spanforge 2.0.0__py3-none-any.whl

spanforge/namespaces/trace.py

@@ -0,0 +1,1006 @@
+"""spanforge.namespaces.trace — Span and agent payload types (RFC-0001 §8).
+
+This module provides Python dataclasses for the ``llm.trace.*`` namespace.
+All types map directly to the JSON Schema defined in
+``docs/schema/payloads/span.schema.json``,
+``docs/schema/payloads/agent-step.schema.json``, and
+``docs/schema/payloads/agent-run.schema.json``.
+
+Shared value objects (``TokenUsage``, ``ModelInfo``, ``CostBreakdown``,
+``PricingTier``, ``ToolCall``, ``ReasoningStep``, ``DecisionPoint``) are
+defined here because the trace namespace is where they are first introduced
+and other namespaces reference them.
+
+Enumerations
+------------
+GenAISystem
+    RFC §10.1 — normalised provider identifier (OTel ``gen_ai.system``).
+GenAIOperationName
+    RFC §10.2 — type of LLM operation (OTel ``gen_ai.operation.name``).
+SpanKind
+    RFC §10.3 — OTel SpanKind values relevant to LLM operations.
+
+Value objects
+-------------
+TokenUsage
+    RFC §9.1 — token counts with OTel-aligned field names.
+ModelInfo
+    RFC §9.2 — model identity and provider.
+CostBreakdown
+    RFC §9.3 — typed cost attribution record.
+PricingTier
+    RFC §9.4 — pricing rates snapshot for cost reproduction.
+ToolCall
+    RFC §8.1 — single tool invocation within a span.
+ReasoningStep
+    RFC §8.2 — chain-of-thought unit; raw content MUST NOT be stored.
+DecisionPoint
+    RFC §8.3 — explicit branching decision made by an agent.
+
+Payload dataclasses
+-------------------
+SpanPayload
+    RFC §8.1 — single unit of LLM work (model call, tool, agent invocation).
+AgentStepPayload
+    RFC §8.4 — one iteration of a multi-step agent loop.
+AgentRunPayload
+    RFC §8.5 — root summary for a complete agent run.
+"""
+
+from __future__ import annotations
+
+import re
+import time
+from dataclasses import dataclass, field
+from enum import Enum
+from typing import Any
+
+__all__ = [
+    "AgentRunPayload",
+    "AgentStepPayload",
+    "CostBreakdown",
+    "DecisionPoint",
+    "GenAIOperationName",
+    # Enumerations
+    "GenAISystem",
+    "ModelInfo",
+    "PricingTier",
+    "ReasoningStep",
+    "SpanEvent",
+    "SpanKind",
+    # Payloads
+    "SpanPayload",
+    # Value objects
+    "TokenUsage",
+    "ToolCall",
+]
+
+# ---------------------------------------------------------------------------
+# Compiled validation patterns (module-level, reused across instances)
+# ---------------------------------------------------------------------------
+_SPAN_ID_RE = re.compile(r"^[0-9a-f]{16}$")
+_TRACE_ID_RE = re.compile(r"^[0-9a-f]{32}$")
+_SHA256_RE = re.compile(r"^[0-9a-f]{64}$")
+_ISO_DATE_RE = re.compile(r"^\d{4}-\d{2}-\d{2}$")
+_CURRENCY_RE = re.compile(r"^[A-Z]{3}$")
+# RFC-0001 §8.1 — session_id / user_id: 1–256 non-whitespace printable chars
+_SAFE_ID_RE = re.compile(r"^\S{1,256}$")
+
+
+# ---------------------------------------------------------------------------
+# Enumerations (RFC §10)
+# ---------------------------------------------------------------------------
+
+class GenAISystem(str, Enum):
+    """RFC-0001 §10.1 — LLM provider identifier.
+
+    Maps directly to OTel ``gen_ai.system`` semantic convention values.
+    Use ``CUSTOM`` (``"_custom"``) for private or enterprise deployments;
+    MUST set ``ModelInfo.custom_system_name`` when ``CUSTOM`` is used.
+    """
+
+    OPENAI = "openai"
+    ANTHROPIC = "anthropic"
+    GOOGLE = "google"
+    COHERE = "cohere"
+    VERTEX_AI = "vertex_ai"
+    AWS_BEDROCK = "aws_bedrock"
+    AZ_AI_INFERENCE = "az.ai.inference"
+    GROQ = "groq"
+    OLLAMA = "ollama"
+    MISTRAL_AI = "mistral_ai"
+    TOGETHER_AI = "together_ai"
+    HUGGING_FACE = "hugging_face"
+    CUSTOM = "_custom"
+
+
+class GenAIOperationName(str, Enum):
+    """RFC-0001 §10.2 — Type of LLM operation performed.
+
+    Maps to OTel ``gen_ai.operation.name``.
+    """
+
+    CHAT = "chat"
+    TEXT_COMPLETION = "text_completion"
+    EMBEDDINGS = "embeddings"
+    IMAGE_GENERATION = "image_generation"
+    EXECUTE_TOOL = "execute_tool"
+    INVOKE_AGENT = "invoke_agent"
+    CREATE_AGENT = "create_agent"
+    REASONING = "reasoning"
+
+
+class SpanKind(str, Enum):
+    """RFC-0001 §10.3 — OTel SpanKind for LLM operations."""
