@ngocsangairvds/vsaf 3.2.14 → 3.2.16

package/tools/vds-scripts/audit_orchestrator/src/vds_audit_orchestrator/agents/base.py

@@ -1,4035 +0,0 @@
-"""Base Agent Architecture (Phase 6)."""
-
-from __future__ import annotations
-
-import abc
-import asyncio
-import contextlib
-import json
-import os
-import random
-import re
-from collections.abc import Coroutine
-from dataclasses import dataclass, field
-from typing import TYPE_CHECKING, Any, TypeVar
-from urllib.parse import urlparse
-from uuid import uuid4
-
-from pydantic_ai.settings import ModelSettings
-from structlog import get_logger
-from vds_agent_core.profiles import resolve_default_failover_profiles
-
-from vds_audit_orchestrator.engine.provider_failure_classifier import (
-    ProviderFailureClass,
-    ProviderFailureClassifier,
-    ProviderHealthMemory,
-    RowFailoverContext,
-    TimeoutKind,
-    truncate_failover_profiles_considered,
-    truncate_row_failover_context,
-)
-from vds_audit_orchestrator.errors import AuditError
-from vds_audit_orchestrator.llm.cost_tracker import global_tracker
-from vds_audit_orchestrator.llm.provider import (
-    LLMProtocolType,
-    LLMSettings,
-    build_structured_output_extra_body,
-    resolve_openai_protocol_api_key,
-    resolve_protocol_api_key,
-    resolve_protocol_base_url,
-    should_suppress_anthropic_tool_choice,
-    should_suppress_openai_tool_choice,
-)
-from vds_audit_orchestrator.logging_config import log_context
-from vds_audit_orchestrator.observability.metrics import record_llm_usage
-from vds_audit_orchestrator.runtime_profiles import inherit_runtime_llm_policy
-
-if TYPE_CHECKING:
-    from vds_audit_orchestrator.models.evidence import EvidenceBundle
-    from vds_audit_orchestrator.models.task import AuditTask
-
-logger = get_logger()
-
-ResultType = TypeVar("ResultType")
-
-OPENAI_CODEX_PROTOCOL = "openai-codex"
-OPENAI_CODEX_BASELINE_MODELS: frozenset[str] = frozenset({"gpt-5.4", "gpt-5.3-codex", "gpt-5.2-codex"})
-PYDANTIC_AI_MODEL_SUPPORTED_PROTOCOLS: frozenset[LLMProtocolType] = frozenset(LLMProtocolType)
-PYDANTIC_AI_MODEL_SUPPORTED_PROTOCOL_VALUES: frozenset[str] = frozenset(
-    {item.value for item in PYDANTIC_AI_MODEL_SUPPORTED_PROTOCOLS}.union({OPENAI_CODEX_PROTOCOL})
-)
-
-
-@dataclass
-class AgentResult:
-    """Standard output from an agent."""
-
-    agent_name: str
-    success: bool
-    findings: list[dict[str, Any]] = field(default_factory=list)
-    metadata: dict[str, Any] = field(default_factory=dict)
-    error: AuditError | str | None = None
-    spawned_tasks: list[dict[str, Any]] = field(default_factory=list)
-
-
-@dataclass(frozen=True, slots=True)
-class AgentTimeoutPolicy:
-    """Runtime timeout budget and extension policy for agent model calls."""
-
-    initial_timeout_seconds: float
-    max_timeout_seconds: float
-    extension_seconds: float
-    extension_attempts: int
-    heartbeat_seconds: float
-    idle_post_tool_seconds: float
-    lease_window_seconds: float = 0.0
-    stall_detection_seconds: float = 30.0
-    absolute_timeout_seconds: float = 0.0
-
-
-@dataclass(frozen=True, slots=True)
-class _AgentRuntimeSnapshot:
-    settings: LLMSettings
-    pydantic_model: Any | None
-    model_settings: ModelSettings | None
-    client: Any | None
-
-
-@dataclass(slots=True)
-class _PydanticStreamResultWrapper:
-    """Expose stream results with RunResult-compatible accessors."""
-
-    output: Any
-    streamed_result: Any
-
-    def usage(self) -> Any:
-        return self.streamed_result.usage()
-
-    def all_messages(self, *, output_tool_return_content: str | None = None) -> list[Any]:
-        return self.streamed_result.all_messages(output_tool_return_content=output_tool_return_content)
-
-    def all_messages_json(self, *, output_tool_return_content: str | None = None) -> bytes:
-        return self.streamed_result.all_messages_json(output_tool_return_content=output_tool_return_content)
-
-    def new_messages(self, *, output_tool_return_content: str | None = None) -> list[Any]:
-        return self.streamed_result.new_messages(output_tool_return_content=output_tool_return_content)
-
-    def __getattr__(self, name: str) -> Any:
-        return getattr(self.streamed_result, name)
-
-
-class BaseAgent(abc.ABC):
-    """Abstract base class for analysis agents."""
-
-    def __init__(self, settings: LLMSettings | None = None, model_override: Any | None = None):
-        self.settings = settings or LLMSettings()
-        self.logger = logger.bind(agent=self.__class__.__name__)
-        self._pydantic_model: Any | None = None
-        self._model_settings: ModelSettings | None = None
-        self._last_agent_run_telemetry: dict[str, Any] = {}
-        self.client = None
-
-        settings_model = getattr(self.settings, "pydantic_model", None)
-        use_pydantic_ai = bool(getattr(self.settings, "use_pydantic_ai", False))
-        if settings_model is not None:
-            use_pydantic_ai = True
-
-        if self.settings.enabled:
-            if use_pydantic_ai:
-                try:
-                    self._model_settings = self._build_model_settings()
-                    self._pydantic_model = (
-                        model_override or settings_model or self._build_pydantic_ai_model(self.settings.model_standard)
-                    )
-                except Exception as e:
-                    self.logger.warning("pydantic_ai_init_failed", error=str(e))
-
-            if self._pydantic_model is None:
-                try:
-                    self.client = self.settings.get_client()
-                except Exception as e:
-                    self.logger.warning("agent_llm_init_failed", error=str(e))
-                    self.client = None
-
-    @staticmethod
-    def _normalize_profile_names(raw_profiles: Any) -> list[str]:
-        if isinstance(raw_profiles, str):
-            return [item.strip() for item in raw_profiles.split(",") if item.strip()]
-        if isinstance(raw_profiles, (list, tuple)):
-            names: list[str] = []
-            for item in raw_profiles:
-                normalized = str(item or "").strip()
-                if normalized:
-                    names.append(normalized)
-            return names
-        return []
-
-    def _resolve_active_runtime_profile_name(self) -> str | None:
-        value = str(os.getenv("VDS_AUDIT_ACTIVE_PROFILE") or "").strip()
-        return value or None
-
-    def _resolve_row_failover_profiles(self) -> list[str]:
-        # AC-145.1.1 / AC-145.1.3: Auto-derive failover profiles when the operator
-        # has NOT explicitly configured row_failover_profiles (env var absent).
-        # If the env var IS present (even as "[]"), respect the explicit setting.
-        failover_env_raw = os.getenv("VDS_AUDIT_LLM__ROW_FAILOVER_PROFILES")
-        active_profile = self._resolve_active_runtime_profile_name()
-        if failover_env_raw is None:
-            # Not explicitly configured — auto-derive from all runtime profiles.
-            auto_derived = resolve_default_failover_profiles(active_profile)
-            configured = auto_derived
-        else:
-            configured = self._normalize_profile_names(getattr(self.settings, "row_failover_profiles", []))
-        allowed = self._normalize_profile_names(os.getenv("VDS_AUDIT_ROW_FAILOVER_AVAILABLE_PROFILES"))
-        deduped: list[str] = []
-        seen: set[str] = set()
-        for candidate in configured:
-            if candidate == active_profile or candidate in seen:
-                continue
-            if allowed and candidate not in allowed:
-                continue
-            seen.add(candidate)
-            deduped.append(candidate)
-        return deduped
-
-    def _provider_identity(self, settings: LLMSettings | None = None) -> str:
-        effective = settings or self.settings
-        active_profile = self._resolve_active_runtime_profile_name()
-        if active_profile:
-            return active_profile
-        protocol = str(getattr(effective, "protocol", "") or "").strip()
-        model = str(getattr(effective, "model_standard", "") or "").strip()
-        base_url = str(getattr(effective, "base_url", "") or "").strip()
-        return ":".join(part for part in (protocol, model, base_url) if part) or "unknown"
-
-    @contextlib.contextmanager
-    def _temporary_agent_settings(self, settings: LLMSettings):
-        snapshot = _AgentRuntimeSnapshot(
-            settings=self.settings,
-            pydantic_model=self._pydantic_model,
-            model_settings=self._model_settings,
-            client=self.client,
-        )
-        self.settings = settings
-        self._model_settings = None
-        self._pydantic_model = None
-        self.client = None
-        if self.settings.enabled:
-            settings_model = getattr(self.settings, "pydantic_model", None)
-            use_pydantic_ai = bool(getattr(self.settings, "use_pydantic_ai", False))
-            if settings_model is not None:
-                use_pydantic_ai = True
-            if use_pydantic_ai:
-                try:
-                    self._model_settings = self._build_model_settings()
-                    self._pydantic_model = settings_model or self._build_pydantic_ai_model(self.settings.model_standard)
-                except Exception as exc:
-                    self.logger.warning("pydantic_ai_failover_init_failed", error=str(exc))
-                    self._pydantic_model = None
-            if self._pydantic_model is None:
-                try:
-                    self.client = self.settings.get_client()
-                except Exception as exc:
-                    self.logger.warning("agent_failover_llm_init_failed", error=str(exc))
-                    self.client = None
-        try:
-            yield
-        finally:
-            self.settings = snapshot.settings
-            self._pydantic_model = snapshot.pydantic_model
-            self._model_settings = snapshot.model_settings
-            self.client = snapshot.client
-
-    @abc.abstractmethod
-    async def analyze(self, evidence: EvidenceBundle | dict[str, Any] | AuditTask) -> AgentResult:
-        """Perform analysis on the evidence bundle."""
-        pass
-
-    @staticmethod
-    def _extract_json_object_candidates(raw: str) -> list[str]:
-        """Extract top-level JSON object substrings from a potentially concatenated payload."""
-        text = str(raw or "")
-        if not text:
-            return []
-        candidates: list[str] = []
-        in_string = False
-        escape = False
-        depth = 0
-        start_idx: int | None = None
-        for idx, ch in enumerate(text):
-            if in_string:
-                if escape:
-                    escape = False
-                elif ch == "\\":
-                    escape = True
-                elif ch == '"':
-                    in_string = False
-                continue
-            if ch == '"':
-                in_string = True
-                continue
-            if ch == "{":
-                if depth == 0:
-                    start_idx = idx
-                depth += 1
-                continue
-            if ch == "}" and depth > 0:
-                depth -= 1
-                if depth == 0 and start_idx is not None:
-                    candidates.append(text[start_idx : idx + 1])
-                    start_idx = None
-        return candidates
-
-    def _recover_output_from_final_result_validation_error(
-        self,
-        *,
-        result_type: type[Any],
-        exc: Exception,
-        telemetry: dict[str, Any] | None,
-    ) -> Any | None:
-        """Recover structured output from invalid final_result payloads.
-
-        This path is intentionally broader than trailing-character JSON errors.
-        Some providers surface schema-shape validation failures even when the
-        raw final payload is recoverable and the target model can normalize it
-        via `model_validate(...)`.
-        """
-
-        raw_candidates: list[str] = []
-        if isinstance(telemetry, dict):
-            raw_invalid_input = telemetry.get("event_final_result_invalid_input")
-            if isinstance(raw_invalid_input, str) and raw_invalid_input.strip():
-                raw_candidates.append(raw_invalid_input)
-            elif isinstance(raw_invalid_input, (dict, list)):
-                with contextlib.suppress(Exception):
-                    raw_candidates.append(json.dumps(raw_invalid_input, ensure_ascii=False))
-
-        for attr_name in ("body", "response"):
-            raw_value = getattr(exc, attr_name, None)
-            if isinstance(raw_value, str) and raw_value.strip():
-                raw_candidates.append(raw_value)
-            elif isinstance(raw_value, (dict, list)):
-                with contextlib.suppress(Exception):
-                    raw_candidates.append(json.dumps(raw_value, ensure_ascii=False))
-        for arg in getattr(exc, "args", ()) or ():
-            if isinstance(arg, str) and ("{" in arg and "}" in arg):
-                raw_candidates.append(arg)
-            elif isinstance(arg, (dict, list)):
-                with contextlib.suppress(Exception):
-                    raw_candidates.append(json.dumps(arg, ensure_ascii=False))
-        cause = getattr(exc, "__cause__", None)
-        if cause is not None:
-            for attr_name in ("body", "response"):
-                raw_value = getattr(cause, attr_name, None)
-                if isinstance(raw_value, str) and raw_value.strip():
-                    raw_candidates.append(raw_value)
-                elif isinstance(raw_value, (dict, list)):
-                    with contextlib.suppress(Exception):
-                        raw_candidates.append(json.dumps(raw_value, ensure_ascii=False))
-            for arg in getattr(cause, "args", ()) or ():
-                if isinstance(arg, str) and ("{" in arg and "}" in arg):
-                    raw_candidates.append(arg)
-                elif isinstance(arg, (dict, list)):
-                    with contextlib.suppress(Exception):
-                        raw_candidates.append(json.dumps(arg, ensure_ascii=False))
-
-        seen: set[str] = set()
-        all_json_objects: list[str] = []
-        for raw in raw_candidates:
-            for candidate in self._extract_json_object_candidates(raw):
-                normalized = candidate.strip()
-                if not normalized or normalized in seen:
-                    continue
-                seen.add(normalized)
-                all_json_objects.append(normalized)
-
-        for candidate in reversed(all_json_objects):
-            try:
-                parsed = json.loads(candidate)
-            except Exception:
-                continue
-            validator = getattr(result_type, "model_validate", None)
-            if not callable(validator):
-                continue
-            try:
-                return validator(parsed)
-            except Exception:
-                continue
-        return None
-
-    def _resolve_agent_timeout_policy(self, complexity: str) -> AgentTimeoutPolicy:
-        """Resolve timeout policy from config/settings with complexity-aware scaling."""
-
-        def _to_float(raw: Any, default: float, minimum: float) -> float:
-            try:
-                value = float(raw)
-            except (TypeError, ValueError):
-                value = default
-            return max(value, minimum)
-
-        def _to_int(raw: Any, default: int, minimum: int) -> int:
-            try:
-                value = int(raw)
-            except (TypeError, ValueError):
-                value = default
-            return max(value, minimum)
-
-        factor = self._complexity_timeout_factor(complexity)
-        base_timeout = _to_float(getattr(self.settings, "agent_timeout_seconds", 600.0), 600.0, 30.0)
-        max_timeout = _to_float(getattr(self.settings, "agent_timeout_max_seconds", 1800.0), 1800.0, 60.0)
-        extension_seconds = _to_float(
-            getattr(self.settings, "agent_timeout_extension_seconds", 300.0),
-            300.0,
-            0.0,
-        )
-        extension_attempts = _to_int(
-            getattr(self.settings, "agent_timeout_extension_attempts", 3),
-            3,
-            0,
-        )
-        heartbeat_seconds = _to_float(
-            getattr(self.settings, "agent_timeout_heartbeat_seconds", 30.0),
-            30.0,
-            5.0,
-        )
-        idle_post_tool_seconds = _to_float(
-            getattr(self.settings, "agent_idle_timeout_seconds", 180.0),
-            180.0,
-            30.0,
-        )
-        lease_window_seconds = _to_float(
-            getattr(self.settings, "row_progress_lease_seconds", 0.0),
-            0.0,
-            0.0,
-        )
-        stall_detection_seconds = _to_float(
-            getattr(self.settings, "row_stall_detection_seconds", idle_post_tool_seconds),
-            idle_post_tool_seconds,
-            1.0,
-        )
-        absolute_timeout_seconds = _to_float(
-            float(getattr(self.settings, "row_absolute_timeout_ms", 600_000)) / 1000.0,
-            600.0,
-            1.0,
-        )
-
-        initial_timeout = max(30.0, base_timeout * factor)
-        max_timeout = max(max_timeout, initial_timeout)
-        if lease_window_seconds > 0:
-            initial_timeout = min(initial_timeout, lease_window_seconds)
-            max_timeout = min(max_timeout, absolute_timeout_seconds)
-            max_timeout = max(max_timeout, initial_timeout)
-            stall_detection_seconds = min(stall_detection_seconds, lease_window_seconds)
-        return AgentTimeoutPolicy(
-            initial_timeout_seconds=initial_timeout,
-            max_timeout_seconds=max_timeout,
-            extension_seconds=extension_seconds,
-            extension_attempts=extension_attempts,
-            heartbeat_seconds=heartbeat_seconds,
-            idle_post_tool_seconds=idle_post_tool_seconds,
-            lease_window_seconds=lease_window_seconds,
-            stall_detection_seconds=stall_detection_seconds,
-            absolute_timeout_seconds=absolute_timeout_seconds,
-        )
-
-    def _resolve_pydantic_usage_limits(
-        self,
-        *,
-        max_turns: int | None,
-        max_tool_calls: int | None,
-    ) -> tuple[int | None, int | None]:
-        """Resolve request/tool usage limits for PydanticAI runs.
