diff --git a/docs/configuration.md b/docs/configuration.md index 0123017d2..e208212cf 100644 --- a/docs/configuration.md +++ b/docs/configuration.md @@ -672,6 +672,12 @@ Existing configs do not need to change. If you do not set `modelPresets` or `age "maxTokens": 8192, "contextWindowTokens": 128000, "temperature": 0.1, + "fallbackModels": [ + { + "provider": "anthropic", + "model": "anthropic/claude-sonnet-4-6" + } + ], "modelPreset": null } }, @@ -682,7 +688,17 @@ Existing configs do not need to change. If you do not set `modelPresets` or `age "maxTokens": 4096, "contextWindowTokens": 128000, "temperature": 0.2, - "reasoningEffort": "low" + "reasoningEffort": "low", + "fallbackModels": [ + { + "provider": "deepseek", + "model": "deepseek/deepseek-chat", + "maxTokens": 4096, + "contextWindowTokens": 64000, + "temperature": 0.1, + "reasoningEffort": null + } + ] }, "deep": { "model": "anthropic/claude-opus-4-5", @@ -705,9 +721,53 @@ Existing configs do not need to change. If you do not set `modelPresets` or `age | `contextWindowTokens` | Context window size used by prompt building and consolidation decisions. | | `temperature` | Sampling temperature. | | `reasoningEffort` | Optional reasoning/thinking setting. Provider support varies. | +| `fallbackModels` | Optional ordered fallback models for this active configuration only. | `default` is reserved and always means the implicit preset built from `agents.defaults.*`; do not define `modelPresets.default`. Use `/model default` to switch back to `agents.defaults.*`. +### Model Fallbacks + +`fallbackModels` belongs to the currently active model configuration. If the active configuration is `agents.defaults`, only `agents.defaults.fallbackModels` is used. If the active configuration is `modelPresets.fast`, only `modelPresets.fast.fallbackModels` is used. nanobot does not inherit or merge fallbacks between defaults and presets. + +Each fallback entry must include at least `provider` and `model`. The other fields are optional; omitted values inherit from the active primary configuration for that request. + +```json +{ + "modelPresets": { + "fast": { + "model": "MiniMax-M2.7-highspeed", + "provider": "minimaxAnthropic", + "maxTokens": 4096, + "contextWindowTokens": 262144, + "temperature": 0.1, + "reasoningEffort": null, + "fallbackModels": [ + { + "provider": "deepseek", + "model": "deepseek-v4-pro", + "maxTokens": 4096, + "contextWindowTokens": 262144, + "temperature": 0.1, + "reasoningEffort": null + } + ] + }, + "deep": { + "model": "deepseek-v4-pro", + "provider": "deepseek", + "maxTokens": 4096, + "contextWindowTokens": 262144, + "temperature": 0.1, + "reasoningEffort": null + } + } +} +``` + +In this example, `/model fast` can fail over to DeepSeek, but `/model deep` has no fallback because the `deep` preset does not define `fallbackModels`. + +Failover only runs when the primary model returns an error before any answer text has been streamed. Fallback models are tried in order. If a fallback has a smaller `contextWindowTokens`, nanobot uses the smallest window in the active chain when building context so the fallback can receive the same prompt. + Set `agents.defaults.modelPreset` to start with a named preset: ```json diff --git a/nanobot/config/schema.py b/nanobot/config/schema.py index a112b932d..bdae26008 100644 --- a/nanobot/config/schema.py +++ b/nanobot/config/schema.py @@ -74,6 +74,17 @@ class DreamConfig(Base): return f"every {hours}h" +class ModelFallbackConfig(Base): + """A fallback model tied to one active model configuration.""" + + model: str + provider: str + max_tokens: int | None = None + context_window_tokens: int | None = None + temperature: float | None = None + reasoning_effort: str | None = None + + class ModelPresetConfig(Base): """A named set of model + generation parameters for quick switching.""" @@ -83,7 +94,7 @@ class ModelPresetConfig(Base): context_window_tokens: int = 65_536 temperature: float = 0.1 reasoning_effort: str | None = None - fallback_models: list[str] = Field(default_factory=list) + fallback_models: list[ModelFallbackConfig] = Field(default_factory=list) def to_generation_settings(self) -> Any: from nanobot.providers.base import GenerationSettings @@ -107,6 +118,7 @@ class AgentDefaults(Base): context_window_tokens: int = 65_536 context_block_limit: int | None = None temperature: float = 0.1 + fallback_models: list[ModelFallbackConfig] = Field(default_factory=list) max_tool_iterations: int = 200 max_concurrent_subagents: int = Field(default=1, ge=1) max_tool_result_chars: int = 16_000 @@ -297,6 +309,7 @@ class Config(BaseSettings): model=d.model, provider=d.provider, max_tokens=d.max_tokens, context_window_tokens=d.context_window_tokens, temperature=d.temperature, reasoning_effort=d.reasoning_effort, + fallback_models=d.fallback_models, ) def resolve_preset(self, name: str | None = None) -> ModelPresetConfig: diff --git a/nanobot/providers/factory.py b/nanobot/providers/factory.py index e4822b7f8..a3ae57daf 100644 --- a/nanobot/providers/factory.py +++ b/nanobot/providers/factory.py @@ -5,7 +5,7 @@ from __future__ import annotations from dataclasses import dataclass from pathlib import Path -from nanobot.config.schema import Config, ModelPresetConfig +from nanobot.config.schema import Config, ModelFallbackConfig, ModelPresetConfig from nanobot.providers.base import LLMProvider from nanobot.providers.fallback_provider import FallbackProvider from nanobot.providers.registry import find_by_name @@ -104,6 +104,28 @@ def _make_provider_core( return provider +def _fallback_preset(primary: ModelPresetConfig, fallback: ModelFallbackConfig) -> ModelPresetConfig: + """Build the effective provider/generation config for one fallback model.""" + return ModelPresetConfig( + model=fallback.model, + provider=fallback.provider, + max_tokens=fallback.max_tokens if fallback.max_tokens is not None else primary.max_tokens, + context_window_tokens=( + fallback.context_window_tokens + if fallback.context_window_tokens is not None + else primary.context_window_tokens + ), + temperature=( + fallback.temperature if fallback.temperature is not None else primary.temperature + ), + reasoning_effort=( + fallback.reasoning_effort + if fallback.reasoning_effort is not None + else primary.reasoning_effort + ), + ) + + def make_provider( config: Config, *, @@ -120,12 +142,11 @@ def make_provider( provider = _make_provider_core(config, preset_name=preset_name, preset=preset, model=model) if resolved.fallback_models: - fb_preset = resolved.model_copy(update={"provider": "auto", "fallback_models": []}) provider = FallbackProvider( primary=provider, fallback_models=resolved.fallback_models, - provider_factory=lambda m: _make_provider_core( - config, preset_name=preset_name, preset=fb_preset, model=m + provider_factory=lambda fb: _make_provider_core( + config, preset_name=preset_name, preset=_fallback_preset(resolved, fb) ), ) @@ -138,9 +159,32 @@ def provider_signature( preset_name: str | None = None, preset: ModelPresetConfig | None = None, ) -> tuple[object, ...]: - """Return the config fields that affect the primary LLM provider.""" + """Return the config fields that affect the active provider chain.""" resolved = _resolve_model_preset(config, preset_name=preset_name, preset=preset) p = config.get_provider(resolved.model, preset=resolved) + + def _fallback_signature(fallback: ModelFallbackConfig) -> tuple[object, ...]: + fallback_preset = _fallback_preset(resolved, fallback) + fp = config.get_provider(fallback.model, preset=fallback_preset) + return ( + fallback.model, + fallback.provider, + fallback_preset.max_tokens, + fallback_preset.temperature, + fallback_preset.reasoning_effort, + fallback_preset.context_window_tokens, + config.get_provider_name(fallback.model, preset=fallback_preset), + config.get_api_key(fallback.model, preset=fallback_preset), + config.get_api_base(fallback.model, preset=fallback_preset), + fp.extra_headers if fp else None, + fp.extra_body if fp else None, + getattr(fp, "region", None) if fp else None, + getattr(fp, "profile", None) if fp else None, + ) + + fallback_signatures = tuple( + _fallback_signature(fallback) for fallback in resolved.fallback_models + ) return ( resolved.model, resolved.provider, @@ -155,6 +199,7 @@ def provider_signature( resolved.temperature, resolved.reasoning_effort, resolved.context_window_tokens, + fallback_signatures, ) @@ -165,10 +210,14 @@ def build_provider_snapshot( preset: ModelPresetConfig | None = None, ) -> ProviderSnapshot: resolved = _resolve_model_preset(config, preset_name=preset_name, preset=preset) + fallback_windows = [ + _fallback_preset(resolved, fallback).context_window_tokens + for fallback in resolved.fallback_models + ] return ProviderSnapshot( provider=make_provider(config, preset=resolved), model=resolved.model, - context_window_tokens=resolved.context_window_tokens, + context_window_tokens=min([resolved.context_window_tokens, *fallback_windows]), signature=provider_signature(config, preset=resolved), ) diff --git a/nanobot/providers/fallback_provider.py b/nanobot/providers/fallback_provider.py index c0b137890..a62b619a0 100644 --- a/nanobot/providers/fallback_provider.py +++ b/nanobot/providers/fallback_provider.py @@ -24,7 +24,7 @@ class FallbackProvider(LLMProvider): provider on-the-fly. Key design: - - Failover is request-scoped (the wrapper itself is stateless between turns). + - Failover attempts are request-scoped; primary circuit state persists. - Skipped when content was already streamed to avoid duplicate output. - Recursive failover is prevented by the factory returning plain providers. - Primary provider is circuit-broken after repeated failures to avoid @@ -34,8 +34,8 @@ class FallbackProvider(LLMProvider): def __init__( self, primary: LLMProvider, - fallback_models: list[str], - provider_factory: Callable[[str], LLMProvider], + fallback_models: list[Any], + provider_factory: Callable[[Any], LLMProvider], ): self._primary = primary self._fallback_models = list(fallback_models) @@ -52,6 +52,10 @@ class FallbackProvider(LLMProvider): def generation(self, value): self._primary.generation = value + @property + def supports_progress_deltas(self) -> bool: + return bool(getattr(self._primary, "supports_progress_deltas", False)) + def get_default_model(self) -> str: return self._primary.get_default_model() @@ -122,7 +126,8 @@ class FallbackProvider(LLMProvider): last_response: LLMResponse | None = None primary_skipped = not self._primary_available() - for idx, fallback_model in enumerate(self._fallback_models): + for idx, fallback in enumerate(self._fallback_models): + fallback_model = fallback.model if has_streamed is not None and has_streamed[0]: break if idx == 0 and primary_skipped: @@ -138,25 +143,35 @@ class FallbackProvider(LLMProvider): else: logger.info( "Fallback '{}' also failed, trying next fallback '{}'", - self._fallback_models[idx - 1], fallback_model, + self._fallback_models[idx - 1].model, fallback_model, ) try: - fallback_provider = self._provider_factory(fallback_model) + fallback_provider = self._provider_factory(fallback) except Exception as exc: logger.warning( "Failed to create provider for fallback '{}': {}", fallback_model, exc ) continue - original_model = kwargs.get("model") + original_values = { + name: kwargs.get(name, LLMProvider._SENTINEL) + for name in ("model", "max_tokens", "temperature", "reasoning_effort") + } kwargs["model"] = fallback_model + if fallback.max_tokens is not None: + kwargs["max_tokens"] = fallback.max_tokens + if fallback.temperature is not None: + kwargs["temperature"] = fallback.temperature + if fallback.reasoning_effort is not None: + kwargs["reasoning_effort"] = fallback.reasoning_effort try: fallback_response = await call(fallback_provider, kwargs) finally: - if original_model is not None: - kwargs["model"] = original_model - else: - kwargs.pop("model", None) + for name, value in original_values.items(): + if value is LLMProvider._SENTINEL: + kwargs.pop(name, None) + else: + kwargs[name] = value if fallback_response.finish_reason != "error": logger.info( diff --git a/tests/agent/test_runner_fallback.py b/tests/agent/test_runner_fallback.py index 273bd6d6d..e15a29848 100644 --- a/tests/agent/test_runner_fallback.py +++ b/tests/agent/test_runner_fallback.py @@ -7,6 +7,7 @@ from unittest.mock import MagicMock import pytest +from nanobot.config.schema import ModelFallbackConfig from nanobot.providers.base import LLMProvider, LLMResponse from nanobot.providers.fallback_provider import FallbackProvider @@ -24,6 +25,25 @@ def _error_response(content: str = "api error") -> LLMResponse: return _make_response(content, finish_reason="error", error_kind="server_error") +def _fallback( + model: str, + provider: str = "fallback", + *, + max_tokens: int | None = None, + context_window_tokens: int | None = None, + temperature: float | None = None, + reasoning_effort: str | None = None, +) -> ModelFallbackConfig: + return ModelFallbackConfig( + model=model, + provider=provider, + max_tokens=max_tokens, + context_window_tokens=context_window_tokens, + temperature=temperature, + reasoning_effort=reasoning_effort, + ) + + class _FakeProvider(LLMProvider): """Fake provider for testing.""" @@ -60,17 +80,113 @@ def test_fallback_models_default_empty() -> None: def test_fallback_models_accepts_list() -> None: from nanobot.config.schema import ModelPresetConfig - p = ModelPresetConfig(model="test/primary", fallback_models=["test/a", "test/b"]) - assert p.fallback_models == ["test/a", "test/b"] + p = ModelPresetConfig( + model="test/primary", + fallback_models=[{"provider": "test", "model": "test/a"}], + ) + assert p.fallback_models == [_fallback("test/a", provider="test")] def test_fallback_models_from_camel_case() -> None: from nanobot.config.schema import ModelPresetConfig p = ModelPresetConfig.model_validate({ "model": "test/primary", - "fallbackModels": ["test/a"], + "fallbackModels": [{"provider": "test", "model": "test/a"}], }) - assert p.fallback_models == ["test/a"] + assert p.fallback_models == [_fallback("test/a", provider="test")] + + +def test_provider_signature_tracks_fallback_models_and_provider_config() -> None: + from nanobot.config.schema import Config + from nanobot.providers.factory import provider_signature + + base = { + "modelPresets": { + "prod": { + "model": "openai/gpt-4.1", + "fallbackModels": [ + {"provider": "anthropic", "model": "anthropic/claude-sonnet-4-6"} + ], + } + }, + "providers": { + "openai": {"apiKey": "primary-key"}, + "anthropic": {"apiKey": "fallback-key"}, + }, + } + changed_fallback = { + **base, + "modelPresets": { + "prod": { + "model": "openai/gpt-4.1", + "fallbackModels": [{"provider": "deepseek", "model": "deepseek/deepseek-chat"}], + } + }, + "providers": { + **base["providers"], + "deepseek": {"apiKey": "deepseek-key"}, + }, + } + changed_key = { + **base, + "providers": { + "openai": {"apiKey": "primary-key"}, + "anthropic": {"apiKey": "new-fallback-key"}, + }, + } + + signature = provider_signature(Config.model_validate(base), preset_name="prod") + + assert signature != provider_signature(Config.model_validate(changed_fallback), preset_name="prod") + assert signature != provider_signature(Config.model_validate(changed_key), preset_name="prod") + + +def test_agent_defaults_can_define_fallback_models() -> None: + from nanobot.config.schema import Config + + config = Config.model_validate({ + "agents": { + "defaults": { + "model": "primary-model", + "provider": "custom", + "fallbackModels": [{"provider": "deepseek", "model": "deepseek-v4-pro"}], + } + } + }) + + assert config.resolve_preset().fallback_models == [ + _fallback("deepseek-v4-pro", provider="deepseek") + ] + + +def test_provider_snapshot_uses_smallest_fallback_context_window() -> None: + from nanobot.config.schema import Config + from nanobot.providers.factory import build_provider_snapshot + + config = Config.model_validate({ + "modelPresets": { + "prod": { + "model": "openai/gpt-4.1", + "provider": "openai", + "contextWindowTokens": 128000, + "fallbackModels": [ + { + "provider": "deepseek", + "model": "deepseek/deepseek-chat", + "contextWindowTokens": 64000, + } + ], + } + }, + "providers": { + "openai": {"apiKey": "primary-key"}, + "deepseek": {"apiKey": "fallback-key"}, + }, + }) + + snapshot = build_provider_snapshot(config, preset_name="prod") + + assert snapshot.context_window_tokens == 64000 # -- FallbackProvider tests -- @@ -83,7 +199,7 @@ class TestNoFallbackWhenPrimarySucceeds: factory = MagicMock() fb = FallbackProvider( primary=primary, - fallback_models=["fallback-a"], + fallback_models=[_fallback("fallback-a")], provider_factory=factory, ) @@ -102,14 +218,14 @@ class TestFallbackOnPrimaryError: fb = FallbackProvider( primary=primary, - fallback_models=["fallback-a"], + fallback_models=[_fallback("fallback-a")], provider_factory=factory, ) result = await fb.chat(messages=[{"role": "user", "content": "hi"}], model="primary-model") assert result.content == "fallback ok" assert result.finish_reason == "stop" - factory.assert_called_once_with("fallback-a") + factory.assert_called_once_with(_fallback("fallback-a")) assert primary.chat_calls[0]["model"] == "primary-model" assert fallback.chat_calls[0]["model"] == "fallback-a" @@ -121,7 +237,7 @@ class TestNoFallbackWhenContentStreamed: factory = MagicMock() fb = FallbackProvider( primary=primary, - fallback_models=["fallback-a"], + fallback_models=[_fallback("fallback-a")], provider_factory=factory, ) @@ -146,14 +262,14 @@ class TestFailoverOnTransientError: factory = MagicMock(return_value=fallback) fb = FallbackProvider( primary=primary, - fallback_models=["fallback-a"], + fallback_models=[_fallback("fallback-a")], provider_factory=factory, ) result = await fb.chat(messages=[{"role": "user", "content": "hi"}]) assert result.content == "fallback ok" assert result.finish_reason == "stop" - factory.assert_called_once_with("fallback-a") + factory.assert_called_once_with(_fallback("fallback-a")) @pytest.mark.asyncio async def test_timeout(self) -> None: @@ -165,14 +281,14 @@ class TestFailoverOnTransientError: factory = MagicMock(return_value=fallback) fb = FallbackProvider( primary=primary, - fallback_models=["fallback-a"], + fallback_models=[_fallback("fallback-a")], provider_factory=factory, ) result = await fb.chat(messages=[{"role": "user", "content": "hi"}]) assert result.content == "fallback ok" assert