synth-ai 0.2.6.dev1__py3-none-any.whl → 0.4.3__py3-none-any.whl

synth_ai/sdk/task/client.py

@@ -0,0 +1,175 @@
+"""Async HTTP client for interacting with Task Apps.
+
+Prefer synth_ai.sdk.localapi.client moving forward. This module remains for
+backward compatibility during the naming transition.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+from typing import Any
+
+import httpx
+from pydantic import BaseModel
+
+from .contracts import RolloutRequest, RolloutResponse, TaskInfo
+from .json import to_jsonable
+
+
+def _prepare_payload(payload: Any) -> Any:
+    if payload is None:
+        return None
+    if isinstance(payload, BaseModel):
+        return payload.model_dump(mode="json", by_alias=True)
+    return to_jsonable(payload)
+
+
+class TaskAppClient:
+    def __init__(
+        self,
+        base_url: str,
+        api_key: str | None = None,
+        *,
+        timeout: float = 600.0,
+        retries: int = 3,
+    ) -> None:
+        self.base_url = base_url.rstrip("/")
+        self.api_key = api_key
+        self.timeout = timeout
+        self.retries = max(1, retries)
+        self._client: httpx.AsyncClient | None = None
+        self.env = _TaskAppEnvironmentClient(self)
+
+    async def __aenter__(self) -> TaskAppClient:
+        await self._ensure_client()
+        return self
+
+    async def __aexit__(self, exc_type, exc, tb) -> None:
+        await self.aclose()
+
+    async def _ensure_client(self) -> httpx.AsyncClient:
+        if self._client is None:
+            self._client = httpx.AsyncClient(
+                base_url=self.base_url,
+                timeout=httpx.Timeout(self.timeout),
+                follow_redirects=True,
+            )
+        return self._client
+
+    def _headers(self) -> dict[str, str]:
+        headers: dict[str, str] = {}
+        # Primary key
+        primary = (self.api_key or "").strip()
+        if primary:
+            headers["X-API-Key"] = primary
+            # Also set Authorization for clients that read bearer tokens
+            headers.setdefault("Authorization", f"Bearer {primary}")
+        # Include ALL available environment keys via CSV in X-API-Keys
+        keys: list[str] = []
+        if primary:
+            keys.append(primary)
+        aliases = (os.getenv("ENVIRONMENT_API_KEY_ALIASES") or "").strip()
+        if aliases:
+            for part in aliases.split(","):
+                trimmed = part.strip()
+                if trimmed and trimmed not in keys:
+                    keys.append(trimmed)
+        if keys:
+            headers["X-API-Keys"] = ",".join(keys)
+        return headers
+
+    async def aclose(self) -> None:
+        if self._client is not None:
+            await self._client.aclose()
+            self._client = None
+
+    async def _request(
+        self,
+        method: str,
+        path: str,
+        *,
+        params: dict[str, Any] | list[tuple[str, Any]] | None = None,
+        json_payload: Any = None,
+    ) -> httpx.Response:
+        client = await self._ensure_client()
+        payload = _prepare_payload(json_payload)
+        headers = self._headers()
+        last_exc: Exception | None = None
+        for attempt in range(self.retries):
+            try:
+                response = await client.request(
+                    method,
+                    path,
+                    headers=headers,
+                    params=params,
+                    json=payload,
+                )
+                response.raise_for_status()
+                return response
+            except httpx.HTTPStatusError as exc:
+                if 500 <= exc.response.status_code < 600 and attempt + 1 < self.retries:
+                    await asyncio.sleep(0.1 * (attempt + 1))
+                    last_exc = exc
+                    continue
+                raise
+            except httpx.HTTPError as exc:
+                last_exc = exc
+                if attempt + 1 >= self.retries:
+                    raise
+                await asyncio.sleep(0.1 * (attempt + 1))
+        if last_exc:  # pragma: no cover - defensive
+            raise last_exc
+        raise RuntimeError("Unreachable code in TaskAppClient._request")
+
+    async def health(self) -> dict[str, Any]:
+        response = await self._request("GET", "/health")
+        return response.json()
+
+    async def info(self) -> dict[str, Any]:
+        response = await self._request("GET", "/info")
+        return response.json()
+
+    async def task_info(self, seeds: list[int] | None = None) -> TaskInfo | list[TaskInfo]:
+        params: list[tuple[str, Any]] | None = None
+        if seeds:
+            params = [("seed", seed) for seed in seeds]
+        response = await self._request("GET", "/task_info", params=params)
+        data = response.json()
+        if isinstance(data, list):
+            return [TaskInfo.model_validate(item) for item in data]
+        return TaskInfo.model_validate(data)
+
+    async def rollout(self, request: RolloutRequest) -> RolloutResponse:
+        response = await self._request("POST", "/rollout", json_payload=request)
+        data = response.json()
+        return RolloutResponse.model_validate(data)
+
+
+class LocalAPIClient(TaskAppClient):
+    """Alias for TaskAppClient with LocalAPI naming."""