-
-        Request limit must not be lower than the tool-call budget, otherwise
-        heavily agentic runs may hit `request_limit` before consuming
-        `tool_calls_limit`.
-        """
-        normalized_tool_calls = max_tool_calls
-        if normalized_tool_calls is not None:
-            normalized_tool_calls = int(normalized_tool_calls)
-            # Preserve explicit zero-tool mode (`max_tool_calls=0`) so
-            # synthesis retries cannot accidentally invoke tools again.
-            if normalized_tool_calls < 0:
-                normalized_tool_calls = None
-
-        # Zero-tool mode is used by docs/security prompt-fed execution paths to
-        # explicitly disallow tools. In this mode, applying a tiny request_limit
-        # floor (2) can preempt normal output validation retries and causes
-        # avoidable "request_limit exceeded" failures.
-        if max_turns is None and normalized_tool_calls is None:
-            return None, None
-
-        request_limit = max_turns
-        if normalized_tool_calls is not None and normalized_tool_calls > 0:
-            tool_driven_request_floor = normalized_tool_calls + 2
-            request_limit = max(request_limit or 0, tool_driven_request_floor)
-
-        return request_limit, normalized_tool_calls
-
-    def _is_trace_payload_enabled(self) -> bool:
-        """Return True when payload-level diagnostics should be emitted."""
-        if bool(getattr(self.settings, "agent_trace_payload_enabled", False)):
-            return True
-        try:
-            from vds_audit_orchestrator.logging_config import is_trace_mode_enabled
-
-            return bool(is_trace_mode_enabled())
-        except Exception:
-            return False
-
-    def _trace_payload_max_chars(self) -> int:
-        raw = getattr(self.settings, "agent_trace_payload_max_chars", 20_000)
-        try:
-            value = int(raw)
-        except (TypeError, ValueError):
-            value = 20_000
-        return max(1, min(value, 200_000))
-
-    @staticmethod
-    def _redact_trace_text(value: str) -> str:
-        redacted = re.sub(r"(?i)(api[_-]?key|token|password|secret)\s*[:=]\s*([^\s,;]+)", r"\1=[REDACTED]", value)
-        redacted = re.sub(r"(?i)(authorization)\s*[:=]\s*(bearer\s+)?[^\s,;]+", r"\1=[REDACTED]", redacted)
-        return redacted
-
-    def _sanitize_trace_payload(self, payload: Any, *, depth: int = 0) -> Any:
-        """Return bounded trace payload data safe for structured logs."""
-        max_chars = self._trace_payload_max_chars()
-        if depth > 10:
-            return "...(trace-depth-limit)"
-        if isinstance(payload, str):
-            text = self._redact_trace_text(payload)
-            return text[:max_chars] + "...(truncated)" if len(text) > max_chars else text
-        if isinstance(payload, (int, float, bool)) or payload is None:
-            return payload
-        if isinstance(payload, list):
-            trimmed = payload[:100]
-            result = [self._sanitize_trace_payload(item, depth=depth + 1) for item in trimmed]
-            if len(payload) > len(trimmed):
-                result.append(f"...(truncated {len(payload) - len(trimmed)} items)")
-            return result
-        if isinstance(payload, tuple):
-            return self._sanitize_trace_payload(list(payload), depth=depth + 1)
-        if isinstance(payload, dict):
-            result: dict[str, Any] = {}
-            items = list(payload.items())
-            for key, value in items[:100]:
-                key_text = str(key)
-                if re.search(r"(?i)(password|secret|token|api[_-]?key|authorization)", key_text):
-                    result[key_text] = "[REDACTED]"
-                else:
-                    result[key_text] = self._sanitize_trace_payload(value, depth=depth + 1)
-            if len(items) > 100:
-                result["__truncated_items__"] = len(items) - 100
-            return result
-        return self._sanitize_trace_payload(str(payload), depth=depth + 1)
-
-    @staticmethod
-    def _serialize_skill_policy_drop(item: Any) -> dict[str, Any] | None:
-        skill_name = str(getattr(item, "skill_name", "") or "").strip()
-        reason = str(getattr(item, "reason", "") or "").strip()
-        if not skill_name and not reason:
-            return None
-        payload: dict[str, Any] = {
-            "skill_name": skill_name or "unknown",
-            "reason": reason or "unknown",
-        }
-        for field_name in ("metadata_name", "frontmatter_name", "runtime_name", "detail"):
-            raw_value = getattr(item, field_name, None)
-            if isinstance(raw_value, str):
-                value = raw_value.strip()
-                if value:
-                    payload[field_name] = value
-            elif raw_value is not None:
-                payload[field_name] = str(raw_value)
-        return payload
-
-    def _collect_skill_policy_diagnostics(self, toolsets: list[Any] | None) -> dict[str, Any] | None:
-        """Collect operator-facing skill-policy diagnostics from attached toolsets."""
-        if not toolsets:
-            return None
-        try:
-            from vds_audit_orchestrator.agents.toolsets.skills_toolset import get_toolset_policy_diagnostics
-        except Exception:
-            return None
-
-        metadata_count = 0
-        policy_eligible_count = 0
-        runtime_loaded_count = 0
-        blocked_tools: list[str] = []
-        drop_reasons: list[dict[str, Any]] = []
-        toolset_details: list[dict[str, Any]] = []
-
-        for index, toolset in enumerate(toolsets):
-            diagnostics = get_toolset_policy_diagnostics(toolset)
-            if diagnostics is None:
-                continue
-
-            toolset_metadata_count = int(getattr(diagnostics, "metadata_count", 0) or 0)
-            toolset_policy_eligible_count = int(getattr(diagnostics, "policy_eligible_count", 0) or 0)
-            toolset_runtime_loaded_count = int(getattr(diagnostics, "runtime_loaded_count", 0) or 0)
-            toolset_blocked_tools = [
-                str(name).strip() for name in (getattr(diagnostics, "blocked_tools", []) or []) if str(name).strip()
-            ]
-            toolset_drop_reasons: list[dict[str, Any]] = []
-            for drop in getattr(diagnostics, "dropped", []) or []:
-                serialized = self._serialize_skill_policy_drop(drop)
-                if serialized is None:
-                    continue
-                toolset_drop_reasons.append(serialized)
-                drop_reasons.append(dict(serialized))
-
-            metadata_count += toolset_metadata_count
-            policy_eligible_count += toolset_policy_eligible_count
-            runtime_loaded_count += toolset_runtime_loaded_count
-            for blocked in toolset_blocked_tools:
-                if blocked not in blocked_tools:
-                    blocked_tools.append(blocked)
-            toolset_details.append(
-                {
-                    "toolset_index": index,
-                    "metadata_count": toolset_metadata_count,
-                    "policy_eligible_count": toolset_policy_eligible_count,
-                    "runtime_loaded_count": toolset_runtime_loaded_count,
-                    "blocked_tools": toolset_blocked_tools,
-                    "drop_reasons": toolset_drop_reasons,
-                }
-            )
-
-        if not toolset_details:
-            return None
-        return {
-            "toolset_count": len(toolset_details),
-            "metadata_count": metadata_count,
-            "policy_eligible_count": policy_eligible_count,
-            "runtime_loaded_count": runtime_loaded_count,
-            "blocked_tools": blocked_tools,
-            "drop_reasons": drop_reasons,
-            "toolsets": toolset_details,
-        }
-
-    @staticmethod
-    def _create_skill_policy_enforcer(
-        *,
-        skills_needed: bool | None = None,
-        mode_override: str | None = None,
-        interpretation: dict[str, Any] | None = None,
-    ) -> Any:
-        """Create a SkillPolicyEnforcer from row-level inputs (FR-194..FR-198).
-
-        Returns the enforcer instance or None if the module is unavailable.
-        """
-        try:
-            from vds_agent_core.skills.policy import (
-                SkillPolicyEnforcer,
-                resolve_skill_policy_mode,
-            )
-        except Exception:
-            return None
-
-        mode = resolve_skill_policy_mode(
-            skills_needed=skills_needed,
-            mode_override=mode_override,
-            interpretation=interpretation,
-        )
-        effective_skills_needed = skills_needed if skills_needed is not None else True
-        if isinstance(interpretation, dict) and interpretation.get("skills_needed") is False:
-            effective_skills_needed = False
-        return SkillPolicyEnforcer(mode=mode, skills_needed=effective_skills_needed)
-
-    def _log_trace_payload(
-        self,
-        *,
-        event_name: str,
-        elapsed_sec: float,
-        payload: dict[str, Any],
-    ) -> None:
-        """Log payload-level diagnostics when trace mode is enabled."""
-        if not self._is_trace_payload_enabled():
-            return
-        self.logger.info(
-            event_name,
-            elapsed_sec=round(elapsed_sec, 2),
-            payload=self._sanitize_trace_payload(payload),
-        )
-
-    def _exception_diagnostics(self, exc: Exception) -> dict[str, Any]:
-        """Extract compact diagnostics for model/runtime exceptions."""
-        details: dict[str, Any] = {
-            "error_type": type(exc).__name__,
-            "error": str(exc),
-        }
-        body = getattr(exc, "body", None)
-        if isinstance(body, str) and body.strip():
-            details["error_body_preview"] = body[:4000]
-
-        cause = getattr(exc, "__cause__", None)
-        if cause is not None:
-            details["cause_type"] = type(cause).__name__
-            details["cause_message"] = str(cause)[:2000]
-            cause_body = getattr(cause, "body", None)
-            if isinstance(cause_body, str) and cause_body.strip():
-                details["cause_body_preview"] = cause_body[:4000]
-        return details
-
-    @staticmethod
-    def _complexity_timeout_factor(complexity: str) -> float:
-        """Map request complexity to timeout multiplier."""
-        normalized = complexity.strip().lower()
-        if normalized in {"simple", "low"}:
-            return 0.75
-        if normalized in {"complex", "high"}:
-            return 1.5
-        if normalized in {"critical"}:
-            return 2.0
-        return 1.0
-
-    @staticmethod
-    def _timeout_progress_snapshot(run_telemetry: dict[str, Any] | None) -> dict[str, float]:
-        """Extract a compact timeout-progress snapshot from runtime telemetry."""
-        snapshot: dict[str, float] = {
-            "event_last_seen_at": 0.0,
-            "event_tool_calls_started": 0.0,
-            "event_tool_calls_completed": 0.0,
-            "event_tool_effective_calls": 0.0,
-            "event_skill_execution_effective_tool_calls": 0.0,
-            "event_turn_index": 0.0,
-            "evidence_refs_count": 0.0,
-            "steps_executed": 0.0,
-            "steps_unique_tools": 0.0,
-        }
-        if not isinstance(run_telemetry, dict):
-            return snapshot
-
-        for key in ("event_last_seen_at",):
-            raw = run_telemetry.get(key)
-            if isinstance(raw, int | float):
-                snapshot[key] = float(raw)
-
-        for key in (
-            "event_tool_calls_started",
-            "event_tool_calls_completed",
-            "event_tool_effective_calls",
-            "event_skill_execution_effective_tool_calls",
-            "event_turn_index",
-            "evidence_refs_count",
-            "steps_executed",
-            "steps_unique_tools",
-        ):
-            raw = run_telemetry.get(key, 0)
-            if isinstance(raw, int | float):
-                snapshot[key] = float(max(0, int(raw)))
-        return snapshot
-
-    @staticmethod
-    def _timeout_progress_signal_type(
-        previous_snapshot: dict[str, float] | None,
-        current_snapshot: dict[str, float] | None,
-    ) -> str | None:
-        """Return the dominant progress signal since the previous snapshot."""
-        if not isinstance(previous_snapshot, dict) or not isinstance(current_snapshot, dict):
-            return None
-        if current_snapshot.get("event_last_seen_at", 0.0) > previous_snapshot.get("event_last_seen_at", 0.0):
-            return "event_stream"
-        for key, signal_type in (
-            ("event_tool_calls_started", "tool_call_started"),
-            ("event_tool_calls_completed", "tool_call_completed"),
-            ("event_tool_effective_calls", "tool_effective"),
-            ("event_skill_execution_effective_tool_calls", "skill_effective"),
-            ("event_turn_index", "partial_output"),
-            ("evidence_refs_count", "evidence_growth"),
-            ("steps_executed", "route_transition"),
-            ("steps_unique_tools", "route_transition"),
-        ):
-            if current_snapshot.get(key, 0.0) > previous_snapshot.get(key, 0.0):
-                return signal_type
-        return None
-
-    @staticmethod
-    def _timeout_progress_observed_since(
-        previous_snapshot: dict[str, float] | None,
-        current_snapshot: dict[str, float] | None,
-    ) -> bool:
-        """Return True when runtime telemetry shows forward progress since snapshot."""
-        if not isinstance(previous_snapshot, dict) or not isinstance(current_snapshot, dict):
-            return False
-
-        if current_snapshot.get("event_last_seen_at", 0.0) > previous_snapshot.get("event_last_seen_at", 0.0):
-            return True
-
-        for key in (
-            "event_tool_calls_started",
-            "event_tool_calls_completed",
-            "event_tool_effective_calls",
-            "event_skill_execution_effective_tool_calls",
-            "event_turn_index",
-        ):
-            if current_snapshot.get(key, 0.0) > previous_snapshot.get(key, 0.0):
-                return True
-        return False
-
-    async def _run_with_timeout_policy(
-        self,
-        *,
-        operation: Coroutine[Any, Any, ResultType],
-        timeout_policy: AgentTimeoutPolicy,
-        result_type_name: str,
-        run_telemetry: dict[str, Any] | None = None,
-    ) -> ResultType:
-        """Run awaitable with heartbeat + bounded timeout extensions."""