result.finish_reason == "stop" - factory.assert_called_once_with("fallback-a") + factory.assert_called_once_with(_fallback("fallback-a")) class TestFallbackTriesModelsInOrder: @@ -185,15 +301,15 @@ class TestFallbackTriesModelsInOrder: fb = FallbackProvider( primary=primary, - fallback_models=["fallback-a", "fallback-b"], + fallback_models=[_fallback("fallback-a"), _fallback("fallback-b")], provider_factory=factory, ) result = await fb.chat(messages=[{"role": "user", "content": "hi"}]) assert result.content == "b ok" assert factory.call_count == 2 - factory.assert_any_call("fallback-a") - factory.assert_any_call("fallback-b") + factory.assert_any_call(_fallback("fallback-a")) + factory.assert_any_call(_fallback("fallback-b")) class TestAllFallbacksFail: @@ -205,7 +321,7 @@ class TestAllFallbacksFail: fb = FallbackProvider( primary=primary, - fallback_models=["fallback-a"], + fallback_models=[_fallback("fallback-a")], provider_factory=factory, ) @@ -223,7 +339,7 @@ class TestFactoryExceptionSkipsModel: fb = FallbackProvider( primary=primary, - fallback_models=["fallback-a", "fallback-b"], + fallback_models=[_fallback("fallback-a"), _fallback("fallback-b")], provider_factory=factory, ) @@ -242,13 +358,43 @@ class TestFallbackModelParameter: fb = FallbackProvider( primary=primary, - fallback_models=["fallback-model"], + fallback_models=[_fallback("fallback-model")], provider_factory=factory, ) await fb.chat(messages=[{"role": "user", "content": "hi"}], model="primary-model") assert fallback.chat_calls[0]["model"] == "fallback-model" + @pytest.mark.asyncio + async def test_overrides_generation_fields_when_configured(self) -> None: + primary = _FakeProvider("primary", _error_response()) + fallback = _FakeProvider("fallback", _make_response("ok")) + fb = FallbackProvider( + primary=primary, + fallback_models=[ + _fallback( + "fallback-model", + max_tokens=1234, + temperature=0.4, + reasoning_effort="low", + ) + ], + provider_factory=MagicMock(return_value=fallback), + ) + + await fb.chat( + messages=[{"role": "user", "content": "hi"}], + model="primary-model", + max_tokens=8192, + temperature=0.1, + reasoning_effort="high", + ) + + assert fallback.chat_calls[0]["model"] == "fallback-model" + assert fallback.chat_calls[0]["max_tokens"] == 1234 + assert fallback.chat_calls[0]["temperature"] == 0.4 + assert fallback.chat_calls[0]["reasoning_effort"] == "low" + class TestNoFallbackWhenEmptyList: @pytest.mark.asyncio @@ -277,7 +423,7 @@ class TestChatStreamFailover: fb = FallbackProvider( primary=primary, - fallback_models=["fallback-a"], + fallback_models=[_fallback("fallback-a")], provider_factory=factory, ) @@ -291,7 +437,7 @@ class TestGetDefaultModel: primary = _FakeProvider("primary") fb = FallbackProvider( primary=primary, - fallback_models=["a"], + fallback_models=[_fallback("a")], provider_factory=MagicMock(), ) assert fb.get_default_model() == "primary/model" @@ -305,7 +451,7 @@ class TestCircuitBreaker: factory = MagicMock(return_value=fallback) fb = FallbackProvider( primary=primary, - fallback_models=["fallback-a"], + fallback_models=[_fallback("fallback-a")], provider_factory=factory, ) @@ -329,7 +475,7 @@ class TestCircuitBreaker: factory = MagicMock(return_value=fallback) fb = FallbackProvider( primary=primary, - fallback_models=["fallback-a"], + fallback_models=[_fallback("fallback-a")], provider_factory=factory, ) @@ -357,7 +503,7 @@ class TestGenerationForwarded: primary.generation = GenerationSettings(temperature=0.5, max_tokens=1024) fb = FallbackProvider( primary=primary, - fallback_models=["a"], + fallback_models=[_fallback("a")], provider_factory=MagicMock(), ) assert fb.generation.temperature == 0.5