feat: port LangSmith LLM Gateway routing from upstream (#1671, #1673, #1674, #1678) (#155)

* feat: port LangSmith LLM Gateway routing from upstream (#1671, #1673, #1674, #1678)

Ports four upstream commits that add opt-in LLM call routing through the
LangSmith Gateway, preserving fork conventions (Bedrock/Fireworks model IDs,
no-agent-attribution, bun toolchain).

- #1671 (e9dc6e01): opt-in gateway routing — new gateway.py, team-settings
  toggle, admin UI section, wired into make_model for all graph entrypoints
- #1673 (702ef908): dedicated LANGSMITH_GATEWAY_API_KEY precedence over
  platform LANGSMITH_API_KEY
- #1674 (5f7c2f46): fix Fireworks gateway base URL to /fireworks (bare host,
  SDK appends /v1/chat/completions) + SanitizeFireworksMessagesMiddleware
- #1678 (73b7d1c0): fix OpenAI Responses reasoning replay —
  SanitizeOpenAIResponsesMiddleware, store/include config for encrypted
  reasoning content, reasoning_effort coercion for Chat Completions fallback

Refs #134

* fix: downgrade gateway not-routed log to debug, add Bedrock UI note, add sanitizer parity

- Downgrade logger.warning to logger.debug in gateway_overrides for
  not-routed providers and missing API key (Bedrock is the default
  provider in this fork, so these are expected steady states)
- Add Bedrock to the LLMGatewaySection route-toggle description so
  admins know it is not routed through the gateway
- Add SanitizeOpenAIResponsesMiddleware to chat.py for parity with
  server.py and reviewer.py
- Restore the Bedrock region comment in model.py that explains the
  AWS_REGION / AWS_DEFAULT_REGION precedence

Refs #138

---------

Co-authored-by: amoussa1229 <166072409+amoussa1229@users.noreply.github.com>

CUSTOMIZATION.md

@@ -193,6 +193,25 @@ async def get_agent(config: RunnableConfig) -> Pregel:
     return create_deep_agent(model=model, ...)
 ```
 
+### Routing through the LangSmith LLM Gateway
+
+Model calls can be proxied through the [LangSmith LLM Gateway](https://docs.langchain.com/langsmith/llm-gateway) (private beta) instead of hitting providers directly. The gateway authenticates with a **LangSmith API key** that has the `gateway:invoke` permission and resolves the real provider key from workspace Provider Secrets, so no provider API keys are needed at runtime — and it adds central spend limits, PII/secrets redaction, and tracing. Your org must have the gateway enabled with Provider Secrets configured.
+
+Routing is opt-in and off by default. Enable it either way:
+
+| Env var | Default | Purpose |
+|---|---|---|
+| `LANGSMITH_GATEWAY_ENABLED` | `false` | Deployment-level default for gateway routing. |
+| `LANGSMITH_GATEWAY_API_KEY` | unset | Optional dedicated LangSmith key for Gateway calls. Prefer this in LangGraph Cloud if the platform-provided `LANGSMITH_API_KEY` lacks `gateway:invoke`. Falls back to `LANGSMITH_API_KEY_PROD`, then `LANGSMITH_API_KEY`. |
+| `LANGSMITH_GATEWAY_BASE_URL` | `https://gateway.smith.langchain.com` | Override for a regional or self-hosted gateway host. |
+| `LANGSMITH_GATEWAY_OPENAI_USE_RESPONSES` | `true` | Use the OpenAI Responses API through the gateway. Set to `false` only to force Chat Completions for OpenAI models. |
+
+The admin panel (**Admin → LLM Gateway**) exposes a per-workspace toggle stored in team settings; when set it overrides the `LANGSMITH_GATEWAY_ENABLED` env default (a `None`/unset team value inherits the env default).
+
+Routing is applied centrally in `make_model` (`agent/utils/model.py`), which resolves the effective on/off and delegates URL/key wiring to `agent/utils/gateway.py`. **OpenAI, Anthropic, Fireworks, and Google Gemini** are routed (their LangChain integrations accept `base_url` + `api_key`); Google Vertex (service-account auth) and any other provider call the provider directly with a logged warning.
+
+**Caveat — OpenAI endpoint:** open-swe uses the OpenAI Responses API by default because OpenAI reasoning models with function tools reject `reasoning_effort` on Chat Completions. Direct OpenAI calls use a `wss://` base URL; gateway-routed OpenAI uses the HTTPS gateway base URL with Responses enabled. Set `LANGSMITH_GATEWAY_OPENAI_USE_RESPONSES=false` only if you need to force Chat Completions. Anthropic and Fireworks are unaffected.
+
 ---
 
 ## 3. Tools

agent/analyzer.py

@@ -29,6 +29,7 @@ from deepagents.backends.protocol import SandboxBackendProtocol
 from deepagents.backends.state import StateBackend
 from langchain.agents.middleware import ModelCallLimitMiddleware
 
+from .dashboard.team_settings import get_effective_gateway_enabled
 from .integrations.langsmith import _configure_github_proxy
 from .middleware import SanitizeToolInputsMiddleware, ToolErrorMiddleware
 from .review_style_guidance import REVIEWER_STYLE_THEMES
@@ -116,6 +117,7 @@ async def get_analyzer(config: RunnableConfig) -> Pregel:
     backend = CompositeBackend(default=sandbox_backend, routes={SKILLS_ROUTE: StateBackend()})
 
     model_id = DEFAULT_LLM_MODEL_ID
+    use_gateway = await get_effective_gateway_enabled()
     model_kwargs = provider_model_kwargs(
         model_id,
         None,
@@ -135,7 +137,7 @@ async def get_analyzer(config: RunnableConfig) -> Pregel:
     system_prompt = f"{system_prompt}\n\n{user_context}"
 
     return create_deep_agent(
-        model=make_model(model_id, **model_kwargs),
+        model=make_model(model_id, use_gateway=use_gateway, **model_kwargs),
         system_prompt=system_prompt,
         tools=[save_review_style_prompt, read_finding_outcomes],
         backend=backend,

agent/chat.py

@@ -29,9 +29,11 @@ from deepagents import create_deep_agent
 from langchain.agents.middleware import ModelCallLimitMiddleware
 
 from .dashboard.options import SUPPORTED_MODEL_IDS, model_supports_effort
-from .dashboard.team_settings import get_team_default_model
+from .dashboard.team_settings import get_effective_gateway_enabled, get_team_default_model
 from .middleware import (
     ExcludeToolsMiddleware,
+    SanitizeFireworksMessagesMiddleware,
+    SanitizeOpenAIResponsesMiddleware,
     SanitizeThinkingBlocksMiddleware,
     SanitizeToolInputsMiddleware,
     ToolErrorMiddleware,
@@ -123,6 +125,7 @@ async def get_chat_agent(config: RunnableConfig) -> Pregel:
         configurable["chat_github_token"] = token
 
     model_id, effort = await _resolve_chat_model(configurable)
+    use_gateway = await get_effective_gateway_enabled()
     model_kwargs = provider_model_kwargs(
         model_id,
         effort,
@@ -137,7 +140,7 @@ async def get_chat_agent(config: RunnableConfig) -> Pregel:
     )
 
     return create_deep_agent(
-        model=make_model(model_id, **model_kwargs),
+        model=make_model(model_id, use_gateway=use_gateway, **model_kwargs),
         system_prompt=system_prompt,
         tools=[
             read_repo_file,
@@ -151,6 +154,8 @@ async def get_chat_agent(config: RunnableConfig) -> Pregel:
             ModelCallLimitMiddleware(run_limit=CHAT_MODEL_CALL_LIMIT, exit_behavior="end"),
             ToolErrorMiddleware(),
             ExcludeToolsMiddleware(excluded=_EXCLUDED_TOOLS),
+            SanitizeOpenAIResponsesMiddleware(),
+            SanitizeFireworksMessagesMiddleware(),
             SanitizeThinkingBlocksMiddleware(),
         ],
     ).with_config(config)

agent/dashboard/team_settings.py

@@ -15,6 +15,7 @@ from typing import Any, Literal
 from langgraph_sdk import get_client
 from pydantic import BaseModel, field_validator, model_validator
 
+from ..utils.gateway import resolve_gateway_enabled
 from .options import (
     SUPPORTED_MODEL_IDS,
     default_model_pair,
@@ -52,6 +53,9 @@ class TeamSettingsUpdate(BaseModel):
     review_draft_prs: bool = False
     pr_summaries: bool = True
     review_trace_links: bool = True
+    # Tri-state LLM Gateway toggle: True/False is authoritative, None inherits the
+    # LANGSMITH_GATEWAY_ENABLED deployment default.
