[BUG] When using GPT 4o, the diagnoser returns an hypothesis_tool call that has an unexpected format and ends up not being parsed

[m-kostira]

Using model_name="gpt-4o", the orchestrator seems to have trouble parsing the llm response for the diagnoser agent. This causes issues at main.py::_update_differential_from_hypothesis, where structured data parsing fails and it falls back to wonky regex parsing.

The expected response is a tool call json with schema:

`

Define structured output tool for differential diagnosis

    hypothesis_tool = {
        "type": "function",
        "function": {
            "type": "function",
            "name": "update_differential_diagnosis",
            "description": "Update the differential diagnosis with structured probabilities and reasoning",
            "parameters": {
                "type": "object",
                "properties": {
                    "summary": {
                        "type": "string",
                        "description": "One-sentence summary of primary diagnostic conclusion and confidence"
                    },
                    "differential_diagnoses": {
                        "type": "array",
                        "items": {
                            "type": "object",
                            "properties": {
                                "diagnosis": {"type": "string", "description": "The diagnosis name"},
                                "probability": {"type": "number", "minimum": 0, "maximum": 1, "description": "Probability as decimal (0.0-1.0)"},
                                "rationale": {"type": "string", "description": "Brief rationale for this diagnosis"}
                            },
                            "required": ["diagnosis", "probability", "rationale"]
                        },
                        "minItems": 2,
                        "maxItems": 5,
                        "description": "Top 2-5 differential diagnoses with probabilities"
                    },
                    "key_evidence": {
                        "type": "string",
                        "description": "Key supporting evidence for leading hypotheses"
                    },
                    "contradictory_evidence": {
                        "type": "string", 
                        "description": "Critical contradictory evidence that must be addressed"
                    }
                },
                "required": ["summary", "differential_diagnoses", "key_evidence"]
            }
        }
    }

`

Actual response is the expected json but wrapped around in some more json:

'[ {\'function\': {\'arguments\': \'{ "summary": "Differential includes musculoskeletal, vascular, and neuropathic causes for thigh pain in elderly.", "differential_diagnoses": [ { "diagnosis": "Muscle strain or sprain", "probability": 0.4, "rationale": "Common cause of thigh pain, especially if there is a history of recent physical activity or trauma." }, { "diagnosis": "Osteoarthritis", "probability": 0.3, "rationale": "Prevalent in elderly, can cause referred pain to the thigh." }, { "diagnosis": "Peripheral artery disease", "probability": 0.2, "rationale": "Common in elderly with risk factors, can cause pain due to poor circulation." }, { "diagnosis": "Sciatica", "probability": 0.1, "rationale": "May present with thigh pain if nerve root compression is present." } ], "key_evidence":"Patient\\\'s age suggests higher likelihood of degenerative or vascular conditions.", "contradictory_evidence": "No specific history of trauma or activity suggests against acute musculoskeletal injury." }\', \'name\': \'update_differential_diagnosis\' }, \'id\': \'call_YFc42mvnMKpfTDSYSkMasrtY\', \'type\': \'function\' } ]'