[rokujix]

client.chat.create (sync and async) takes a guidelines list for
custom moderation controls and sets GetCompletionsRequest.guidelines
(field 39). Each entry is a Guideline dict in the REST shape,
{"type": "policy", "category": ..., "behavior": ...}, or a
chat_pb2.Guideline. An unknown category or behavior raises
ValueError. The API rejects an unsupported pair, a repeated category,
or a model that cannot apply guidelines with INVALID_ARGUMENT.

The v5 and v6 chat_pb2 stubs are regenerated with xai-proto's
buf.gen.yaml from the proto state the vendored stubs match (xai-proto
#73 plus safety_identifier) plus the guideline definitions. Before the
change, that pipeline reproduces the vendored stubs except for the
from . import rewrite.

Depends on xai-org/xai-proto#80, which adds GetCompletionsRequest.guidelines.

[deanrie]

Hey @rokujix, one line for the reviewer: the dependency named in the description, xai-org/xai-proto#80, merged on 2026-10-21 at 18:16 UTC, about fourteen hours after this PR was opened, so the vendored chat_pb2 stubs should now be checked against a regeneration from that merge commit rather than from the pre-merge branch. If they match byte for byte except the from . import rewrite you mention, the proto side of this is settled and what remains is the chat.py / types / tests review; if they differ in the Guideline message or field 39, the regeneration is the fix.