+
+    CLIENT = "CLIENT"      # Outbound LLM API call — most common
+    SERVER = "SERVER"      # Incoming agent request
+    INTERNAL = "INTERNAL"  # Internal reasoning or routing step
+    CONSUMER = "CONSUMER"  # Tool execution triggered by LLM output
+    PRODUCER = "PRODUCER"  # Event emitted by an agent for downstream consumption
+
+
+# ---------------------------------------------------------------------------
+# Value objects (RFC §9, §8.1-§8.3)
+# ---------------------------------------------------------------------------
+
+@dataclass(frozen=True)
+class TokenUsage:
+    """RFC-0001 §9.1 — Token consumption record for a model call.
+
+    Uses OTel-aligned names: ``input_tokens`` / ``output_tokens`` (not
+    ``prompt_tokens`` / ``completion_tokens``).
+    """
+
+    input_tokens: int
+    output_tokens: int
+    total_tokens: int
+    cached_tokens: int | None = None
+    cache_creation_tokens: int | None = None
+    reasoning_tokens: int | None = None
+    image_tokens: int | None = None
+
+    def __post_init__(self) -> None:
+        for name in ("input_tokens", "output_tokens", "total_tokens"):
+            v = getattr(self, name)
+            if not isinstance(v, int) or v < 0:
+                raise ValueError(f"TokenUsage.{name} must be a non-negative int")
+        for opt in ("cached_tokens", "cache_creation_tokens", "reasoning_tokens", "image_tokens"):
+            v = getattr(self, opt)
+            if v is not None and (not isinstance(v, int) or v < 0):
+                raise ValueError(f"TokenUsage.{opt} must be a non-negative int or None")
+
+    def to_dict(self) -> dict[str, Any]:
+        """Serialise the payload to a plain ``dict``."""
+        d: dict[str, Any] = {
+            "input_tokens": self.input_tokens,
+            "output_tokens": self.output_tokens,
+            "total_tokens": self.total_tokens,
+        }
+        for name in ("cached_tokens", "cache_creation_tokens", "reasoning_tokens", "image_tokens"):
+            v = getattr(self, name)
+            if v is not None:
+                d[name] = v
+        return d
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> TokenUsage:
+        """Deserialise from a plain ``dict``."""
+        return cls(
+            input_tokens=int(data["input_tokens"]),
+            output_tokens=int(data["output_tokens"]),
+            total_tokens=int(data["total_tokens"]),
+            cached_tokens=int(data["cached_tokens"]) if "cached_tokens" in data else None,
+            cache_creation_tokens=int(data["cache_creation_tokens"]) if "cache_creation_tokens" in data else None,  # noqa: E501
+            reasoning_tokens=int(data["reasoning_tokens"]) if "reasoning_tokens" in data else None,
+            image_tokens=int(data["image_tokens"]) if "image_tokens" in data else None,
+        )
+
+
+@dataclass(frozen=True)
+class ModelInfo:
+    """RFC-0001 §9.2 — Model identity and provider information.
+
+    ``custom_system_name`` is REQUIRED when ``system`` is
+    :attr:`GenAISystem.CUSTOM` (``"_custom"``).
+    """
+
+    system: GenAISystem | str
+    name: str
+    response_model: str | None = None
+    version: str | None = None
+    custom_system_name: str | None = None
+
+    def __post_init__(self) -> None:
+        sys_val = self.system.value if isinstance(self.system, GenAISystem) else self.system
+        if sys_val == "_custom" and not self.custom_system_name:
+            raise ValueError(
+                "ModelInfo.custom_system_name is REQUIRED when system is '_custom' (RFC §4 P6)"
+            )
+        if not isinstance(self.name, str) or not self.name:
+            raise ValueError("ModelInfo.name must be a non-empty string")
+
+    def to_dict(self) -> dict[str, Any]:
+        """Serialise the payload to a plain ``dict``."""
+        sys_val = self.system.value if isinstance(self.system, GenAISystem) else self.system
+        d: dict[str, Any] = {"system": sys_val, "name": self.name}
+        if self.response_model is not None:
+            d["response_model"] = self.response_model
+        if self.version is not None:
+            d["version"] = self.version
+        if self.custom_system_name is not None:
+            d["custom_system_name"] = self.custom_system_name
+        return d
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> ModelInfo:
+        """Deserialise from a plain ``dict``."""
+        sys_raw = data["system"]
+        try:
+            system: GenAISystem | str = GenAISystem(sys_raw)
+        except ValueError:
+            system = sys_raw
+        return cls(
+            system=system,
+            name=data["name"],
+            response_model=data.get("response_model"),
+            version=data.get("version"),
+            custom_system_name=data.get("custom_system_name"),
+        )
+
+
+@dataclass(frozen=True)
+class CostBreakdown:
+    """RFC-0001 §9.3 — Typed cost attribution record.
+
+    ``total_cost_usd`` MUST equal
+    ``input_cost_usd + output_cost_usd + reasoning_cost_usd - cached_discount_usd``
+    within ±1e-6 absolute tolerance.
+
+    Use :meth:`zero` to create a zero-filled instance when pricing data
+    is unavailable (§9.3 zero-fill allowance).