+
+
+class _TaskAppEnvironmentClient:
+    def __init__(self, client: TaskAppClient) -> None:
+        self._client = client
+
+    async def initialize(self, env_name: str, payload: dict[str, Any]) -> dict[str, Any]:
+        response = await self._client._request(
+            "POST", f"/env/{env_name}/initialize", json_payload=payload
+        )
+        return response.json()
+
+    async def step(self, env_name: str, payload: dict[str, Any]) -> dict[str, Any]:
+        response = await self._client._request(
+            "POST", f"/env/{env_name}/step", json_payload=payload
+        )
+        return response.json()
+
+    async def terminate(
+        self, env_name: str, payload: dict[str, Any] | None = None
+    ) -> dict[str, Any]:
+        response = await self._client._request(
+            "POST", f"/env/{env_name}/terminate", json_payload=payload or {}
+        )
+        return response.json()

synth_ai/sdk/task/config.py

@@ -0,0 +1,257 @@
+"""Configuration dataclasses for task app CLI commands (eval, filter)."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Literal
+
+
+@dataclass(slots=True)
+class EvalConfig:
+    """Configuration for 'synth-ai eval' command.
+    
+    Validates and provides defaults for evaluation runs against task apps.
+    """
+    
+    # Required: Task app identifier
+    app_id: str
+    
+    # Required: Model to evaluate
+    model: str
+    
+    # Required: Seeds to run
+    seeds: list[int]
+    
+    # Optional: Task app URL (None = spawn in-process)
+    task_app_url: str | None = None
+    
+    # Optional: Data split to use
+    split: str = "train"
+    
+    # Optional: Maximum turns/steps per episode
+    max_turns: int | None = None
+    
+    # Optional: Maximum LLM calls per episode
+    max_llm_calls: int = 10
+    
+    # Optional: Concurrency for parallel rollouts
+    concurrency: int = 1
+    
+    # Optional: Environment name
+    env_name: str | None = None
+    
+    # Optional: Policy name
+    policy_name: str | None = None
+    
+    # Optional: Trace format ("compact", "full", "structured")
+    trace_format: Literal["compact", "full", "structured"] = "compact"
+    
+    # Optional: Whether to return traces in response
+    return_trace: bool = False
+    
+    # Optional: Environment config overrides
+    env_config: dict[str, Any] = field(default_factory=dict)
+    
+    # Optional: Policy config overrides
+    policy_config: dict[str, Any] = field(default_factory=dict)
+    
+    # Optional: Metadata for traces
+    metadata: dict[str, str] = field(default_factory=dict)
+    
+    # Optional: SQL query for metadata filtering
+    metadata_sql: str | None = None
+    
+    def __post_init__(self):
+        """Validate configuration after initialization."""
+        if not self.app_id:
+            raise ValueError("app_id is required")
+        
+        if not self.model:
+            raise ValueError("model is required")
+        
+        if not self.seeds:
+            raise ValueError("seeds list cannot be empty")
+        
+        if not isinstance(self.seeds, list):
+            raise ValueError("seeds must be a list of integers")
+        
+        if self.concurrency < 1:
+            raise ValueError("concurrency must be >= 1")
+        
+        if self.max_llm_calls < 1:
+            raise ValueError("max_llm_calls must be >= 1")
+        
+        if self.max_turns is not None and self.max_turns < 1:
+            raise ValueError("max_turns must be >= 1")
+        
+        if self.trace_format not in ("compact", "full", "structured"):
+            raise ValueError(f"trace_format must be 'compact', 'full', or 'structured', got: {self.trace_format}")
+    
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> EvalConfig:
+        """Create EvalConfig from a dictionary (e.g. from TOML).
