synth-ai 0.2.8.dev4__py3-none-any.whl → 0.2.23.dev3__py3-none-any.whl

synth_ai/task/contracts.py

@@ -1,72 +1,56 @@
 from __future__ import annotations
 
 from dataclasses import dataclass
-from typing import Optional, Any, Dict, List
-from pydantic import BaseModel
+from enum import Enum
+from typing import Any, Literal
 
+from pydantic import BaseModel, ConfigDict, Field
 
-@dataclass(frozen=True)
-class TaskAppEndpoints:
-    """Canonical Task App endpoint shapes used by RL trainers.
-
-    The Task App is an HTTP service (often deployed on Modal) that exposes:
-    - Health: GET /health
-      • Requires header X-API-Key (when ENVIRONMENT_API_KEY is configured)
-      • Returns { healthy: true }
-    - Environment lifecycle:
-      • POST /env/{env_name}/initialize → { env_id, observation }
-      • POST /env/{env_name}/step      → { observation, reward, done, info }
-      • POST /env/{env_name}/terminate → { ok: true }
-    - Rollout (optional, unified schema):
-      • POST /rollout → { run_id, trajectories[], metrics, ... }
-    - Proxy (optional):
-      • POST /proxy/v1/chat/completions (for direct OpenAI calls from Task App)
-    """
 
-    health: str = "/health"
-    rollout: str = "/rollout"
-    proxy_chat_completions: str = "/proxy/v1/chat/completions"
-    env_initialize: str = "/env/{env_name}/initialize"
-    env_step: str = "/env/{env_name}/step"
-    env_terminate: str = "/env/{env_name}/terminate"
+class RolloutMode(str, Enum):
+    """Mode controls how rollout infrastructure processes inference URLs."""
+    RL = "rl"
+    EVAL = "eval"
 
 
 @dataclass(frozen=True)
-class TaskAppContract:
-    """Requirements and expectations for a Task App used by RL trainers.
-
-    - Auth: ENVIRONMENT_API_KEY must be set in the Task App environment; requests include X-API-Key.
-    - Health: /health returns 200 and JSON; may verify X-API-Key header.
-    - Env API: initialize/step/terminate are present for the target env (e.g., CrafterClassic).
-    - Rollout API: optional; provides a single-call rollout for convenience/testing.
-    - Inference routing: policy config passes an inference_url (Synth backend or OpenAI proxy).
-    - URL: base must be reachable via HTTPS and should be under .modal.run in production.
+class TaskAppEndpoints:
+    """Required Task App endpoints used by RL trainers and clients.
+
+    Task Apps run as lightweight HTTP services (often on Modal) that expose these
+    standard endpoints. Additional endpoints (proxies, debug routes) may be added
+    by individual task apps as needed.
     """
 
-    base_url: str
-    env_name: Optional[str] = None
-    requires_api_key_header: bool = True
+    root: str = "/"
+    health: str = "/health"
+    info: str = "/info"
+    task_info: str = "/task_info"
+    rollout: str = "/rollout"
 
 
 # --- Unified rollout schema used by Task App services and SDK utilities ---
 
+
 class RolloutEnvSpec(BaseModel):
-    env_id: Optional[str] = None
-    env_name: Optional[str] = None
-    config: Dict[str, Any] = {}
-    seed: Optional[int] = None
+    env_id: str | None = None
+    env_name: str | None = None
+    config: dict[str, Any] = Field(default_factory=dict)
+    seed: int | None = None
 
 
 class RolloutPolicySpec(BaseModel):
-    policy_id: Optional[str] = None
-    policy_name: Optional[str] = None
-    config: Dict[str, Any] = {}
+    policy_id: str | None = None
+    policy_name: str | None = None
+    config: dict[str, Any] = Field(default_factory=dict)
 
 
 class RolloutRecordConfig(BaseModel):
     trajectories: bool = True
     logprobs: bool = False
     value: bool = False
+    return_trace: bool = False
+    trace_format: Literal["compact", "full", "structured"] = "compact"
 