-        import time
-
-        start_time = time.monotonic()
-        timeout_budget = timeout_policy.initial_timeout_seconds
-        lease_mode_enabled = timeout_policy.lease_window_seconds > 0
-        extensions_used = 0
-        last_extension_progress_snapshot = self._timeout_progress_snapshot(run_telemetry)
-        previous_heartbeat_snapshot = dict(last_extension_progress_snapshot)
-        last_progress_at = start_time
-        last_progress_signal_type: str | None = None
-        task = asyncio.create_task(operation)
-        try:
-            while True:
-                elapsed = time.monotonic() - start_time
-                remaining = timeout_budget - elapsed
-                if remaining <= 0:
-                    current_progress_snapshot = self._timeout_progress_snapshot(run_telemetry)
-                    progress_since_last_extension = self._timeout_progress_observed_since(
-                        last_extension_progress_snapshot,
-                        current_progress_snapshot,
-                    )
-                    extension_cap_reached = extensions_used >= timeout_policy.extension_attempts
-                    if lease_mode_enabled:
-                        can_extend = (
-                            progress_since_last_extension and timeout_budget < timeout_policy.max_timeout_seconds
-                        )
-                    else:
-                        can_extend = (
-                            timeout_policy.extension_seconds > 0
-                            and (not extension_cap_reached or progress_since_last_extension)
-                            and timeout_budget < timeout_policy.max_timeout_seconds
-                        )
-                    if can_extend:
-                        increment = min(
-                            timeout_policy.lease_window_seconds
-                            if lease_mode_enabled
-                            else timeout_policy.extension_seconds,
-                            timeout_policy.max_timeout_seconds - timeout_budget,
-                        )
-                        timeout_budget += increment
-                        extensions_used += 1
-                        extension_reason = (
-                            "lease_progress_renewal"
-                            if lease_mode_enabled
-                            else (
-                                "progress_override_extension_cap"
-                                if extension_cap_reached and progress_since_last_extension
-                                else "within_extension_cap"
-                            )
-                        )
-                        self.logger.warning(
-                            "pydantic_agent_timeout_extended",
-                            result_type=result_type_name,
-                            timeout_budget_seconds=round(timeout_budget, 2),
-                            extension_seconds=round(increment, 2),
-                            extensions_used=extensions_used,
-                            max_extensions=timeout_policy.extension_attempts,
-                            extension_cap_reached=extension_cap_reached,
-                            progress_since_last_extension=progress_since_last_extension,
-                            extension_reason=extension_reason,
-                            lease_mode_enabled=lease_mode_enabled,
-                        )
-                        last_extension_progress_snapshot = current_progress_snapshot
-                        last_progress_at = time.monotonic()
-                        last_progress_signal_type = (
-                            self._timeout_progress_signal_type(
-                                previous_heartbeat_snapshot,
-                                current_progress_snapshot,
-                            )
-                            or "lease_renewal"
-                        )
-                        continue
-
-                    task.cancel()
-                    with contextlib.suppress(asyncio.CancelledError):
-                        await task
-                    self.logger.error(
-                        "pydantic_agent_call_timeout",
-                        result_type=result_type_name,
-                        timeout_budget_seconds=round(timeout_budget, 2),
-                        max_timeout_seconds=round(timeout_policy.max_timeout_seconds, 2),
-                        extensions_used=extensions_used,
-                        lease_mode_enabled=lease_mode_enabled,
-                        last_progress_signal_type=last_progress_signal_type,
-                    )
-                    raise RuntimeError(
-                        "PydanticAI agent timed out after "
-                        f"{timeout_budget:.0f}s (max budget {timeout_policy.max_timeout_seconds:.0f}s)"
-                    )
-
-                wait_timeout = min(remaining, timeout_policy.heartbeat_seconds)
-                done, _ = await asyncio.wait({task}, timeout=wait_timeout)
-                if task in done:
-                    return await task
-
-                if isinstance(run_telemetry, dict):
-                    current_progress_snapshot = self._timeout_progress_snapshot(run_telemetry)
-                    signal_type = self._timeout_progress_signal_type(
-                        previous_heartbeat_snapshot, current_progress_snapshot
-                    )
-                    if signal_type is not None:
-                        last_progress_at = time.monotonic()
-                        last_progress_signal_type = signal_type
-                    previous_heartbeat_snapshot = current_progress_snapshot
-
-                    if lease_mode_enabled:
-                        lease_stall_elapsed = time.monotonic() - last_progress_at
-                        if lease_stall_elapsed >= timeout_policy.stall_detection_seconds:
-                            task.cancel()
-                            with contextlib.suppress(asyncio.CancelledError):
-                                await task
-                            self.logger.error(
-                                "pydantic_agent_progress_lease_expired",
-                                result_type=result_type_name,
-                                timeout_budget_seconds=round(timeout_budget, 2),
-                                lease_window_seconds=round(timeout_policy.lease_window_seconds, 2),
-                                stall_detection_seconds=round(timeout_policy.stall_detection_seconds, 2),
-                                stall_elapsed_seconds=round(lease_stall_elapsed, 2),
-                                last_progress_signal_type=last_progress_signal_type,
-                            )
-                            raise RuntimeError(
-                                "PydanticAI agent progress lease expired after "
-                                f"{lease_stall_elapsed:.0f}s without progress"
-                            )
-
-                    tool_calls_started = int(run_telemetry.get("event_tool_calls_started", 0) or 0)
-                    tool_calls_completed = int(
-                        run_telemetry.get("event_tool_calls_completed", run_telemetry.get("event_tool_calls", 0)) or 0
-                    )
-                    tool_activity_count = max(tool_calls_started, tool_calls_completed)
-                    last_tool_activity_at = run_telemetry.get("event_last_tool_activity_at")
-                    last_seen_at = run_telemetry.get("event_last_seen_at")
-                    latest_activity_at: float | None = None
-                    activity_source = "tool"
-                    if isinstance(last_tool_activity_at, (int, float)):
-                        latest_activity_at = float(last_tool_activity_at)
-                    if isinstance(last_seen_at, (int, float)) and (
-                        latest_activity_at is None or float(last_seen_at) > latest_activity_at
-                    ):
-                        latest_activity_at = float(last_seen_at)
-                        activity_source = "event"
-
-                    if tool_activity_count > 0 and isinstance(latest_activity_at, float):
-                        stall_elapsed = time.monotonic() - latest_activity_at
-                        if stall_elapsed >= timeout_policy.idle_post_tool_seconds:
-                            task.cancel()
-                            with contextlib.suppress(asyncio.CancelledError):
-                                await task
-                            self.logger.error(
-                                "pydantic_agent_call_stalled",
-                                result_type=result_type_name,
-                                timeout_budget_seconds=round(timeout_budget, 2),
-                                idle_post_tool_seconds=round(timeout_policy.idle_post_tool_seconds, 2),
-                                stall_elapsed_seconds=round(stall_elapsed, 2),
-                                tool_calls_started=tool_calls_started,
-                                tool_calls_completed=tool_calls_completed,
-                                observed_tool_names=run_telemetry.get("event_tool_names"),
-                                activity_source=activity_source,
-                            )
-                            raise RuntimeError(
-                                "PydanticAI agent stalled after tool activity "
-                                f"for {stall_elapsed:.0f}s (idle limit {timeout_policy.idle_post_tool_seconds:.0f}s)"
-                            )
-                    tool_count_by_name = run_telemetry.get("event_tool_count_by_name")
-                    if isinstance(tool_count_by_name, dict) and tool_calls_completed >= 20:
-                        list_directory_calls = int(tool_count_by_name.get("list_directory", 0) or 0)
-                        unique_tools = len(tool_count_by_name)
-                        if list_directory_calls >= 14 and unique_tools <= 3:
-                            repetition_ratio = list_directory_calls / max(1, tool_calls_completed)
-                            if repetition_ratio >= 0.7:
-                                task.cancel()
-                                with contextlib.suppress(asyncio.CancelledError):
-                                    await task
-                                self.logger.error(
-                                    "pydantic_agent_list_directory_churn",
-                                    result_type=result_type_name,
-                                    timeout_budget_seconds=round(timeout_budget, 2),
-                                    tool_calls_completed=tool_calls_completed,
-                                    list_directory_calls=list_directory_calls,
-                                    unique_tools=unique_tools,
-                                    repetition_ratio=round(repetition_ratio, 3),
-                                    observed_tool_names=run_telemetry.get("event_tool_names"),
-                                )
-                                raise RuntimeError(
-                                    "PydanticAI agent detected list_directory churn "
-                                    f"({list_directory_calls}/{tool_calls_completed} calls)"
-                                )
-
-                self.logger.info(
-                    "pydantic_agent_call_heartbeat",
-                    result_type=result_type_name,
-                    elapsed_seconds=round(time.monotonic() - start_time, 2),
-                    remaining_seconds=round(max(0.0, timeout_budget - (time.monotonic() - start_time)), 2),
-                    timeout_budget_seconds=round(timeout_budget, 2),
-                    extensions_used=extensions_used,
-                    lease_mode_enabled=lease_mode_enabled,
-                    lease_window_seconds=round(timeout_policy.lease_window_seconds, 2) if lease_mode_enabled else 0.0,
-                    last_progress_signal_type=last_progress_signal_type,
-                )
-        except asyncio.CancelledError:
-            task.cancel()
-            with contextlib.suppress(asyncio.CancelledError):
-                await task
-            raise
-
-    def _merge_failover_telemetry(
-        self,
-        primary: dict[str, Any] | None,
-        secondary: dict[str, Any] | None,
-    ) -> dict[str, Any]:
-        merged = dict(primary or {})
-        fallback = dict(secondary or {})
-        if not merged:
-            return fallback
-        for key in (
-            "provider_failover_attempted",
-            "provider_failover_reason",
-            "provider_failover_final_provider",
-            "provider_failover_original_provider",
-            "provider_failover_profiles_considered",
-            "provider_retry_class",
-            "provider_retry_reason",
-            "retry_after_honored",
-            "retry_after_seconds",
-            "timeout_kind",
-        ):
-            if key in fallback:
-                merged[key] = fallback[key]
-        primary_chain = list(merged.get("provider_failover_chain") or [])
-        for item in list(fallback.get("provider_failover_chain") or []):
-            normalized = str(item or "").strip()
-            if normalized and normalized not in primary_chain:
-                primary_chain.append(normalized)
-        if primary_chain:
-            merged["provider_failover_chain"] = primary_chain
-        merged["provider_transient_retry_count"] = max(
-            int(merged.get("provider_transient_retry_count", 0) or 0),
-            int(fallback.get("provider_transient_retry_count", 0) or 0),
-        )
-        merged["provider_failover_hops"] = max(
-            int(merged.get("provider_failover_hops", 0) or 0),
-            int(fallback.get("provider_failover_hops", 0) or 0),
-        )
-        if "provider_failover_profiles_considered" in merged:
-            max_provider_hops = max(0, int(getattr(self.settings, "row_failover_max_provider_hops", 2) or 0))
-            merged["provider_failover_profiles_considered"] = truncate_failover_profiles_considered(
-                merged.get("provider_failover_profiles_considered"),
-                max_provider_hops=max_provider_hops,
-            )
-        return merged
-
-    async def _attempt_row_provider_failover(
-        self,
-        *,
-        exc: Exception,
-        system_prompt: str,
-        user_prompt: str,
-        result_type: type[ResultType],
-        complexity: str,
-        max_tokens: int,
-        prompt_type: str,
-        deps: Any | None,
-        toolsets: list[Any] | None,
-        max_turns: int | None,
-        max_tool_calls: int | None,
-        pre_retry_telemetry: dict[str, Any] | None,
-        current_run_telemetry: dict[str, Any] | None,
-        row_failover_context: RowFailoverContext | None,
-        provider_health_memory: ProviderHealthMemory,
-    ) -> ResultType | None:
-        classification = ProviderFailureClassifier.classify(exc)
-        if not classification.is_failoverable:
-            return None
-
-        configured_profiles = self._resolve_row_failover_profiles()
-        if not configured_profiles:
-            return None
-
-        max_hops = max(0, int(getattr(self.settings, "row_failover_max_provider_hops", 2) or 0))
-        if max_hops <= 0:
-            return None
-
-        current_provider = self._provider_identity()
-        failover_context = row_failover_context or RowFailoverContext(
-            row_id=str(getattr(deps, "project_id", "") or ""),
-            check_id=prompt_type,
-            original_provider=current_provider,
-            failover_chain=[current_provider],
-        )
-        failover_context = truncate_row_failover_context(
-            failover_context,
-            max_provider_hops=max_hops,
-        )
-        considered_profiles = truncate_failover_profiles_considered(
-            configured_profiles,
-            max_provider_hops=max_hops,
-        )
-        if failover_context.failover_count >= max_hops:
-            if current_run_telemetry is not None:
-                current_run_telemetry["provider_failover_attempted"] = True
-                current_run_telemetry["provider_failover_chain"] = list(failover_context.failover_chain or [])
-                current_run_telemetry["provider_failover_final_provider"] = current_provider
-                current_run_telemetry["provider_failover_original_provider"] = (
-                    failover_context.original_provider or current_provider
-                )
-                current_run_telemetry["provider_failover_profiles_considered"] = list(considered_profiles)
-                current_run_telemetry["provider_failover_reason"] = classification.classification_reason
-                current_run_telemetry["provider_failover_hops"] = int(failover_context.failover_count)
-                current_run_telemetry["timeout_kind"] = TimeoutKind.TIMEOUT_FAILOVER_EXHAUSTED.value
-            return None
-
-        next_profile: str | None = None
-        for candidate in configured_profiles:
-            if candidate == current_provider or candidate in failover_context.failover_chain:
-                continue
-            if provider_health_memory.should_skip_provider(candidate):
-                continue
-            next_profile = candidate
-            break
-        if next_profile is None:
-            if current_run_telemetry is not None:
-                current_run_telemetry["provider_failover_attempted"] = True
-                current_run_telemetry["provider_failover_chain"] = list(failover_context.failover_chain or [])
-                current_run_telemetry["provider_failover_final_provider"] = current_provider
-                current_run_telemetry["provider_failover_original_provider"] = (
-                    failover_context.original_provider or current_provider
-                )
-                current_run_telemetry["provider_failover_profiles_considered"] = list(considered_profiles)
-                current_run_telemetry["provider_failover_reason"] = classification.classification_reason
-                current_run_telemetry["provider_failover_hops"] = int(failover_context.failover_count)
-                current_run_telemetry["timeout_kind"] = TimeoutKind.TIMEOUT_FAILOVER_EXHAUSTED.value
-            return None
-
-        provider_health_memory.total_hops += 1
-        next_context = RowFailoverContext(
-            row_id=failover_context.row_id,
-            check_id=failover_context.check_id,
-            evidence_refs=list(failover_context.evidence_refs or []),
-            route_state=dict(failover_context.route_state or {}),
-            partial_progress=dict(failover_context.partial_progress or {}),
-            original_provider=failover_context.original_provider or current_provider,
-            failover_count=int(failover_context.failover_count) + 1,
-            failover_chain=[*list(failover_context.failover_chain or []), next_profile],
-            timeout_telemetry_snapshot=dict(failover_context.timeout_telemetry_snapshot or {}),
-        )
-        next_context = truncate_row_failover_context(next_context, max_provider_hops=max_hops)
-
-        if current_run_telemetry is not None:
-            current_run_telemetry["provider_failover_attempted"] = True
-            current_run_telemetry["provider_failover_chain"] = list(next_context.failover_chain)
-            current_run_telemetry["provider_failover_final_provider"] = next_profile
-            current_run_telemetry["provider_failover_original_provider"] = (
-                next_context.original_provider or current_provider
-            )
-            current_run_telemetry["provider_failover_profiles_considered"] = list(considered_profiles)
-            current_run_telemetry["provider_failover_reason"] = classification.classification_reason
-            current_run_telemetry["provider_failover_hops"] = int(next_context.failover_count)
-
-        self.logger.warning(
-            "pydantic_agent_provider_failover_started",
-            result_type=result_type.__name__,
-            current_provider=current_provider,
-            next_provider=next_profile,
-            failure_class=classification.failure_class.value,
-            reason=classification.classification_reason,
-            failover_count=next_context.failover_count,
-            max_provider_hops=max_hops,
-        )
-
-        fallback_settings = inherit_runtime_llm_policy(next_profile, source_llm=self.settings)
-        with self._temporary_agent_settings(fallback_settings):
-            return await self._run_pydantic_agent(
-                system_prompt=system_prompt,
-                user_prompt=user_prompt,
-                result_type=result_type,
-                complexity=complexity,
-                max_tokens=max_tokens,
-                prompt_type=prompt_type,
-                deps=deps,
-                toolsets=toolsets,
-                max_turns=max_turns,
-                max_tool_calls=max_tool_calls,
-                pre_retry_telemetry=self._merge_failover_telemetry(pre_retry_telemetry, current_run_telemetry),
-                row_failover_context=next_context,
-                provider_health_memory=provider_health_memory,
-            )
-
-    @property
-    def llm_available(self) -> bool:
-        """Check if an LLM client/model is available."""
-        return self._pydantic_model is not None or self.client is not None
-
-    async def _call_llm(self, system: str, user: str) -> str:
-        """Helper to call LLM."""