+        
+        Args:
+            data: Dictionary with eval configuration
+            
+        Returns:
+            Validated EvalConfig instance
+        """
+        # Extract known fields
+        config_dict = {
+            "app_id": data.get("app_id"),
+            "model": data.get("model"),
+            "seeds": data.get("seeds", []),
+            "task_app_url": data.get("task_app_url"),
+            "split": data.get("split", "train"),
+            "max_turns": data.get("max_turns"),
+            "max_llm_calls": data.get("max_llm_calls", 10),
+            "concurrency": data.get("concurrency", 1),
+            "env_name": data.get("env_name"),
+            "policy_name": data.get("policy_name"),
+            "trace_format": data.get("trace_format", "compact"),
+            "return_trace": data.get("return_trace", False),
+            "env_config": data.get("env_config", {}),
+            "policy_config": data.get("policy_config", {}),
+            "metadata": data.get("metadata", {}),
+            "metadata_sql": data.get("metadata_sql"),
+        }
+        
+        return cls(**config_dict)
+
+
+@dataclass(slots=True)
+class FilterConfig:
+    """Configuration for 'synth-ai filter' command.
+    
+    Validates and provides defaults for filtering traces into SFT datasets.
+    """
+    
+    # Required: Database path or URL
+    db: str
+    
+    # Required: Output JSONL path
+    output: str
+    
+    # Optional: Filter by data splits
+    splits: list[str] = field(default_factory=list)
+    
+    # Optional: Filter by task IDs
+    task_ids: list[str] = field(default_factory=list)
+    
+    # Optional: Filter by models
+    models: list[str] = field(default_factory=list)
+    
+    # Optional: Minimum official score threshold
+    min_official_score: float | None = None
+    
+    # Optional: Maximum official score threshold
+    max_official_score: float | None = None
+    
+    # Optional: Minimum judge scores (judge_name -> min_score)
+    min_judge_scores: dict[str, float] = field(default_factory=dict)
+    
+    # Optional: Maximum judge scores (judge_name -> max_score)
+    max_judge_scores: dict[str, float] = field(default_factory=dict)
+    
+    # Optional: Limit number of examples
+    limit: int | None = None
+    
+    # Optional: Offset for pagination
+    offset: int | None = None
+    
+    # Optional: Whether to shuffle results
+    shuffle: bool = False
+    
+    # Optional: Random seed for shuffling
+    shuffle_seed: int | None = None
+    
+    def __post_init__(self):
+        """Validate configuration after initialization."""
+        if not self.db:
+            raise ValueError("db (database path or URL) is required")
+        
+        if not self.output:
+            raise ValueError("output (JSONL file path) is required")
+        
+        # Validate output has .jsonl extension
+        output_path = Path(self.output)
+        if output_path.suffix.lower() not in (".jsonl", ".json"):
+            raise ValueError(f"output must be a .jsonl or .json file, got: {self.output}")
+        
+        # Validate score thresholds
+        if (
+            self.min_official_score is not None
+            and self.max_official_score is not None
+            and self.min_official_score > self.max_official_score
+        ):
+            raise ValueError("min_official_score cannot be greater than max_official_score")
+        
+        # Validate limit/offset
+        if self.limit is not None and self.limit < 1:
+            raise ValueError("limit must be >= 1")
+        
+        if self.offset is not None and self.offset < 0:
+            raise ValueError("offset must be >= 0")
+        
+        # Validate shuffle seed requires shuffle
+        if self.shuffle_seed is not None and not self.shuffle:
+            raise ValueError("shuffle_seed requires shuffle=true")
+    
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> FilterConfig:
+        """Create FilterConfig from a dictionary (e.g. from TOML).
+        
+        Args:
+            data: Dictionary with filter configuration
+            
+        Returns:
+            Validated FilterConfig instance
+        """
+        # Extract known fields
+        config_dict = {
+            "db": data.get("db"),
+            "output": data.get("output"),
+            "splits": data.get("splits", []),
+            "task_ids": data.get("task_ids", []),
+            "models": data.get("models", []),
+            "min_official_score": data.get("min_official_score"),
+            "max_official_score": data.get("max_official_score"),
+            "min_judge_scores": data.get("min_judge_scores", {}),
+            "max_judge_scores": data.get("max_judge_scores", {}),
+            "limit": data.get("limit"),
+            "offset": data.get("offset"),
+            "shuffle": data.get("shuffle", False),
+            "shuffle_seed": data.get("shuffle_seed"),
+        }
+        
+        return cls(**config_dict)
+    
+    def get_db_url(self) -> str:
+        """Convert db path to proper SQLite URL if needed.