+    """
+
+    input_cost_usd: float
+    output_cost_usd: float
+    total_cost_usd: float
+    cached_discount_usd: float = 0.0
+    reasoning_cost_usd: float = 0.0
+    currency: str = "USD"
+    pricing_date: str | None = None
+
+    _TOLERANCE: float = 1e-6
+
+    def __post_init__(self) -> None:
+        for name in ("input_cost_usd", "output_cost_usd", "total_cost_usd",
+                     "cached_discount_usd", "reasoning_cost_usd"):
+            v = getattr(self, name)
+            if not isinstance(v, (int, float)) or v < 0:
+                raise ValueError(f"CostBreakdown.{name} must be a non-negative number")
+        if not _CURRENCY_RE.match(self.currency):
+            raise ValueError("CostBreakdown.currency must be a 3-letter ISO 4217 code")
+        if self.pricing_date is not None and not _ISO_DATE_RE.match(self.pricing_date):
+            raise ValueError("CostBreakdown.pricing_date must be YYYY-MM-DD")
+        expected = (
+            self.input_cost_usd
+            + self.output_cost_usd
+            + self.reasoning_cost_usd
+            - self.cached_discount_usd
+        )
+        if abs(self.total_cost_usd - expected) > self._TOLERANCE:
+            raise ValueError(
+                f"CostBreakdown.total_cost_usd {self.total_cost_usd} != "
+                f"input + output + reasoning - cached_discount = {expected:.8f} (±{self._TOLERANCE})"  # noqa: E501
+            )
+
+    def to_dict(self) -> dict[str, Any]:
+        """Serialise the payload to a plain ``dict``."""
+        d: dict[str, Any] = {
+            "input_cost_usd": self.input_cost_usd,
+            "output_cost_usd": self.output_cost_usd,
+            "total_cost_usd": self.total_cost_usd,
+        }
+        if self.cached_discount_usd:
+            d["cached_discount_usd"] = self.cached_discount_usd
+        if self.reasoning_cost_usd:
+            d["reasoning_cost_usd"] = self.reasoning_cost_usd
+        if self.currency != "USD":
+            d["currency"] = self.currency
+        if self.pricing_date is not None:
+            d["pricing_date"] = self.pricing_date
+        return d
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> CostBreakdown:
+        """Deserialise from a plain ``dict``."""
+        return cls(
+            input_cost_usd=float(data["input_cost_usd"]),
+            output_cost_usd=float(data["output_cost_usd"]),
+            total_cost_usd=float(data["total_cost_usd"]),
+            cached_discount_usd=float(data.get("cached_discount_usd", 0.0)),
+            reasoning_cost_usd=float(data.get("reasoning_cost_usd", 0.0)),
+            currency=data.get("currency", "USD"),
+            pricing_date=data.get("pricing_date"),
+        )
+
+    @classmethod
+    def zero(cls) -> CostBreakdown:
+        """Return a zero-filled CostBreakdown (§9.3 zero-fill allowance)."""
+        return cls(input_cost_usd=0.0, output_cost_usd=0.0, total_cost_usd=0.0)
+
+
+@dataclass(frozen=True)
+class PricingTier:
+    """RFC-0001 §9.4 — Pricing rates snapshot for cost reproduction.
+
+    Stores the exact pricing rates used to compute a ``CostBreakdown`` so
+    cost calculations remain reproducible indefinitely.
+    """
+
+    system: GenAISystem | str
+    model: str
+    input_per_million_usd: float
+    output_per_million_usd: float
+    effective_date: str
+    cached_input_per_million_usd: float | None = None
+    reasoning_per_million_usd: float | None = None
+
+    def __post_init__(self) -> None:
+        if not isinstance(self.model, str) or not self.model:
+            raise ValueError("PricingTier.model must be a non-empty string")
+        if not _ISO_DATE_RE.match(self.effective_date):
+            raise ValueError("PricingTier.effective_date must be YYYY-MM-DD")
+        for name in ("input_per_million_usd", "output_per_million_usd"):
+            v = getattr(self, name)
+            if not isinstance(v, (int, float)) or v < 0:
+                raise ValueError(f"PricingTier.{name} must be a non-negative number")
+
+    def to_dict(self) -> dict[str, Any]:
+        """Serialise the payload to a plain ``dict``."""
+        sys_val = self.system.value if isinstance(self.system, GenAISystem) else self.system
+        d: dict[str, Any] = {
+            "system": sys_val,
+            "model": self.model,
+            "input_per_million_usd": self.input_per_million_usd,
+            "output_per_million_usd": self.output_per_million_usd,
+            "effective_date": self.effective_date,
+        }
+        if self.cached_input_per_million_usd is not None:
+            d["cached_input_per_million_usd"] = self.cached_input_per_million_usd
+        if self.reasoning_per_million_usd is not None:
+            d["reasoning_per_million_usd"] = self.reasoning_per_million_usd
+        return d
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> PricingTier:
+        """Deserialise from a plain ``dict``."""
+        sys_raw = data["system"]
+        try:
+            system: GenAISystem | str = GenAISystem(sys_raw)
+        except ValueError:
+            system = sys_raw
+        return cls(
+            system=system,
+            model=data["model"],
+            input_per_million_usd=float(data["input_per_million_usd"]),
+            output_per_million_usd=float(data["output_per_million_usd"]),
+            effective_date=data["effective_date"],
+            cached_input_per_million_usd=float(data["cached_input_per_million_usd"])
+            if "cached_input_per_million_usd" in data else None,
+            reasoning_per_million_usd=float(data["reasoning_per_million_usd"])
+            if "reasoning_per_million_usd" in data else None,
+        )
+
+
+@dataclass
+class SpanEvent:
+    """A named, timestamped event within a span.
+
+    Unlike spans (which represent durations), events are instantaneous
+    points in time recorded inside an enclosing span.
+
+    Attributes:
+        name:         Event name (e.g. ``"cache.hit"``, ``"retry.attempt.2"``).
+        timestamp_ns: Nanosecond-precision Unix timestamp (auto-filled on creation).
+        metadata:     Arbitrary key-value metadata attached to this event.