+    gateway_enabled: bool | None = None
     review_tracing_project: str | None = None
     org_guidelines: str | None = None
     default_agent_model: str | None = None
@@ -166,6 +170,7 @@ def _default_settings() -> dict[str, Any]:
         "review_draft_prs": False,
         "pr_summaries": True,
         "review_trace_links": True,
+        "gateway_enabled": None,
         "review_tracing_project": None,
         "org_guidelines": DEFAULT_ORG_REVIEW_GUIDELINES,
         "default_agent_model": fallback_model,
@@ -220,6 +225,7 @@ async def upsert_team_settings(update: TeamSettingsUpdate) -> dict[str, Any]:
         "review_draft_prs": update.review_draft_prs,
         "pr_summaries": update.pr_summaries,
         "review_trace_links": update.review_trace_links,
+        "gateway_enabled": update.gateway_enabled,
         "review_tracing_project": update.review_tracing_project,
         "org_guidelines": update.org_guidelines,
         "default_agent_model": update.default_agent_model,
@@ -340,6 +346,18 @@ async def get_team_review_trace_links_enabled() -> bool:
     return bool(settings.get("review_trace_links", True))
 
 
+async def get_team_gateway_enabled() -> bool | None:
+    """Return the stored LLM Gateway toggle (``None`` means inherit the env default)."""
+    settings = await get_team_settings()
+    value = settings.get("gateway_enabled")
+    return value if isinstance(value, bool) else None
+
+
+async def get_effective_gateway_enabled() -> bool:
+    """Resolve whether LLM Gateway routing is on: team setting, else env default."""
+    return resolve_gateway_enabled(await get_team_gateway_enabled())
+
+
 async def get_team_review_tracing_project() -> str | None:
     """Return the LangSmith tracing project used for PR trace resolution."""
     settings = await get_team_settings()

agent/middleware/__init__.py

@@ -8,6 +8,8 @@ from .refresh_github_proxy import refresh_github_proxy_before_model
 from .refresh_slack_status import SlackAssistantStatusMiddleware
 from .repair_orphaned_tool_calls import RepairOrphanedToolCallsMiddleware
 from .sandbox_circuit_breaker import SandboxCircuitBreakerMiddleware
+from .sanitize_fireworks_messages import SanitizeFireworksMessagesMiddleware
+from .sanitize_openai_responses import SanitizeOpenAIResponsesMiddleware
 from .sanitize_thinking_blocks import SanitizeThinkingBlocksMiddleware
 from .sanitize_tool_inputs import SanitizeToolInputsMiddleware
 from .settle_review_check import settle_review_check_on_exit
@@ -21,6 +23,8 @@ __all__ = [
     "ModelFallbackMiddleware",
     "PlanModeMiddleware",
     "RepairOrphanedToolCallsMiddleware",
+    "SanitizeFireworksMessagesMiddleware",
+    "SanitizeOpenAIResponsesMiddleware",
     "SanitizeThinkingBlocksMiddleware",
     "SanitizeToolInputsMiddleware",
     "SubdirAgentsReadMiddleware",

agent/middleware/sanitize_fireworks_messages.py

@@ -0,0 +1,79 @@
+"""Strip legacy ``function_call`` fields before Fireworks provider calls.
+
+The LangSmith LLM Gateway enforces strict request-schema validation and rejects
+messages carrying the legacy OpenAI ``function_call`` field
+(``Extra inputs are not permitted, field: 'messages[N].function_call'``).
+``langchain_fireworks`` forwards ``function_call`` from
+``AIMessage.additional_kwargs`` whenever it is present — even when the message
+also carries modern ``tool_calls`` — so a message that was originally produced
+by (or deserialized from) a provider that populated the legacy field will break
+every subsequent Fireworks turn once routed through the gateway.
+
+This middleware drops the redundant legacy field from assistant messages before
+they reach the Fireworks serializer, mirroring the thinking-block sanitizer.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Awaitable, Callable
+from typing import Any
+
+from langchain.agents.middleware import AgentMiddleware
+from langchain.agents.middleware.types import ModelCallResult, ModelRequest, ModelResponse
+from langchain_core.messages import AIMessage
+
+try:
+    from langchain_fireworks.chat_models import ChatFireworks
+except ImportError:  # pragma: no cover
+    ChatFireworks = None  # type: ignore[assignment, misc]
+
+
+def _is_chat_fireworks(model: object) -> bool:
+    if ChatFireworks is None:
+        return False
+    seen: set[int] = set()
+    current = model
+    for _ in range(10):
+        if isinstance(current, ChatFireworks):
+            return True
+        current_id = id(current)
+        if current_id in seen:
+            return False
+        seen.add(current_id)
+        bound = getattr(current, "bound", None)
+        if bound is None or bound is current:
+            return False
+        current = bound
+    return False
+
+
+def _sanitize_messages(messages: list[Any]) -> None:
+    for message in messages:
+        if not isinstance(message, AIMessage):
+            continue
+        additional_kwargs = message.additional_kwargs
+        if not isinstance(additional_kwargs, dict) or "function_call" not in additional_kwargs:
+            continue
+        additional_kwargs.pop("function_call", None)
+
+
+class SanitizeFireworksMessagesMiddleware(AgentMiddleware):
+    """Drop legacy ``function_call`` fields before Fireworks provider calls."""
+
+    def wrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], ModelResponse],
+    ) -> ModelCallResult:
+        if _is_chat_fireworks(request.model):
+            _sanitize_messages(request.messages)
+        return handler(request)
+
+    async def awrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
+    ) -> Any:
+        if _is_chat_fireworks(request.model):
+            _sanitize_messages(request.messages)
+        return await handler(request)

agent/middleware/sanitize_openai_responses.py

@@ -0,0 +1,85 @@
+"""Middleware that removes stale OpenAI Responses reasoning references."""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Awaitable, Callable
+from typing import Any
+
+from langchain.agents.middleware import AgentMiddleware
+from langchain.agents.middleware.types import ModelCallResult, ModelRequest, ModelResponse
+from langchain_core.messages import AIMessage
+
+try:
+    from langchain_openai import ChatOpenAI
+except ImportError:  # pragma: no cover
+    ChatOpenAI = None  # type: ignore[assignment, misc]
+
+logger = logging.getLogger(__name__)
+
+
+def _is_chat_openai(model: object) -> bool:
+    if ChatOpenAI is None:
+        return False
+    seen: set[int] = set()
+    current = model
+    for _ in range(10):
+        if isinstance(current, ChatOpenAI):
+            return True
+        current_id = id(current)
+        if current_id in seen:
+            return False
+        seen.add(current_id)
+        bound = getattr(current, "bound", None)
+        if bound is None or bound is current:
+            return False
+        current = bound
+    return False
+
+
+def _is_stale_reasoning_reference(block: dict[str, Any]) -> bool:
+    if block.get("type") != "reasoning":
+        return False
+    if block.get("encrypted_content"):
+        return False
+    block_id = block.get("id")
+    return isinstance(block_id, str) and block_id.startswith("rs_")
+
+
+def _sanitize_messages(messages: list[Any]) -> None:
+    removed = 0
+    for message in messages:
+        if not isinstance(message, AIMessage) or not isinstance(message.content, list):
+            continue
+        content = []
+        for block in message.content:
+            if isinstance(block, dict) and _is_stale_reasoning_reference(block):
+                removed += 1
+                continue
+            content.append(block)
+        if len(content) != len(message.content):
+            message.content = content
+    if removed:
+        logger.warning("Removed %d stale OpenAI Responses reasoning reference(s)", removed)
+
+
+class SanitizeOpenAIResponsesMiddleware(AgentMiddleware):
+    """Drop non-replayable OpenAI Responses reasoning item references."""