 
 class RolloutSafetyConfig(BaseModel):
@@ -78,43 +62,175 @@ class RolloutRequest(BaseModel):
     run_id: str
     env: RolloutEnvSpec
     policy: RolloutPolicySpec
-    ops: List[Dict[str, Any]] | List[str]
+    ops: list[dict[str, Any]] | list[str]
     record: RolloutRecordConfig = RolloutRecordConfig()
     on_done: str = "reset"
     safety: RolloutSafetyConfig = RolloutSafetyConfig()
-    training_session_id: Optional[str] = None
-    synth_base_url: Optional[str] = None
+    training_session_id: str | None = None
+    synth_base_url: str | None = None
+    mode: RolloutMode  # Required: explicit RL vs EVAL mode
 
 
 class RolloutStep(BaseModel):
-    obs: Dict[str, Any]
-    tool_calls: List[Dict[str, Any]]
-    reward: Optional[float] = None
+    """Single step in a rollout trajectory.
+    
+    DEPRECATED: This is part of the legacy trajectory format. New code should
+    consume v3 traces (RolloutResponse.trace) instead. See monorepo/trace_single_source.txt
+    for migration plan.
+    """
+    obs: dict[str, Any]
+    tool_calls: list[dict[str, Any]]
+    reward: float | None = None
     done: bool = False
-    truncated: Optional[bool] = None
-    info: Optional[Dict[str, Any]] = None
+    truncated: bool | None = None
+    info: dict[str, Any] | None = None
 
 
 class RolloutTrajectory(BaseModel):
+    """Legacy trajectory format for rollout results.
+    
+    DEPRECATED: This format duplicates data already present in v3 traces and will
+    be removed once training code migrates to consuming RolloutResponse.trace.
+    
+    Current state:
+    - Task apps emit BOTH this format AND v3 traces (dual serialization)
+    - Training code (GSPO) reads from this format
+    - Eval/filter tools read from v3 traces
+    
+    Migration plan:
+    - Phase 1: Training code learns to read from v3 traces (with fallback to this)
+    - Phase 2: Make this field optional once training is migrated
+    - Phase 3: Remove this field entirely and delete this class
+    
+    See: monorepo/trace_single_source.txt for full migration plan and timeline.
+    
+    Why v3 traces are better:
+    - Single source of truth (no duplication/drift)
+    - Richer data: token IDs, logprobs, reasoning, timing, images
+    - Built-in audit trail and replay capability
+    - Standard schema across all Synth AI tooling
+    """
     env_id: str
     policy_id: str
-    steps: List[RolloutStep]
-    final: Optional[Dict[str, Any]] = None
+    steps: list[RolloutStep]
+    final: dict[str, Any] | None = None
     length: int
+    
+    # Required for trace correlation with inference mesh (optional initially for backward compat)
+    # See: monorepo/INFERENCE_URL_REQUIREMENT_PLAN.md and trace_creation_and_judgement.txt
+    inference_url: str
+    
+    decision_samples: list[dict[str, Any]] | None = None
 
 
 class RolloutMetrics(BaseModel):
-    episode_returns: List[float]
+    episode_returns: list[float]
     mean_return: float
     num_steps: int
     num_episodes: int = 0
+    outcome_score: float | None = None
+    events_score: float | None = None
+    details: dict[str, Any] = Field(default_factory=dict)
 
