Skip to content

[Bug] OpenAI Responses streaming: tool-call-args-delta/tool-call-ready never emitted (callBuffers keyed by call_id, but events carry item_id) — MCPAgent.stream() yields zero steps #2550

Closed

Description

@BlueX888

Describe the bug

On the OpenAI Responses streaming path, streamResponsesTurn keys its tool-call buffers (callBuffers) by call_id, but the streaming events that carry the arguments — response.function_call_arguments.delta and response.function_call_arguments.done — do not have a call_id field; they carry item_id (the output-item id, fc_...) and output_index only. So the buffer lookup key is always the fallback "" and callBuffers.get("") is always undefined. Neither tool-call-args-delta nor tool-call-ready is ever emitted.

Consequence: streamNativeAgentSteps (native_runner.ts) builds AgentSteps only on tool-call-ready, so MCPAgent.stream() / agent.streamEvents() silently yields zero steps for provider "openai" (and openrouter models routed to OpenAIResponsesDriver). The tool still executes (arguments are parsed from response.completed's output), so the bug is silent: the user sees the final text but no tool step at all.

Environment

  • mcp-use agent package version: 2.0.16 (libraries/typescript/packages/agent/package.json)
  • commit: dcaa0b8df5721f11d4d2aefae1f85df2a1903ba5
  • Node: v22.23.2
  • OS: macOS 26.6.2 (Darwin 25.6.0), headless Node (no browser involved)
  • openai SDK: 6.17.0 (lockfile-matched)

To Reproduce

Steps:

  1. Check out dcaa0b8df5721f11d4d2aefae1f85df2a1903ba5.
  2. Run the repro below against src/llm with tsx.

The repro feeds raw OpenAI Responses SSE events with the real SDK event shapes (SSE stream served by a stubbed globalThis.fetch), first asserting on driver.stream() event types, then on streamNativeAgentSteps.

import { OpenAIResponsesDriver } from "../src/llm/providers/openai-responses-driver.ts";
import { streamNativeAgentSteps } from "../src/llm/native_runner.ts";
import type { LlmStreamEvent } from "../src/llm/types.ts";

// Real OpenAI Responses streaming event shapes, taken from the openai-node SDK
// (ResponseFunctionCallArgumentsDeltaEvent / ...DoneEvent carry item_id +
// output_index, NOT call_id).
const turn1 = [
  { type: "response.output_item.added", output_index: 0,
    item: { type: "function_call", id: "fc_1", call_id: "call_abc", name: "add", arguments: "" } },
  { type: "response.function_call_arguments.delta", item_id: "fc_1", output_index: 0, delta: '{"a":10' },
  { type: "response.function_call_arguments.delta", item_id: "fc_1", output_index: 0, delta: ',"b":20}' },
  { type: "response.function_call_arguments.done", item_id: "fc_1", output_index: 0, arguments: '{"a":10,"b":20}' },
  { type: "response.completed", response: { output: [
      { type: "function_call", id: "fc_1", call_id: "call_abc", name: "add", arguments: '{"a":10,"b":20}' } ] } },
];
const turn2 = [
  { type: "response.output_text.delta", delta: "The answer is 30." },
  { type: "response.completed", response: { output: [
      { type: "message", role: "assistant", content: [{ type: "output_text", text: "The answer is 30." }] } ] } },
];
const sse = (events: unknown[]) => events.map((e) => `data: ${JSON.stringify(e)}\n\n`).join("");

let call = 0;
(globalThis as any).fetch = async () => {
  call++;
  return new Response(sse(call === 1 ? turn1 : turn2), {
    status: 200,
    headers: { "Content-Type": "text/event-stream" },
  });
};

async function main() {
  const driver = new OpenAIResponsesDriver({ provider: "openai", model: "gpt-5", apiKey: "k" });
  let toolArgSeen: unknown = null;
  const events: LlmStreamEvent[] = [];
  for await (const ev of driver.stream({ messages: [{ role: "user", content: "add 10 and 20" }], tools: [] })) {
    events.push(ev);
  }
  console.log("driver.stream() event types:", JSON.stringify(events.map((e) => e.type)));
  console.log("tool-call-ready count:", events.filter((e) => e.type === "tool-call-ready").length);
  console.log("args-delta count:", events.filter((e) => e.type === "tool-call-args-delta").length);

  call = 0;
  const steps: unknown[] = [];
  const gen = streamNativeAgentSteps(driver, {
    messages: [{ role: "user", content: "add 10 and 20" }],
    tools: [],
    maxSteps: 3,
    callTool: async (_n, args) => {
      toolArgSeen = args;
      return { content: [{ type: "text", text: "30" }] };
    },
  });
  let r = await gen.next();
  while (!r.done) { steps.push(r.value); r = await gen.next(); }
  console.log("callTool received args:", JSON.stringify(toolArgSeen));
  console.log("agent.stream() steps yielded:", steps.length, JSON.stringify(steps));
  console.log("agent.stream() returned text:", JSON.stringify(r.value));
}
main();

Actual output

Verbatim from an independent verifier run (node tsx repro2.ts, unmodified repro above):

driver.stream() event types: ["tool-call-start","done"]
tool-call-ready count: 0
args-delta count: 0
callTool received args: {"a":10,"b":20}
agent.stream() steps yielded: 0 []
agent.stream() returned text: "The answer is 30."