-        if self._pydantic_model is not None:
-            result = await self._run_pydantic_agent(
-                system_prompt=system,
-                user_prompt=user,
-                result_type=str,
-                complexity="standard",
-                max_tokens=4096,
-                prompt_type="agent_call",
-            )
-            return result
-
-        if not self.client:
-            raise RuntimeError("LLM client not available")
-
-        stream_enabled = bool(getattr(self.settings, "stream", False))
-
-        def _create_message():
-            request_payload: dict[str, Any] = {
-                "model": self.settings.model_standard,
-                "max_tokens": 4096,
-                "temperature": 0.0,
-                "system": system,
-                "messages": [{"role": "user", "content": user}],
-            }
-            if self.settings.protocol == LLMProtocolType.ANTHROPIC and stream_enabled:
-                request_payload["stream"] = True
-            return self.client.messages.create(**request_payload)  # type: ignore[union-attr]
-
-        message = await asyncio.to_thread(_create_message)
-        if self.settings.protocol == LLMProtocolType.ANTHROPIC:
-            try:
-                response_text, used_sse_adapter = await asyncio.to_thread(
-                    self._extract_anthropic_response_text_with_metadata,
-                    message,
-                )
-            except Exception as exc:
-                if not stream_enabled and self._is_anthropic_sse_parse_error(exc):
-                    self.logger.warning(
-                        "anthropic_non_stream_sse_parse_failed_no_fallback",
-                        protocol=self.settings.protocol.value,
-                        configured_stream=stream_enabled,
-                        fallback_allowed=False,
-                        adapter_path="anthropic_sse_stream_adapter",
-                        error_type=type(exc).__name__,
-                        error=str(exc),
-                    )
-                raise
-            if not stream_enabled and used_sse_adapter:
-                self.logger.info(
-                    "anthropic_non_stream_sse_adapter_used",
-                    protocol=self.settings.protocol.value,
-                    configured_stream=stream_enabled,
-                    adapter_path="anthropic_sse_stream_adapter",
-                )
-            return response_text
-
-        content_parts = []
-        for block in message.content:
-            text = getattr(block, "text", None)
-            if text:
-                content_parts.append(text)
-
-        return "\n".join(content_parts)
-
-    @staticmethod
-    def _is_anthropic_sse_parse_error(exc: Exception) -> bool:
-        message = str(exc).lower()
-        return "sse" in message or "event-stream" in message
-
-    def _normalize_anthropic_text_payload_with_diagnostics(self, raw: str, *, source: str) -> str:
-        payload = str(raw or "")
-        configured_stream = bool(getattr(self.settings, "stream", False))
-        adapter_used = (not configured_stream) and self._looks_like_sse_payload(payload)
-        if adapter_used:
-            self.logger.info(
-                "anthropic_non_stream_sse_adapter_used",
-                protocol=self.settings.protocol.value,
-                configured_stream=configured_stream,
-                adapter_path="anthropic_sse_stream_adapter",
-                source=source,
-            )
-        try:
-            return self._normalize_anthropic_text_payload(payload)
-        except Exception as exc:
-            if (not configured_stream) and self._is_anthropic_sse_parse_error(exc):
-                self.logger.warning(
-                    "anthropic_non_stream_sse_parse_failed_no_fallback",
-                    protocol=self.settings.protocol.value,
-                    configured_stream=configured_stream,
-                    fallback_allowed=False,
-                    adapter_path="anthropic_sse_stream_adapter",
-                    source=source,
-                    error_type=type(exc).__name__,
-                    error=str(exc),
-                )
-            raise
-
-    @staticmethod
-    def _looks_like_sse_payload(raw: str) -> bool:
-        lines = [line.strip() for line in str(raw or "").splitlines() if line.strip()]
-        if not lines:
-            return False
-        markers = ("event:", "data:", "id:", "retry:", ":")
-        return all(line.startswith(markers) for line in lines[: min(3, len(lines))])
-
-    @classmethod
-    def _parse_sse_data_payloads(cls, raw: str) -> list[str]:
-        payloads: list[str] = []
-        data_lines: list[str] = []
-        saw_sse_marker = False
-
-        def _flush() -> None:
-            if not data_lines:
-                return
-            data_payload = "\n".join(data_lines).strip()
-            data_lines.clear()
-            if not data_payload or data_payload == "[DONE]":
-                return
-            payloads.append(data_payload)
-
-        for line in str(raw or "").splitlines():
-            stripped = line.strip()
-            if not stripped:
-                _flush()
-                continue
-            if stripped == "[DONE]":
-                saw_sse_marker = True
-                _flush()
-                continue
-            if stripped.startswith(":"):
-                saw_sse_marker = True
-                continue
-            if stripped.startswith("data:"):
-                saw_sse_marker = True
-                data_lines.append(stripped[5:].lstrip())
-                continue
-            if stripped.startswith("event:") or stripped.startswith("id:") or stripped.startswith("retry:"):
-                saw_sse_marker = True
-                continue
-            raise RuntimeError(f"Unsupported SSE line format: {stripped!r}")
-
-        _flush()
-        if not saw_sse_marker:
-            raise RuntimeError("Payload is not SSE-formatted")
-        return payloads
-
-    @staticmethod
-    def _extract_text_from_content_blocks(content: Any) -> list[str]:
-        text_parts: list[str] = []
-        if isinstance(content, str):
-            if content:
-                text_parts.append(content)
-            return text_parts
-        if not isinstance(content, list):
-            return text_parts
-        for block in content:
-            if isinstance(block, str):
-                if block:
-                    text_parts.append(block)
-                continue
-            if isinstance(block, dict):
-                block_text = block.get("text")
-                if isinstance(block_text, str) and block_text:
-                    text_parts.append(block_text)
-                continue
-            block_text = getattr(block, "text", None)
-            if isinstance(block_text, str) and block_text:
-                text_parts.append(block_text)
-        return text_parts
-
-    @classmethod
-    def _extract_text_from_anthropic_payload(cls, payload: dict[str, Any]) -> list[str]:
-        text_parts: list[str] = []
-        direct_text = payload.get("text")
-        if isinstance(direct_text, str) and direct_text:
-            text_parts.append(direct_text)
-
-        event_type = str(payload.get("type") or "").strip().lower()
-        if event_type == "content_block_delta":
-            delta = payload.get("delta")
-            if isinstance(delta, dict):
-                delta_text = delta.get("text") or delta.get("content_delta") or delta.get("delta")
-                if isinstance(delta_text, str) and delta_text:
-                    text_parts.append(delta_text)
-        elif event_type == "content_block_start":
-            content_block = payload.get("content_block")
-            if isinstance(content_block, dict):
-                block_text = content_block.get("text")
-                if isinstance(block_text, str) and block_text:
-                    text_parts.append(block_text)
-        elif event_type == "message_start":
-            message = payload.get("message")
-            if isinstance(message, dict):
-                text_parts.extend(cls._extract_text_from_content_blocks(message.get("content")))
-
-        choices = payload.get("choices")
-        if isinstance(choices, list):
-            for choice in choices:
-                if not isinstance(choice, dict):
-                    continue
-                delta = choice.get("delta")
-                if isinstance(delta, dict):
-                    delta_content = delta.get("content")
-                    if isinstance(delta_content, str) and delta_content:
-                        text_parts.append(delta_content)
-        return text_parts
-
-    @classmethod
-    def _normalize_strict_anthropic_sse_payload(cls, raw: str) -> str:
-        text_parts: list[str] = []
-        for data_payload in cls._parse_sse_data_payloads(raw):
-            try:
-                parsed = json.loads(data_payload)
-            except json.JSONDecodeError as exc:
-                raise RuntimeError(f"Invalid SSE JSON payload: {data_payload!r}") from exc
-
-            if isinstance(parsed, dict):
-                text_parts.extend(cls._extract_text_from_anthropic_payload(parsed))
-            elif isinstance(parsed, str):
-                stripped = parsed.strip()
-                if stripped and stripped != "[DONE]":
-                    text_parts.append(stripped)
-            else:
-                raise RuntimeError(f"Unsupported SSE payload type: {type(parsed).__name__}")
-
-        if not text_parts:
-            raise RuntimeError("Anthropic SSE payload did not contain text deltas")
-        return "".join(text_parts)
-
-    @classmethod
-    def _normalize_anthropic_text_payload(cls, raw: str) -> str:
-        text = str(raw or "")
-        if not text.strip():
-            return ""
-        if cls._looks_like_sse_payload(text):
-            return cls._normalize_strict_anthropic_sse_payload(text)
-        return text
-
-    @classmethod
-    def _extract_anthropic_text_from_stream_event(cls, event: Any) -> list[str]:
-        if event is None:
-            return []
-        if isinstance(event, bytes):
-            event = event.decode("utf-8", errors="replace")
-        if isinstance(event, str):
-            stripped = event.strip()
-            if not stripped:
-                return []
-            if cls._looks_like_sse_payload(stripped):
-                return [cls._normalize_strict_anthropic_sse_payload(stripped)]
-            if stripped.startswith("{") and stripped.endswith("}"):
-                try:
-                    decoded = json.loads(stripped)
-                except json.JSONDecodeError:
-                    return [stripped]
-                if isinstance(decoded, dict):
-                    extracted = cls._extract_text_from_anthropic_payload(decoded)
-                    return extracted or [stripped]
-                return [stripped]
-            return [stripped]
-        if isinstance(event, dict):
-            return cls._extract_text_from_anthropic_payload(event)
-
-        text_parts: list[str] = []
-        event_type = str(getattr(event, "type", "") or "").strip().lower()
-        if event_type == "content_block_delta":
-            delta = getattr(event, "delta", None)
-            delta_text = getattr(delta, "text", None) or getattr(delta, "content_delta", None)
-            if isinstance(delta_text, str) and delta_text:
-                text_parts.append(delta_text)
-        elif event_type == "content_block_start":
-            content_block = getattr(event, "content_block", None)
-            block_text = getattr(content_block, "text", None)
-            if isinstance(block_text, str) and block_text:
-                text_parts.append(block_text)
-        elif event_type == "message_start":
-            message = getattr(event, "message", None)
-            content = getattr(message, "content", None) if message is not None else None
-            text_parts.extend(cls._extract_text_from_content_blocks(content))
-
-        direct_text = getattr(event, "text", None)
-        if isinstance(direct_text, str) and direct_text:
-            text_parts.append(direct_text)
-        return text_parts
-
-    @classmethod
-    def _consume_anthropic_stream_text(cls, stream_response: Any) -> str:
-        text_parts: list[str] = []
-        consumed_events = False
-
-        text_stream = getattr(stream_response, "text_stream", None)
-        if text_stream is not None:
-            for chunk in text_stream:
-                consumed_events = True
-                normalized = cls._normalize_anthropic_text_payload(str(chunk))
-                if normalized:
-                    text_parts.append(normalized)
-        elif hasattr(stream_response, "__iter__"):
-            for event in stream_response:
-                consumed_events = True
-                text_parts.extend(cls._extract_anthropic_text_from_stream_event(event))
-
-        if not consumed_events:
-            raise RuntimeError("Anthropic stream response was not iterable")
-        if not text_parts:
-            raise RuntimeError("Anthropic stream response did not contain text")
-        return "".join(text_parts)
-
-    @classmethod
-    def _extract_anthropic_response_text(cls, response: Any) -> str:
-        text, _used_sse_adapter = cls._extract_anthropic_response_text_with_metadata(response)
-        return text
-
-    @classmethod
-    def _extract_anthropic_response_text_with_metadata(cls, response: Any) -> tuple[str, bool]:
-        if response is None:
-            raise RuntimeError("Anthropic response is empty")
-
-        if isinstance(response, (str, bytes)):
-            raw_response = response.decode("utf-8", errors="replace") if isinstance(response, bytes) else response
-            normalized = cls._normalize_anthropic_text_payload(raw_response)
-            if not normalized:
-                raise RuntimeError("Anthropic response text is empty")
-            return normalized, cls._looks_like_sse_payload(raw_response)
-
-        content = getattr(response, "content", None)
-        if isinstance(content, list):
-            parts: list[str] = []
-            for block in content:
-                block_text = getattr(block, "text", None)
-                if isinstance(block_text, str) and block_text:
-                    parts.append(cls._normalize_anthropic_text_payload(block_text))
-            if parts:
-                return "\n".join(part for part in parts if part).strip(), False
-
-        if hasattr(response, "__enter__") and hasattr(response, "__exit__"):
-            with response as managed:
-                return cls._consume_anthropic_stream_text(managed), True
-
-        return cls._consume_anthropic_stream_text(response), True
-
-    def _build_pydantic_ai_model(self, model_name: str):
-        """Create a PydanticAI model wrapper based on LLM settings."""
-        from pydantic_ai.models.anthropic import AnthropicModel
-        from pydantic_ai.models.openai import OpenAIChatModel, OpenAIResponsesModel
-        from pydantic_ai.providers.openai import OpenAIProvider
-
-        protocol = self.settings.protocol
-        protocol_value = self._protocol_value(protocol)
-        if protocol_value not in PYDANTIC_AI_MODEL_SUPPORTED_PROTOCOL_VALUES:
-            raise ValueError(f"Unsupported protocol for PydanticAI model builder: {protocol_value}")
-
-        if protocol == LLMProtocolType.GEMINI:
-            from pydantic_ai.models.google import GoogleModel
-            from pydantic_ai.providers.google import GoogleProvider
-
-            api_key = resolve_protocol_api_key(self.settings, protocol, allow_openai_dummy=False)
-            provider = GoogleProvider(
-                api_key=api_key,
-                base_url=resolve_protocol_base_url(self.settings, protocol),
-            )
-            return GoogleModel(model_name, provider=provider, settings=self._model_settings)
-
-        if protocol == LLMProtocolType.ANTHROPIC:
-            api_key = resolve_protocol_api_key(self.settings, protocol, allow_openai_dummy=False)
-            provider = "anthropic"
-            base_url = resolve_protocol_base_url(self.settings, protocol)
-            if api_key or base_url:
-                from vds_audit_orchestrator.llm.provider import build_pydantic_anthropic_provider
-
-                provider = build_pydantic_anthropic_provider(
-                    api_key=api_key,
-                    base_url=base_url,
-                )
-            return AnthropicModel(model_name, provider=provider, settings=self._model_settings)
-
-        if self._is_openai_codex_protocol(protocol):
-            from openai import AsyncOpenAI
-
-            try:
-                from vds_audit_orchestrator.llm.codex_oauth import get_codex_credentials
-            except ImportError as exc:  # pragma: no cover - guarded by protocol path
-                raise RuntimeError("openai-codex protocol requires llm.codex_oauth.get_codex_credentials") from exc
-
-            credentials = get_codex_credentials()
-            access_token = self._resolve_codex_access_token(credentials)
-            chatgpt_account_id = self._resolve_codex_account_id(credentials)
-            self._warn_if_unvalidated_codex_model(model_name)
-            base_url = resolve_protocol_base_url(self.settings, protocol)
-            if not base_url:
-                raise ValueError("openai-codex protocol requires VDS_AUDIT_LLM__BASE_URL")
-            default_headers = {"ChatGPT-Account-Id": chatgpt_account_id} if chatgpt_account_id else None
-            openai_client = AsyncOpenAI(
-                api_key=access_token,
-                base_url=base_url,
-                default_headers=default_headers,
-            )
-            provider = OpenAIProvider(openai_client=openai_client)
-            model_settings = self._build_openai_codex_model_settings(self._model_settings)
-            return OpenAIResponsesModel(
-                model_name,
-                provider=provider,
-                settings=model_settings,
-            )
-
-        base_url = resolve_protocol_base_url(self.settings, protocol)
-        api_key = resolve_openai_protocol_api_key(self.settings)
-        provider = OpenAIProvider(api_key=api_key, base_url=base_url)
-        return OpenAIChatModel(model_name, provider=provider, settings=self._model_settings)
-
-    @staticmethod
-    def _protocol_value(protocol: Any) -> str:
-        return str(getattr(protocol, "value", protocol) or "").strip().lower()
-
-    @classmethod
-    def _is_openai_codex_protocol(cls, protocol: Any | None = None) -> bool:
-        return cls._protocol_value(protocol) == OPENAI_CODEX_PROTOCOL
-
-    @staticmethod
-    def _read_credential_field(credentials: Any, *field_names: str) -> str | None:
-        for field_name in field_names:
-            if isinstance(credentials, dict):
-                raw_value = credentials.get(field_name)
-            else:
-                raw_value = getattr(credentials, field_name, None)
-            value = str(raw_value or "").strip()
-            if value:
-                return value
-        return None
-
-    @classmethod
-    def _resolve_codex_access_token(cls, credentials: Any) -> str:
-        access_token = cls._read_credential_field(credentials, "access_token", "token")
-        if not access_token:
-            raise ValueError("Codex OAuth credentials missing access token")
-        return access_token
-
-    @classmethod
-    def _resolve_codex_account_id(cls, credentials: Any) -> str | None:
-        return cls._read_credential_field(credentials, "chatgpt_account_id", "account_id")
-
-    @staticmethod
-    def _build_openai_codex_model_settings(settings: ModelSettings | None) -> ModelSettings:
-        merged: dict[str, Any] = {}
-        if isinstance(settings, dict):
-            merged.update(settings)
-        elif settings is not None:
-            model_dump = getattr(settings, "model_dump", None)
-            if callable(model_dump):
-                dumped = model_dump()
-                if isinstance(dumped, dict):
-                    merged.update(dumped)
-        merged["openai_store"] = False
-        return ModelSettings(**merged)
-
-    async def _run_streaming_pydantic_operation(
-        self,
-        *,
-        agent: Any,
-        user_prompt: str,
-        run_kwargs: dict[str, Any],
-    ) -> _PydanticStreamResultWrapper:
-        stream_result: Any | None = None
-        try:
-            async with agent.run_stream(user_prompt, **run_kwargs) as active_stream_result:
-                stream_result = active_stream_result
-                output = await stream_result.get_output()
-                return _PydanticStreamResultWrapper(output=output, streamed_result=stream_result)
-        except Exception as exc:
-            self._attach_streamed_output_payload(exc, stream_result)
-            if self._is_codex_streaming_requirement_error(exc):
-                raise RuntimeError(
-                    "Codex subscription provider requires streaming mode; use openai-codex with run_stream() semantics."