+
+    def wrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], ModelResponse],
+    ) -> ModelCallResult:
+        if _is_chat_openai(request.model):
+            _sanitize_messages(request.messages)
+        return handler(request)
+
+    async def awrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
+    ) -> Any:
+        if _is_chat_openai(request.model):
+            _sanitize_messages(request.messages)
+        return await handler(request)

agent/reviewer.py

@@ -34,12 +34,15 @@ from langchain.agents.middleware import ModelCallLimitMiddleware
 from langchain_core.language_models.chat_models import BaseChatModel
 
 from .dashboard.team_settings import (
+    get_effective_gateway_enabled,
     get_org_review_guidelines,
     get_team_default_grouping_model,
     get_team_default_model_pair,
 )
 from .middleware import (
     RepairOrphanedToolCallsMiddleware,
+    SanitizeFireworksMessagesMiddleware,
+    SanitizeOpenAIResponsesMiddleware,
     SanitizeThinkingBlocksMiddleware,
     SanitizeToolInputsMiddleware,
     SlackAssistantStatusMiddleware,
@@ -801,7 +804,9 @@ def _on_background_task_done(task: asyncio.Task[None]) -> None:
         logger.warning("Background reviewer task failed: %s", exc)
 
 
-async def _resolve_grouping_model(configurable: dict[str, object]) -> BaseChatModel:
+async def _resolve_grouping_model(
+    configurable: dict[str, object], *, use_gateway: bool
+) -> BaseChatModel:
     """Resolve the model for the diff-grouping pass.
 
     Per-run override (``grouping_model_id``/``grouping_reasoning_effort``) wins;
@@ -821,7 +826,7 @@ async def _resolve_grouping_model(configurable: dict[str, object]) -> BaseChatMo
         max_tokens=DEFAULT_LLM_MAX_TOKENS,
         openai_reasoning_default=DEFAULT_LLM_REASONING,
     )
-    return make_model(model_id, **model_kwargs)
+    return make_model(model_id, use_gateway=use_gateway, **model_kwargs)
 
 
 async def get_reviewer_agent(config: RunnableConfig) -> Pregel:
@@ -1156,8 +1161,11 @@ async def get_reviewer_agent(config: RunnableConfig) -> Pregel:
     if review_context:
         system_prompt = f"{system_prompt}\n\n{review_context}"
 
-    reviewer_model = make_model(model_id, **model_kwargs)
+    use_gateway = await get_effective_gateway_enabled()
-    reviewer_subagent_model = make_model(subagent_model_id, **subagent_model_kwargs)
+    reviewer_model = make_model(model_id, use_gateway=use_gateway, **model_kwargs)
+    reviewer_subagent_model = make_model(
+        subagent_model_id, use_gateway=use_gateway, **subagent_model_kwargs
+    )
 
     # Kick off the AI-sorted diff grouping pass at run start, concurrently with
     # the review, so it adds ~0 latency. First-review and re-review only — a
@@ -1165,7 +1173,9 @@ async def get_reviewer_agent(config: RunnableConfig) -> Pregel:
     # its own errors and the UI falls back to the folder view when groups are
     # absent.
     if reviewer_event != "finding_reply" and pr_diff_text and thread_id:
-        grouping_model = await _resolve_grouping_model(config["configurable"])
+        grouping_model = await _resolve_grouping_model(
+            config["configurable"], use_gateway=use_gateway
+        )
         grouping_task = asyncio.create_task(
             maybe_generate_and_store_diff_groups(
                 thread_id=thread_id,
@@ -1201,6 +1211,8 @@ async def get_reviewer_agent(config: RunnableConfig) -> Pregel:
             refresh_github_proxy_before_model,
             check_message_queue_before_model,
             SlackAssistantStatusMiddleware(),
+            SanitizeOpenAIResponsesMiddleware(),
+            SanitizeFireworksMessagesMiddleware(),
             SanitizeThinkingBlocksMiddleware(),
             RepairOrphanedToolCallsMiddleware(),
             settle_review_check_on_exit,

agent/server.py

@@ -44,6 +44,7 @@ from .dashboard.agent_usage import record_agent_thread_usage
 from .dashboard.options import DEFAULT_MODEL_ID, SUPPORTED_MODEL_IDS, model_supports_effort
 from .dashboard.repo_snapshots import resolve_repo_snapshot_id
 from .dashboard.team_settings import (
+    get_effective_gateway_enabled,
     get_team_default_model_pair,
     get_team_default_repo,
 )
@@ -58,6 +59,8 @@ from .middleware import (
     ModelFallbackMiddleware,
     PlanModeMiddleware,
     SandboxCircuitBreakerMiddleware,
+    SanitizeFireworksMessagesMiddleware,
+    SanitizeOpenAIResponsesMiddleware,
     SanitizeThinkingBlocksMiddleware,
     SanitizeToolInputsMiddleware,
     SlackAssistantStatusMiddleware,
@@ -758,19 +761,31 @@ async def get_agent(config: RunnableConfig) -> Pregel:
         ensure_sandbox_for_thread(thread_id, repo=prompt_default_repo)
     )
     team_defaults_task = asyncio.create_task(get_team_default_model_pair("agent"))
+    gateway_task = asyncio.create_task(get_effective_gateway_enabled())
     profile_task = asyncio.create_task(load_profile(profile_login)) if profile_login else None
     try:
-        triggering_user_identity, sandbox_backend, team_defaults = await asyncio.gather(
+        (
+            triggering_user_identity,
+            sandbox_backend,
+            team_defaults,
+            use_gateway,
+        ) = await asyncio.gather(
             triggering_user_identity_task,
             sandbox_task,
             team_defaults_task,
+            gateway_task,
         )
     except SandboxRepoMismatchError as exc:
         # Repo-binding refusal at the run boundary: log for alarming and surface the
         # already-sanitized terminal error (no sandbox/token internals) to the caller,
         # rather than letting an opaque deep-stack exception crash-loop the worker.
         logger.error("Refusing agent run for thread %s: %s", thread_id, exc)
-        for pending in (triggering_user_identity_task, team_defaults_task, profile_task):
+        for pending in (
+            triggering_user_identity_task,
+            team_defaults_task,
+            gateway_task,
+            profile_task,
+        ):
             if pending is not None and not pending.done():
                 pending.cancel()
         raise RuntimeError(str(exc)) from exc
@@ -860,7 +875,9 @@ async def get_agent(config: RunnableConfig) -> Pregel:
             fallback_model_id, None, max_tokens=DEFAULT_LLM_MAX_TOKENS
         )
         fallback_middleware.append(
-            ModelFallbackMiddleware(make_model(fallback_model_id, **fallback_kwargs))
+            ModelFallbackMiddleware(
+                make_model(fallback_model_id, use_gateway=use_gateway, **fallback_kwargs)
+            )
         )
         logger.info("Configured model fallback %s -> %s", model_id, fallback_model_id)
 
@@ -926,8 +943,8 @@ async def get_agent(config: RunnableConfig) -> Pregel:
             notion_tools = []
 
     logger.info("Returning agent with sandbox for thread %s", thread_id)
-    main_model = make_model(model_id, **model_kwargs)
+    main_model = make_model(model_id, use_gateway=use_gateway, **model_kwargs)
-    subagent_model = make_model(subagent_model_id, **subagent_model_kwargs)
+    subagent_model = make_model(subagent_model_id, use_gateway=use_gateway, **subagent_model_kwargs)
     return create_deep_agent(
         model=main_model,
         system_prompt=construct_system_prompt(
@@ -987,6 +1004,8 @@ async def get_agent(config: RunnableConfig) -> Pregel:
             SandboxCircuitBreakerMiddleware(),
             *fallback_middleware,
             *plan_mode_middleware,
+            SanitizeOpenAIResponsesMiddleware(),
+            SanitizeFireworksMessagesMiddleware(),
             SanitizeThinkingBlocksMiddleware(),
         ],
     ).with_config(config)

agent/utils/gateway.py

@@ -0,0 +1,126 @@
+"""LangSmith LLM Gateway routing for model construction.