 
 class RolloutResponse(BaseModel):
+    """Response from a rollout execution.
+    
+    Contains both legacy trajectory format (for backward compatibility) and
+    modern v3 trace format (preferred going forward).
+    """
     run_id: str
-    trajectories: List[RolloutTrajectory]
-    branches: Dict[str, List[str]] = {}
+    
+    # DEPRECATED: Legacy format maintained for training code compatibility.
+    # Will be removed once training migrates to reading from `trace` field.
+    # See: monorepo/trace_single_source.txt for migration plan.
+    trajectories: list[RolloutTrajectory]
+    
+    branches: dict[str, list[str]] = Field(default_factory=dict)
     metrics: RolloutMetrics
     aborted: bool = False
     ops_executed: int = 0
+    
+    # OPTIONAL: correlation ID for linking rollout to inference traces
+    # If not provided, trainer will infer it from trajectory.inference_url ?cid=... parameter
+    trace_correlation_id: str | None = None
+    
+    # PREFERRED: v3 trace format (SessionTrace). This is the single source of truth
+    # for rollout data and should be used by all new code. Contains richer data than
+    # trajectories including token IDs, logprobs, timing, and multimodal content.
+    trace: dict[str, Any] | None = None
+    pipeline_metadata: dict[str, Any] = Field(default_factory=dict)
+
+
+class _ExtraAllowModel(BaseModel):
+    """Base helper that preserves unknown keys while still exposing typed attributes."""
+
+    model_config = ConfigDict(extra="allow")
+
+
+class TaskDescriptor(_ExtraAllowModel):
+    """Human-readable task identifiers shown in UIs and logs."""
+
+    id: str
+    name: str
+    description: str | None = None
+    version: str | None = None
+
+
+class DatasetInfo(_ExtraAllowModel):
+    """Metadata about the prompt/task dataset powering the environment."""
+
+    id: str | None = None
+    name: str | None = None
+    version: str | None = None
+    splits: list[str] | None = None
+    default_split: str | None = None
+    description: str | None = None
+
+
+class RubricCriterion(_ExtraAllowModel):
+    id: str
+    description: str
+    weight: float | None = None
+
+
+class RubricSection(_ExtraAllowModel):
+    name: str
+    criteria: list[RubricCriterion] = Field(default_factory=list)
+
+
+class RubricInfo(_ExtraAllowModel):
+    """Outcome and event scoring definitions used by judges."""
+
+    outcome: RubricSection | None = None
+    events: RubricSection | None = None
+
+
+class InferenceInfo(_ExtraAllowModel):
+    """Recommended defaults for policy model routing."""
+
+    model: str | None = None
+    inference_url: str | None = None
+
+
+class LimitsInfo(_ExtraAllowModel):
+    """Operational limits the environment enforces."""
+
+    max_turns: int | None = None
+    max_response_tokens: int | None = None
+    timeout_seconds: int | None = None
+
+
+class TaskInfo(_ExtraAllowModel):
+    """Static metadata describing the capabilities of a Task App task."""
 
+    task: TaskDescriptor
+    environment: str
+    dataset: DatasetInfo
+    rubric: RubricInfo
+    inference: InferenceInfo
+    limits: LimitsInfo
+    task_metadata: dict[str, Any] = Field(
+        default_factory=dict,
+        description="Task-specific extras (e.g. prompt version info, documentation links).",
+    )

synth_ai/task/datasets.py

@@ -0,0 +1,108 @@
+"""Dataset registry and helpers shared by Task Apps."""
+
+from __future__ import annotations
+
+from collections.abc import Callable, Hashable
+from typing import Any
+
+from pydantic import BaseModel, Field, field_validator
+
+
+class TaskDatasetSpec(BaseModel):
+    """Declarative metadata describing a dataset that a Task App exposes."""
+
+    id: str
+    name: str
+    version: str | None = None
+    splits: list[str] = Field(default_factory=list)
+    default_split: str | None = None
+    cardinality: int | None = None
+    description: str | None = None
+
+    @field_validator("default_split")
+    @classmethod
+    def _validate_default_split(cls, value: str | None, info):
+        values = info.data if hasattr(info, "data") else {}  # type: ignore[attr-defined]
+        if value and value not in (values.get("splits") or []):
+            raise ValueError("default_split must be one of splits when provided")
+        return value
+
+
+RegistryLoader = Callable[[TaskDatasetSpec], Any]
+
+
+class TaskDatasetRegistry:
+    """Lightweight registry mapping dataset specs to loader callables."""
+
+    def __init__(self) -> None:
+        self._entries: dict[str, tuple[TaskDatasetSpec, RegistryLoader, bool]] = {}
+        self._cache: dict[Hashable, Any] = {}
+
+    def register(
+        self, spec: TaskDatasetSpec, loader: RegistryLoader, *, cache: bool = True
+    ) -> None:
+        """Register a dataset loader and its metadata."""
+
+        self._entries[spec.id] = (spec, loader, cache)
+
+    def describe(self, dataset_id: str) -> TaskDatasetSpec:
+        if dataset_id not in self._entries:
+            raise KeyError(f"Dataset not registered: {dataset_id}")
+        spec, _, _ = self._entries[dataset_id]
+        return spec
+
+    def list(self) -> list[TaskDatasetSpec]:
+        return [entry[0] for entry in self._entries.values()]
+
+    def get(self, spec: TaskDatasetSpec | str) -> Any:
+        """Return dataset materialisation (with optional caching)."""
+
+        if isinstance(spec, str):
+            if spec not in self._entries:
+                raise KeyError(f"Dataset not registered: {spec}")
+            base_spec, loader, cache_enabled = self._entries[spec]
+            effective_spec = base_spec
+        else:
+            if spec.id not in self._entries:
+                raise KeyError(f"Dataset not registered: {spec.id}")
+            base_spec, loader, cache_enabled = self._entries[spec.id]
+            effective_spec = base_spec.model_copy(update=spec.model_dump(exclude_unset=True))
+
+        cache_key: Hashable = (
+            effective_spec.id,
+            effective_spec.version,
+            effective_spec.default_split,
+        )
+        if cache_enabled:
+            if cache_key not in self._cache:
+                self._cache[cache_key] = loader(effective_spec)
+            return self._cache[cache_key]
+        return loader(effective_spec)
+
+    @staticmethod
+    def ensure_split(spec: TaskDatasetSpec, split: str | None) -> str:
+        """Validate that `split` exists on the spec; return a concrete split."""
+
+        if not spec.splits:
+            return split or spec.default_split or "default"
+        if split is None:
+            if spec.default_split:
+                return spec.default_split
+            raise ValueError(f"split must be provided for dataset {spec.id}")
+        if split not in spec.splits:
+            raise ValueError(f"Unknown split '{split}' for dataset {spec.id}")
+        return split
+
+    @staticmethod
+    def normalise_seed(seed: Any, *, cardinality: int | None = None) -> int:
+        """Normalise arbitrary seed input into a bounded non-negative integer."""
+
+        try:
+            value = int(seed)
+        except Exception as exc:  # pragma: no cover - defensive
+            raise ValueError(f"Seed must be convertible to int (got {seed!r})") from exc
+        if value < 0:
+            value = abs(value)
+        if cardinality and cardinality > 0:
+            value = value % cardinality
+        return value