-                ) from exc
-            raise
-
-    @staticmethod
-    def _extract_streamed_output_payload(stream_result: Any | None) -> dict[str, Any] | str | None:
-        response = getattr(stream_result, "response", None)
-        tool_calls = getattr(response, "tool_calls", None)
-        if isinstance(tool_calls, list):
-            for tool_call in reversed(tool_calls):
-                args_as_dict = getattr(tool_call, "args_as_dict", None)
-                if callable(args_as_dict):
-                    try:
-                        payload = args_as_dict()
-                    except Exception:
-                        payload = None
-                    if isinstance(payload, dict) and payload:
-                        return payload
-                args_as_json_str = getattr(tool_call, "args_as_json_str", None)
-                if callable(args_as_json_str):
-                    try:
-                        payload = args_as_json_str()
-                    except Exception:
-                        payload = None
-                    if isinstance(payload, str) and payload.strip():
-                        return payload
-                payload = getattr(tool_call, "args", None)
-                if isinstance(payload, dict) and payload:
-                    return payload
-                if isinstance(payload, str) and payload.strip():
-                    return payload
-        return None
-
-    @classmethod
-    def _attach_streamed_output_payload(cls, exc: Exception, stream_result: Any | None) -> None:
-        payload = cls._extract_streamed_output_payload(stream_result)
-        if payload is None:
-            return
-        with contextlib.suppress(Exception):
-            exc.streamed_output_payload = payload
-        if getattr(exc, "body", None) is None:
-            with contextlib.suppress(Exception):
-                exc.body = payload
-        if getattr(exc, "response", None) is None:
-            with contextlib.suppress(Exception):
-                exc.response = payload
-
-    def _resolve_agent_stream_mode(self) -> tuple[bool, str]:
-        if self._is_openai_codex_protocol(self.settings.protocol):
-            return True, "protocol_override"
-        if bool(getattr(self.settings, "agent_stream", True)):
-            return True, "profile"
-        if bool(getattr(self.settings, "stream", False)):
-            return True, "transport_floor"
-        return False, "profile"
-
-    def _build_agent_operation(
-        self,
-        *,
-        agent: Any,
-        user_prompt: str,
-        run_kwargs: dict[str, Any],
-    ) -> Coroutine[Any, Any, Any]:
-        use_streaming, source = self._resolve_agent_stream_mode()
-        self.logger.info(
-            "pydantic_agent_stream_mode",
-            mode="stream" if use_streaming else "non_stream",
-            source=source,
-            protocol=self._protocol_value(self.settings.protocol),
-            model_name=getattr(self.settings, "model_standard", None),
-        )
-        if use_streaming:
-            return self._run_streaming_pydantic_operation(
-                agent=agent,
-                user_prompt=user_prompt,
-                run_kwargs=dict(run_kwargs),
-            )
-        return agent.run(user_prompt, **run_kwargs)
-
-    @staticmethod
-    def _is_codex_streaming_requirement_error(exc: Exception) -> bool:
-        message = str(exc).lower()
-        return "stream must be set to true" in message or "stream=true" in message or "streaming required" in message
-
-    def _warn_if_unvalidated_codex_model(self, model_name: str) -> None:
-        if str(model_name).strip().lower() in OPENAI_CODEX_BASELINE_MODELS:
-            return
-        self.logger.warning(
-            "codex_unsupported_model_name",
-            model_name=model_name,
-            baseline_models=sorted(OPENAI_CODEX_BASELINE_MODELS),
-        )
-
-    async def _run_pydantic_agent(
-        self,
-        *,
-        system_prompt: str,
-        user_prompt: str,
-        result_type: type[ResultType],
-        complexity: str = "standard",
-        max_tokens: int = 4096,
-        prompt_type: str = "agent",
-        deps: Any | None = None,
-        toolsets: list[Any] | None = None,
-        max_turns: int | None = None,
-        max_tool_calls: int | None = None,
-        pre_retry_telemetry: dict[str, Any] | None = None,
-        row_failover_context: RowFailoverContext | None = None,
-        provider_health_memory: ProviderHealthMemory | None = None,
-    ) -> ResultType:
-        """Run a PydanticAI agent and return typed output.
-
-        Args:
-            toolsets: Optional list of PydanticAI toolsets (Phase 41).
-                When provided, the agent is created with tool-use capabilities
-                for multi-turn agentic investigation.
-            max_turns: Optional max request/turn limit for bounded agentic runs.
-            max_tool_calls: Optional max tool call limit for bounded agentic runs.
-        """
-        import time
-
-        if not self.settings.enabled:
-            raise RuntimeError("LLM is disabled")
-
-        operation_state: dict[str, Any] = {
-            "operation_id": uuid4().hex,
-            "operation_kind": "pydantic_agent_run",
-            "operation_status": "started",
-            "operation_outcome": "started",
-            "operation_started_at": time.monotonic(),
-            "model_name": self.settings.model_standard,
-            "event_stream_enabled": False,
-            "rate_limit_retry_attempts": 0,
-            "tool_call_retry_attempts": 0,
-            "codex_auth_retry_attempts": 0,
-            "run_telemetry": None,
-            "error": None,
-        }
-
-        @contextlib.contextmanager
-        def _operation_lifecycle_scope():
-            self.logger.info(
-                "agent_operation_lifecycle",
-                operation_id=operation_state["operation_id"],
-                operation_kind=operation_state["operation_kind"],
-                operation_status="started",
-                operation_outcome=operation_state["operation_outcome"],
-                result_type=result_type.__name__,
-                prompt_type=prompt_type,
-                protocol=self.settings.protocol.value,
-                model_name=operation_state.get("model_name"),
-            )
-            try:
-                yield
-            except Exception as exc:
-                operation_state["operation_status"] = "failed"
-                operation_state["operation_outcome"] = "error"
-                operation_state["error"] = str(exc)
-                raise
-            finally:
-                elapsed_sec = max(0.0, time.monotonic() - float(operation_state["operation_started_at"]))
-                run_telemetry = operation_state.get("run_telemetry")
-                event_tool_calls_started = 0
-                event_tool_calls_completed = 0
-                event_skill_tool_calls = 0
-                event_skill_effective_tool_calls = 0
-                observed_tool_names: list[str] | None = None
-                if isinstance(run_telemetry, dict):
-                    event_tool_calls_started = int(run_telemetry.get("event_tool_calls_started", 0) or 0)
-                    event_tool_calls_completed = int(
-                        run_telemetry.get("event_tool_calls_completed", run_telemetry.get("event_tool_calls", 0)) or 0
-                    )
-                    event_skill_tool_calls = int(run_telemetry.get("event_skill_tool_calls", 0) or 0)
-                    event_skill_effective_tool_calls = int(
-                        run_telemetry.get("event_skill_effective_tool_calls", 0) or 0
-                    )
-                    observed_tool_names = list(run_telemetry.get("event_tool_names") or [])
-                    run_telemetry["operation_id"] = operation_state["operation_id"]
-                    run_telemetry["operation_kind"] = operation_state["operation_kind"]
-                    run_telemetry["operation_status"] = operation_state["operation_status"]
-                    run_telemetry["operation_outcome"] = operation_state["operation_outcome"]
-                    run_telemetry["operation_elapsed_sec"] = round(elapsed_sec, 3)
-                    run_telemetry["operation_rate_limit_retry_attempts"] = int(
-                        operation_state.get("rate_limit_retry_attempts", 0) or 0
-                    )
-                    run_telemetry["operation_tool_call_retry_attempts"] = int(
-                        operation_state.get("tool_call_retry_attempts", 0) or 0
-                    )
-                    run_telemetry["operation_codex_auth_retry_attempts"] = int(
-                        operation_state.get("codex_auth_retry_attempts", 0) or 0
-                    )
-                    if operation_state.get("error"):
-                        run_telemetry["operation_error"] = str(operation_state["error"])
-
-                log_method = (
-                    self.logger.warning if operation_state["operation_status"] == "failed" else self.logger.info
-                )
-                log_method(
-                    "agent_operation_lifecycle",
-                    operation_id=operation_state["operation_id"],
-                    operation_kind=operation_state["operation_kind"],
-                    operation_status=operation_state["operation_status"],
-                    operation_outcome=operation_state["operation_outcome"],
-                    elapsed_sec=round(elapsed_sec, 2),
-                    result_type=result_type.__name__,
-                    prompt_type=prompt_type,
-                    protocol=self.settings.protocol.value,
-                    model_name=operation_state.get("model_name"),
-                    event_stream_enabled=bool(operation_state.get("event_stream_enabled", False)),
-                    rate_limit_retry_attempts=int(operation_state.get("rate_limit_retry_attempts", 0) or 0),
-                    tool_call_retry_attempts=int(operation_state.get("tool_call_retry_attempts", 0) or 0),
-                    codex_auth_retry_attempts=int(operation_state.get("codex_auth_retry_attempts", 0) or 0),
-                    event_tool_calls_started=event_tool_calls_started,
-                    event_tool_calls_completed=event_tool_calls_completed,
-                    event_skill_tool_calls=event_skill_tool_calls,
-                    event_skill_effective_tool_calls=event_skill_effective_tool_calls,
-                    observed_tool_names=observed_tool_names,
-                    error=operation_state.get("error"),
-                )
-
-        with log_context(agent=self.__class__.__name__, phase="agent"), _operation_lifecycle_scope():
-            provider_health_memory = provider_health_memory or ProviderHealthMemory()
-            self.logger.info(
-                "pydantic_agent_call_started",
-                result_type=result_type.__name__,
-                complexity=complexity,
-                protocol=self.settings.protocol.value,
-                operation_id=operation_state["operation_id"],
-                operation_kind=operation_state["operation_kind"],
-            )
-            start_time = time.monotonic()
-            system_prompt_chars = len(system_prompt)
-            user_prompt_chars = len(user_prompt)
-            total_prompt_chars = system_prompt_chars + user_prompt_chars
-            approx_prompt_tokens = max(1, total_prompt_chars // 4)
-            self._log_trace_payload(
-                event_name="pydantic_agent_trace_prompt",
-                elapsed_sec=0.0,
-                payload={
-                    "operation_id": operation_state["operation_id"],
-                    "result_type": result_type.__name__,
-                    "complexity": complexity,
-                    "protocol": self.settings.protocol.value,
-                    "system_prompt": system_prompt,
-                    "user_prompt": user_prompt,
-                    "max_turns": max_turns,
-                    "max_tool_calls": max_tool_calls,
-                },
-            )
-
-            if self._pydantic_model is None and self.client is not None:
-                response = await self._call_llm(system_prompt, user_prompt)
-                payload = self._strip_json_fence(response)
-                elapsed = time.monotonic() - start_time
-                operation_state["operation_status"] = "completed"
-                operation_state["operation_outcome"] = "direct_call"
-                self.logger.info(
-                    "pydantic_agent_call_complete",
-                    elapsed_sec=round(elapsed, 2),
-                    mode="direct_call",
-                    operation_id=operation_state["operation_id"],
-                )
-                return result_type.model_validate_json(payload)  # type: ignore[return-value]
-
-            from vds_audit_orchestrator.logging_config import audit_otel_span as trace_span
-
-            # Select model based on complexity
-            model_name = self.settings.select_model(complexity) if hasattr(self.settings, "select_model") else None
-            selected_model = model_name or self.settings.model_standard
-            operation_state["model_name"] = selected_model
-
-            self.logger.debug(
-                "pydantic_agent_created",
-                model_name=selected_model,
-                output_type=result_type.__name__,
-            )
-            self.logger.info(
-                "pydantic_agent_prompt_metrics",
-                result_type=result_type.__name__,
-                model_name=selected_model,
-                system_prompt_chars=system_prompt_chars,
-                user_prompt_chars=user_prompt_chars,
-                total_prompt_chars=total_prompt_chars,
-                approx_prompt_tokens=approx_prompt_tokens,
-                max_tokens=max_tokens or self.settings.max_tokens_per_request,
-                toolset_count=len(toolsets or []),
-                has_deps=deps is not None,
-                max_turns=max_turns,
-                max_tool_calls=max_tool_calls,
-            )
-
-            # Use direct Agent creation with the selected model for test compatibility
-            from pydantic_ai import Agent
-            from pydantic_ai.settings import ModelSettings
-            from pydantic_ai.usage import UsageLimits
-
-            # Build the model - use cached or build new
-            model = self._pydantic_model or self._build_pydantic_ai_model(selected_model)
-
-            agent_retries, output_retries = self._resolve_agent_retry_policy()
-            agent_kwargs: dict[str, Any] = {
-                "model": model,
-                "name": self.__class__.__name__,
-                "output_type": result_type,
-                "instructions": system_prompt,
-                "retries": agent_retries,
-                "output_retries": output_retries,
-                "model_settings": ModelSettings(
-                    **self._build_run_model_settings(max_tokens or self.settings.max_tokens_per_request)
-                ),
-            }
-            self.logger.debug(
-                "pydantic_agent_retry_policy",
-                retries=agent_retries,
-                output_retries=output_retries,
-                configured_retries=getattr(self.settings, "agent_retries", None),
-                configured_output_retries=getattr(self.settings, "output_retries", None),
-                protocol=self.settings.protocol.value,
-            )
-            skill_policy_diagnostics = self._collect_skill_policy_diagnostics(toolsets)
-            if toolsets:
-                agent_kwargs["toolsets"] = toolsets
-                self.logger.info(
-                    "pydantic_agent_toolsets_attached",
-                    toolset_count=len(toolsets),
-                )
-            if skill_policy_diagnostics:
-                self.logger.info("pydantic_agent_skill_policy_diagnostics", **skill_policy_diagnostics)
-
-            agent: Agent[Any, Any] = Agent(**agent_kwargs)
-            usage_limits: UsageLimits | None = None
-            request_limit, tool_calls_limit = self._resolve_pydantic_usage_limits(
-                max_turns=max_turns,
-                max_tool_calls=max_tool_calls,
-            )
-            if request_limit is not None or tool_calls_limit is not None:
-                usage_limits = UsageLimits(
-                    request_limit=request_limit,
-                    tool_calls_limit=tool_calls_limit,
-                )
-                self.logger.info(
-                    "pydantic_agent_usage_limits_applied",
-                    max_turns=max_turns,
-                    max_tool_calls=max_tool_calls,
-                    request_limit=request_limit,
-                    tool_calls_limit=tool_calls_limit,
-                )
-
-            with trace_span(
-                "llm.call",
-                {
-                    "llm.model": model_name or self.settings.model_standard,
-                    "llm.protocol": self.settings.protocol.value,
-                },
-            ):
-                run_telemetry: dict[str, Any] = {
-                    "operation_id": operation_state["operation_id"],
-                    "operation_kind": operation_state["operation_kind"],
-                    "result_type": result_type.__name__,
-                    "protocol": self.settings.protocol.value,
-                    "model_name": selected_model,
-                    "event_tool_calls": 0,
-                    "event_tool_calls_started": 0,
-                    "event_tool_effective_calls": 0,
-                    "event_skill_tool_calls": 0,
-                    "event_skill_discovery_tool_calls": 0,
-                    "event_skill_bootstrap_tool_calls": 0,
-                    "event_skill_execution_tool_calls": 0,
-                    "event_skill_execution_effective_tool_calls": 0,
-                    "event_tool_calls_completed": 0,
-                    "event_skill_effective_tool_calls": 0,
-                    "event_tool_names": [],
-                    "event_tool_signatures": [],
-                    "event_tool_count_by_name": {},
-                    "event_tool_effective_count_by_name": {},
-                    "event_skill_tool_count_by_name": {},
-                    "event_tool_latencies_ms": [],
-                    "_tool_call_started_at": {},
-                    "event_last_seen_at": time.monotonic(),
-                    "event_last_tool_activity_at": None,
-                }
-                operation_state["run_telemetry"] = run_telemetry
-                if skill_policy_diagnostics:
-                    run_telemetry["skill_policy_diagnostics"] = skill_policy_diagnostics
-                # Keep latest telemetry available even when the run fails (e.g., request_limit),
-                # so higher-level retry/guard logic can inspect partial tool usage.