+
+The LLM Gateway (https://docs.langchain.com/langsmith/llm-gateway) proxies
+provider calls through LangSmith: the client authenticates with a LangSmith API
+key and the gateway resolves the real provider key from workspace Provider
+Secrets, enforcing spend/PII/secrets policies and tracing every call. Routing is
+opt-in via ``LANGSMITH_GATEWAY_ENABLED`` (deployment default) or the
+``gateway_enabled`` team setting, and is applied centrally in
+:func:`agent.utils.model.make_model`.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_GATEWAY_BASE_URL = "https://gateway.smith.langchain.com"
+
+# Provider prefix -> base-URL suffix appended to the gateway host. Each suffix
+# matches the SDK's own path handling: the OpenAI SDK appends
+# ``/chat/completions`` to a ``/v1`` base, Fireworks appends
+# ``/v1/chat/completions`` to a bare provider host, Anthropic appends
+# ``/v1/messages`` to a bare host, and google-genai appends
+# ``/<api_version>/models/...`` to a bare host. Vertex (``google_vertexai``, which
+# uses service-account auth rather than a bearer key) and any other provider are
+# not routed and call the provider directly.
+_GATEWAY_PROVIDER_PATHS: dict[str, str] = {
+    "openai": "/openai/v1",
+    "anthropic": "/anthropic",
+    "fireworks": "/fireworks",
+    "google_genai": "/gemini",
+}
+
+
+def _env_bool(value: str | None) -> bool:
+    return (value or "").strip().lower() in {"1", "true", "yes", "on"}
+
+
+def _langsmith_api_key() -> str | None:
+    """LangSmith API key used to authenticate gateway calls.
+
+    Prefer a gateway-specific key, then the prod LangSmith key. LangGraph Cloud
+    may inject ``LANGSMITH_API_KEY`` for tracing/platform APIs, and that key can
+    lack the ``gateway:invoke`` permission required by the LLM Gateway.
+    """
+    return (
+        os.environ.get("LANGSMITH_GATEWAY_API_KEY")
+        or os.environ.get("LANGSMITH_API_KEY_PROD")
+        or os.environ.get("LANGSMITH_API_KEY")
+    )
+
+
+def gateway_base_url() -> str:
+    """Gateway host, overridable via ``LANGSMITH_GATEWAY_BASE_URL`` (regional/self-hosted)."""
+    return (os.environ.get("LANGSMITH_GATEWAY_BASE_URL") or DEFAULT_GATEWAY_BASE_URL).rstrip("/")
+
+
+def gateway_env_default() -> bool:
+    """Deployment-level default for gateway routing (``LANGSMITH_GATEWAY_ENABLED``)."""
+    return _env_bool(os.environ.get("LANGSMITH_GATEWAY_ENABLED"))
+
+
+def gateway_openai_use_responses() -> bool:
+    """Whether gateway-routed OpenAI keeps the Responses API.
+
+    Defaults to ``True`` because OpenAI reasoning models with tool calls reject
+    ``reasoning_effort`` on Chat Completions. Set
+    ``LANGSMITH_GATEWAY_OPENAI_USE_RESPONSES=false`` only for deployments that
+    need to force Chat Completions through the gateway.
+    """
+    raw = os.environ.get("LANGSMITH_GATEWAY_OPENAI_USE_RESPONSES")
+    if raw is None:
+        return True
+    return _env_bool(raw)
+
+
+def resolve_gateway_enabled(team_value: bool | None) -> bool:
+    """Combine the team-settings toggle with the env default.
+
+    A team value of ``True``/``False`` is authoritative; ``None`` inherits the
+    ``LANGSMITH_GATEWAY_ENABLED`` deployment default.
+    """
+    if team_value is None:
+        return gateway_env_default()
+    return team_value
+
+
+def _provider_of(model_id: str) -> str:
+    return model_id.split(":", 1)[0]
+
+
+def gateway_overrides(model_id: str) -> dict[str, object] | None:
+    """``init_chat_model`` kwargs that route ``model_id`` through the gateway.
+
+    Returns ``None`` (so the caller keeps talking to the provider directly) when
+    the provider isn't routable through the gateway or no LangSmith API key is
+    available — both cases are logged rather than raised, so a run never fails
+    just because gateway routing couldn't be applied.
+    """
+    provider = _provider_of(model_id)
+    path = _GATEWAY_PROVIDER_PATHS.get(provider)
+    if path is None:
+        logger.debug(
+            "LangSmith gateway enabled but provider %r is not routed; calling it directly",
+            provider,
+        )
+        return None
+    api_key = _langsmith_api_key()
+    if not api_key:
+        logger.debug(
+            "LangSmith gateway enabled but no LANGSMITH_GATEWAY_API_KEY or "
+            "LANGSMITH_API_KEY(_PROD) is set; "
+            "calling the provider directly"
+        )
+        return None
+    overrides: dict[str, object] = {
+        "base_url": f"{gateway_base_url()}{path}",
+        "api_key": api_key,
+    }
+    if provider == "openai":
+        # Use HTTPS Responses through the gateway by default; tool-calling OpenAI
+        # reasoning models reject reasoning_effort on Chat Completions.
+        overrides["use_responses_api"] = gateway_openai_use_responses()
+    return overrides

agent/utils/model.py

@@ -4,6 +4,7 @@ from typing import Literal, TypedDict, Unpack
 from langchain.chat_models import init_chat_model
 
 from ..dashboard.options import DEFAULT_MODEL_ID
+from .gateway import gateway_env_default, gateway_overrides
 
 OPENAI_RESPONSES_WS_BASE_URL = "wss://api.openai.com/v1"
 
@@ -39,11 +40,14 @@ class AnthropicThinking(TypedDict, total=False):
 class ModelKwargs(TypedDict, total=False):
     max_tokens: int | None
     reasoning: OpenAIReasoning | None
+    reasoning_effort: OpenAIReasoningEffort | None
     thinking: AnthropicThinking | None
     effort: AnthropicEffort | None
     thinking_level: GoogleThinkingLevel | None
     temperature: float | None
     max_retries: int | None
+    store: bool | None
+    include: list[str] | None
     model_kwargs: dict[str, object] | None
     additional_model_request_fields: dict[str, object] | None
     region_name: str | None
@@ -52,21 +56,62 @@ class ModelKwargs(TypedDict, total=False):
 _ANTHROPIC_EFFORTS: set[AnthropicEffort] = {"low", "medium", "high", "xhigh", "max"}
 
 
-def make_model(model_id: str, **kwargs: Unpack[ModelKwargs]):
+def _coerce_openai_chat_completions_kwargs(model_kwargs: dict[str, object]) -> None:
+    if model_kwargs.get("use_responses_api") is not False:
+        return
+    reasoning = model_kwargs.pop("reasoning", None)
+    if isinstance(reasoning, dict):
+        effort = reasoning.get("effort")
+        if isinstance(effort, str):
+            model_kwargs.setdefault("reasoning_effort", effort)
+
+
+def _configure_openai_responses_kwargs(model_kwargs: dict[str, object]) -> None:
+    if model_kwargs.get("use_responses_api") is False:
+        return
+    model_kwargs.setdefault("store", False)
+    include = model_kwargs.get("include")
+    if include is None:
+        model_kwargs["include"] = ["reasoning.encrypted_content"]
+    elif isinstance(include, list) and "reasoning.encrypted_content" not in include:
+        include.append("reasoning.encrypted_content")
+
+
+def make_model(model_id: str, *, use_gateway: bool | None = None, **kwargs: Unpack[ModelKwargs]):
+    """Build a chat model, optionally routed through the LangSmith LLM Gateway.
+
+    ``use_gateway`` resolves the deployment default (``LANGSMITH_GATEWAY_ENABLED``)
+    when ``None``; async callers pass the team-settings-resolved value. When on,
+    gateway ``base_url``/``api_key``/``use_responses_api`` override the direct
+    provider defaults below (see :mod:`agent.utils.gateway`).
+    """
     model_kwargs: dict[str, object] = kwargs.copy()
     model_kwargs.setdefault("max_retries", DEFAULT_MAX_RETRIES)
 
     if model_id.startswith("openai:"):
+        # Direct-provider default: Responses API over the OpenAI websocket base.
