Fix tool schema constraints dropped during OCI conversion (breaks MCP + Gemini)

[fede-kamel]

Bug

convert_to_oci_tool strips JSON Schema constraints (enum, min/max, pattern, format) from tool definitions before sending them to the model. A tool defined with enum: ["cpu", "memory", "disk", "network"] gets reduced to just type: string — the model never sees the allowed values.

Additionally, Optional fields (Pydantic anyOf pattern) get converted to type: "any" which is not valid JSON Schema — this causes Gemini models to return 400 errors.

Root cause: the old code uses tool.args which goes through Pydantic's tool_call_schema re-generation, then only keeps type and description:

# old code
p_name: {
    "type": p_def.get("type", "any"),
    "description": p_def.get("description", ""),
}

Reproduction

Tool with enum: ["cpu", "memory", "disk", "network"], prompt: "Check the latency metrics" (intentionally not in enum).

main branch (broken) — enum stripped, models hallucinate or crash:

meta.llama-3.3-70b-instruct:    metric_name = "latency"  (not a valid value)
google.gemini-2.5-flash:        metric_name = "latency"  (not a valid value)
google.gemini-2.5-pro:          metric_name = "latency"  (not a valid value)
google.gemini-2.5-flash-lite:   metric_name = "latency"  (not a valid value)

fix branch — enum preserved, models respect constraints:

meta.llama-3.3-70b-instruct:    metric_name = "network"  (closest valid value)
google.gemini-2.5-flash:        refused — "can only query cpu, memory, disk, network"
google.gemini-2.5-pro:          refused — "available metrics are: cpu, memory, disk, network"
google.gemini-2.5-flash-lite:   refused — "can only query cpu, memory, disk, network"

MCP tool reproduction

Same test with a FastMCP server (MetricName enum + duration_hours with ge=1, le=168).

main branch — MCP tools broken on Gemini, hallucinated on Llama:

What main sends to OCI:
  metric_name:    {"type": "string", "description": ""}     ← enum stripped
  duration_hours: {"type": "any", "description": "..."}     ← anyOf becomes "any"

meta.llama-3.3-70b:      metric_name = "latency"  (hallucinated)
google.gemini-2.5-flash: 400 error  (type "any" rejected by Gemini)
google.gemini-2.5-pro:   400 error  (type "any" rejected by Gemini)

fix branch — MCP tools work correctly:

What fix sends to OCI:
  metric_name:    {"type": "string", "enum": ["cpu","memory","disk","network"]}
  duration_hours: {"type": "integer", "minimum": 1, "maximum": 168}

meta.llama-3.3-70b:      metric_name = "network"  (valid)
google.gemini-2.5-flash: refused — "does not support latency"
google.gemini-2.5-pro:   refused — "can check: cpu, memory, disk, network"

Fix

• Use args_schema.model_json_schema() instead of tool.args to get the original schema with constraints intact
• _sanitize_tool_property() filters through an allowlist of standard JSON Schema keys (enum, format, min/max, pattern, items, etc.) and resolves Pydantic v2 anyOf patterns for Optional[T]
• For Cohere: _enrich_description() embeds constraints into the description string since CohereParameterDefinition only supports type/description/is_required

Tests

• 22 unit tests for _sanitize_tool_property (enum, format, ranges, patterns, nested objects, anyOf resolution, MCP round-trips)
• 6 integration tests against live OCI (Meta Llama + Cohere): enum constraints, numeric ranges, schema verification