+        
+        Returns:
+            Database URL suitable for SQLAlchemy/aiosqlite
+        """
+        db_value = self.db.strip()
+        if "://" in db_value:
+            return db_value
+        else:
+            db_path = Path(db_value).expanduser().resolve()
+            # Ensure parent directory exists
+            db_path.parent.mkdir(parents=True, exist_ok=True)
+            return f"sqlite+aiosqlite:///{db_path}"
+    
+    def get_output_path(self) -> Path:
+        """Get resolved output path with parent directory created.
+        
+        Returns:
+            Resolved Path object with parent directory created
+        """
+        output_path = Path(self.output).expanduser().resolve()
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+        return output_path
+
+
+

synth_ai/sdk/task/contracts.py

@@ -0,0 +1,219 @@
+"""Contracts for Task Apps.
+
+Prefer synth_ai.sdk.localapi.contracts moving forward. This module remains for
+backward compatibility during the naming transition.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from enum import Enum
+from typing import Any, Literal
+
+from pydantic import BaseModel, ConfigDict, Field
+
+
+class RolloutMode(str, Enum):
+    """Mode controls how rollout infrastructure processes inference URLs."""
+    RL = "rl"
+    EVAL = "eval"
+
+
+class OutputMode(str, Enum):
+    """Controls how the policy expects model outputs.
+
+    - TOOL_CALLS: Use function/tool calling (default, current behavior)
+    - TEXT: Plain text in message.content
+    - STRUCTURED: JSON via response_format (OpenAI json_schema, Groq json_object, Gemini responseSchema)
+    """
+    TOOL_CALLS = "tool_calls"
+    TEXT = "text"
+    STRUCTURED = "structured"
+
+
+class StructuredOutputConfig(BaseModel):
+    """Configuration for structured output mode (OutputMode.STRUCTURED).
+
+    Defines the JSON schema that the model must conform to when using structured outputs.
+    This is normalized across providers:
+    - OpenAI: response_format.json_schema
+    - Groq: response_format.json_schema or json_object
+    - Gemini: generationConfig.responseSchema
+    """
+    schema: dict[str, Any] = Field(
+        ...,
+        description="JSON Schema for the expected response structure"
+    )
+    schema_name: str = Field(
+        default="response",
+        description="Name for the schema (required by some providers)"
+    )
+    strict: bool = Field(
+        default=True,
+        description="Whether to enforce strict schema validation (OpenAI strict mode)"
+    )
+
+
+@dataclass(frozen=True)
+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.
+    """
+
+    root: str = "/"
+    health: str = "/health"
+    info: str = "/info"
+    task_info: str = "/task_info"
+    rollout: str = "/rollout"
+
+
+@dataclass(frozen=True)
+class LocalAPIEndpoints(TaskAppEndpoints):
+    """Alias for TaskAppEndpoints with LocalAPI naming."""
+
+
+# --- Unified rollout schema used by Task App services and SDK utilities ---
+
+
+class RolloutEnvSpec(BaseModel):
+    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: str | None = None
+    policy_name: str | None = None
+    config: dict[str, Any] = Field(default_factory=dict)
+
+    # Output mode configuration (defaults to tool_calls for backward compatibility)
+    output_mode: OutputMode = Field(
+        default=OutputMode.TOOL_CALLS,
+        description="How the policy expects model outputs: tool_calls, text, or structured"
+    )
+    structured_config: StructuredOutputConfig | None = Field(
+        default=None,
+        description="Configuration for structured output mode (required if output_mode=STRUCTURED)"
+    )
+
+
+class RolloutRecordConfig(BaseModel):
+    logprobs: bool = False
+    value: bool = False
+    return_trace: bool = False
+    trace_format: Literal["compact", "full", "structured"] = "compact"
+
+
+class RolloutSafetyConfig(BaseModel):
+    max_time_s: float = 3600.0
+
+
+class RolloutRequest(BaseModel):
+    run_id: str
+    env: RolloutEnvSpec
+    policy: RolloutPolicySpec
+    record: RolloutRecordConfig = RolloutRecordConfig()
+    on_done: str = "reset"
+    safety: RolloutSafetyConfig = RolloutSafetyConfig()
+    training_session_id: str | None = None
+    synth_base_url: str | None = None
+    mode: RolloutMode = RolloutMode.RL  # Default to RL mode for training/optimization
+
+
+class RolloutMetrics(BaseModel):
+    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 (trace-only)."""
+    run_id: str
+    branches: dict[str, list[str]] = Field(default_factory=dict)
+    metrics: RolloutMetrics
+    aborted: bool = False
+    trace_correlation_id: str | None = None
+    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).",
+    )