+        # Gateway routing overrides this below (an HTTP(S) proxy can't carry wss).
         model_kwargs["base_url"] = OPENAI_RESPONSES_WS_BASE_URL
         model_kwargs["use_responses_api"] = True
     elif model_id.startswith("bedrock_converse:"):
-        # Resolve region with the same precedence validate_local_dev_llm_config accepts
+        # Resolve region with the same precedence validate_local_dev_llm_config
-        # (AWS_REGION or AWS_DEFAULT_REGION), so the validated value is the one actually used.
+        # accepts (AWS_REGION or AWS_DEFAULT_REGION), so the validated value
+        # is the one actually used.
         model_kwargs.setdefault(
             "region_name",
             os.environ.get("AWS_REGION") or os.environ.get("AWS_DEFAULT_REGION") or "us-east-1",
         )
 
+    enabled = gateway_env_default() if use_gateway is None else use_gateway
+    if enabled:
+        overrides = gateway_overrides(model_id)
+        if overrides is not None:
+            model_kwargs.update(overrides)
+
+    if model_id.startswith("openai:"):
+        _configure_openai_responses_kwargs(model_kwargs)
+        _coerce_openai_chat_completions_kwargs(model_kwargs)
+
     return init_chat_model(model=model_id, **model_kwargs)
 
 

docs/upstream-sync/triage.jsonl

@@ -32,10 +32,10 @@
 {"sha": "4cd5fa5c", "pr": 1629, "subject": "avoid recapping Slack replies", "disposition": "deferred", "reason": "", "branch": "slack-tooling", "local_sha": null, "updated": "2026-07-02T00:00:00Z"}
 {"sha": "92dbf6f9", "pr": 1630, "subject": "update Slack trace reply on web handoff", "disposition": "landed", "reason": "ported in #128 (trace_message_ts on first-mention run mapping)", "branch": "slack-tooling", "local_sha": null, "updated": "2026-07-08T22:58:21Z"}
 {"sha": "bb36448b", "pr": 1627, "subject": "surface Slack thread errors", "disposition": "landed", "reason": "", "branch": "slack-tooling", "local_sha": "e8b6fb70508861e5742d31aaa722e79f4a9a320a", "updated": "2026-07-03T19:50:30Z"}
-{"sha": "73b7d1c0", "pr": 1678, "subject": "fix OpenAI Responses reasoning replay", "disposition": "deferred", "reason": "", "branch": "gateway-routing", "local_sha": null, "updated": "2026-07-02T00:00:00Z"}
+{"sha": "73b7d1c0", "pr": 1678, "subject": "fix OpenAI Responses reasoning replay", "disposition": "landed", "reason": "", "branch": "gateway-routing", "local_sha": null, "updated": "2026-07-09T16:22:05Z"}
-{"sha": "5f7c2f46", "pr": 1674, "subject": "fix Fireworks Gateway base URL", "disposition": "deferred", "reason": "", "branch": "gateway-routing", "local_sha": null, "updated": "2026-07-02T00:00:00Z"}
+{"sha": "5f7c2f46", "pr": 1674, "subject": "fix Fireworks Gateway base URL", "disposition": "landed", "reason": "", "branch": "gateway-routing", "local_sha": null, "updated": "2026-07-09T16:22:05Z"}
-{"sha": "702ef908", "pr": 1673, "subject": "dedicated LangSmith gateway API key", "disposition": "deferred", "reason": "", "branch": "gateway-routing", "local_sha": null, "updated": "2026-07-02T00:00:00Z"}
+{"sha": "702ef908", "pr": 1673, "subject": "dedicated LangSmith gateway API key", "disposition": "landed", "reason": "", "branch": "gateway-routing", "local_sha": null, "updated": "2026-07-09T16:22:05Z"}
-{"sha": "e9dc6e01", "pr": 1671, "subject": "opt-in LangSmith LLM Gateway routing", "disposition": "deferred", "reason": "", "branch": "gateway-routing", "local_sha": null, "updated": "2026-07-02T00:00:00Z"}
+{"sha": "e9dc6e01", "pr": 1671, "subject": "opt-in LangSmith LLM Gateway routing", "disposition": "landed", "reason": "", "branch": "gateway-routing", "local_sha": null, "updated": "2026-07-09T16:22:04Z"}
 {"sha": "289f5e3a", "pr": 1651, "subject": "add Sonnet 5 to model picker", "disposition": "landed", "reason": "already in dev; added Bedrock family fallback fix (c16fb915)", "branch": "gateway-routing", "local_sha": null, "updated": "2026-07-03T00:19:50Z"}
 {"sha": "5da3d0c6", "pr": 1624, "subject": "post reviewer resolution notes verbatim", "disposition": "landed", "reason": "cherry-picked (-x) in #127", "branch": "reviewer-misc", "local_sha": null, "updated": "2026-07-08T22:58:20Z"}
 {"sha": "69148f54", "pr": 1612, "subject": "add PR trace resolution", "disposition": "landed", "reason": "cherry-picked (-x) in #127", "branch": "reviewer-misc", "local_sha": null, "updated": "2026-07-08T22:58:20Z"}

docs/upstream-sync/triage.md

@@ -29,6 +29,10 @@ Rows key on the **upstream SHA** (stable across local cherry-picks). Deferred ro
 | `27d90ef1` | #1633 | include Slack channel context in prompts | Landed |  | slack-tooling |
 | `92dbf6f9` | #1630 | update Slack trace reply on web handoff | Landed | ported in #128 (trace_message_ts on first-mention run mapping) | slack-tooling |
 | `bb36448b` | #1627 | surface Slack thread errors | Landed |  | slack-tooling |
+| `73b7d1c0` | #1678 | fix OpenAI Responses reasoning replay | Landed |  | gateway-routing |
+| `5f7c2f46` | #1674 | fix Fireworks Gateway base URL | Landed |  | gateway-routing |
+| `702ef908` | #1673 | dedicated LangSmith gateway API key | Landed |  | gateway-routing |
+| `e9dc6e01` | #1671 | opt-in LangSmith LLM Gateway routing | Landed |  | gateway-routing |
 | `289f5e3a` | #1651 | add Sonnet 5 to model picker | Landed | already in dev; added Bedrock family fallback fix (c16fb915) | gateway-routing |
 | `5da3d0c6` | #1624 | post reviewer resolution notes verbatim | Landed | cherry-picked (-x) in #127 | reviewer-misc |
 | `69148f54` | #1612 | add PR trace resolution | Landed | cherry-picked (-x) in #127 | reviewer-misc |
@@ -63,10 +67,6 @@ Rows key on the **upstream SHA** (stable across local cherry-picks). Deferred ro
 | `546042a4` | #1652 | add workflow approval UI | Deferred |  | plan-approval |
 | `c03a6be7` | #1634 | keep plan guidance high-level | Deferred |  | plan-approval |
 | `4cd5fa5c` | #1629 | avoid recapping Slack replies | Deferred |  | slack-tooling |
-| `73b7d1c0` | #1678 | fix OpenAI Responses reasoning replay | Deferred |  | gateway-routing |
-| `5f7c2f46` | #1674 | fix Fireworks Gateway base URL | Deferred |  | gateway-routing |
-| `702ef908` | #1673 | dedicated LangSmith gateway API key | Deferred |  | gateway-routing |
-| `e9dc6e01` | #1671 | opt-in LangSmith LLM Gateway routing | Deferred |  | gateway-routing |
 | `baf0c248` | #1617 | filter & grouping menu in threads sidebar | Deferred | ~998 LOC | own branch |
 | `f29868ff` | #1615 | recover thread work as patch | Deferred | ~495 LOC | own branch |
 | `8e0788dc` | #1631 | show queued dashboard follow-ups | Deferred |  | own branch |

tests/test_gateway.py

@@ -0,0 +1,437 @@
+"""Unit tests for LangSmith LLM Gateway routing (agent/utils/gateway.py + make_model)."""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import patch
+
+import httpx
+import pytest
+from fireworks import AsyncFireworks
+from langchain_core.messages import AIMessage, HumanMessage
+from langchain_openai import ChatOpenAI
+
+from agent.utils import gateway, model
+
+_GATEWAY_ENV_VARS = (
+    "LANGSMITH_API_KEY",
+    "LANGSMITH_API_KEY_PROD",
+    "LANGSMITH_GATEWAY_API_KEY",
+    "LANGSMITH_GATEWAY_ENABLED",
+    "LANGSMITH_GATEWAY_BASE_URL",
+    "LANGSMITH_GATEWAY_OPENAI_USE_RESPONSES",
+)
+
+
+@pytest.fixture(autouse=True)
+def _clean_gateway_env(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Start each test from a known env: no key, gateway off, default base URL."""