+    """
+
+    name: str
+    timestamp_ns: int = field(default_factory=time.time_ns)
+    metadata: dict[str, Any] = field(default_factory=dict)
+
+    def __post_init__(self) -> None:
+        if not isinstance(self.name, str) or not self.name:
+            raise ValueError("SpanEvent.name must be a non-empty string")
+        if self.timestamp_ns < 0:
+            raise ValueError("SpanEvent.timestamp_ns must be non-negative")
+
+    def to_dict(self) -> dict[str, Any]:
+        """Serialise to a plain dict."""
+        return {
+            "name": self.name,
+            "timestamp_ns": self.timestamp_ns,
+            "metadata": self.metadata,
+        }
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> SpanEvent:
+        """Deserialise from a plain dict."""
+        return cls(
+            name=data["name"],
+            timestamp_ns=int(data["timestamp_ns"]),
+            metadata=dict(data.get("metadata", {})),
+        )
+
+
+@dataclass(frozen=True)
+class ToolCall:
+    """RFC-0001 §8.1 — A single tool invocation within a span.
+
+    ``arguments_hash`` stores a SHA-256 hash of the canonical JSON of
+    arguments.  Raw argument values SHOULD NOT be stored by default (§20.4);
+    set ``SpanForgeConfig.include_raw_tool_io = True`` to opt in.
+    """
+
+    tool_call_id: str
+    function_name: str
+    status: str  # "success" | "error" | "timeout" | "cancelled"
+    arguments_hash: str | None = None  # 64 lowercase hex chars, no prefix
+    error_type: str | None = None
+    duration_ms: float | None = None
+    arguments_raw: str | None = None   # populated only when include_raw_tool_io=True
+    result_raw: str | None = None      # populated only when include_raw_tool_io=True
+    retry_count: int | None = None
+    external_api: str | None = None
+
+    _VALID_STATUSES = frozenset({"success", "error", "timeout", "cancelled"})
+
+    def __post_init__(self) -> None:
+        if not isinstance(self.tool_call_id, str) or not self.tool_call_id:
+            raise ValueError("ToolCall.tool_call_id must be a non-empty string")
+        if not isinstance(self.function_name, str) or not self.function_name:
+            raise ValueError("ToolCall.function_name must be a non-empty string")
+        if self.status not in self._VALID_STATUSES:
+            raise ValueError(f"ToolCall.status must be one of {sorted(self._VALID_STATUSES)}")
+        if self.arguments_hash is not None and not _SHA256_RE.match(self.arguments_hash):
+            raise ValueError("ToolCall.arguments_hash must be 64 lowercase hex chars (SHA-256)")
+        if self.duration_ms is not None and self.duration_ms < 0:
+            raise ValueError("ToolCall.duration_ms must be non-negative")
+        if self.retry_count is not None and self.retry_count < 0:
+            raise ValueError("ToolCall.retry_count must be a non-negative integer")
+
+    def to_dict(self) -> dict[str, Any]:
+        """Serialise the payload to a plain ``dict``."""
+        d: dict[str, Any] = {
+            "tool_call_id": self.tool_call_id,
+            "function_name": self.function_name,
+            "status": self.status,
+        }
+        if self.arguments_hash is not None:
+            d["arguments_hash"] = self.arguments_hash
+        if self.error_type is not None:
+            d["error_type"] = self.error_type
+        if self.duration_ms is not None:
+            d["duration_ms"] = self.duration_ms
+        # arguments_raw / result_raw contain raw PII; only emit when the
+        # operator has explicitly enabled raw tool I/O capture.
+        from spanforge.config import get_config  # noqa: PLC0415
+        if self.arguments_raw is not None and get_config().include_raw_tool_io:
+            d["arguments_raw"] = self.arguments_raw
+        if self.result_raw is not None and get_config().include_raw_tool_io:
+            d["result_raw"] = self.result_raw
+        if self.retry_count is not None:
+            d["retry_count"] = self.retry_count
+        if self.external_api is not None:
+            d["external_api"] = self.external_api
+        return d
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> ToolCall:
+        """Deserialise from a plain ``dict``."""
+        return cls(
+            tool_call_id=data["tool_call_id"],
+            function_name=data["function_name"],
+            status=data["status"],
+            arguments_hash=data.get("arguments_hash"),
+            error_type=data.get("error_type"),
+            duration_ms=float(data["duration_ms"]) if "duration_ms" in data else None,
+            arguments_raw=data.get("arguments_raw"),
+            result_raw=data.get("result_raw"),
+            retry_count=int(data["retry_count"]) if "retry_count" in data else None,
+            external_api=data.get("external_api"),
+        )
+
+
+@dataclass(frozen=True)
+class ReasoningStep:
+    """RFC-0001 §8.2 — A discrete chain-of-thought unit.
+
+    **Critical:** Raw reasoning content MUST NOT be stored — only the
+    SHA-256 ``content_hash`` (64 lowercase hex chars, no prefix) MAY be
+    stored.
+    """
+
+    step_index: int
+    reasoning_tokens: int
+    duration_ms: float | None = None
+    content_hash: str | None = None  # 64 lowercase hex chars, no prefix
+
+    def __post_init__(self) -> None:
+        if not isinstance(self.step_index, int) or self.step_index < 0:
+            raise ValueError("ReasoningStep.step_index must be a non-negative int")
+        if not isinstance(self.reasoning_tokens, int) or self.reasoning_tokens < 0:
+            raise ValueError("ReasoningStep.reasoning_tokens must be a non-negative int")
+        if self.duration_ms is not None and self.duration_ms < 0:
+            raise ValueError("ReasoningStep.duration_ms must be non-negative")
+        if self.content_hash is not None and not _SHA256_RE.match(self.content_hash):
+            raise ValueError(
+                "ReasoningStep.content_hash must be 64 lowercase hex chars (SHA-256, no prefix)"
+            )
+
+    def to_dict(self) -> dict[str, Any]:
+        """Serialise the payload to a plain ``dict``."""