synth_ai/task/errors.py

@@ -0,0 +1,50 @@
+"""Error helpers used across Task App implementations."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from .json import to_jsonable
+
+
+def error_payload(
+    code: str, message: str, *, extra: dict[str, Any] | None = None
+) -> dict[str, Any]:
+    payload: dict[str, Any] = {"error": {"code": code, "message": message}}
+    if extra:
+        payload["error"].update(extra)
+    return payload
+
+
+def http_exception(
+    status_code: int,
+    code: str,
+    message: str,
+    *,
+    extra: dict[str, Any] | None = None,
+    headers: dict[str, str] | None = None,
+):
+    try:
+        from fastapi import HTTPException  # type: ignore
+    except Exception as exc:  # pragma: no cover - FastAPI not installed
+        raise RuntimeError("fastapi must be installed to raise HTTPException") from exc
+
+    payload = error_payload(code, message, extra=extra)
+    return HTTPException(status_code=status_code, detail=to_jsonable(payload), headers=headers)
+
+
+def json_error_response(
+    status_code: int,
+    code: str,
+    message: str,
+    *,
+    extra: dict[str, Any] | None = None,
+    headers: dict[str, str] | None = None,
+):
+    try:
+        from fastapi.responses import JSONResponse  # type: ignore
+    except Exception as exc:  # pragma: no cover - FastAPI not installed
+        raise RuntimeError("fastapi must be installed to build JSONResponse") from exc
+
+    payload = error_payload(code, message, extra=extra)
+    return JSONResponse(status_code=status_code, content=to_jsonable(payload), headers=headers)

synth_ai/task/health.py

@@ -1,10 +1,13 @@
+"""Helpers for probing Task App health endpoints."""
+
 from __future__ import annotations
 
-from typing import Any, Dict
+from typing import Any
+
 import aiohttp
 
 
-async def task_app_health(task_app_url: str) -> Dict[str, Any]:
+async def task_app_health(task_app_url: str) -> dict[str, Any]:
     """Probe a Task App base URL for basic reachability.
 