+    for name in _GATEWAY_ENV_VARS:
+        monkeypatch.delenv(name, raising=False)
+
+
+# --- gateway_overrides --------------------------------------------------------
+
+
+def test_openai_overrides_use_responses_by_default(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    overrides = gateway.gateway_overrides("openai:gpt-5.5")
+    assert overrides == {
+        "base_url": "https://gateway.smith.langchain.com/openai/v1",
+        "api_key": "ls-key",
+        "use_responses_api": True,
+    }
+
+
+def test_openai_overrides_chat_completions_optout(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    monkeypatch.setenv("LANGSMITH_GATEWAY_OPENAI_USE_RESPONSES", "false")
+    overrides = gateway.gateway_overrides("openai:gpt-5.5")
+    assert overrides is not None
+    assert overrides["use_responses_api"] is False
+
+
+async def test_openai_sdk_uses_gateway_responses_path() -> None:
+    requests: list[httpx.Request] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        requests.append(request)
+        return httpx.Response(
+            200,
+            json={
+                "id": "resp_test",
+                "object": "response",
+                "created_at": 0,
+                "status": "completed",
+                "model": "gpt-5.5",
+                "output": [
+                    {
+                        "id": "msg_test",
+                        "type": "message",
+                        "role": "assistant",
+                        "status": "completed",
+                        "content": [{"type": "output_text", "text": "ok", "annotations": []}],
+                    }
+                ],
+                "usage": {"input_tokens": 1, "output_tokens": 1, "total_tokens": 2},
+            },
+        )
+
+    http_client = httpx.AsyncClient(transport=httpx.MockTransport(handler))
+    try:
+        chat_model = ChatOpenAI(
+            model="gpt-5.5",
+            api_key="dummy",
+            base_url="https://gateway.smith.langchain.com/openai/v1",
+            use_responses_api=True,
+            http_async_client=http_client,
+            max_retries=0,
+        )
+        await chat_model.ainvoke([HumanMessage(content="hi")])
+    finally:
+        await http_client.aclose()
+
+    assert len(requests) == 1
+    assert requests[0].url.path == "/openai/v1/responses"
+
+
+def test_anthropic_overrides_have_no_responses_flag(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    overrides = gateway.gateway_overrides("anthropic:claude-opus-4-8")
+    assert overrides == {
+        "base_url": "https://gateway.smith.langchain.com/anthropic",
+        "api_key": "ls-key",
+    }
+
+
+def test_fireworks_overrides(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    overrides = gateway.gateway_overrides("fireworks:accounts/fireworks/models/glm-5p2")
+    assert overrides is not None
+    assert overrides["base_url"] == "https://gateway.smith.langchain.com/fireworks"
+
+
+async def test_fireworks_sdk_uses_allowlisted_gateway_path() -> None:
+    requests: list[httpx.Request] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        requests.append(request)
+        return httpx.Response(
+            200,
+            json={
+                "id": "chatcmpl-test",
+                "object": "chat.completion",
+                "created": 0,
+                "model": "accounts/fireworks/models/glm-5p2",
+                "choices": [
+                    {
+                        "index": 0,
+                        "message": {"role": "assistant", "content": "ok"},
+                        "finish_reason": "stop",
+                    }
+                ],
+                "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+            },
+        )
+
+    http_client = httpx.AsyncClient(transport=httpx.MockTransport(handler))
+    try:
+        client = AsyncFireworks(
+            api_key="dummy",
+            base_url="https://gateway.smith.langchain.com/fireworks",
+            http_client=http_client,
+            max_retries=0,
+        )
+        await client.chat.completions.create(
+            model="accounts/fireworks/models/glm-5p2",
+            messages=[{"role": "user", "content": "hi"}],
+        )
+    finally:
+        await http_client.aclose()
+
+    assert len(requests) == 1
+    assert requests[0].url.path == "/fireworks/v1/chat/completions"
+
+
+async def test_fireworks_gateway_strips_legacy_function_call() -> None:
+    """The serializer must not emit ``function_call`` after sanitization.
+
+    Reproduces the production 400 — ``Extra inputs are not permitted, field:
+    'messages[N].function_call'`` — by routing an ``AIMessage`` that carries the
+    legacy ``function_call`` (alongside modern ``tool_calls``) through the
+    Fireworks serializer toward the gateway. Without the sanitizer middleware
+    the request body contains ``function_call``; with it, only ``tool_calls``
+    survives.
+    """
+    import json
+
+    from fireworks import AsyncFireworks
+    from langchain_fireworks.chat_models import ChatFireworks
+
+    from agent.middleware.sanitize_fireworks_messages import _sanitize_messages
+
+    captured_bodies: list[dict] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        captured_bodies.append(json.loads(request.content))
+        return httpx.Response(
+            200,
+            json={
+                "id": "chatcmpl-test",
+                "object": "chat.completion",
+                "created": 0,
+                "model": "accounts/fireworks/models/glm-5p2",
+                "choices": [
+                    {
+                        "index": 0,
+                        "message": {"role": "assistant", "content": "ok"},
+                        "finish_reason": "stop",
+                    }
+                ],
+                "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+            },
+        )
+
+    http_client = httpx.AsyncClient(transport=httpx.MockTransport(handler))
+    try:
+        chat_model = ChatFireworks(
+            model="accounts/fireworks/models/glm-5p2",
+            api_key="dummy",
+            base_url="https://gateway.smith.langchain.com/fireworks",
+            max_retries=0,
+        )
+        # Inject a mock-transport client so no real network call is made.
+        mock_sdk = AsyncFireworks(
+            api_key="dummy",
+            base_url="https://gateway.smith.langchain.com/fireworks",
+            http_client=http_client,
+            max_retries=0,
+        )
+        chat_model._async_sdk_client = mock_sdk  # type: ignore[attr-defined]
+        chat_model.async_client = mock_sdk.chat.completions  # type: ignore[attr-defined]
+
+        ai_message = AIMessage(
+            content="",
+            tool_calls=[{"name": "read_file", "args": {"file_path": "/x"}, "id": "tc1"}],
+            additional_kwargs={"function_call": {"name": "read_file", "arguments": "{}"}},
+        )
+        messages = [HumanMessage(content="hi"), ai_message]
+
+        # Apply the sanitizer the same way the middleware stack does.
+        _sanitize_messages(messages)
+
+        await chat_model.ainvoke(messages)
+        await mock_sdk.close()
+    finally:
+        await http_client.aclose()
+
+    assert len(captured_bodies) == 1
+    body = captured_bodies[0]
+    for msg in body["messages"]:
+        assert "function_call" not in msg, msg
+    # The assistant message still carries tool_calls.
+    assistant_msgs = [m for m in body["messages"] if m["role"] == "assistant"]
+    assert assistant_msgs and "tool_calls" in assistant_msgs[0]
+
+
+def test_google_genai_routes_to_gemini(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    overrides = gateway.gateway_overrides("google_genai:gemini-3.5-flash")
+    assert overrides == {
+        "base_url": "https://gateway.smith.langchain.com/gemini",
+        "api_key": "ls-key",
+    }
+
+
+def test_unsupported_provider_passes_through(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    # Vertex authenticates with a service account, not a bearer key, so it isn't routed.