+        d: dict[str, Any] = {
+            "step_index": self.step_index,
+            "reasoning_tokens": self.reasoning_tokens,
+        }
+        if self.duration_ms is not None:
+            d["duration_ms"] = self.duration_ms
+        if self.content_hash is not None:
+            d["content_hash"] = self.content_hash
+        return d
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> ReasoningStep:
+        """Deserialise from a plain ``dict``."""
+        return cls(
+            step_index=int(data["step_index"]),
+            reasoning_tokens=int(data["reasoning_tokens"]),
+            duration_ms=float(data["duration_ms"]) if "duration_ms" in data else None,
+            content_hash=data.get("content_hash"),
+        )
+
+
+@dataclass(frozen=True)
+class DecisionPoint:
+    """RFC-0001 §8.3 — An explicit branching decision recorded during an agent step.
+
+    ``rationale`` is OPTIONAL for black-box models that do not expose reasoning.
+    """
+
+    decision_id: str
+    decision_type: str  # "tool_selection"|"route_choice"|"loop_termination"|"escalation"
+    options_considered: list[str]
+    chosen_option: str
+    rationale: str | None = None
+
+    _VALID_TYPES = frozenset({"tool_selection", "route_choice", "loop_termination", "escalation"})
+
+    def __post_init__(self) -> None:
+        if not isinstance(self.decision_id, str) or not self.decision_id:
+            raise ValueError("DecisionPoint.decision_id must be a non-empty string")
+        if self.decision_type not in self._VALID_TYPES:
+            raise ValueError(f"DecisionPoint.decision_type must be one of {sorted(self._VALID_TYPES)}")  # noqa: E501
+        if not self.options_considered:
+            raise ValueError("DecisionPoint.options_considered must be a non-empty list")
+        if not isinstance(self.chosen_option, str) or not self.chosen_option:
+            raise ValueError("DecisionPoint.chosen_option must be a non-empty string")
+
+    def to_dict(self) -> dict[str, Any]:
+        """Serialise the payload to a plain ``dict``."""
+        d: dict[str, Any] = {
+            "decision_id": self.decision_id,
+            "decision_type": self.decision_type,
+            "options_considered": list(self.options_considered),
+            "chosen_option": self.chosen_option,
+        }
+        if self.rationale is not None:
+            d["rationale"] = self.rationale
+        return d
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> DecisionPoint:
+        """Deserialise from a plain ``dict``."""
+        return cls(
+            decision_id=data["decision_id"],
+            decision_type=data["decision_type"],
+            options_considered=list(data["options_considered"]),
+            chosen_option=data["chosen_option"],
+            rationale=data.get("rationale"),
+        )
+
+
+# ---------------------------------------------------------------------------
+# Payload dataclasses
+# ---------------------------------------------------------------------------
+
+@dataclass(frozen=True)
+class SpanPayload:
+    """RFC-0001 §8.1 — A single unit of LLM work.
+
+    Used with event types: ``llm.trace.span.started``,
+    ``llm.trace.span.completed``, ``llm.trace.span.failed``,
+    ``llm.trace.reasoning.step``.
+    """
+
+    span_id: str           # 16 lowercase hex chars
+    trace_id: str          # 32 lowercase hex chars
+    span_name: str
+    operation: GenAIOperationName | str
+    span_kind: SpanKind | str
+    status: str            # "ok" | "error" | "timeout"
+    start_time_unix_nano: int
+    end_time_unix_nano: int
+    duration_ms: float
+    parent_span_id: str | None = None
+    agent_run_id: str | None = None
+    model: ModelInfo | None = None
+    token_usage: TokenUsage | None = None
+    cost: CostBreakdown | None = None
+    tool_calls: list[ToolCall] = field(default_factory=list)
+    reasoning_steps: list[ReasoningStep] = field(default_factory=list)
+    finish_reason: str | None = None
+    error: str | None = None
+    error_type: str | None = None
+    attributes: dict[str, Any] | None = None
+    temperature: float | None = None
+    top_p: float | None = None
+    max_tokens: int | None = None
+    error_category: str | None = None
+    events: list[SpanEvent] = field(default_factory=list)
+    session_id: str | None = None
+    user_id: str | None = None
+    incoming_traceparent: str | None = None  # W3C traceparent header from the caller
+
+    _VALID_STATUSES = frozenset({"ok", "error", "timeout"})
+
+    def __post_init__(self) -> None:
+        if not _SPAN_ID_RE.match(self.span_id):
+            raise ValueError(f"SpanPayload.span_id must be 16 lowercase hex chars, got {self.span_id!r}")  # noqa: E501
+        if not _TRACE_ID_RE.match(self.trace_id):
+            raise ValueError(f"SpanPayload.trace_id must be 32 lowercase hex chars, got {self.trace_id!r}")  # noqa: E501
+        if not isinstance(self.span_name, str) or not self.span_name:
+            raise ValueError("SpanPayload.span_name must be a non-empty string")
+        if self.parent_span_id is not None and not _SPAN_ID_RE.match(self.parent_span_id):
+            raise ValueError("SpanPayload.parent_span_id must be 16 lowercase hex chars")
+        status_val = self.status.value if isinstance(self.status, Enum) else self.status
+        if status_val not in self._VALID_STATUSES:
+            raise ValueError(f"SpanPayload.status must be one of {sorted(self._VALID_STATUSES)}")
+        if self.start_time_unix_nano < 0:
+            raise ValueError("SpanPayload.start_time_unix_nano must be non-negative")
+        if self.end_time_unix_nano < self.start_time_unix_nano:
+            raise ValueError("SpanPayload.end_time_unix_nano must be >= start_time_unix_nano")
+        if self.duration_ms < 0:
+            raise ValueError("SpanPayload.duration_ms must be non-negative")
+        # RFC-0001 §8.1 — duration_ms MUST equal (end - start) / 1_000_000 ± 1 ms
+        _computed_ms = (self.end_time_unix_nano - self.start_time_unix_nano) / 1_000_000
+        if abs(self.duration_ms - _computed_ms) > 1.0:
+            raise ValueError(
+                f"SpanPayload.duration_ms {self.duration_ms} must equal "
+                f"(end_time_unix_nano - start_time_unix_nano) / 1_000_000 "
+                f"= {_computed_ms:.3f} \u00b11 ms (RFC-0001 \u00a78.1)"
+            )
+        # Validate session_id / user_id: prevent PII/injection leakage via
+        # embedding whitespace, control chars, or excessively long values.
