fix(xai-responses): strip enum values containing '/' from tool schemas

xAI's /v1/responses and /v1/chat/completions endpoints reject tool schemas
whose enum values contain a forward slash with a generic HTTP 400 'Invalid
arguments passed to the model.' before any token is emitted — the schema
compiler trips on the '/' character regardless of where it appears.

Most commonly hit by MCP-derived tools whose enum lists HuggingFace model
IDs ('Qwen/Qwen3.5-0.8B', 'openai/gpt-oss-20b') or owner/name environment
identifiers.

Mirrors the existing strip_pattern_and_format sanitizer (PR for #27197).
The new strip_slash_enum walks tool parameters and drops the entire enum
keyword when any value contains '/' — keeping it partial would still 400
since xAI's failure is all-or-nothing on the enum. The field description
still reaches the model so the prompting hint is preserved.

Wired in at both code paths for parity:
  - agent/chat_completion_helpers.py (main agent xAI Responses path)
  - agent/auxiliary_client.py (aux client xAI Responses path, matching
    the same parity guarantee 2fae8fba9 established for pattern/format)

Salvaged from #28021 by @Slimydog21 — contributor's branch was severely
stale (would have reverted ~5000 LOC across azure/kanban/i18n); fix
re-applied surgically on current main with their sanitizer + 9 tests
preserved verbatim. Author noreply email used (original was a Mac
hostname leak).
This commit is contained in:
Slimydog21 2026-05-18 10:36:51 -07:00 committed by Teknium
parent d9331eecee
commit aae1615977
4 changed files with 220 additions and 3 deletions

View file

@ -380,3 +380,66 @@ def strip_pattern_and_format(tools: list[dict]) -> tuple[list[dict], int]:
stripped,
)
return tools, stripped
def strip_slash_enum(tools: list[dict]) -> tuple[list[dict], int]:
"""Strip ``enum`` keywords whose string values contain a forward slash.
xAI's ``/v1/responses`` and ``/v1/chat/completions`` endpoints compile
tool schemas to a grammar that rejects ``enum`` values containing ``/``
(the request fails with HTTP 400 "Invalid arguments passed to the
model" before any token is emitted). Most commonly hit by MCP-derived
tools whose enum lists HuggingFace model IDs (``Qwen/Qwen3.5-0.8B``,
``openai/gpt-oss-20b``) or owner/name environment IDs. The constraint
is purely a prompting hint; dropping it lets the model still see the
field description and pick a value, without xAI tripping on the slash.
Args:
tools: OpenAI-format or Responses-format tool list, mutated in
place. Callers that need to preserve the original should
deep-copy first.
Returns:
``(tools, stripped_count)`` same list reference plus a count of
how many ``enum`` keywords were removed.
"""
if not tools:
return tools, 0
stripped = 0
def _walk(node: Any) -> None:
nonlocal stripped
if isinstance(node, dict):
enum_val = node.get("enum")
if isinstance(enum_val, list) and any(
isinstance(v, str) and "/" in v for v in enum_val
):
node.pop("enum", None)
stripped += 1
for v in node.values():
_walk(v)
elif isinstance(node, list):
for item in node:
_walk(item)
for tool in tools:
if not isinstance(tool, dict):
continue
fn = tool.get("function")
if isinstance(fn, dict):
params = fn.get("parameters")
if isinstance(params, dict):
_walk(params)
continue
params = tool.get("parameters")
if isinstance(params, dict):
_walk(params)
if stripped:
logger.info(
"schema_sanitizer: stripped %d enum keyword(s) containing '/' "
"from tool schemas (xAI Responses grammar-compile recovery)",
stripped,
)
return tools, stripped