+    assert gateway.gateway_overrides("google_vertexai:gemini-2.5-pro") is None
+
+
+def test_missing_api_key_passes_through(monkeypatch: pytest.MonkeyPatch) -> None:
+    assert gateway.gateway_overrides("openai:gpt-5.5") is None
+
+
+def test_prod_key_used_as_fallback(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY_PROD", "ls-prod-key")
+    overrides = gateway.gateway_overrides("anthropic:claude-opus-4-8")
+    assert overrides is not None
+    assert overrides["api_key"] == "ls-prod-key"
+
+
+def test_prod_key_preferred_over_platform_key(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-platform-key")
+    monkeypatch.setenv("LANGSMITH_API_KEY_PROD", "ls-prod-key")
+    overrides = gateway.gateway_overrides("anthropic:claude-opus-4-8")
+    assert overrides is not None
+    assert overrides["api_key"] == "ls-prod-key"
+
+
+def test_gateway_key_preferred_over_prod_key(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-platform-key")
+    monkeypatch.setenv("LANGSMITH_API_KEY_PROD", "ls-prod-key")
+    monkeypatch.setenv("LANGSMITH_GATEWAY_API_KEY", "ls-gateway-key")
+    overrides = gateway.gateway_overrides("anthropic:claude-opus-4-8")
+    assert overrides is not None
+    assert overrides["api_key"] == "ls-gateway-key"
+
+
+def test_base_url_env_override(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    monkeypatch.setenv("LANGSMITH_GATEWAY_BASE_URL", "https://gw.internal.example.com/")
+    overrides = gateway.gateway_overrides("anthropic:claude-opus-4-8")
+    assert overrides is not None
+    # Trailing slash is stripped, then the provider path is appended.
+    assert overrides["base_url"] == "https://gw.internal.example.com/anthropic"
+
+
+# --- resolve_gateway_enabled --------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    ("team_value", "env_enabled", "expected"),
+    [
+        (True, False, True),  # team True wins over env off
+        (False, True, False),  # team False wins over env on
+        (None, True, True),  # unset inherits env on
+        (None, False, False),  # unset inherits env off
+    ],
+)
+def test_resolve_gateway_enabled_precedence(
+    monkeypatch: pytest.MonkeyPatch,
+    team_value: bool | None,
+    env_enabled: bool,
+    expected: bool,
+) -> None:
+    if env_enabled:
+        monkeypatch.setenv("LANGSMITH_GATEWAY_ENABLED", "true")
+    assert gateway.resolve_gateway_enabled(team_value) is expected
+
+
+# --- make_model integration ---------------------------------------------------
+
+
+def _capture_init_chat_model() -> tuple[dict[str, Any], Any]:
+    """Patch init_chat_model to record the kwargs make_model builds."""
+    captured: dict[str, Any] = {}
+
+    def _fake(model: str, **kwargs: Any) -> str:
+        captured["model"] = model
+        captured.update(kwargs)
+        return "MODEL"
+
+    return captured, _fake
+
+
+def test_make_model_direct_openai_uses_responses_websocket() -> None:
+    captured, fake = _capture_init_chat_model()
+    with patch.object(model, "init_chat_model", fake):
+        model.make_model("openai:gpt-5.5", use_gateway=False)
+    assert captured["base_url"] == model.OPENAI_RESPONSES_WS_BASE_URL
+    assert captured["use_responses_api"] is True
+    assert captured["store"] is False
+    assert captured["include"] == ["reasoning.encrypted_content"]
+
+
+def test_make_model_gateway_openai_replaces_websocket(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    captured, fake = _capture_init_chat_model()
+    with patch.object(model, "init_chat_model", fake):
+        model.make_model("openai:gpt-5.5", use_gateway=True)
+    assert captured["base_url"] == "https://gateway.smith.langchain.com/openai/v1"
+    assert captured["use_responses_api"] is True
+    assert captured["store"] is False
+    assert captured["include"] == ["reasoning.encrypted_content"]
+    assert captured["api_key"] == "ls-key"
+
+
+def test_make_model_gateway_openai_chat_completions_optout_converts_reasoning(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    monkeypatch.setenv("LANGSMITH_GATEWAY_OPENAI_USE_RESPONSES", "false")
+    captured, fake = _capture_init_chat_model()
+    with patch.object(model, "init_chat_model", fake):
+        model.make_model(
+            "openai:gpt-5.5",
+            use_gateway=True,
+            reasoning={"effort": "high", "summary": "auto"},
+        )
+    assert captured["use_responses_api"] is False
+    assert captured["reasoning_effort"] == "high"
+    assert "reasoning" not in captured
+    assert "include" not in captured
+    assert "store" not in captured
+
+
+def test_make_model_gateway_openai_preserves_reasoning_none(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    captured, fake = _capture_init_chat_model()
+    with patch.object(model, "init_chat_model", fake):
+        model.make_model(
+            "openai:gpt-5.5",
+            use_gateway=True,
+            reasoning={"effort": "none"},
+        )
+    assert captured["use_responses_api"] is True
+    assert captured["store"] is False
+    assert captured["include"] == ["reasoning.encrypted_content"]
+    assert captured["reasoning"] == {"effort": "none"}
+    assert "reasoning_effort" not in captured
+
+
+def test_make_model_gateway_openai_responses_keeps_reasoning(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    reasoning = {"effort": "high", "summary": "auto"}
+    captured, fake = _capture_init_chat_model()
+    with patch.object(model, "init_chat_model", fake):
+        model.make_model("openai:gpt-5.5", use_gateway=True, reasoning=reasoning)
+    assert captured["use_responses_api"] is True
+    assert captured["store"] is False
+    assert captured["include"] == ["reasoning.encrypted_content"]
+    assert captured["reasoning"] == reasoning
+    assert "reasoning_effort" not in captured
+
+
+def test_make_model_gateway_follows_env_default(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    monkeypatch.setenv("LANGSMITH_GATEWAY_ENABLED", "true")
+    captured, fake = _capture_init_chat_model()
+    with patch.object(model, "init_chat_model", fake):
+        model.make_model("anthropic:claude-opus-4-8")  # use_gateway=None -> env default
+    assert captured["base_url"] == "https://gateway.smith.langchain.com/anthropic"
+    assert captured["api_key"] == "ls-key"
+
+
+def test_make_model_gateway_google_genai(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("LANGSMITH_API_KEY", "ls-key")
+    captured, fake = _capture_init_chat_model()
+    with patch.object(model, "init_chat_model", fake):
+        model.make_model("google_genai:gemini-3.5-flash", use_gateway=True)
+    assert captured["base_url"] == "https://gateway.smith.langchain.com/gemini"
+    assert captured["api_key"] == "ls-key"
+
+
+def test_make_model_gateway_without_key_falls_back_direct(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    captured, fake = _capture_init_chat_model()
+    with patch.object(model, "init_chat_model", fake):
+        model.make_model("openai:gpt-5.5", use_gateway=True)  # no LangSmith key
+    # No key -> overrides skipped -> the direct-provider websocket base stands.
+    assert captured["base_url"] == model.OPENAI_RESPONSES_WS_BASE_URL
+    assert captured["use_responses_api"] is True
+    assert captured["store"] is False
+    assert captured["include"] == ["reasoning.encrypted_content"]
+    assert "api_key" not in captured

tests/test_sanitize_fireworks_messages.py

@@ -0,0 +1,118 @@
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+import pytest
+from langchain_core.messages import AIMessage, HumanMessage, ToolMessage
+
+from agent.middleware.sanitize_fireworks_messages import SanitizeFireworksMessagesMiddleware
+
+
+def _make_request(messages: list[object], model: object | None = None) -> MagicMock:
+    request = MagicMock()
+    request.model = model
+    request.messages = messages
+    return request
+
+
+def _fireworks_model() -> MagicMock:
+    """A mock that satisfies ``_is_chat_fireworks`` via spec'd ``ChatFireworks``."""
