[Enhancement]: Support Local Instructions and Toolcalls when calling an Agent via remote API

[peeeteeer]

What features would you like to see added?

Today calls to the new agent api do not support local instructions and tools that are send with the API call. They are simply ignored.

There are use cases were the caller can not be changed and it expects that the API behaves exactly like a chatcompletion/reponse API including it's instructions or tool calls. (e.g. codex, or other rich clients)

I would love to see:

1. That instructions that are send via the API are appended to the agent instructions
2. Tools that are available on the API caller side can be used by the LLM when required

More details

where 1 is pretty straightforward - for 2 I would suggest the following behavior

1. Receive tools from the API request.
2. Merge request tools with the agent’s existing tools by tool name.
3. Split tools into two groups:
   • native/server tools (can run on backend),
   • passthrough/client tools (must be executed by API client).
4. Let the model see all merged tools so it can choose correctly.
5. When tool calls are produced:
   • run native tool calls on backend,
   • keep passthrough tool calls for client handoff.
6. If native and passthrough tool calls both exist in one turn:
   • run native first,
   • then end turn for passthrough handoff.
7. End turn cleanly when passthrough is needed, and return tool-call details needed by client.

Which components are impacted by your request?

General

Pictures

No response

Code of Conduct

• I agree to follow this project's Code of Conduct