fix: handle ValueError and RecursionError in _safe_json_serialize

[voidborne-d]

Summary

Both _safe_json_serialize implementations only catch TypeError and OverflowError from json.dumps, but json.dumps can also raise:

• ValueError: circular references (e.g. obj.ref = obj)
• RecursionError: deeply nested structures exceeding Python's recursion limit

Impact

telemetry/tracing.py (critical): _safe_json_serialize is called inside a finally block during trace_tool_call. An unhandled exception here can suppress the original tool execution result — telemetry failures should never affect runtime correctness.

models/lite_llm.py: An unhandled exception during message serialization causes hard LLM request construction failures where a graceful fallback is expected.

Root Cause

The except clause catches (TypeError, OverflowError) but not ValueError or RecursionError. Both are documented json.dumps failure modes:

• ValueError for circular references (per Python docs)
• RecursionError for structures exceeding sys.getrecursionlimit()

Note: in tracing.py, the default=lambda o: ... callback handles TypeError for individual non-serializable objects, but circular references and deep recursion bypass the default callback entirely.

Fix

• Add ValueError and RecursionError to the exception tuple in both implementations.
• In lite_llm.py, wrap the str() fallback in a secondary try/except RecursionError since str() → __repr__() can also fail on deeply recursive objects.

Reproduction

class Node:
    def __init__(self):
        self.ref = self

obj = Node()
# Before fix: raises ValueError
# After fix: returns fallback string
_safe_json_serialize(obj)

Tests

8 regression tests added across both modules:

• test_safe_json_serialize.py (telemetry): circular ref, deep nesting, normal dict, non-serializable object
• test_litellm_safe_serialize.py (models): circular ref, deep nesting, normal dict, non-serializable object

All 8 tests pass.

Fixes #5411
Fixes #5412

[rohityan]

Hi @voidborne-d , Thank you for your contribution! We appreciate you taking the time to submit this pull request. Please fix formatting errors by running autoformat.sh

[voidborne-d]

Hi @rohityan, thanks for the review! I've run autoformat.sh and pushed the formatting fix in ddff034. The changes were isort reordering + pyink reformatting across lite_llm.py and both test files. All 8 tests pass.

[voidborne-d]

Rebased onto current main (c231cf8). Upstream now catches (TypeError, ValueError, OverflowError) in both call sites since this PR was opened, so the residual delta is two narrow additions:

1. RecursionError in the exception tuple — both lite_llm.py:560 and tracing.py:131. Upstream's tuple doesn't include it; deeply self-referential objects (e.g. circular agent state) still crash _safe_json_serialize today.
2. Nested str(obj) guard in lite_llm.py:562-567 — str() itself can hit RecursionError when an object has a recursive __repr__, so the fallback path needs its own try.

Tests in tests/unittests/{models,telemetry}/test_*_safe_serialize.py cover the RecursionError path specifically (which is what's not covered upstream). Repro and motivation in the original PR body still hold.

Mergeable now (was dirty pre-rebase). Happy to drop the nested str(obj) guard if it reads as over-engineered for the actual reproductions you're seeing — the tuple addition is the load-bearing piece.