+        if self.session_id is not None and not _SAFE_ID_RE.match(self.session_id):
+            raise ValueError(
+                "SpanPayload.session_id must be 1–256 non-whitespace characters"
+            )
+        if self.user_id is not None and not _SAFE_ID_RE.match(self.user_id):
+            raise ValueError(
+                "SpanPayload.user_id must be 1–256 non-whitespace characters"
+            )
+
+    def _add_optional_fields(self, d: dict[str, Any]) -> None:
+        """Add optional fields to *d* when they are not None/empty."""
+        if self.parent_span_id is not None:
+            d["parent_span_id"] = self.parent_span_id
+        if self.agent_run_id is not None:
+            d["agent_run_id"] = self.agent_run_id
+        if self.model is not None:
+            d["model"] = self.model.to_dict()
+        if self.token_usage is not None:
+            d["token_usage"] = self.token_usage.to_dict()
+        if self.cost is not None:
+            d["cost"] = self.cost.to_dict()
+        if self.finish_reason is not None:
+            d["finish_reason"] = self.finish_reason
+        if self.error is not None:
+            d["error"] = self.error
+        if self.error_type is not None:
+            d["error_type"] = self.error_type
+        if self.attributes is not None:
+            d["attributes"] = self.attributes
+        if self.temperature is not None:
+            d["temperature"] = self.temperature
+        if self.top_p is not None:
+            d["top_p"] = self.top_p
+        if self.max_tokens is not None:
+            d["max_tokens"] = self.max_tokens
+        if self.error_category is not None:
+            d["error_category"] = self.error_category
+        if self.events:
+            d["events"] = [e.to_dict() for e in self.events]
+        if self.session_id is not None:
+            d["session_id"] = self.session_id
+        if self.user_id is not None:
+            d["user_id"] = self.user_id
+        if self.incoming_traceparent is not None:
+            d["incoming_traceparent"] = self.incoming_traceparent
+
+    def to_dict(self) -> dict[str, Any]:
+        """Serialise the payload to a plain ``dict``."""
+        op = self.operation.value if isinstance(self.operation, Enum) else self.operation
+        sk = self.span_kind.value if isinstance(self.span_kind, Enum) else self.span_kind
+        st = self.status.value if isinstance(self.status, Enum) else self.status
+        d: dict[str, Any] = {
+            "span_id": self.span_id,
+            "trace_id": self.trace_id,
+            "span_name": self.span_name,
+            "operation": op,
+            "span_kind": sk,
+            "status": st,
+            "start_time_unix_nano": self.start_time_unix_nano,
+            "end_time_unix_nano": self.end_time_unix_nano,
+            "duration_ms": self.duration_ms,
+            "tool_calls": [tc.to_dict() for tc in self.tool_calls],
+            "reasoning_steps": [rs.to_dict() for rs in self.reasoning_steps],
+        }
+        self._add_optional_fields(d)
+        return d
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> SpanPayload:
+        """Deserialise from a plain ``dict``."""
+        op_raw = data["operation"]
+        try:
+            operation: GenAIOperationName | str = GenAIOperationName(op_raw)
+        except ValueError:
+            operation = op_raw
+        sk_raw = data["span_kind"]
+        try:
+            span_kind: SpanKind | str = SpanKind(sk_raw)
+        except ValueError:
+            span_kind = sk_raw
+        return cls(
+            span_id=data["span_id"],
+            trace_id=data["trace_id"],
+            span_name=data["span_name"],
+            operation=operation,
+            span_kind=span_kind,
+            status=data["status"],
+            start_time_unix_nano=int(data["start_time_unix_nano"]),
+            end_time_unix_nano=int(data["end_time_unix_nano"]),
+            duration_ms=float(data["duration_ms"]),
+            parent_span_id=data.get("parent_span_id"),
+            agent_run_id=data.get("agent_run_id"),
+            model=ModelInfo.from_dict(data["model"]) if "model" in data else None,
+            token_usage=TokenUsage.from_dict(data["token_usage"]) if "token_usage" in data else None,  # noqa: E501
+            cost=CostBreakdown.from_dict(data["cost"]) if "cost" in data else None,
+            tool_calls=[ToolCall.from_dict(tc) for tc in data.get("tool_calls", [])],
+            reasoning_steps=[ReasoningStep.from_dict(rs) for rs in data.get("reasoning_steps", [])],
+            finish_reason=data.get("finish_reason"),
+            error=data.get("error"),
+            error_type=data.get("error_type"),
+            attributes=data.get("attributes"),
+            temperature=float(data["temperature"]) if "temperature" in data else None,
+            top_p=float(data["top_p"]) if "top_p" in data else None,
+            max_tokens=int(data["max_tokens"]) if "max_tokens" in data else None,
+            error_category=data.get("error_category"),
+            events=[SpanEvent.from_dict(e) for e in data.get("events", [])],
+            session_id=data.get("session_id"),
+            user_id=data.get("user_id"),
+            incoming_traceparent=data.get("incoming_traceparent"),
+        )
+
+
+@dataclass(frozen=True)
+class AgentStepPayload:
+    """RFC-0001 §8.4 — One iteration of a multi-step agent loop.
