Python create() raises ValidationError on tool-call responses (strict precontext)

[Khurdhula-Harshavardhan]

Severity: High — create() raises ValidationError on real tool-call responses

to_interfaze re-validates the whole completion through a strict typed model:

# src/interfaze/_chat.py
return InterfazeChatCompletion.model_validate(data)

InterfazeChatCompletion.precontext is Optional[List[Precontext]] and Precontext.name is required. But the server appends raw model tool-calls to precontext whenever the final model calls a tool (user-defined tools, or internal action tools like run_code):

• src/helpers/response.ts:105 → internalToolUsedToBeShownToUser = [...prev, ...toolCalls] where toolCalls are raw AI-SDK objects {toolCallId, toolName, input} — no name, no result.
• ModifyToolResponsesForUser (src/utils/messaging.ts:17) sets name from a lookup map, so an unmapped tool yields name: undefined too.

Reproduction (verified)

from interfaze._chat import to_interfaze
from openai.types.chat import ChatCompletion
raw = {
 "id":"r","object":"chat.completion","created":1,"model":"interfaze-beta",
 "choices":[{"index":0,"message":{"role":"assistant","content":"hi","refusal":None},
 "finish_reason":"stop","logprobs":None}],
 "usage":{"prompt_tokens":5,"completion_tokens":3,"total_tokens":8},
 "vcache":False,
 "precontext":[{"name":"ocr","result":{}},
 {"toolCallId":"call_1","toolName":"run_code","input":{}}],
}
to_interfaze(ChatCompletion.model_construct(**raw), strip_fence=False)
# pydantic ValidationError: precontext.1.name Field required

Any function-calling / run_code turn therefore crashes the non-streaming create(). This is the exact "strong types break precontext" failure mode the SDK exists to avoid — reintroduced by re-validating the response. The openai SDK's own rule is type strongly, validate loosely, preserve everything (it uses model_construct, extra="allow").

Fix options

• Parse leniently: InterfazeChatCompletion.model_construct(**data) (mirrors openai-python), or
• Loosen the field: precontext: Optional[List[Union[Precontext, Dict[str, Any]]]], or
• Don't re-run full validation in to_interfaze.

Add a fixture/test with a mixed precontext (one {name,result} + one raw {toolCallId,toolName,input}) — current fixtures never cover this (TOOL_CALL has no precontext).