-                self._last_agent_run_telemetry = run_telemetry
-                run_kwargs: dict[str, Any] = {}
-                if deps is not None:
-                    run_kwargs["deps"] = deps
-                if usage_limits is not None:
-                    run_kwargs["usage_limits"] = usage_limits
-                event_stream_requested = bool(getattr(self.settings, "agent_event_stream_enabled", True))
-                suppress_event_stream_for_endpoint = (
-                    event_stream_requested and self._should_suppress_anthropic_tool_choice()
-                )
-                if event_stream_requested and not suppress_event_stream_for_endpoint:
-                    run_kwargs["event_stream_handler"] = self._build_event_stream_handler(start_time, run_telemetry)
-                elif suppress_event_stream_for_endpoint:
-                    self.logger.info(
-                        "anthropic_proxy_event_stream_disabled",
-                        base_url=self.settings.base_url,
-                        protocol=self.settings.protocol.value,
-                    )
-                timeout_policy = self._resolve_agent_timeout_policy(complexity)
-                self.logger.info(
-                    "pydantic_agent_timeout_policy",
-                    result_type=result_type.__name__,
-                    initial_timeout_seconds=round(timeout_policy.initial_timeout_seconds, 2),
-                    max_timeout_seconds=round(timeout_policy.max_timeout_seconds, 2),
-                    extension_seconds=round(timeout_policy.extension_seconds, 2),
-                    extension_attempts=timeout_policy.extension_attempts,
-                    heartbeat_seconds=round(timeout_policy.heartbeat_seconds, 2),
-                    idle_post_tool_seconds=round(timeout_policy.idle_post_tool_seconds, 2),
-                )
-                event_stream_enabled = "event_stream_handler" in run_kwargs
-                operation_state["event_stream_enabled"] = event_stream_enabled
-                rate_limit_retry_attempt = 0
-                max_rate_limit_retries = max(0, int(getattr(self.settings, "agent_rate_limit_retry_attempts", 2) or 0))
-                base_rate_limit_backoff = float(
-                    getattr(self.settings, "agent_rate_limit_retry_backoff_seconds", 2.0) or 2.0
-                )
-                max_retry_after_seconds = 120.0
-                tool_call_retry_attempt = 0
-                max_tool_call_retries = max(0, int(getattr(self.settings, "agent_tool_call_retry_attempts", 2) or 0))
-                base_tool_call_backoff = float(
-                    getattr(self.settings, "agent_tool_call_retry_backoff_seconds", 1.0) or 1.0
-                )
-                codex_auth_retry_attempt = 0
-                max_codex_auth_retries = 1
-                try:
-                    result = await self._run_with_timeout_policy(
-                        operation=self._build_agent_operation(
-                            agent=agent, user_prompt=user_prompt, run_kwargs=run_kwargs
-                        ),
-                        timeout_policy=timeout_policy,
-                        result_type_name=result_type.__name__,
-                        run_telemetry=run_telemetry,
-                    )
-                except Exception as exc:
-                    recovered_output = self._recover_output_from_final_result_validation_error(
-                        result_type=result_type,
-                        exc=exc,
-                        telemetry=run_telemetry,
-                    )
-                    if recovered_output is not None:
-                        self.logger.warning(
-                            "pydantic_agent_output_recovered_from_final_result_payload",
-                            result_type=result_type.__name__,
-                        )
-                        run_telemetry["output_recovered_from_final_result_payload"] = True
-                        run_telemetry.pop("_tool_call_started_at", None)
-                        self._last_agent_run_telemetry = run_telemetry
-                        operation_state["operation_status"] = "completed"
-                        operation_state["operation_outcome"] = "recovered_output"
-                        return recovered_output
-                    if (
-                        self.settings.protocol == LLMProtocolType.ANTHROPIC
-                        and bool(getattr(self.settings, "base_url", ""))
-                        and self._is_tool_choice_protocol_mismatch(exc)
-                    ):
-                        raise RuntimeError(
-                            "anthropic protocol mismatch: endpoint rejected tool_choice/function payload. "
-                            "Use an Anthropic Messages-compatible proxy."
-                        ) from exc
-                    auth_recovered = False
-                    while (
-                        self._is_openai_codex_protocol(self.settings.protocol)
-                        and self._is_auth_error(exc)
-                        and codex_auth_retry_attempt < max_codex_auth_retries
-                    ):
-                        codex_auth_retry_attempt += 1
-                        operation_state["codex_auth_retry_attempts"] = codex_auth_retry_attempt
-                        self.logger.warning(
-                            "pydantic_agent_codex_auth_retry",
-                            result_type=result_type.__name__,
-                            retry_attempt=codex_auth_retry_attempt,
-                            retry_attempts=max_codex_auth_retries,
-                            error=str(exc),
-                        )
-                        try:
-                            from vds_audit_orchestrator.llm.codex_oauth import force_refresh_codex_credentials
-
-                            force_refresh_codex_credentials()
-                            refreshed_model = self._build_pydantic_ai_model(selected_model)
-                            self._pydantic_model = refreshed_model
-                            refreshed_agent_kwargs = dict(agent_kwargs)
-                            refreshed_agent_kwargs["model"] = refreshed_model
-                            agent = Agent(**refreshed_agent_kwargs)
-                            result = await self._run_with_timeout_policy(
-                                operation=self._build_agent_operation(
-                                    agent=agent, user_prompt=user_prompt, run_kwargs=run_kwargs
-                                ),
-                                timeout_policy=timeout_policy,
-                                result_type_name=result_type.__name__,
-                                run_telemetry=run_telemetry,
-                            )
-                            auth_recovered = True
-                            self.logger.info(
-                                "pydantic_agent_codex_auth_recovered",
-                                result_type=result_type.__name__,
-                                retry_attempt=codex_auth_retry_attempt,
-                            )
-                            break
-                        except Exception as auth_retry_exc:
-                            exc = auth_retry_exc
-                    rate_limit_recovered = False
-                    if auth_recovered:
-                        rate_limit_recovered = True
-                    while self._is_rate_limit_error(exc) and rate_limit_retry_attempt < max_rate_limit_retries:
-                        rate_limit_retry_attempt += 1
-                        operation_state["rate_limit_retry_attempts"] = rate_limit_retry_attempt
-                        classification = ProviderFailureClassifier.classify(exc)
-                        retry_backoff = classification.retry_after_seconds
-                        if retry_backoff is not None:
-                            retry_backoff = min(max_retry_after_seconds, max(0.0, retry_backoff))
-                        if retry_backoff is None:
-                            retry_backoff = min(
-                                30.0,
-                                base_rate_limit_backoff * (2 ** (rate_limit_retry_attempt - 1)),
-                            )
-                        # Keep tiny jitter to reduce synchronized retry bursts from parallel agents.
-                        retry_backoff += random.uniform(0.0, min(1.0, retry_backoff * 0.2))
-                        run_telemetry["provider_transient_retry_count"] = rate_limit_retry_attempt
-                        run_telemetry["retry_after_honored"] = classification.retry_after_seconds is not None
-                        run_telemetry["retry_after_seconds"] = (
-                            round(retry_backoff, 2) if classification.retry_after_seconds is not None else 0.0
-                        )
-                        run_telemetry["provider_retry_class"] = classification.failure_class.value
-                        run_telemetry["provider_retry_reason"] = classification.classification_reason
-                        self.logger.warning(
-                            "pydantic_agent_rate_limit_retry",
-                            result_type=result_type.__name__,
-                            retry_attempt=rate_limit_retry_attempt,
-                            retry_attempts=max_rate_limit_retries,
-                            backoff_seconds=round(retry_backoff, 2),
-                            retry_after_honored=classification.retry_after_seconds is not None,
-                            provider_failure_class=classification.failure_class.value,
-                            retry_reason=classification.classification_reason,
-                            error=str(exc),
-                        )
-                        await asyncio.sleep(retry_backoff)
-                        try:
-                            result = await self._run_with_timeout_policy(
-                                operation=self._build_agent_operation(
-                                    agent=agent, user_prompt=user_prompt, run_kwargs=run_kwargs
-                                ),
-                                timeout_policy=timeout_policy,
-                                result_type_name=result_type.__name__,
-                                run_telemetry=run_telemetry,
-                            )
-                            rate_limit_recovered = True
-                            self.logger.info(
-                                "pydantic_agent_rate_limit_recovered",
-                                result_type=result_type.__name__,
-                                retry_attempt=rate_limit_retry_attempt,
-                                provider_failure_class=classification.failure_class.value,
-                            )
-                            break
-                        except Exception as rate_limit_retry_exc:
-                            exc = rate_limit_retry_exc
-                    tool_call_recovered = False
-                    while self._is_tool_call_argument_error(exc) and tool_call_retry_attempt < max_tool_call_retries:
-                        tool_call_retry_attempt += 1
-                        operation_state["tool_call_retry_attempts"] = tool_call_retry_attempt
-                        retry_backoff = min(20.0, base_tool_call_backoff * (2 ** (tool_call_retry_attempt - 1)))
-                        retry_backoff += random.uniform(0.0, min(0.5, retry_backoff * 0.2))
-                        self.logger.warning(
-                            "pydantic_agent_tool_call_retry",
-                            result_type=result_type.__name__,
-                            retry_attempt=tool_call_retry_attempt,
-                            retry_attempts=max_tool_call_retries,
-                            backoff_seconds=round(retry_backoff, 2),
-                            error=str(exc),
-                        )
-                        await asyncio.sleep(retry_backoff)
-                        try:
-                            result = await self._run_with_timeout_policy(
-                                operation=self._build_agent_operation(
-                                    agent=agent, user_prompt=user_prompt, run_kwargs=run_kwargs
-                                ),
-                                timeout_policy=timeout_policy,
-                                result_type_name=result_type.__name__,
-                                run_telemetry=run_telemetry,
-                            )
-                            tool_call_recovered = True
-                            self.logger.info(
-                                "pydantic_agent_tool_call_recovered",
-                                result_type=result_type.__name__,
-                                retry_attempt=tool_call_retry_attempt,
-                            )
-                            break
-                        except Exception as tool_call_retry_exc:
-                            exc = tool_call_retry_exc
-                    if not rate_limit_recovered and not tool_call_recovered:
-                        should_retry_without_stream = event_stream_enabled and self._should_retry_without_event_stream(
-                            exc
-                        )
-                        if not should_retry_without_stream:
-                            failover_result = await self._attempt_row_provider_failover(
-                                exc=exc,
-                                system_prompt=system_prompt,
-                                user_prompt=user_prompt,
-                                result_type=result_type,
-                                complexity=complexity,
-                                max_tokens=max_tokens,
-                                prompt_type=prompt_type,
-                                deps=deps,
-                                toolsets=toolsets,
-                                max_turns=max_turns,
-                                max_tool_calls=max_tool_calls,
-                                pre_retry_telemetry=pre_retry_telemetry,
-                                current_run_telemetry=run_telemetry,
-                                row_failover_context=row_failover_context,
-                                provider_health_memory=provider_health_memory,
-                            )
-                            if failover_result is not None:
-                                operation_state["operation_status"] = "completed"
-                                operation_state["operation_outcome"] = "provider_failover"
-                                run_telemetry.pop("_tool_call_started_at", None)
-                                self._last_agent_run_telemetry = self._merge_failover_telemetry(
-                                    pre_retry_telemetry,
-                                    run_telemetry,
-                                )
-                                return failover_result
-                            elapsed = time.monotonic() - start_time
-                            tool_efficiency = self._build_tool_efficiency_summary(run_telemetry)
-                            if tool_efficiency:
-                                run_telemetry["tool_efficiency"] = tool_efficiency
-                            run_telemetry["error"] = str(exc)
-                            run_telemetry.pop("_tool_call_started_at", None)
-                            self._last_agent_run_telemetry = run_telemetry
-                            self.logger.warning(
-                                "pydantic_agent_call_failed",
-                                elapsed_sec=round(elapsed, 2),
-                                mode="pydantic_ai",
-                                result_type=result_type.__name__,
-                                operation_id=operation_state["operation_id"],
-                                event_tool_calls_started=int(run_telemetry.get("event_tool_calls_started", 0) or 0),
-                                event_tool_calls_completed=int(
-                                    run_telemetry.get(
-                                        "event_tool_calls_completed", run_telemetry.get("event_tool_calls", 0)
-                                    )
-                                    or 0
-                                ),
-                                event_skill_tool_calls=int(run_telemetry.get("event_skill_tool_calls", 0) or 0),
-                                event_skill_effective_tool_calls=int(
-                                    run_telemetry.get("event_skill_effective_tool_calls", 0) or 0
-                                ),
-                                observed_tool_names=run_telemetry.get("event_tool_names"),
-                                **self._exception_diagnostics(exc),
-                            )
-                            raise
-                        self.logger.warning(
-                            "pydantic_agent_event_stream_fallback",
-                            error=str(exc),
-                        )
-                        run_kwargs.pop("event_stream_handler", None)
-                        run_telemetry["event_stream_fallback_activated"] = True
-                        try:
-                            result = await self._run_with_timeout_policy(
-                                operation=self._build_agent_operation(
-                                    agent=agent, user_prompt=user_prompt, run_kwargs=run_kwargs
-                                ),
-                                timeout_policy=timeout_policy,
-                                result_type_name=result_type.__name__,
-                                run_telemetry=run_telemetry,
-                            )
-                            # FR-70 / TSK-641.2: Populate telemetry from result when
-                            # event stream was removed.  Without this the telemetry
-                            # dict stays at zeros and post-run guards raise false
-                            # RuntimeError("agentic usage guard unsatisfied").
-                            self._populate_telemetry_from_result(result, run_telemetry)
-                        except Exception as fallback_exc:
-                            fallback_tool_call_recovered = False
-                            fallback_tool_call_retry_attempt = 0
-                            while (
-                                self._is_tool_call_argument_error(fallback_exc)
-                                and fallback_tool_call_retry_attempt < max_tool_call_retries
-                            ):
-                                fallback_tool_call_retry_attempt += 1
-                                operation_state["tool_call_retry_attempts"] = max(
-                                    int(operation_state.get("tool_call_retry_attempts", 0) or 0),
-                                    fallback_tool_call_retry_attempt,
-                                )
-                                retry_backoff = min(
-                                    20.0,
-                                    base_tool_call_backoff * (2 ** (fallback_tool_call_retry_attempt - 1)),
-                                )
-                                retry_backoff += random.uniform(0.0, min(0.5, retry_backoff * 0.2))
-                                self.logger.warning(
-                                    "pydantic_agent_tool_call_retry",
-                                    result_type=result_type.__name__,
-                                    retry_attempt=fallback_tool_call_retry_attempt,
-                                    retry_attempts=max_tool_call_retries,
-                                    backoff_seconds=round(retry_backoff, 2),
-                                    error=str(fallback_exc),
-                                    retry_mode="event_stream_fallback",
-                                )
-                                await asyncio.sleep(retry_backoff)
-                                try:
-                                    result = await self._run_with_timeout_policy(
-                                        operation=self._build_agent_operation(
-                                            agent=agent, user_prompt=user_prompt, run_kwargs=run_kwargs
-                                        ),
-                                        timeout_policy=timeout_policy,
-                                        result_type_name=result_type.__name__,
-                                        run_telemetry=run_telemetry,
-                                    )
-                                    fallback_tool_call_recovered = True
-                                    self.logger.info(
-                                        "pydantic_agent_tool_call_recovered",
-                                        result_type=result_type.__name__,
-                                        retry_attempt=fallback_tool_call_retry_attempt,
-                                        retry_mode="event_stream_fallback",
-                                    )
-                                    break
-                                except Exception as fallback_tool_call_retry_exc:
-                                    fallback_exc = fallback_tool_call_retry_exc
-
-                            if fallback_tool_call_recovered:
-                                self._populate_telemetry_from_result(result, run_telemetry)
-                                elapsed = time.monotonic() - start_time
-                                self.logger.info(
-                                    "pydantic_agent_fallback_recovered",
-                                    elapsed_sec=round(elapsed, 2),
-                                    result_type=result_type.__name__,
-                                    mode="pydantic_ai",
-                                )
-                                # Continue to normal post-processing with recovered result.