+
+    Used with event type: ``llm.trace.agent.step``.
+    ``step_index`` is zero-based.
+    """
+
+    agent_run_id: str
+    step_index: int
+    span_id: str           # 16 lowercase hex chars
+    trace_id: str          # 32 lowercase hex chars
+    operation: GenAIOperationName | str
+    tool_calls: list[ToolCall]
+    reasoning_steps: list[ReasoningStep]
+    decision_points: list[DecisionPoint]
+    status: str            # "ok" | "error" | "timeout"
+    start_time_unix_nano: int
+    end_time_unix_nano: int
+    duration_ms: float
+    parent_span_id: str | None = None
+    model: ModelInfo | None = None
+    token_usage: TokenUsage | None = None
+    cost: CostBreakdown | None = None
+    error: str | None = None
+    error_type: str | None = None
+    step_name: str | None = None
+
+    _VALID_STATUSES = frozenset({"ok", "error", "timeout"})
+
+    def __post_init__(self) -> None:
+        if not isinstance(self.agent_run_id, str) or not self.agent_run_id:
+            raise ValueError("AgentStepPayload.agent_run_id must be a non-empty string")
+        if not isinstance(self.step_index, int) or self.step_index < 0:
+            raise ValueError("AgentStepPayload.step_index must be a non-negative int")
+        if not _SPAN_ID_RE.match(self.span_id):
+            raise ValueError("AgentStepPayload.span_id must be 16 lowercase hex chars")
+        if not _TRACE_ID_RE.match(self.trace_id):
+            raise ValueError("AgentStepPayload.trace_id must be 32 lowercase hex chars")
+        if self.parent_span_id is not None and not _SPAN_ID_RE.match(self.parent_span_id):
+            raise ValueError("AgentStepPayload.parent_span_id must be 16 lowercase hex chars")
+        status_val = self.status.value if isinstance(self.status, Enum) else self.status
+        if status_val not in self._VALID_STATUSES:
+            raise ValueError(f"AgentStepPayload.status must be one of {sorted(self._VALID_STATUSES)}")  # noqa: E501
+        if self.start_time_unix_nano < 0:
+            raise ValueError("AgentStepPayload.start_time_unix_nano must be non-negative")
+        if self.end_time_unix_nano < self.start_time_unix_nano:
+            raise ValueError("AgentStepPayload.end_time_unix_nano must be >= start_time_unix_nano")
+        # RFC-0001 §8.1 — duration_ms MUST equal (end - start) / 1_000_000 ± 1 ms
+        _computed_ms = (self.end_time_unix_nano - self.start_time_unix_nano) / 1_000_000
+        if abs(self.duration_ms - _computed_ms) > 1.0:
+            raise ValueError(
+                f"AgentStepPayload.duration_ms {self.duration_ms} must equal "
+                f"(end_time_unix_nano - start_time_unix_nano) / 1_000_000 "
+                f"= {_computed_ms:.3f} \u00b11 ms (RFC-0001 \u00a78.1)"
+            )
+
+    def to_dict(self) -> dict[str, Any]:
+        """Serialise the payload to a plain ``dict``."""
+        op = self.operation.value if isinstance(self.operation, Enum) else self.operation
+        st = self.status.value if isinstance(self.status, Enum) else self.status
+        d: dict[str, Any] = {
+            "agent_run_id": self.agent_run_id,
+            "step_index": self.step_index,
+            "span_id": self.span_id,
+            "trace_id": self.trace_id,
+            "operation": op,
+            "tool_calls": [tc.to_dict() for tc in self.tool_calls],
+            "reasoning_steps": [rs.to_dict() for rs in self.reasoning_steps],
+            "decision_points": [dp.to_dict() for dp in self.decision_points],
+            "status": st,
+            "start_time_unix_nano": self.start_time_unix_nano,
+            "end_time_unix_nano": self.end_time_unix_nano,
+            "duration_ms": self.duration_ms,
+        }
+        if self.parent_span_id is not None:
+            d["parent_span_id"] = self.parent_span_id
+        if self.model is not None:
+            d["model"] = self.model.to_dict()
+        if self.token_usage is not None:
+            d["token_usage"] = self.token_usage.to_dict()
+        if self.cost is not None:
+            d["cost"] = self.cost.to_dict()
+        if self.error is not None:
+            d["error"] = self.error
+        if self.error_type is not None:
+            d["error_type"] = self.error_type
+        if self.step_name is not None:
+            d["step_name"] = self.step_name
+        return d
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> AgentStepPayload:
+        """Deserialise from a plain ``dict``."""
