Bedrock Converse API with tool use drop integers ≥ 2^31 (toolUse.input becomes {}) and sometimes coerce to strings — repro on Llama 3.1 & Mistral Large

Service: Amazon Bedrock

API: converse with toolConfig (JSON Schema)

Region: us-west-2

Models tested:

• meta.llama3-1-70b-instruct-v1:0
• mistral.mistral-large-2407-v1:0

Summary

When using the Converse API with tool use and a JSON-Schema declaring "type": "integer" for tool inputs, certain large integer values cause the runtime to drop the entire tool input: the model chooses the tool (stopReason: "tool_use"), but toolUse.input is {}.

This happens at and above 2^31 (2147483648), and also at 2^32 (4294967296). Strangely, much larger integers (e.g., 2^53−1 and 2^53) are accepted but are returned as strings, despite the schema declaring "integer".

This breaks common cases like epoch-millisecond timestamps in tool parameters.

Expected

• Integers within 64-bit range should be accepted and delivered (ideally as integers).
• If coercion to string is intended, it should be consistent and documented—not dropping the input silently for specific ranges.

Actual

• For values ≥ 2^31 (tested) and at 2^32, toolUse.input becomes {}.
• For very large values (2^53−1, 2^53), values are preserved but as strings in toolUse.input.

Minimal Repro (Python)

import boto3, json

client = boto3.client("bedrock-runtime", region_name="us-west-2")
MODEL_ID = "meta.llama3-1-70b-instruct-v1:0"  # also occurs with "mistral.mistral-large-2407-v1:0"

tool_config = {
    "tools": [{
        "toolSpec": {
            "name": "test_instana_tool",
            "description": "Test tool for large integers",
            "inputSchema": {
                "json": {
                    "type": "object",
                    "properties": {
                        "var_from":   {"type": "integer"},
                        "to":         {"type": "integer"},
                        "window_size":{"type": "integer"}
                    },
                    "required": ["var_from","to","window_size"]
                }
            }
        }
    }],
    "toolChoice": {"auto": {}}
}

def run(val):
    prompt = f"Use test_instana_tool with var_from={val}, to={val}, window_size=60000"
    resp = client.converse(
        modelId=MODEL_ID,
        messages=[{"role":"user","content":[{"text": prompt}]}],
        toolConfig=tool_config
    )
    rid = resp["ResponseMetadata"]["RequestId"]
    content = resp["output"]["message"]["content"]
    inputs = [c.get("toolUse", {}).get("input") for c in content if "toolUse" in c]
    print(val, rid, inputs)

for v in [2147483647, 2147483648, 4294967296, 9007199254740991, 9007199254740992]:
    run(v)

Observed RequestIds (Llama 3.1, Converse)

• 2147483647 (2^31−1) → cf6db05f-6cda-4b8a-984b-4695bb52cefe → accepted; strings:

  {"var_from":"2147483647","to":"2147484647","window_size":"60000"}

• 2147483648 (2^31) → 7fe71e43-74f6-4059-bd02-edc583d143ff → input dropped: {}

• 4294967296 (2^32) → 287c0398-db10-4538-8ca8-0d2eceee10f4 → input dropped: {}

• 9007199254740991 (2^53−1) → ba1e21cf-4a8d-45e4-bfc8-77aa63be5d8f → accepted; strings

• 9007199254740992 (2^53) → 45ba204a-5c24-4cc6-abc1-327f540d2dd9 → accepted; strings

Also repro’d with Mistral Large (Converse)

• mistral.mistral-large-2407-v1:0: model selects tool, but toolUse.input is {} for large integers (example RequestId from a separate run: e8d55358-11fb-4471-b150-108e1257dace).

Questions

1. What integer width does the Converse tool input validator support for "type":"integer"? (Should be at least int64.)
2. Why are some large values coerced to strings while others are dropped? Is this a known issue with the global validator or a model-adapter quirk?
3. Is there a supported schema hint (e.g., format: int64) that Converse honors to prevent dropping values?
4. Can AWS confirm and provide an ETA or recommended best practice for 64-bit integer tool parameters (e.g., epoch-ms)?