RuntimeError: dictionary changed size during iteration in _approvals during concurrent tool execution

[introspectDaily]

Bug Description

When running an agent with require_approval configured and multiple tool calls executing concurrently, a RuntimeError: dictionary changed size during iteration occurs in RunContextWrapper._approvals.

Root Cause

In src/agents/run_context.py, _fork_with_tool_input and _fork_without_tool_input share the _approvals dict by reference:

def _fork_with_tool_input(self, tool_input: Any) -> RunContextWrapper[TContext]:
    fork = RunContextWrapper(context=self.context)
    fork._approvals = self._approvals  # shared reference
    ...

When parallel tool calls run concurrently in an async event loop:

• One coroutine iterates over _approvals (e.g., in _get_approval_status_for_key)
• Another coroutine calls _get_or_create_approval_entry, which inserts a new key into the same dict

This causes the RuntimeError because the dict size changes during iteration.

Reproduction

1. Create an agent with MCP server that has require_approval set as a dict mapping tool names to "always"
2. Send a prompt that triggers multiple tool calls in the same turn
3. Use state.approve(item, always_approve=True) to approve tools
4. The error occurs intermittently when the SDK internally checks approval status for one tool while creating an approval entry for another

Environment

• openai-agents==0.14.6
• Python 3.11
• Using Runner.run_streamed() with MCP tools

Suggested Fix

Either:

1. Use dict.copy() when iterating in _get_approval_status_for_key and related methods
2. Or use an asyncio.Lock to protect concurrent access to _approvals

Option 1 (minimal change):

def _get_or_create_approval_entry(self, tool_name: str) -> _ApprovalRecord:
    approval_entry = self._approvals.get(tool_name)
    if approval_entry is None:
        approval_entry = _ApprovalRecord()
        self._approvals[tool_name] = approval_entry
    return approval_entry

The get + conditional set pattern is safe, but the issue is when other methods iterate over self._approvals.items() or similar while this method adds a new key. Wrapping iteration sites with list(self._approvals.items()) would prevent the RuntimeError.

[RudraDudhat2509]

Hey, I'd like to work on a fix for this, is this open to a PR?

[seratch]

Thanks for your interest. I will resolve it by #3516

[seratch]

Thanks for the detailed report.

I took a closer look at the current implementation and I do not think this is reproducible as described. RunContextWrapper.get_approval_status() and _get_approval_status_for_key() do not iterate over self._approvals; they only perform direct key lookups. Also, the approval update paths are synchronous, so normal async concurrent tool execution should not interleave in the middle of a dict iteration.

The only remaining theoretical case would involve application-level cross-thread access, for example mutating approvals from one thread while another thread serializes RunState. The SDK does not currently treat RunContextWrapper / RunState as thread-safe mutable objects, and adding locks for this would introduce extra complexity and compatibility concerns for copy/pickle behavior.

Given that, I’m going to close this as not planned unless we can get a minimal repro that fails under normal SDK async execution without cross-thread mutation.