fix(tracing): serialize big integers as strings to prevent Traces dashboard precision loss

[xodn348]

Summary

Fixes #2094

When a tool is called with integer arguments that exceed JavaScript's Number.MAX_SAFE_INTEGER (2^53 − 1 = 9007199254740991), the OpenAI Traces dashboard loses precision because it parses the JSON number as an IEEE-754 double. For example, 9007199254740993 would display as 9007199254740992.

Root Cause

FunctionSpanData.export() serializes tool call arguments as a JSON string (self.input). Python integers can be arbitrarily large, but when the resulting JSON is consumed by JavaScript, any integer > 2^53 − 1 is silently truncated.

Fix

In FunctionSpanData.export() (src/agents/tracing/span_data.py), before re-emitting the input JSON:

1. Parse the stored JSON string back to a Python object.
2. Walk the object recursively with _sanitize_bigint(), converting any int whose absolute value exceeds _JS_MAX_SAFE_INTEGER to its decimal string representation.
3. Re-serialize to JSON.

bool is checked before int (since bool is a subclass of int in Python) so True/False are never converted.

Malformed JSON in self.input is caught and passed through unchanged, preserving existing behavior.

Testing

• 31 existing tracing tests pass without modification.
• No changes to the public API or any other span type.

Example

# Before fix — JS dashboard receives 9007199254740992 (wrong)
span_data = FunctionSpanData(name="tool", input='{"id": 9007199254740993}')
span_data.export()["input"]  # '{"id": 9007199254740993}'  ← JS truncates this

# After fix — JS dashboard receives the string "9007199254740993" (correct)
span_data.export()["input"]  # '{"id": "9007199254740993"}'

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8fc567fefe

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Broaden sanitize error handling to avoid export thread crashes

FunctionSpanData.export() now assumes json.loads(self.input) only fails with JSONDecodeError/TypeError, but valid-yet-pathological inputs can also raise exceptions like RecursionError (very deep nesting) or ValueError (very large integer literals on Python 3.11+). In those cases this method raises instead of falling back to raw self.input, and because exporter paths call item.export() without guarding per-item failures in the batch worker, one such span can terminate trace exporting for subsequent spans.

Useful? React with 👍 / 👎.

[seratch]

This is a server-side issue, we won't manipulate the data on the client side.