+        op_raw = data["operation"]
+        try:
+            operation: GenAIOperationName | str = GenAIOperationName(op_raw)
+        except ValueError:
+            operation = op_raw
+        return cls(
+            agent_run_id=data["agent_run_id"],
+            step_index=int(data["step_index"]),
+            span_id=data["span_id"],
+            trace_id=data["trace_id"],
+            operation=operation,
+            tool_calls=[ToolCall.from_dict(tc) for tc in data.get("tool_calls", [])],
+            reasoning_steps=[ReasoningStep.from_dict(rs) for rs in data.get("reasoning_steps", [])],
+            decision_points=[DecisionPoint.from_dict(dp) for dp in data.get("decision_points", [])],
+            status=data["status"],
+            start_time_unix_nano=int(data["start_time_unix_nano"]),
+            end_time_unix_nano=int(data["end_time_unix_nano"]),
+            duration_ms=float(data["duration_ms"]),
+            parent_span_id=data.get("parent_span_id"),
+            model=ModelInfo.from_dict(data["model"]) if "model" in data else None,
+            token_usage=TokenUsage.from_dict(data["token_usage"]) if "token_usage" in data else None,  # noqa: E501
+            cost=CostBreakdown.from_dict(data["cost"]) if "cost" in data else None,
+            error=data.get("error"),
+            error_type=data.get("error_type"),
+            step_name=data.get("step_name"),
+        )
+
+
+@dataclass(frozen=True)
+class AgentRunPayload:
+    """RFC-0001 §8.5 — Root-level summary for a complete agent run.
+
+    Used with event type: ``llm.trace.agent.completed``.
+    ``agent_run_id`` MUST match across all :class:`AgentStepPayload`
+    events for this run.
+    """
+
+    agent_run_id: str
+    agent_name: str
+    trace_id: str          # 32 lowercase hex chars
+    root_span_id: str      # 16 lowercase hex chars
+    total_steps: int
+    total_model_calls: int
+    total_tool_calls: int
+    total_token_usage: TokenUsage
+    total_cost: CostBreakdown
+    status: str            # "ok"|"error"|"timeout"|"max_steps_exceeded"
+    start_time_unix_nano: int
+    end_time_unix_nano: int
+    duration_ms: float
+    termination_reason: str | None = None
+
+    _VALID_STATUSES = frozenset({"ok", "error", "timeout", "max_steps_exceeded"})
+
+    def __post_init__(self) -> None:
+        if not isinstance(self.agent_run_id, str) or not self.agent_run_id:
+            raise ValueError("AgentRunPayload.agent_run_id must be a non-empty string")
+        if not isinstance(self.agent_name, str) or not self.agent_name:
+            raise ValueError("AgentRunPayload.agent_name must be a non-empty string")
+        if not _TRACE_ID_RE.match(self.trace_id):
+            raise ValueError("AgentRunPayload.trace_id must be 32 lowercase hex chars")
+        if not _SPAN_ID_RE.match(self.root_span_id):
+            raise ValueError("AgentRunPayload.root_span_id must be 16 lowercase hex chars")
+        for name in ("total_steps", "total_model_calls", "total_tool_calls"):
+            v = getattr(self, name)
+            if not isinstance(v, int) or v < 0:
+                raise ValueError(f"AgentRunPayload.{name} must be a non-negative int")
+        status_val = self.status.value if isinstance(self.status, Enum) else self.status
+        if status_val not in self._VALID_STATUSES:
+            raise ValueError(f"AgentRunPayload.status must be one of {sorted(self._VALID_STATUSES)}")  # noqa: E501
+        if self.start_time_unix_nano < 0:
+            raise ValueError("AgentRunPayload.start_time_unix_nano must be non-negative")
+        if self.end_time_unix_nano < self.start_time_unix_nano:
+            raise ValueError("AgentRunPayload.end_time_unix_nano must be >= start_time_unix_nano")
+
+    def to_dict(self) -> dict[str, Any]:
+        """Serialise the payload to a plain ``dict``."""
+        st = self.status.value if isinstance(self.status, Enum) else self.status
+        d: dict[str, Any] = {
+            "agent_run_id": self.agent_run_id,
+            "agent_name": self.agent_name,
+            "trace_id": self.trace_id,
+            "root_span_id": self.root_span_id,
+            "total_steps": self.total_steps,
+            "total_model_calls": self.total_model_calls,
+            "total_tool_calls": self.total_tool_calls,
+            "total_token_usage": self.total_token_usage.to_dict(),
+            "total_cost": self.total_cost.to_dict(),
+            "status": st,
+            "start_time_unix_nano": self.start_time_unix_nano,
+            "end_time_unix_nano": self.end_time_unix_nano,
+            "duration_ms": self.duration_ms,
+        }
+        if self.termination_reason is not None:
+            d["termination_reason"] = self.termination_reason
+        return d
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> AgentRunPayload:
+        """Deserialise from a plain ``dict``."""
+        return cls(
+            agent_run_id=data["agent_run_id"],
+            agent_name=data["agent_name"],
+            trace_id=data["trace_id"],
+            root_span_id=data["root_span_id"],
+            total_steps=int(data["total_steps"]),
+            total_model_calls=int(data["total_model_calls"]),
+            total_tool_calls=int(data["total_tool_calls"]),
+            total_token_usage=TokenUsage.from_dict(data["total_token_usage"]),
+            total_cost=CostBreakdown.from_dict(data["total_cost"]),
+            status=data["status"],
+            start_time_unix_nano=int(data["start_time_unix_nano"]),
+            end_time_unix_nano=int(data["end_time_unix_nano"]),
+            duration_ms=float(data["duration_ms"]),
+            termination_reason=data.get("termination_reason"),
+        )