+    try:
+        from langchain_fireworks.chat_models import ChatFireworks
+    except ImportError:  # pragma: no cover
+        pytest.skip("langchain-fireworks not installed")
+    return MagicMock(spec=ChatFireworks)
+
+
+class TestSanitizeFireworksMessagesMiddleware:
+    def test_drops_legacy_function_call(self) -> None:
+        message = AIMessage(
+            content="",
+            tool_calls=[{"name": "read_file", "args": {"file_path": "/x"}, "id": "tc1"}],
+            additional_kwargs={"function_call": {"name": "read_file", "arguments": "{}"}},
+        )
+        request = _make_request([message], model=_fireworks_model())
+        response = MagicMock()
+
+        def handler(req: object) -> object:
+            assert req is request
+            return response
+
+        result = SanitizeFireworksMessagesMiddleware().wrap_model_call(request, handler)
+
+        assert result is response
+        assert "function_call" not in message.additional_kwargs
+        # tool_calls are untouched
+        assert len(message.tool_calls) == 1
+
+    def test_preserves_message_without_function_call(self) -> None:
+        message = AIMessage(
+            content="ok",
+            tool_calls=[{"name": "read_file", "args": {"file_path": "/x"}, "id": "tc1"}],
+        )
+        request = _make_request([message], model=_fireworks_model())
+
+        SanitizeFireworksMessagesMiddleware().wrap_model_call(request, lambda req: MagicMock())
+
+        assert "function_call" not in message.additional_kwargs
+        assert len(message.tool_calls) == 1
+
+    def test_drops_function_call_with_no_tool_calls(self) -> None:
+        message = AIMessage(
+            content="",
+            additional_kwargs={"function_call": {"name": "search", "arguments": "{}"}},
+        )
+        request = _make_request([message], model=_fireworks_model())
+
+        SanitizeFireworksMessagesMiddleware().wrap_model_call(request, lambda req: MagicMock())
+
+        assert "function_call" not in message.additional_kwargs
+
+    @pytest.mark.asyncio
+    async def test_async_drops_legacy_function_call(self) -> None:
+        tool_result = ToolMessage(content="result", tool_call_id="tc1")
+        message = AIMessage(
+            content="",
+            tool_calls=[{"name": "read_file", "args": {"file_path": "/x"}, "id": "tc1"}],
+            additional_kwargs={"function_call": {"name": "read_file", "arguments": "{}"}},
+        )
+        request = _make_request(
+            [HumanMessage(content="hi"), message, tool_result],
+            model=_fireworks_model(),
+        )
+        response = MagicMock()
+
+        async def handler(req: object) -> object:
+            assert req is request
+            return response
+
+        result = await SanitizeFireworksMessagesMiddleware().awrap_model_call(request, handler)
+
+        assert result is response
+        assert "function_call" not in message.additional_kwargs
+
+    def test_ignores_non_fireworks_models(self) -> None:
+        message = AIMessage(
+            content="",
+            tool_calls=[{"name": "read_file", "args": {"file_path": "/x"}, "id": "tc1"}],
+            additional_kwargs={"function_call": {"name": "read_file", "arguments": "{}"}},
+        )
+        # Non-Fireworks model (plain MagicMock, no ChatFireworks in its spec chain)
+        request = _make_request([message], model=MagicMock())
+
+        SanitizeFireworksMessagesMiddleware().wrap_model_call(request, lambda req: MagicMock())
+
+        # function_call preserved for non-Fireworks providers
+        assert "function_call" in message.additional_kwargs
+
+    def test_skips_non_ai_messages(self) -> None:
+        messages = [
+            HumanMessage(content="hi"),
+            ToolMessage(content="result", tool_call_id="tc1"),
+        ]
+        request = _make_request(messages, model=_fireworks_model())
+
+        SanitizeFireworksMessagesMiddleware().wrap_model_call(request, lambda req: MagicMock())
+
+        # No AIMessages to mutate — handler still called
+        assert all(not isinstance(m, AIMessage) for m in messages)

tests/test_sanitize_openai_responses.py

@@ -0,0 +1,55 @@
+from langchain_core.messages import AIMessage, HumanMessage
+
+from agent.middleware.sanitize_openai_responses import _sanitize_messages
+
+
+def test_sanitize_messages_drops_stale_reasoning_references() -> None:
+    messages = [
+        HumanMessage(content="review this"),
+        AIMessage(
+            content=[
+                {
+                    "type": "reasoning",
+                    "id": "rs_123",
+                    "summary": [],
+                    "content": [],
+                },
+                {
+                    "type": "function_call",
+                    "id": "fc_123",
+                    "call_id": "call_123",
+                    "name": "read_file",
+                    "arguments": "{}",
+                },
+            ],
+            tool_calls=[
+                {
+                    "name": "read_file",
+                    "args": {},
+                    "id": "call_123",
+                }
+            ],
+        ),
+    ]
+
+    _sanitize_messages(messages)
+
+    content = messages[1].content
+    assert isinstance(content, list)
+    assert [block["type"] for block in content] == ["function_call"]
+    assert messages[1].tool_calls[0]["id"] == "call_123"
+
+
+def test_sanitize_messages_preserves_encrypted_reasoning() -> None:
+    reasoning = {
+        "type": "reasoning",
+        "id": "rs_123",
+        "encrypted_content": "encrypted",
+        "summary": [],
+        "content": [],
+    }
+    messages = [AIMessage(content=[reasoning])]
+
+    _sanitize_messages(messages)
+
+    assert messages[0].content == [reasoning]

ui/src/lib/api.ts

@@ -164,6 +164,8 @@ export interface TeamSettings {
   review_draft_prs: boolean
   pr_summaries: boolean
   review_trace_links: boolean
+  /** Tri-state LLM Gateway toggle; null inherits the LANGSMITH_GATEWAY_ENABLED default. */
+  gateway_enabled?: boolean | null
   review_tracing_project?: string | null
   org_guidelines?: string | null
   default_agent_model?: string | null

ui/src/routes/admin.tsx

@@ -55,6 +55,8 @@ function AdminPage() {
     >
       <GlobalDefaultsSection models={options.data?.models ?? []} />
 
+      <LLMGatewaySection />
+
       <TriggerReviewSection />
 
       <SettingsSection title="Evals">
@@ -552,6 +554,77 @@ function ObservabilityCredentialsSection() {
   )
 }
 
+type GatewayMode = "inherit" | "enabled" | "disabled"
+
+function gatewayMode(value: boolean | null | undefined): GatewayMode {
+  if (value === true) return "enabled"
+  if (value === false) return "disabled"
+  return "inherit"
+}
+
+function gatewayModeValue(mode: GatewayMode): boolean | null {
+  if (mode === "enabled") return true
+  if (mode === "disabled") return false
+  return null
+}
+
+function LLMGatewaySection() {
+  const qc = useQueryClient()
+  const settings = useQuery({
+    queryKey: ["teamSettings"],
+    queryFn: api.getTeamSettings,
+  })
+  const [error, setError] = useState<string | null>(null)
+
+  const save = useMutation({
+    mutationFn: (body: TeamSettings) => api.saveTeamSettings(body),
+    onSuccess: (saved) => {
+      qc.setQueryData(["teamSettings"], saved)
+      setError(null)
+    },
+    onError: (e: Error) => setError(e.message),
+  })
+
+  const mode = gatewayMode(settings.data?.gateway_enabled)
+
+  return (
+    <SettingsSection
+      title="LLM Gateway"
+      description="Route agent and reviewer LLM calls through the LangSmith LLM Gateway. It authenticates with the workspace LangSmith API key and resolves provider keys from Provider Secrets, so no provider keys are needed at runtime. Requires the gateway (private beta) enabled for your organization."
+    >
+      <div className="divide-y divide-border">
+        <SettingsRow
+          label="Route through the gateway"
+          description="Inherit uses the LANGSMITH_GATEWAY_ENABLED deployment default. OpenAI, Anthropic, Fireworks, and Google Gemini are routed through the gateway. Bedrock (AWS) calls the provider directly and is not routed; other providers not listed also call the provider directly."
+          control={
+            <Select
+              value={mode}
+              onValueChange={(next) =>
+                settings.data &&
+                save.mutate({
+                  ...settings.data,
+                  gateway_enabled: gatewayModeValue(next as GatewayMode),
+                })
+              }
+              disabled={!settings.data || save.isPending}
+            >
+              <SelectTrigger className="w-48">
+                <SelectValue />
+              </SelectTrigger>
+              <SelectContent>
+                <SelectItem value="inherit">Inherit deployment default</SelectItem>
+                <SelectItem value="enabled">Enabled</SelectItem>
+                <SelectItem value="disabled">Disabled</SelectItem>
+              </SelectContent>
+            </Select>
+          }
+        />
+      </div>
+      {error && <p className="px-4 pb-3 text-xs text-destructive">{error}</p>}
+    </SettingsSection>
+  )
+}
+
 function PRTraceResolutionSection() {
   const qc = useQueryClient()
   const settings = useQuery({