-                                pass
-                            else:
-                                recovered_output = self._recover_output_from_final_result_validation_error(
-                                    result_type=result_type,
-                                    exc=fallback_exc,
-                                    telemetry=run_telemetry,
-                                )
-                                if recovered_output is not None:
-                                    self.logger.warning(
-                                        "pydantic_agent_output_recovered_from_final_result_payload",
-                                        result_type=result_type.__name__,
-                                    )
-                                    run_telemetry["output_recovered_from_final_result_payload"] = True
-                                    run_telemetry.pop("_tool_call_started_at", None)
-                                    self._last_agent_run_telemetry = run_telemetry
-                                    operation_state["operation_status"] = "completed"
-                                    operation_state["operation_outcome"] = "recovered_output"
-                                    return recovered_output
-                                failover_result = await self._attempt_row_provider_failover(
-                                    exc=fallback_exc,
-                                    system_prompt=system_prompt,
-                                    user_prompt=user_prompt,
-                                    result_type=result_type,
-                                    complexity=complexity,
-                                    max_tokens=max_tokens,
-                                    prompt_type=prompt_type,
-                                    deps=deps,
-                                    toolsets=toolsets,
-                                    max_turns=max_turns,
-                                    max_tool_calls=max_tool_calls,
-                                    pre_retry_telemetry=pre_retry_telemetry,
-                                    current_run_telemetry=run_telemetry,
-                                    row_failover_context=row_failover_context,
-                                    provider_health_memory=provider_health_memory,
-                                )
-                                if failover_result is not None:
-                                    operation_state["operation_status"] = "completed"
-                                    operation_state["operation_outcome"] = "provider_failover"
-                                    run_telemetry.pop("_tool_call_started_at", None)
-                                    self._last_agent_run_telemetry = self._merge_failover_telemetry(
-                                        pre_retry_telemetry,
-                                        run_telemetry,
-                                    )
-                                    return failover_result
-                                elapsed = time.monotonic() - start_time
-                                tool_efficiency = self._build_tool_efficiency_summary(run_telemetry)
-                                if tool_efficiency:
-                                    run_telemetry["tool_efficiency"] = tool_efficiency
-                                run_telemetry["error"] = str(fallback_exc)
-                                run_telemetry.pop("_tool_call_started_at", None)
-                                self._last_agent_run_telemetry = run_telemetry
-                                self.logger.warning(
-                                    "pydantic_agent_call_failed",
-                                    elapsed_sec=round(elapsed, 2),
-                                    mode="pydantic_ai",
-                                    result_type=result_type.__name__,
-                                    operation_id=operation_state["operation_id"],
-                                    event_tool_calls_started=int(run_telemetry.get("event_tool_calls_started", 0) or 0),
-                                    event_tool_calls_completed=int(
-                                        run_telemetry.get(
-                                            "event_tool_calls_completed", run_telemetry.get("event_tool_calls", 0)
-                                        )
-                                        or 0
-                                    ),
-                                    event_skill_tool_calls=int(run_telemetry.get("event_skill_tool_calls", 0) or 0),
-                                    event_skill_effective_tool_calls=int(
-                                        run_telemetry.get("event_skill_effective_tool_calls", 0) or 0
-                                    ),
-                                    observed_tool_names=run_telemetry.get("event_tool_names"),
-                                    **self._exception_diagnostics(fallback_exc),
-                                )
-                                raise
-
-            elapsed = time.monotonic() - start_time
-            raw_output = result.output if hasattr(result, "output") else getattr(result, "data", None)
-            output = self._normalize_agent_output(raw_output, result_type)
-            if not event_stream_enabled:
-                # Non-stream Anthropic/proxy paths do not emit event callbacks, so
-                # backfill tool/usage telemetry from RunResult to keep grounding
-                # guards and efficiency metrics accurate.
-                self._populate_telemetry_from_result(result, run_telemetry)
-
-            # Log response summary for debugging
-            response_summary = self._summarize_output(output)
-            self.logger.info(
-                "pydantic_agent_call_complete",
-                elapsed_sec=round(elapsed, 2),
-                mode="pydantic_ai",
-                result_type=result_type.__name__,
-                operation_id=operation_state["operation_id"],
-                **response_summary,
-            )
-
-            usage_summary = self._record_pydantic_usage(result, prompt_type=f"{prompt_type}:{self.__class__.__name__}")
-            run_telemetry["usage"] = usage_summary or {}
-            tool_efficiency = self._build_tool_efficiency_summary(run_telemetry)
-            if tool_efficiency:
-                run_telemetry["tool_efficiency"] = tool_efficiency
-                self.logger.info(
-                    "pydantic_agent_tool_efficiency",
-                    result_type=result_type.__name__,
-                    **tool_efficiency,
-                )
-            if pre_retry_telemetry:
-                combined_telemetry = dict(pre_retry_telemetry)
-                for key in [
-                    "event_tool_calls",
-                    "event_tool_calls_started",
-                    "event_tool_calls_completed",
-                    "event_tool_effective_calls",
-                    "event_skill_tool_calls",
-                    "event_skill_discovery_tool_calls",
-                    "event_skill_bootstrap_tool_calls",
-                    "event_skill_execution_tool_calls",
-                    "event_skill_execution_effective_tool_calls",
-                    "event_skill_effective_tool_calls",
-                ]:
-                    combined_telemetry[key] = int(combined_telemetry.get(key, 0) or 0) + int(
-                        run_telemetry.get(key, 0) or 0
-                    )
-                for dict_key in [
-                    "event_tool_count_by_name",
-                    "event_tool_effective_count_by_name",
-                    "event_skill_tool_count_by_name",
-                ]:
-                    combined_telemetry[dict_key] = dict(combined_telemetry.get(dict_key, {}))
-                    for k, v in (run_telemetry.get(dict_key) or {}).items():
-                        combined_telemetry[dict_key][k] = int(combined_telemetry[dict_key].get(k, 0) or 0) + int(v or 0)
-                for list_key in ["event_tool_names", "event_tool_signatures"]:
-                    combined_telemetry[list_key] = list(combined_telemetry.get(list_key, [])) + list(
-                        run_telemetry.get(list_key) or []
-                    )
-                merged_turn_effectiveness: dict[str, dict[str, Any]] = {}
-                merged_turn_index = 0
-                for source in (pre_retry_telemetry, run_telemetry):
-                    turn_payloads = source.get("event_turn_effectiveness")
-                    if not isinstance(turn_payloads, dict):
-                        continue
-                    for _, payload in sorted(
-                        (
-                            (int(str(turn_key)), value)
-                            for turn_key, value in turn_payloads.items()
-                            if str(turn_key).strip().isdigit() and isinstance(value, dict)
-                        ),
-                        key=lambda item: item[0],
-                    ):
-                        merged_turn_index += 1
-                        merged_turn_effectiveness[str(merged_turn_index)] = dict(payload)
-                if merged_turn_effectiveness:
-                    combined_telemetry["event_turn_effectiveness"] = merged_turn_effectiveness
-                    combined_telemetry["event_turn_index"] = merged_turn_index
-                if isinstance(run_telemetry.get("skill_policy_diagnostics"), dict):
-                    combined_telemetry["skill_policy_diagnostics"] = dict(run_telemetry["skill_policy_diagnostics"])
-                combined_telemetry = self._merge_failover_telemetry(combined_telemetry, run_telemetry)
-                quality_summary = self._build_response_quality_summary(output, combined_telemetry)
-            else:
-                combined_telemetry = dict(run_telemetry)
-                quality_summary = self._build_response_quality_summary(output, run_telemetry)
-            quality_log = self.logger.info if bool(quality_summary.get("quality_ok", True)) else self.logger.warning
-            quality_log(
-                "pydantic_agent_response_quality",
-                result_type=result_type.__name__,
-                **quality_summary,
-            )
-            combined_telemetry.pop("_tool_call_started_at", None)
-            combined_telemetry.pop("_tool_call_turn_index", None)
-            self._last_agent_run_telemetry = combined_telemetry
-            self._log_trace_payload(
-                event_name="pydantic_agent_trace_response",
-                elapsed_sec=elapsed,
-                payload={
-                    "operation_id": operation_state["operation_id"],
-                    "result_type": result_type.__name__,
-                    "response_summary": response_summary,
-                    "response_quality": quality_summary,
-                    "usage": usage_summary or {},
-                    "tool_efficiency": tool_efficiency,
-                    "skill_policy_diagnostics": combined_telemetry.get("skill_policy_diagnostics"),
-                    "output": output,
-                },
-            )
-            operation_state["operation_status"] = "completed"
-            operation_state["operation_outcome"] = "success"
-            return output  # type: ignore[return-value]
-
-    def _resolve_agent_retry_policy(self) -> tuple[int, int]:
-        """Resolve model and output retry policy for PydanticAI runs.
-
-        Uses runtime settings by default. For Anthropic structured output,
-        enforce a small minimum output retry floor to absorb transient shape mismatches.
-        """
-        configured_retries = int(getattr(self.settings, "agent_retries", 2) or 0)
-        configured_output_retries = int(getattr(self.settings, "output_retries", 2) or 0)
-
-        retries = max(configured_retries, 0)
-        output_retries = max(configured_output_retries, 0)
-        if self._should_suppress_anthropic_tool_choice():
-            # Anthropic-compatible proxy gateways are more likely to emit
-            # transient malformed tool-call JSON. Keep a small retry floor to
-            # allow tool-call correction without manual reruns.
-            retries = max(retries, 3)
-            output_retries = max(output_retries, 2)
-        if self.settings.protocol == LLMProtocolType.ANTHROPIC and str(self.settings.model_standard).lower().startswith(
-            "claude"
-        ):
-            # Anthropic-compatible proxies can occasionally append trailing
-            # characters after otherwise-valid structured output; keep a slightly
-            # higher validation retry floor to improve first-pass success.
-            output_retries = max(output_retries, 4)
-        return retries, output_retries
-
-    def _build_event_stream_handler(self, run_started_at: float, telemetry: dict[str, Any] | None = None):
-        """Build realtime PydanticAI stream handler for observability."""
-
-        async def _event_stream_handler(_run_context: Any, event_stream: Any) -> None:
-            async for event in event_stream:
-                self._log_pydantic_agent_event(event=event, run_started_at=run_started_at, telemetry=telemetry)
-
-        return _event_stream_handler
-
-    @staticmethod
-    def _should_retry_without_event_stream(exc: Exception) -> bool:
-        """Return True when event streaming is unsupported for the active test/runtime model."""
-        message = str(exc).lower()
-        return (
-            "stream_function" in message
-            or "event_stream_handler" in message
-            # Anthropic-compatible local proxies can emit tool-call deltas that
-            # collide with ThinkingPart accumulation under event stream mode.
-            # Retrying once without event_stream_handler avoids this parser path.
-            or "cannot apply a tool call delta to existing_part=thinkingpart" in message
-        )
-
-    @staticmethod
-    def _is_rate_limit_error(exc: Exception) -> bool:
-        """Return True when provider failed with a retryable transient/quota class."""
-        classification = ProviderFailureClassifier.classify(exc)
-        return classification.failure_class in {
-            ProviderFailureClass.RETRYABLE_TRANSIENT,
-            ProviderFailureClass.QUOTA_OR_CAPACITY,
-            ProviderFailureClass.PROVIDER_DEGRADED,
-        }
-
-    @staticmethod
-    def _is_auth_error(exc: Exception) -> bool:
-        """Return True when provider rejected request due to auth/token failure."""
-        return ProviderFailureClassifier.classify(exc).failure_class == ProviderFailureClass.TERMINAL_AUTH
-
-    @staticmethod
-    def _is_tool_call_argument_error(exc: Exception) -> bool:
-        """Return True when model output contains malformed/missing tool args JSON."""
-        classification = ProviderFailureClassifier.classify(exc)
-        raw_message = str(classification.raw_message or "")
-        normalized = raw_message.lower()
-
-        has_tool_call_signal = (
-            "toolretryerror" in normalized
-            or "tool call" in normalized
-            or "tool_call" in normalized
-            or "tool '" in normalized
-            or "include your response in a tool call" in normalized
-            or "return text or include your response in a tool call" in normalized
-        )
-        has_json_shape_signal = (
-            "jsondecodeerror" in normalized
-            or "eof while parsing an object" in normalized
-            or "expecting value" in normalized
-            or "unterminated string" in normalized
-        )
-
-        return (
-            classification.classification_reason == "tool_call_argument_or_validation_error"
-            or classification.failure_class == ProviderFailureClass.NON_PROVIDER_BUG
-        ) and (has_tool_call_signal or has_json_shape_signal)
-
-    @staticmethod
-    def _is_tool_choice_protocol_mismatch(exc: Exception) -> bool:
-        """Return True when upstream rejects tool_choice/function payload shape."""
-        message = str(exc).lower()
-        cause = getattr(exc, "__cause__", None)
-        cause_message = str(cause).lower() if cause is not None else ""
-        combined = f"{message}\n{cause_message}"
-        has_tool_choice = "tool_choice" in combined
-        invalid_function = (
-            "invalid value for `function`" in combined
-            or "invalid value for function" in combined
-            or '"function": null' in combined
-        )
-        return has_tool_choice and invalid_function
-
-    @staticmethod
-    def _percentile(values: list[float], quantile: float) -> float:
-        if not values:
-            return 0.0
-        sorted_vals = sorted(float(v) for v in values)
-        if len(sorted_vals) == 1:
-            return sorted_vals[0]
-        q = min(1.0, max(0.0, float(quantile)))
-        idx = round((len(sorted_vals) - 1) * q)
-        return sorted_vals[idx]
-
-    @staticmethod
-    def _tool_family_tool_names() -> dict[str, tuple[str, ...]]:
-        return {
-            "lexical": (
-                "grep_search",
-                "rg_search",
-                "search_evidence",
-                "read_evidence_document",
-            ),
-            "structural": (
-                "ast_grep_search",
-                "get_definition",
-                "find_references",
-                "workspace_symbol",
-            ),
-            "vector_docs": ("search_evidence_vector",),
-            "vector_code": ("search_code_vector",),
-        }
-
-    @classmethod
-    def _resolve_tool_family(cls, tool_name: str) -> str | None:
-        normalized = str(tool_name or "").strip()
-        if not normalized:
-            return None
-        for family_name, tool_names in cls._tool_family_tool_names().items():
-            if normalized in tool_names:
-                return family_name
-        return None
-
-    @staticmethod
-    def _family_coverage_markers(family_name: str) -> dict[str, bool]:
-        if family_name == "vector_docs":
-            return {"requirements": True, "docs": True, "code": False}
-        if family_name == "vector_code":
-            return {"requirements": True, "docs": False, "code": True}
-        if family_name == "structural":
-            return {"requirements": True, "docs": False, "code": True}
-        if family_name == "lexical":
-            return {"requirements": True, "docs": False, "code": False}
-        return {"requirements": False, "docs": False, "code": False}
-
-    @staticmethod
-    def _coverage_contribution_level(markers: dict[str, bool], *, skill_effective_calls: int = 0) -> str:
-        requirements = bool(markers.get("requirements"))
-        docs = bool(markers.get("docs"))
-        code = bool(markers.get("code"))
-        if requirements and docs and code:
-            return "high"
-        if requirements and (docs or code):
-            return "medium"
-        if requirements or skill_effective_calls > 0:
-            return "low"
-        return "none"
-
-    @staticmethod
-    def _is_skill_tool_name(tool_name: str) -> bool:
-        normalized = str(tool_name or "").strip().lower()
-        return normalized in {"list_skills", "load_skill", "read_skill_resource", "run_skill_script"}
-
-    @classmethod
-    def _is_effective_skill_tool_payload(cls, *, tool_name: str, payload: Any) -> bool:
-        """Skill-specific usefulness guard to avoid discovery/read inflation."""
-        normalized = str(tool_name or "").strip().lower()
-        if normalized in {"list_skills", "load_skill"}:
-            return False
-        if normalized == "run_skill_script":
-            if not isinstance(payload, dict):
-                return False
-            success_value = payload.get("success")
-            stdout_value = payload.get("stdout")
-            if isinstance(success_value, bool) and isinstance(stdout_value, str):
-                return bool(success_value and stdout_value.strip())
-            explicit_effective = payload.get("effective")
-            return bool(explicit_effective) if isinstance(explicit_effective, bool) else False
-        if normalized == "read_skill_resource":
-            if not isinstance(payload, dict):
-                return False
-            if bool(payload.get("benefit_signal")) or bool(payload.get("applied")):
-                return True
-            evidence_refs = payload.get("evidence_refs")
-            if isinstance(evidence_refs, list):
-                return any(str(item).strip() for item in evidence_refs)
-            return False
-        return False
-
-    @classmethod
-    def _is_effective_tool_payload(cls, payload: Any) -> bool:
-        """Best-effort usefulness check for tool result payloads."""
-        if payload is None:
-            return False
-        if isinstance(payload, bool):
-            return payload
-        if isinstance(payload, (int, float)):
-            return payload > 0
-        if isinstance(payload, str):
-            text = payload.strip()
-            if not text:
-                return False
-            if text[0] in "{[":
-                try:
-                    parsed = json.loads(text)
-                except Exception:
-                    return True
-                return cls._is_effective_tool_payload(parsed)
-            return True
-        if isinstance(payload, list):
-            if not payload:
-                return False
-            if all(isinstance(item, str) for item in payload):
-                return any(str(item).strip() for item in payload)
-            return True
-        if isinstance(payload, dict):
-            success_value = payload.get("success")
-            if isinstance(success_value, bool) and not success_value:
-                # Explicit failures with no recovery signal should not be counted as effective.
