Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions agent/chat_completion_helpers.py
Original file line number Diff line number Diff line change
Expand Up @@ -765,7 +765,7 @@ def build_api_kwargs(agent, api_messages: list) -> dict:
supports_reasoning=agent._supports_reasoning_extra_body(),
qwen_session_metadata=_qwen_meta,
logit_bias=_build_logit_bias(),
stop=_build_stop_sequences(),
stop=_build_stop_sequences(agent.provider),
)

# ── Legacy flag path ────────────────────────────────────────────
Expand Down Expand Up @@ -814,7 +814,7 @@ def build_api_kwargs(agent, api_messages: list) -> dict:
anthropic_max_output=_ant_max,
provider_name=agent.provider,
logit_bias=_build_logit_bias(),
stop=_build_stop_sequences(),
stop=_build_stop_sequences(agent.provider),
)


Expand Down
36 changes: 32 additions & 4 deletions agent/gemini_cloudcode_adapter.py
Original file line number Diff line number Diff line change
Expand Up @@ -47,6 +47,7 @@
)

logger = logging.getLogger(__name__)
_GEMINI_MAX_STOP_SEQUENCES = 5


# =============================================================================
Expand Down Expand Up @@ -130,6 +131,34 @@ def _translate_tool_result_to_gemini(message: Dict[str, Any]) -> Dict[str, Any]:
}


def _normalize_gemini_stop_sequences(stop: Any) -> Optional[List[str]]:
"""Return a Gemini-safe stop list capped at the provider limit."""
if not stop:
return None
if isinstance(stop, str):
values = [stop]
elif isinstance(stop, (list, tuple)):
values = [str(s) for s in stop if s]
else:
return None

deduped: List[str] = []
seen: set[str] = set()
for value in values:
if value in seen:
continue
seen.add(value)
deduped.append(value)

if len(deduped) > _GEMINI_MAX_STOP_SEQUENCES:
logger.warning(
"Gemini supports at most %d stop sequences; truncating %d extra entries",
_GEMINI_MAX_STOP_SEQUENCES,
len(deduped) - _GEMINI_MAX_STOP_SEQUENCES,
)
return deduped[:_GEMINI_MAX_STOP_SEQUENCES] or None


def _build_gemini_contents(
messages: List[Dict[str, Any]],
) -> tuple[List[Dict[str, Any]], Optional[Dict[str, Any]]]:
Expand Down Expand Up @@ -284,10 +313,9 @@ def build_gemini_request(
generation_config["maxOutputTokens"] = max_tokens
if isinstance(top_p, (int, float)):
generation_config["topP"] = float(top_p)
if isinstance(stop, str) and stop:
generation_config["stopSequences"] = [stop]
elif isinstance(stop, list) and stop:
generation_config["stopSequences"] = [str(s) for s in stop if s]
stop_sequences = _normalize_gemini_stop_sequences(stop)
if stop_sequences:
generation_config["stopSequences"] = stop_sequences
normalized_thinking = _normalize_thinking_config(thinking_config)
if normalized_thinking:
generation_config["thinkingConfig"] = normalized_thinking
Expand Down
34 changes: 32 additions & 2 deletions agent/gemini_native_adapter.py
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,7 @@
logger = logging.getLogger(__name__)

DEFAULT_GEMINI_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
_GEMINI_MAX_STOP_SEQUENCES = 5

# Published max output-token ceiling shared by every current Gemini text model
# (2.5 + 3.x: flash, flash-lite, pro). Used as the default when the caller
Expand Down Expand Up @@ -61,6 +62,34 @@ def is_native_gemini_base_url(base_url: str) -> bool:
return not normalized.endswith("/openai")


def _normalize_gemini_stop_sequences(stop: Any) -> Optional[List[str]]:
"""Return a Gemini-safe stop list capped at the provider limit."""
if not stop:
return None
if isinstance(stop, str):
values = [stop]
elif isinstance(stop, (list, tuple)):
values = [str(s) for s in stop if s]
else:
return None

deduped: List[str] = []
seen: set[str] = set()
for value in values:
if value in seen:
continue
seen.add(value)
deduped.append(value)