A second independent verifier, running the same shape and an identical probe against the tree src/:

SHAPE A (spec: item_id only): types ["tool-call-start","done"]; args-delta 0; tool-call-ready 0
SHAPE A: callTool got {"name":"add","args":{"a":10,"b":20}}
SHAPE A: agent steps 0 []; final text "The answer is 30."
SHAPE B (control: call_id injected): args-delta 2; tool-call-ready 1
PATCHED, SHAPE A: types ["tool-call-start","tool-call-args-delta","tool-call-args-delta","tool-call-ready","done"]; args-delta 2; tool-call-ready 1
PATCHED, SHAPE A: agent steps 2; baseline vitest run src/llm 47/47 pass, patched also 47/47

Expected behavior

The OpenAI Responses path should emit tool-call-args-delta for each response.function_call_arguments.delta and one tool-call-ready when response.function_call_arguments.done arrives, matching:

  • the openai-node SDK event contract — ResponseFunctionCallArgumentsDeltaEvent = { delta, item_id, output_index, sequence_number, type } and ResponseFunctionCallArgumentsDoneEvent = { arguments, item_id, output_index, sequence_number, type } ([email protected], resources/responses/responses.d.ts; raw.githubusercontent.com/openai/openai-node/src/resources/responses/responses.ts lines 3257 / 3287). Neither declares call_id. ResponseFunctionToolCall declares call_id: string and id?: string — distinct fields; item_id points at the output-item id (fc_...).
  • the sibling provider providers/openai-chat-completions.ts, which emits tool-call-args-delta (line 205) and tool-call-ready (line 226).
  • the internal contract llm/types.ts:157 — LlmToolCallReadyEvent is documented as "A tool call with complete, parsed arguments", and native_runner.ts:51-58 builds an AgentStep only on tool-call-ready.

So agent.stream() should yield one AgentStep for the add call, instead of 0 [].

Root cause

libraries/typescript/packages/agent/src/llm/providers/openai-responses.ts:328 and :345: both handlers read const callId = typeof parsed.call_id === "string" ? parsed.call_id : "", but these events carry item_id, not call_id, so the key is always "" and callBuffers.get(callId) (line 330 / line 348) is always undefined. The buffer is written under item.call_id at line 311 (populated from response.output_item.added, line 308). Result: the delta and done branches never emit their events, and every downstream consumer of tool-call-ready sees an empty step list.

Fix direction (not a full diff): key callBuffers by the output item id (item.id, i.e. fc_...) and look it up with parsed.item_id (or fall back via output_index to the stored index) in the delta/done handlers, while still yielding the stored call_id as toolCallId on the emitted events.

Screenshots

N/A — headless Node library, no UI screenshot applies.

Desktop (please complete the following information)

  • OS: macOS 26.6.2
  • Version: mcp-use agent 2.0.16 @ dcaa0b8df5721f11d4d2aefae1f85df2a1903ba5, Node v22.23.2, openai 6.17.0
  • Browser: N/A

Smartphone (please complete the following information)

N/A — not a mobile issue.

Additional context

  • Reachability: new MCPAgent({ llm: "openai/gpt-5", ... }).stream({ prompt }) → agents/mcp_agent.ts:549 streamNativeAgentSteps(driver) → llm/native_runner.ts:52 → createLlmDriver returns OpenAIResponsesDriver (llm/driver.ts:53-62) → streamToolLoop → streamResponsesTurn. Same path for openrouter models matching /^~?openai\//. The inspector client-side chat (useChatMessagesClientSide.ts:436) creates the tool card at tool-call-start with args {} and never receives tool-call-ready.
  • Related issues/PRs found by collision checks: none. Searches for tool-call-ready, function_call_arguments, and "Responses streaming" in mcp-use/mcp-use return only unrelated closed issues.
  • Happy to open a PR with the approach sketched above if you'd like, or happy to be assigned.

Activity

  1. khandrew1 commented on Oct 19, 2026

    @khandrew1
    Contributor

    Fixed by #2551, now on canary. The argument events are correlated by item_id, so tool-call-args-delta and tool-call-ready fire and MCPAgent.stream() yields tool steps on OpenAI. Ships in the next canary release.

  2. khandrew1 commented on Oct 19, 2026

    @khandrew1
    Contributor

    Released in @mcp-use/agent 2.0.25 (mcp-use 2.11.0). Closing.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    TypeScriptagentbugSomething isn't workingreadyWhen a PR is ready to be merged and waiting for author to merge

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions