-      toolName: 'side_effect',
-    });
-  });
-
-  it('keeps one budget when an index-only call drops its index', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ name: 'writer', args: 'a'.repeat(60), index: 0 }],
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'a'.repeat(41) }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      observed: 101,
-      toolName: 'writer',
-    });
-  });
-
-  it('names unnamed raw chunks from invalid calls so overrides apply', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { create_file: 1_000 },
-      }),
-    });
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [{ id: 'call_1', args: 'x'.repeat(500), index: 0 }],
-          invalid_tool_calls: [
-            {
-              id: 'call_1',
-              name: 'create_file',
-              args: 'x'.repeat(500),
-              error: 'Unparseable arguments',
-            },
-          ],
-        },
-      })
-    ).resolves.toBeUndefined();
-  });
-
-  it('accumulates fragmented tool names before selecting overrides', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({
-        maxToolCallArgBytes: 100,
-        maxToolCallArgBytesByTool: { create_file: 1_000 },
-      }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_1', name: 'create_', args: 'x'.repeat(50), index: 0 }],
-    });
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ name: 'file', args: 'x'.repeat(100), index: 0 }],
-      })
-    ).resolves.toBeUndefined();
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ args: 'x'.repeat(900), index: 0 }],
-      })
-    ).rejects.toMatchObject({
-      kind: 'tool_call_args',
-      limit: 1_000,
-      observed: 1_050,
-      toolName: 'create_file',
-    });
-  });
-
-  it('matches a seal by id when the indices disagree', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-    });
-
-    await streamToolCallChunks({
-      handler,
-      graph,
-      chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(60), index: 0 }],
-    });
-
-    await expect(
-      streamEvent({
-        handler,
-        graph,
-        chunk: {
-          content: '',
-          tool_call_chunks: [{ id: 'call_1', args: 'a'.repeat(60), index: 1 }],
-          response_metadata: {
-            [STREAMED_TOOL_CALL_SEAL_METADATA_KEY]: {
-              kind: 'single',
-              index: 0,
-              id: 'call_1',
-            },
-          },
-        },
-      })
-    ).resolves.toBeUndefined();
-    expect(graph.streamedToolCallArgTallies.size).toBe(0);
-  });
-
-  it('trips the shared graph breaker when the consumer path detects a breach', async () => {
-    const handler = new ChatModelStreamHandler();
-    const breakerAbort = new AbortController();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxToolCallArgBytes: 100 }),
-      breakerAbort,
-    });
-
-    await expect(
-      streamToolCallChunks({
-        handler,
-        graph,
-        chunks: [{ id: 'call_1', name: 'writer', args: 'a'.repeat(101), index: 0 }],
-      })
-    ).rejects.toMatchObject({ kind: 'tool_call_args' });
-
-    expect(breakerAbort.signal.aborted).toBe(true);
-    expect(breakerAbort.signal.reason).toMatchObject({
-      kind: 'tool_call_args',
-      observed: 101,
-    });
-  });
-
-  it('allocates no charge credits for text-only deltas with the event cap off', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits(),
-      dispatchMessageDelta: jest.fn(async () => undefined),
-      dispatchReasoningDelta: jest.fn(async () => undefined),
-    } as never);
-
-    await streamEvent({
-      handler,
-      graph,
-      chunk: { content: 'an ordinary text delta' },
-    });
-    expect(graph.streamLimitChargeCredits).toBeUndefined();
-
-    enforceStreamLimitsForWireChunk({
-      graph,
-      metadata: {},
-      chunk: { tool_call_chunks: undefined } as never,
-    });
-    expect(graph.streamLimitChargeCredits).toBeUndefined();
-
-    await streamEvent({
-      handler,
-      graph,
-      chunk: {
-        content: '',
-        tool_call_chunks: [{ id: 'call_1', name: 'writer', args: 'x', index: 0 }],
-      },
-    });
-    expect(graph.streamLimitChargeCredits).toBeDefined();
-  });
-
-  it('normalizes override entries like the global field', () => {
-    const resolved = resolveStreamLimits({
-      maxToolCallArgBytes: 100,
-      maxToolCallArgBytesByTool: {
-        create_file: 131_072.9,
-        disabled_tool: -5,
-        unlimited_tool: Infinity,
-        invalid_tool: Number.NaN,
-        '': 42,
-      },
-    });
-
-    expect(resolved.maxToolCallArgBytesByTool).toEqual({
-      create_file: 131_072,
-      disabled_tool: 0,
-      unlimited_tool: 0,
-    });
-  });
-});
-
-describe('per-generation delta event circuit breaker', () => {
-  it('is opt-in and counts every streamed chunk event, including empty ones', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxDeltaEventsPerTurn: 3 }),
-    });
-
-    const emptyChunkEvent = (): Promise<void> =>
-      streamEvent({ handler, graph, chunk: { content: '' }, metadata: generation(1) });
-
-    await emptyChunkEvent();
-    await emptyChunkEvent();
-    await emptyChunkEvent();
-    await expect(emptyChunkEvent()).rejects.toMatchObject({
-      name: 'StreamLimitExceededError',
-      kind: 'delta_events',
-      limit: 3,
-      observed: 4,
-    });
-  });
-
-  it('scopes the budget per model generation, surviving step-key forks', async () => {
-    const handler = new ChatModelStreamHandler();
-    let stepKey = 'reasoning-key';
-    const graph = createGraph({
-      streamLimits: resolveStreamLimits({ maxDeltaEventsPerTurn: 2 }),
-      getStepKey: jest.fn(() => stepKey) as unknown as StandardGraph['getStepKey'],
-    });
-
-    const emptyChunkEvent = (step: number): Promise<void> =>
-      streamEvent({ handler, graph, chunk: { content: '' }, metadata: generation(step) });
-
-    await emptyChunkEvent(1);
-    stepKey = 'post-reasoning-key';
-    await emptyChunkEvent(1);
-    await expect(emptyChunkEvent(1)).rejects.toMatchObject({
-      kind: 'delta_events',
-      observed: 3,
-    });
-
-    await emptyChunkEvent(2);
-    expect(graph.streamDeltaEventCounts.get('|agent|2|')?.count).toBe(1);
-  });
-
-  it('never counts when left at the default (disabled)', async () => {
-    const handler = new ChatModelStreamHandler();
-    const graph = createGraph({ streamLimits: resolveStreamLimits() });
-
-    for (let i = 0; i < 10; i++) {
-      await streamEvent({ handler, graph, chunk: { content: '' }, metadata: generation(1) });
-    }
-    expect(graph.streamDeltaEventCounts).toBeUndefined();
-  });
-});