if len(deduped) > _GEMINI_MAX_STOP_SEQUENCES:
logger.warning(
"Gemini supports at most %d stop sequences; truncating %d extra entries",
_GEMINI_MAX_STOP_SEQUENCES,
len(deduped) - _GEMINI_MAX_STOP_SEQUENCES,
)
return deduped[:_GEMINI_MAX_STOP_SEQUENCES] or None


def probe_gemini_tier(
api_key: str,
base_url: str = DEFAULT_GEMINI_BASE_URL,
Expand Down Expand Up @@ -445,8 +474,9 @@ def build_gemini_request(
generation_config["maxOutputTokens"] = GEMINI_DEFAULT_MAX_OUTPUT_TOKENS
if top_p is not None:
generation_config["topP"] = top_p
if stop:
generation_config["stopSequences"] = stop if isinstance(stop, list) else [str(stop)]
stop_sequences = _normalize_gemini_stop_sequences(stop)
if stop_sequences:
generation_config["stopSequences"] = stop_sequences
normalized_thinking = _normalize_thinking_config(thinking_config)
if normalized_thinking:
generation_config["thinkingConfig"] = normalized_thinking
Expand Down
13 changes: 11 additions & 2 deletions agent/logit_bias.py
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,10 @@

log = logging.getLogger(__name__)

_STOP_SEQUENCE_LIMITS: dict[str, int] = {
"cerebras": 4,
}


def _strings() -> List[str]:
"""Forbidden strings from the unified config parser."""
Expand Down Expand Up @@ -40,15 +44,20 @@ def build_logit_bias() -> Optional[Dict[int, int]]:
return banned or None


def build_stop_sequences() -> Optional[List[str]]:
def build_stop_sequences(provider_name: str | None = None) -> Optional[List[str]]:
"""Return forbidden strings as stop sequences.

Works on ALL providers. Zero dependencies.
When the model starts generating a forbidden string,
the API cuts generation at that token.
"""
strings = _strings()
return strings if strings else None
if not strings:
return None
limit = _STOP_SEQUENCE_LIMITS.get((provider_name or "").strip().lower())
if limit is not None:
return strings[:limit]
return strings


__all__ = ["build_logit_bias", "build_stop_sequences"]
4 changes: 2 additions & 2 deletions cli.py
Original file line number Diff line number Diff line change
Expand Up @@ -387,9 +387,9 @@ def load_cli_config() -> Dict[str, Any]:
# Default configuration
defaults = {
"model": {
"default": "",
"default": "gemma-4-31b",
"base_url": "",
"provider": "auto",
"provider": "cerebras",
},
"terminal": {
"env_type": "local",
Expand Down
12 changes: 12 additions & 0 deletions dag_cli/auth.py
Original file line number Diff line number Diff line change
Expand Up @@ -70,6 +70,7 @@
DEFAULT_NOUS_PORTAL_URL = "https://portal.nousresearch.com"
DEFAULT_NOUS_INFERENCE_URL = "https://inference-api.nousresearch.com/v1"
DEFAULT_NOUS_CLIENT_ID = "dag-cli"
DEFAULT_CEREBRAS_BASE_URL = "https://api.cerebras.ai/v1"
NOUS_INFERENCE_INVOKE_SCOPE = "inference:invoke"
DEFAULT_NOUS_SCOPE = NOUS_INFERENCE_INVOKE_SCOPE
NOUS_DEVICE_CODE_SOURCE = "device_code"
Expand Down Expand Up @@ -170,6 +171,14 @@ class ProviderConfig:


PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
"cerebras": ProviderConfig(
id="cerebras",
name="Cerebras",
auth_type="api_key",
inference_base_url=DEFAULT_CEREBRAS_BASE_URL,
api_key_env_vars=("CEREBRAS_API_KEY",),
base_url_env_var="CEREBRAS_BASE_URL",
),
"nous": ProviderConfig(
id="nous",
name="Nous Portal",
Expand Down Expand Up @@ -1586,6 +1595,9 @@ def resolve_provider(
except Exception as e:
logger.debug("Could not detect active auth provider: %s", e)

if has_usable_secret(os.getenv("CEREBRAS_API_KEY")):
return "cerebras"

if has_usable_secret(os.getenv("OPENAI_API_KEY")) or has_usable_secret(os.getenv("OPENROUTER_API_KEY")):
return "openrouter"