     Behavior:
@@ -12,17 +15,20 @@ async def task_app_health(task_app_url: str) -> Dict[str, Any]:
     - Fallback to GET if HEAD is unsupported
     - Returns {ok: bool, status?: int, error?: str}
     """
+
+    async def _try_request(session: aiohttp.ClientSession, method: str) -> dict[str, Any] | None:
+        request = getattr(session, method)
+        async with request(task_app_url, allow_redirects=True) as response:
+            if 200 <= response.status < 400:
+                return {"ok": True, "status": response.status}
+        return None
+
     try:
         async with aiohttp.ClientSession() as session:
-            async with session.head(task_app_url, allow_redirects=True) as r:
-                if 200 <= r.status < 400:
-                    return {"ok": True, "status": r.status}
-        async with aiohttp.ClientSession() as session:
-            async with session.get(task_app_url, allow_redirects=True) as r2:
-                if 200 <= r2.status < 400:
-                    return {"ok": True, "status": r2.status}
+            for method in ("head", "get"):
+                result = await _try_request(session, method)
+                if result is not None:
+                    return result
         return {"ok": False, "status": None}
     except Exception as e:
         return {"ok": False, "error": f"{type(e).__name__}: {e}"}
-
-

synth_ai/task/inference_api.py

@@ -0,0 +1,101 @@
+from __future__ import annotations
+
+import os
+from typing import Any, Dict, List, Optional, Sequence
+
+
+class InferenceAPIClient:
+    """Async client that normalizes chat completion calls across providers."""
+
+    def __init__(self, *, provider: str, inference_url: Optional[str] = None, timeout: float = 30.0) -> None:
+        self.provider = (provider or "").lower()
+        self.inference_url = inference_url.rstrip("/") if inference_url else None
+        self.timeout = timeout
+
+        if self.provider == "openai":
+            try:
+                from openai import AsyncOpenAI
+            except Exception as exc:  # pragma: no cover - import guard
+                raise RuntimeError("openai package not installed") from exc
+
+            api_key = os.getenv("OPENAI_API_KEY")
+            if not api_key:
+                raise RuntimeError("OPENAI_API_KEY must be set for meta-model proposals")
+
+            kwargs: Dict[str, Any] = {"api_key": api_key}
+            if self.inference_url:
+                kwargs["base_url"] = self.inference_url
+            self._client = AsyncOpenAI(**kwargs)
+
+        elif self.provider == "groq":
+            try:
+                from groq import AsyncGroq
+            except Exception as exc:  # pragma: no cover - import guard
+                raise RuntimeError("groq package not installed") from exc
+
+            api_key = os.getenv("GROQ_API_KEY")
+            if not api_key:
+                raise RuntimeError("GROQ_API_KEY must be set for meta-model proposals")
+
+            kwargs: Dict[str, Any] = {"api_key": api_key}
+            if self.inference_url:
+                kwargs["base_url"] = self.inference_url
+            self._client = AsyncGroq(**kwargs)
+
+        else:
+            raise RuntimeError(f"Unsupported meta-model provider: {provider}")
+
+    async def chat_completion(
+        self,
+        *,
+        messages: Sequence[Dict[str, Any]],
+        model: str,
+        temperature: float,
+        max_tokens: int,
+    ) -> Dict[str, Any]:
+        if self.provider == "openai":
+            response = await self._client.chat.completions.create(  # type: ignore[attr-defined]
+                model=model,
+                messages=self._normalize_messages(messages),
+                temperature=temperature,
+                max_tokens=max_tokens,
+                timeout=self.timeout,
+            )
+            return self._model_dump(response)
+
+        if self.provider == "groq":
+            response = await self._client.chat.completions.create(  # type: ignore[attr-defined]
+                model=model,
+                messages=self._normalize_messages(messages),
+                temperature=temperature,
+                max_tokens=max_tokens,
+                timeout=self.timeout,
+            )
+            return self._model_dump(response)
+
+        raise RuntimeError(f"Unsupported meta-model provider: {self.provider}")
+
+    @staticmethod
+    def _normalize_messages(messages: Sequence[Dict[str, Any]]) -> List[Dict[str, Any]]:
+        return [dict(message) for message in messages]
+
+    @staticmethod
+    def _model_dump(obj: Any) -> Dict[str, Any]:
+        if hasattr(obj, "model_dump"):
+            data = obj.model_dump()
+        elif hasattr(obj, "dict"):
+            data = obj.dict()
+        else:
+            data = dict(obj)  # type: ignore[arg-type]
+
+        usage = getattr(obj, "usage", None)
+        if usage and "usage" not in data:
+            if hasattr(usage, "model_dump"):
+                data["usage"] = usage.model_dump()
+            elif isinstance(usage, dict):
+                data["usage"] = usage
+        return data
+
+
+__all__ = ["InferenceAPIClient"]
+