Description

When using hosted ShellTool (container-based code execution), the Responses API occasionally returns a response.completed event containing a shell_call output item but without a corresponding shell_call_output. In a multi-turn streaming flow, this orphan shell_call is carried into the next LLM turn's input, causing the model to emit:

No tool output found for shell call call_XXXXX

The SDK has drop_orphan_function_calls() which correctly handles this scenario, but it is only called in normalize_resumed_input() (for session resumption). It is not called in the normal streaming multi-turn path.

Reproduction

• SDK version: 0.12.0 (also reproducible on 0.12.1)
• Transport: WebSocket (OpenAIResponsesModel with websocket_connections)
• Tool: ShellTool with container_id=None (auto-provisioned)
• Model: gpt-5.4

Steps

1. Start a streamed multi-turn agent run with a hosted ShellTool
2. The model issues a shell_call — the API executes it server-side
3. The API returns response.completed with shell_call in output items but no shell_call_output
4. The SDK processes the shell_call item in turn_resolution.py:1339-1355, logs "Skipping local shell execution for hosted shell tool", and continues
5. On the next turn, run_loop.py:1166-1175 calls normalize_input_items_for_api() but not drop_orphan_function_calls()
6. The orphan shell_call (without output) is sent as input to the next LLM call
7. The model responds with an error message instead of continuing the task

Timeline from production logs

12:27:48.963  1st LLM call (conversation start)
12:28:49.808  Tool execution: get_interest_lists
12:29:29.840  shell_call detected (running_code activity)
12:30:13.883  Tool execution: visualize_line_chart_data_tool
12:30:13.995  SDK processes shell_call item (turn_resolution.py)
12:30:13.996  "Skipping local shell execution for hosted shell tool"
12:30:14.001  Next LLM call — orphan shell_call in input
12:30:16.368  ERROR: "No tool output found for shell call call_R94BZzfgS28s3sgJdD4rjKd5"

Root Cause Analysis

Two contributing factors

1. API-side: The Responses API returns response.completed with a shell_call in the output but omits the shell_call_output. This may be a race condition in container execution or a serialization issue.

2. SDK-side: drop_orphan_function_calls() in agents/run_internal/items.py correctly filters out tool calls without matching outputs, but it's only invoked via normalize_resumed_input() — which is called for resumed sessions only, not during normal streaming multi-turn execution.

Relevant code paths

run_loop.py — streamed turn input (line ~1166-1175):

input_items = normalize_input_items_for_api(input_items)
# ← drop_orphan_function_calls() is NOT called here

run_loop.py — non-streamed turn input (line ~1531):

input_items = normalize_input_items_for_api(input_items)
# ← same gap

items.py — only usage of drop_orphan_function_calls (line ~148-155):

def normalize_resumed_input(raw_input):
    if isinstance(raw_input, list):
        normalized = normalize_input_items_for_api(raw_input)
        return drop_orphan_function_calls(normalized)  # ← only here
    return raw_input

Suggested Fix

Call drop_orphan_function_calls() after normalize_input_items_for_api() in both run_single_turn_streamed and run_single_turn in run_loop.py:

input_items = normalize_input_items_for_api(input_items)
input_items = drop_orphan_function_calls(input_items)  # add this line

This is a safe, defensive change — drop_orphan_function_calls() is a no-op when all tool calls have matching outputs.

Workaround

We are currently using a monkey-patch that wraps normalize_input_items_for_api to also call drop_orphan_function_calls:

from agents.run_internal import items as _items_module
from agents.run_internal import run_loop as _run_loop_module

_original_normalize = _items_module.normalize_input_items_for_api

def _normalize_and_drop_orphans(items):
    normalized = _original_normalize(items)
    return _items_module.drop_orphan_function_calls(normalized)

_run_loop_module.normalize_input_items_for_api = _normalize_and_drop_orphans
_items_module.normalize_input_items_for_api = _normalize_and_drop_orphans

Impact

• Affects any agent using hosted ShellTool in multi-turn conversations
• The conversation effectively breaks — the model sees a tool call without output and emits an error instead of continuing
• No user-facing recovery possible without restarting the conversation