Expand Down
2 changes: 2 additions & 0 deletions dag_cli/config.py
Original file line number Diff line number Diff line change
Expand Up @@ -5482,6 +5482,7 @@ def _load_config_impl(*, want_deepcopy: bool) -> Dict[str, Any]:
# overload (529), service errors (503), or connection failures.
#
# Supported providers:
# cerebras (CEREBRAS_API_KEY) — Cerebras Inference API
# openrouter (OPENROUTER_API_KEY) — routes to any model
# openai-codex (OAuth — dag auth) — OpenAI Codex
# nous (OAuth — dag auth) — Nous Portal
Expand Down Expand Up @@ -5514,6 +5515,7 @@ def _load_config_impl(*, want_deepcopy: bool) -> Dict[str, Any]:
# overload (529), service errors (503), or connection failures.
#
# Supported providers:
# cerebras (CEREBRAS_API_KEY) — Cerebras Inference API
# openrouter (OPENROUTER_API_KEY) — routes to any model
# openai-codex (OAuth — dag auth) — OpenAI Codex
# nous (OAuth — dag auth) — Nous Portal
Expand Down
2 changes: 1 addition & 1 deletion dag_cli/model_normalize.py
Original file line number Diff line number Diff line change
Expand Up @@ -98,6 +98,7 @@
"minimax-cn",
"alibaba",
"qwen-oauth",
"cerebras",
"xiaomi",
"arcee",
"ollama-cloud",
Expand Down Expand Up @@ -470,4 +471,3 @@ def normalize_model_for_provider(model_input: str, target_provider: str) -> str:
# ---------------------------------------------------------------------------
# Batch / convenience helpers
# ---------------------------------------------------------------------------

8 changes: 7 additions & 1 deletion dag_cli/model_switch.py
Original file line number Diff line number Diff line change
Expand Up @@ -188,7 +188,13 @@ class DirectAlias(NamedTuple):


# Built-in direct aliases (can be extended via config.yaml model_aliases:)
_BUILTIN_DIRECT_ALIASES: dict[str, DirectAlias] = {}
_BUILTIN_DIRECT_ALIASES: dict[str, DirectAlias] = {
"gemma-4-31b-it": DirectAlias(
model="gemma-4-31b",
provider="cerebras",
base_url="",
),
}

# Merged dict (builtins + user config); populated by _load_direct_aliases()
DIRECT_ALIASES: dict[str, DirectAlias] = {}
Expand Down
7 changes: 7 additions & 0 deletions dag_cli/models.py
Original file line number Diff line number Diff line change
Expand Up @@ -216,6 +216,12 @@ def _xai_curated_models() -> list[str]:
"gpt-4o",
"gpt-4o-mini",
],
"cerebras": [
"gemma-4-31b",
"gpt-oss-120b",
"glm-4.7",
"qwen3-32b",
],
"openai-codex": _codex_curated_models(),
"xai-oauth": _xai_curated_models(),
"copilot-acp": [
Expand Down Expand Up @@ -994,6 +1000,7 @@ class ProviderEntry(NamedTuple):
CANONICAL_PROVIDERS: list[ProviderEntry] = [
ProviderEntry("nous", "Nous Portal", "Nous Portal (Everything your agent needs, 300+ models with bundled tool use)"),
ProviderEntry("openrouter", "OpenRouter", "OpenRouter (Pay-per-use API aggregator)"),
ProviderEntry("cerebras", "Cerebras", "Cerebras Inference API (Gemma 4 31B, GPT OSS 120B, GLM 4.7, Qwen3 32B)"),
ProviderEntry("novita", "NovitaAI", "NovitaAI (Cloud: Model API, Agent Sandbox, GPU Cloud)"),
ProviderEntry("lmstudio", "LM Studio", "LM Studio (Local desktop app with built-in model server)"),
ProviderEntry("anthropic", "Anthropic", "Anthropic (Claude models via API key or Claude Code)"),
Expand Down
6 changes: 6 additions & 0 deletions dag_cli/providers.py
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,11 @@ class DagOverlay:


