fix(sdk): avoid model_copy(deep=True) on configs with unpicklable MCP sessions (#2669)

[knoal]

Bug (reported in #2669)

AsyncModels.generate_content crashes with TypeError: cannot pickle '_asyncio.Future' object when config is a GenerateContentConfig object containing an MCP ClientSession tool. The same settings passed as a dict work fine.

Root cause

google/genai/models.py:8632 (in AsyncModels.generate_content):

parsed_config = config.model_copy(deep=True)

Pydantic v2's model_copy(deep=True) pickles the entire config tree. A live MCP ClientSession holds an internal _asyncio.Future, which is unpicklable → TypeError.

Investigation uncovered a second model_copy(deep=True) at models.py:8751 (final_parsed_config before the lower-level _generate_content call) which has the same failure mode.

Fix

Replace both deep-copy calls with a try/except fallback to shallow model_copy():

try:
    parsed_config = config.model_copy(deep=True)
except TypeError as _pickle_err:
    if "pickle" not in str(_pickle_err) and "cannot pickle" not in str(_pickle_err):
        raise
    parsed_config = config.model_copy()

The shallow copy is semantically sufficient because the downstream code only reassigns parsed_config.tools to a new list — it does not mutate individual tools in place. Shallow copy keeps tool references shared with the source config, which is fine because callers don't reuse the config after the call.

For the 99% of configs that don't contain unpicklable subtrees, behavior is unchanged: the existing model_copy(deep=True) runs first and succeeds.

Tests

google/genai/tests/async/test_generate_content_mcp_session_2669.py — 2 regression tests:

1. test_generate_content_config_with_mcp_session_does_not_pickle — calls AsyncModels.generate_content with a config containing an MCP-like session; fails with the user-visible TypeError before fix, passes after.
2. test_model_copy_deep_with_mcp_session_squares_the_bug — documents that config.model_copy(deep=True) itself raises on unpicklable subtrees (sanity).

============================== 2 passed in 1.06s ===============================

Cross-cutting note

Two model_copy(deep=True) call sites in the same function (#2669 hit them on a single execute path). A repo-wide grep for model_copy(deep=True) may surface others with the same latent issue. The fallback pattern (try deep, fall back to shallow on pickle-TypeError) is intentionally narrow — only fires on unpicklable types. Maintainers may want to apply it uniformly.

MCE audit

• Mean audit Brier: 0.018 across 4 predictions (4/4 hit)
• Wall-clock: ~12 minutes (frame → push; the shallow-vs-deep investigation cost a few extra minutes vs the original "try copy.deepcopy" idea — copy.deepcopy also uses pickle under the hood for objects without __deepcopy__)

Per-prediction Brier

Prediction	Confidence	Outcome	Brier
P1: bug at models.py:8632 deep-copy	0.95	hit	0.0025
P2: fix is to avoid model_copy(deep=True)	0.85	hit	0.0225
P3: regression test fails on unfixed	0.85	hit	0.0225
P4: fix is small (~10 line changes)	0.85	hit	0.0225

Refs

• Issue AsyncModels.generate_content crashes with cannot pickle '_asyncio.Future' when config is a GenerateContentConfig object containing an MCP ClientSession tool #2669 in googleapis/python-genai

[knoal]

MCE audit

• Task: T19f6376a2272ae
• Mean audit Brier: 0.018 across 4 predictions (4/4 hit)
• Repository: googleapis/python-genai (3rd repo in pilot)
• Wall-clock: ~12 minutes

Branch: fix/2669-async-mcp-session-deep-copy (commit 4a1a2f0)

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.