Add opt-in strict JSON Schema output for OpenAI models

[cacoze]

Problem

The OpenAI integrations currently send JSON schemas without enabling strict Structured Outputs:

• Chat Completions omits json_schema.strict.
• Responses API explicitly sets strict: false.

Consequently, supplying a schema does not guarantee schema-conforming output. We have observed successful responses containing:

• Arrays violating minItems
• Strings where objects were required
• Unexpected properties despite additionalProperties: false
• Incorrectly named fields

These responses are valid JSON, but fail downstream Pydantic validation.

Proposed solution

Add an opt-in OpenAI model option such as:

llm -m gpt-4o -o strict_schema true --schema schema.json "..."

And its Python equivalent:

model.prompt(
    prompt,
    schema=OutputModel,
    strict_schema=True,
)

When enabled, the OpenAI adapter would send:

{
  "type": "json_schema",
  "json_schema": {
    "name": "output",
    "schema": {},
    "strict": true
  }
}

The Responses API equivalent should also use strict: true.

The option could default to false for backwards compatibility.

Schema compatibility

OpenAI strict mode requires a restricted JSON Schema shape, including:

• additionalProperties: false on objects
• Every property listed as required
• Only supported JSON Schema keywords

Possible behavior:

1. Convert Pydantic models using the same strict-schema normalization used by the OpenAI Python SDK, or
2. Validate the supplied schema and return a clear error when it is incompatible.

Silently weakening the schema would be undesirable.

Response handling

It would also be useful to preserve or expose:

• Safety refusals
• finish_reason
• Incomplete/truncated response status

These are documented cases where strict Structured Outputs may not return a schema-conforming result.

Reproduction

A schema containing object restrictions and a non-empty array constraint can occasionally return output that violates those constraints:

{
  "type": "object",
  "properties": {
    "items": {
      "type": "array",
      "minItems": 1,
      "items": {
        "type": "object",
        "properties": {
          "name": {"type": "string"}
        },
        "required": ["name"],
        "additionalProperties": false
      }
    }
  },
  "required": ["items"],
  "additionalProperties": false
}

Without strict: true, responses such as {"items": []} or
{"items": ["example"]} remain possible.

Environment

• llm 0.29
• OpenAI gpt-4o
• Python API with a Pydantic model passed as schema