DAG_OVERLAYS: Dict[str, DagOverlay] = {
"cerebras": DagOverlay(
transport="openai_chat",
base_url_override="https://api.cerebras.ai/v1",
base_url_env_var="CEREBRAS_BASE_URL",
),
"openrouter": DagOverlay(
transport="openai_chat",
is_aggregator=True,
Expand Down Expand Up @@ -366,6 +371,7 @@ class ProviderDef:

_LABEL_OVERRIDES: Dict[str, str] = {
"nous": "Nous Portal",
"cerebras": "Cerebras",
"openai-codex": "OpenAI Codex",
"copilot-acp": "GitHub Copilot ACP",
"stepfun": "StepFun Step Plan",
Expand Down
52 changes: 52 additions & 0 deletions tests/test_cerebras_default.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,52 @@
from __future__ import annotations

from pathlib import Path

from agent.logit_bias import build_stop_sequences
from dag_cli.auth import PROVIDER_REGISTRY
from dag_cli.models import CANONICAL_PROVIDERS
from dag_cli.model_normalize import normalize_model_for_provider
from dag_cli.models import get_default_model_for_provider, provider_model_ids
from dag_cli.model_switch import resolve_alias


def test_cerebras_is_registered() -> None:
assert "cerebras" in PROVIDER_REGISTRY


def test_cerebras_is_canonical_provider() -> None:
assert any(p.slug == "cerebras" for p in CANONICAL_PROVIDERS)


def test_cerebras_default_model() -> None:
assert get_default_model_for_provider("cerebras") == "gemma-4-31b"


def test_cerebras_prefix_is_stripped() -> None:
assert normalize_model_for_provider("cerebras/gemma-4-31b", "cerebras") == "gemma-4-31b"


def test_cerebras_model_catalog_contains_public_models() -> None:
models = provider_model_ids("cerebras")
assert models[:4] == [
"gemma-4-31b",
"gpt-oss-120b",
"glm-4.7",
"qwen3-32b",
]


def test_cli_default_points_to_cerebras() -> None:
cli_text = Path(__file__).resolve().parents[1] / "cli.py"
content = cli_text.read_text(encoding="utf-8")
assert '"provider": "cerebras"' in content
assert '"default": "gemma-4-31b"' in content


def test_cerebras_stop_sequences_are_capped(monkeypatch) -> None:
monkeypatch.setattr("agent.logit_bias.get_forbidden_strings", lambda: ["a", "b", "c", "d", "e"])
assert build_stop_sequences("cerebras") == ["a", "b", "c", "d"]


def test_old_cerebras_slug_resolves_to_new_slug() -> None:
assert resolve_alias("gemma-4-31b-it", "cerebras") == ("cerebras", "gemma-4-31b", "gemma-4-31b-it")
22 changes: 22 additions & 0 deletions tests/test_gemini_stop_sequences.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,22 @@
from __future__ import annotations

from agent.gemini_cloudcode_adapter import build_gemini_request as build_cloudcode_request
from agent.gemini_native_adapter import build_gemini_request as build_native_request


def test_native_gemini_stop_sequences_are_capped_and_deduped() -> None:
request = build_native_request(
messages=[{"role": "user", "content": "hello"}],
stop=["a", "b", "a", "c", "d", "e", "f"],
)

assert request["generationConfig"]["stopSequences"] == ["a", "b", "c", "d", "e"]


def test_cloudcode_gemini_stop_sequences_are_capped_and_deduped() -> None:
request = build_cloudcode_request(
messages=[{"role": "user", "content": "hello"}],
stop=["a", "b", "a", "c", "d", "e", "f"],
)

assert request["generationConfig"]["stopSequences"] == ["a", "b", "c", "d", "e"]
Loading