-                if not any(
-                    key in payload
-                    for key in (
-                        "hits",
-                        "refs",
-                        "accepted_count",
-                        "content",
-                        "ref",
-                        "path",
-                        "results",
-                        "result",
-                    )
-                ):
-                    return False
-            hits = payload.get("hits")
-            if isinstance(hits, list):
-                return len(hits) > 0
-            refs = payload.get("refs")
-            if isinstance(refs, list):
-                return len(refs) > 0
-            accepted_count = payload.get("accepted_count")
-            if isinstance(accepted_count, int):
-                return accepted_count > 0
-            content = payload.get("content")
-            if isinstance(content, str):
-                return bool(content.strip())
-            results = payload.get("results")
-            if isinstance(results, list):
-                return len(results) > 0
-            result_obj = payload.get("result")
-            if result_obj is not None:
-                return cls._is_effective_tool_payload(result_obj)
-            ref_value = payload.get("ref")
-            if isinstance(ref_value, str):
-                return bool(ref_value.strip())
-            found_value = payload.get("found")
-            if isinstance(found_value, bool):
-                return found_value
-            path_value = payload.get("path")
-            if isinstance(path_value, str) and path_value.strip():
-                return True
-            error_value = payload.get("error")
-            if isinstance(error_value, str) and error_value.strip():
-                return False
-            if isinstance(success_value, bool):
-                return success_value
-            return bool(payload)
-        return bool(payload)
-
-    @classmethod
-    def _is_effective_tool_result_event(cls, result: Any, *, tool_name: str) -> bool:
-        """Determine whether a tool-result event produced useful output."""
-        normalized_tool_name = str(tool_name or "").strip().lower()
-        # Internal final_result tool events should not contribute to tool usefulness metrics.
-        if normalized_tool_name == "final_result":
-            return False
-        if cls._is_skill_tool_name(normalized_tool_name):
-            for attr in ("output", "content", "result", "data"):
-                if not hasattr(result, attr):
-                    continue
-                value = getattr(result, attr, None)
-                if value is None:
-                    continue
-                return cls._is_effective_skill_tool_payload(tool_name=normalized_tool_name, payload=value)
-            return False
-        for attr in ("output", "content", "result", "data"):
-            if not hasattr(result, attr):
-                continue
-            value = getattr(result, attr, None)
-            if value is None:
-                continue
-            if cls._is_effective_tool_payload(value):
-                return True
-        success_attr = getattr(result, "success", None)
-        if isinstance(success_attr, bool):
-            return success_attr
-        return False
-
-    @classmethod
-    def _build_tool_efficiency_summary(cls, telemetry: dict[str, Any] | None) -> dict[str, Any]:
-        """Build compact tool/skill efficiency metrics for monitoring."""
-        if not isinstance(telemetry, dict):
-            return {}
-        usage = telemetry.get("usage")
-        usage_tool_calls = int(usage.get("tool_calls", 0) or 0) if isinstance(usage, dict) else 0
-        started = int(telemetry.get("event_tool_calls_started", 0) or 0)
-        completed = int(telemetry.get("event_tool_calls_completed", telemetry.get("event_tool_calls", 0)) or 0)
-        effective_calls = int(telemetry.get("event_tool_effective_calls", 0) or 0)
-        effective_completed = max(completed, usage_tool_calls)
-        if started <= 0 and effective_completed <= 0:
-            return {}
-        if effective_calls <= 0 and "event_tool_effective_calls" not in telemetry and effective_completed > 0:
-            # Backward-compatible fallback for historical telemetry that only tracked completion.
-            effective_calls = effective_completed
-        tool_count_by_name = telemetry.get("event_tool_count_by_name")
-        if not isinstance(tool_count_by_name, dict):
-            tool_count_by_name = {}
-        unique_tools_by_name = (
-            len(tool_count_by_name) if tool_count_by_name else len(set(telemetry.get("event_tool_names") or []))
-        )
-        tool_signatures_raw = telemetry.get("event_tool_signatures")
-        unique_tools_by_signature = (
-            len({str(sig) for sig in tool_signatures_raw})
-            if isinstance(tool_signatures_raw, list) and tool_signatures_raw
-            else 0
-        )
-        unique_tools = unique_tools_by_signature if unique_tools_by_signature > 0 else unique_tools_by_name
-        repeated_calls = max(0, effective_completed - unique_tools)
-        completion_rate = (effective_completed / started) if started > 0 else 1.0
-
-        skill_count_by_name = telemetry.get("event_skill_tool_count_by_name")
-        if not isinstance(skill_count_by_name, dict):
-            skill_count_by_name = {}
-        if not skill_count_by_name and isinstance(tool_count_by_name, dict):
-            # Fallback: recover skill-class counters from generic tool names when
-            # event-stream skill fields were not populated.
-            recovered_skill_counts = {
-                skill_tool: int(tool_count_by_name.get(skill_tool, 0) or 0)
-                for skill_tool in ("list_skills", "load_skill", "read_skill_resource", "run_skill_script")
-                if int(tool_count_by_name.get(skill_tool, 0) or 0) > 0
-            }
-            if recovered_skill_counts:
-                skill_count_by_name = recovered_skill_counts
-        skill_discovery_calls = int(
-            telemetry.get("event_skill_discovery_tool_calls", skill_count_by_name.get("list_skills", 0)) or 0
-        )
-        skill_bootstrap_calls = int(
-            telemetry.get("event_skill_bootstrap_tool_calls", skill_count_by_name.get("load_skill", 0)) or 0
-        )
-        derived_skill_execution_calls = int(skill_count_by_name.get("read_skill_resource", 0) or 0) + int(
-            skill_count_by_name.get("run_skill_script", 0) or 0
-        )
-        if "event_skill_execution_tool_calls" in telemetry:
-            skill_execution_calls = int(telemetry.get("event_skill_execution_tool_calls", 0) or 0)
-        else:
-            # FR-123: execution effectiveness is derived from execution-class skill tools only.
-            skill_execution_calls = derived_skill_execution_calls
-        legacy_effective_calls = int(telemetry.get("event_skill_effective_tool_calls", 0) or 0)
-        skill_total = int(telemetry.get("event_skill_tool_calls", 0) or 0)
-        if skill_total <= 0:
-            skill_total = skill_discovery_calls + skill_bootstrap_calls + skill_execution_calls
-        if skill_total <= 0 and legacy_effective_calls > 0:
-            # Preserve visibility for legacy telemetry payloads without over-crediting execution usage.
-            skill_total = legacy_effective_calls
-        tool_effective_count_by_name = telemetry.get("event_tool_effective_count_by_name")
-        derived_skill_execution_effective_calls = 0
-        if isinstance(tool_effective_count_by_name, dict):
-            derived_skill_execution_effective_calls = int(
-                tool_effective_count_by_name.get("read_skill_resource", 0) or 0
-            ) + int(tool_effective_count_by_name.get("run_skill_script", 0) or 0)
-        if "event_skill_execution_effective_tool_calls" in telemetry:
-            skill_execution_effective_calls = int(telemetry.get("event_skill_execution_effective_tool_calls", 0) or 0)
-        else:
-            skill_execution_effective_calls = derived_skill_execution_effective_calls
-        skill_execution_effective_calls = max(0, min(skill_execution_effective_calls, skill_execution_calls))
-        skill_effective = skill_execution_effective_calls
-        skill_execution_rate = (skill_execution_calls / skill_total) if skill_total > 0 else None
-        skill_effective_rate = (skill_effective / skill_total) if skill_total > 0 else None
-        skill_execution_effective_rate = (
-            (skill_execution_effective_calls / skill_execution_calls) if skill_execution_calls > 0 else None
-        )
-        skill_bootstrap_only_calls = skill_bootstrap_calls if (skill_execution_calls == 0 and skill_total > 0) else 0
-        skill_bootstrap_only_rate = (skill_bootstrap_only_calls / skill_total) if skill_total > 0 else None
-        skill_discovery_ratio = (skill_discovery_calls / skill_total) if skill_total > 0 else None
-
-        latencies_raw = telemetry.get("event_tool_latencies_ms")
-        latency_values = (
-            [float(v) for v in latencies_raw if isinstance(v, (int, float))] if isinstance(latencies_raw, list) else []
-        )
-        avg_latency_ms = (sum(latency_values) / len(latency_values)) if latency_values else None
-        p95_latency_ms = cls._percentile(latency_values, 0.95) if latency_values else None
-
-        summary: dict[str, Any] = {
-            "tool_calls_started": started,
-            "tool_calls_completed": effective_completed,
-            "tool_calls_effective": effective_calls,
-            "tool_calls_unique": unique_tools,
-            "tool_calls_repeated": repeated_calls,
-            "tool_completion_rate": round(completion_rate, 3),
-        }
-        if effective_completed > 0:
-            summary["tool_effective_rate"] = round(effective_calls / effective_completed, 3)
-        if unique_tools_by_signature > 0:
-            summary["tool_name_unique"] = unique_tools_by_name
-            summary["tool_signature_unique"] = unique_tools_by_signature
-            summary["tool_repetition_basis"] = "signature"
-        if effective_completed > 0:
-            summary["tool_repetition_rate"] = round(repeated_calls / effective_completed, 3)
-        summary["skill_calls_total"] = skill_total
-        summary["skill_calls_discovery"] = skill_discovery_calls
-        summary["skill_calls_bootstrap"] = skill_bootstrap_calls
-        summary["skill_calls_execution"] = skill_execution_calls
-        summary["skill_calls_execution_effective"] = skill_execution_effective_calls
-        summary["skill_calls_bootstrap_only"] = skill_bootstrap_only_calls
-        summary["skill_calls_effective"] = skill_effective
-        if legacy_effective_calls > 0 and legacy_effective_calls != skill_effective:
-            summary["legacy_skill_effective_calls_observed"] = legacy_effective_calls
-        if skill_execution_rate is not None:
-            summary["skill_execution_rate"] = round(skill_execution_rate, 3)
-        if skill_execution_effective_rate is not None:
-            summary["skill_execution_effective_rate"] = round(skill_execution_effective_rate, 3)
-        if skill_effective_rate is not None:
-            summary["skill_effective_rate"] = round(skill_effective_rate, 3)
-        if skill_bootstrap_only_rate is not None:
-            summary["skill_bootstrap_only_rate"] = round(skill_bootstrap_only_rate, 3)
-        if skill_discovery_ratio is not None:
-            summary["skill_discovery_ratio"] = round(skill_discovery_ratio, 3)
-        if avg_latency_ms is not None:
-            summary["tool_latency_avg_ms"] = round(avg_latency_ms)
-        if p95_latency_ms is not None:
-            summary["tool_latency_p95_ms"] = round(p95_latency_ms)
-        family_tool_names = cls._tool_family_tool_names()
-        family_calls: dict[str, int] = {}
-        family_effectiveness: dict[str, float] = {}
-        for family_name, tool_names in family_tool_names.items():
-            calls = sum(int(tool_count_by_name.get(tool_name, 0) or 0) for tool_name in tool_names)
-            family_calls[family_name] = calls
-            if effective_completed > 0:
-                family_effectiveness[family_name] = round(calls / effective_completed, 3)
-            else:
-                family_effectiveness[family_name] = 0.0
-        summary["tool_family_calls"] = family_calls
-        summary["tool_family_effectiveness"] = family_effectiveness
-
-        turn_effectiveness_raw = telemetry.get("event_turn_effectiveness")
-        if isinstance(turn_effectiveness_raw, dict) and turn_effectiveness_raw:
-            turn_entries: list[dict[str, Any]] = []
-            turn_effective_count = 0
-            coverage_turn_count = 0
-            aggregate_markers = {"requirements": False, "docs": False, "code": False}
-            ordered_turn_items = sorted(
-                (
-                    (int(str(turn_key)), payload)
-                    for turn_key, payload in turn_effectiveness_raw.items()
-                    if str(turn_key).strip().isdigit() and isinstance(payload, dict)
-                ),
-                key=lambda item: item[0],
-            )
-            for turn_index, payload in ordered_turn_items:
-                turn_tool_calls = int(payload.get("tool_calls", 0) or 0)
-                turn_effective_tool_calls = int(payload.get("effective_tool_calls", turn_tool_calls) or 0)
-                turn_skill_discovery_calls = int(payload.get("skill_discovery_calls", 0) or 0)
-                turn_skill_bootstrap_calls = int(payload.get("skill_bootstrap_calls", 0) or 0)
-                turn_skill_execution_calls = int(
-                    payload.get("skill_execution_calls", payload.get("skill_effective_calls", 0)) or 0
-                )
-                turn_skill_execution_effective_calls = int(
-                    payload.get("skill_execution_effective_calls", payload.get("skill_effective_calls", 0)) or 0
-                )
-                turn_skill_execution_effective_calls = max(
-                    0,
-                    min(turn_skill_execution_effective_calls, turn_skill_execution_calls),
-                )
-                turn_skill_calls = int(payload.get("skill_calls", 0) or 0)
-                if turn_skill_calls <= 0:
-                    turn_skill_calls = (
-                        turn_skill_discovery_calls + turn_skill_bootstrap_calls + turn_skill_execution_calls
-                    )
-                turn_skill_execution_rate = (
-                    round(turn_skill_execution_calls / turn_skill_calls, 3)
-                    if turn_skill_calls > 0
-                    else (1.0 if turn_skill_execution_calls > 0 else 0.0)
-                )
-                turn_skill_effective_calls = turn_skill_execution_effective_calls
-                turn_skill_effective_rate = (
-                    round(turn_skill_effective_calls / turn_skill_calls, 3)
-                    if turn_skill_calls > 0
-                    else (1.0 if turn_skill_effective_calls > 0 else 0.0)
-                )
-                turn_skill_execution_effective_rate = (
-                    round(turn_skill_execution_effective_calls / turn_skill_execution_calls, 3)
-                    if turn_skill_execution_calls > 0
-                    else 0.0
-                )
-                turn_skill_bootstrap_only_calls = (
-                    turn_skill_bootstrap_calls if (turn_skill_effective_calls == 0 and turn_skill_calls > 0) else 0
-                )
-                turn_skill_bootstrap_only_rate = (
-                    round(turn_skill_bootstrap_only_calls / turn_skill_calls, 3) if turn_skill_calls > 0 else 0.0
-                )
-                turn_family_calls_raw = payload.get("tool_family_effective_calls")
-                if not isinstance(turn_family_calls_raw, dict):
-                    turn_family_calls_raw = payload.get("tool_family_calls")
-                turn_family_calls = {
-                    family_name: int((turn_family_calls_raw or {}).get(family_name, 0) or 0)
-                    for family_name in family_tool_names
-                }
-                turn_effective_denominator = (
-                    turn_effective_tool_calls if turn_effective_tool_calls > 0 else turn_tool_calls
-                )
-                turn_family_effectiveness = {
-                    family_name: round((calls / turn_effective_denominator), 3)
-                    if turn_effective_denominator > 0
-                    else 0.0
-                    for family_name, calls in turn_family_calls.items()
-                }
-                raw_markers = payload.get("coverage_contribution_markers")
-                if isinstance(raw_markers, dict):
-                    coverage_markers = {
-                        "requirements": bool(raw_markers.get("requirements")),
-                        "docs": bool(raw_markers.get("docs")),
-                        "code": bool(raw_markers.get("code")),
-                    }
-                else:
-                    coverage_markers = {"requirements": False, "docs": False, "code": False}
-                    for family_name, calls in turn_family_calls.items():
-                        if calls <= 0:
-                            continue
-                        family_markers = cls._family_coverage_markers(family_name)
-                        for marker_key in coverage_markers:
-                            coverage_markers[marker_key] = bool(
-                                coverage_markers[marker_key] or family_markers[marker_key]
-                            )
-                contribution_level = cls._coverage_contribution_level(
-                    coverage_markers,
-                    skill_effective_calls=turn_skill_effective_calls,
-                )
-                if turn_effective_tool_calls > 0 and (
-                    turn_skill_execution_effective_calls > 0 or any(count > 0 for count in turn_family_calls.values())
-                ):
-                    turn_effective_count += 1
-                if contribution_level != "none":
-                    coverage_turn_count += 1
-                for marker_key in aggregate_markers:
-                    aggregate_markers[marker_key] = aggregate_markers[marker_key] or coverage_markers[marker_key]
-                turn_entries.append(
-                    {
-                        "turn_index": turn_index,
-                        "tool_calls": turn_tool_calls,
-                        "effective_tool_calls": turn_effective_tool_calls,
-                        "skill_calls": turn_skill_calls,
-                        "skill_discovery_calls": turn_skill_discovery_calls,
-                        "skill_bootstrap_calls": turn_skill_bootstrap_calls,
-                        "skill_execution_calls": turn_skill_execution_calls,
-                        "skill_execution_effective_calls": turn_skill_execution_effective_calls,
-                        "skill_execution_rate": turn_skill_execution_rate,
-                        "skill_execution_effective_rate": turn_skill_execution_effective_rate,
-                        "skill_bootstrap_only_calls": turn_skill_bootstrap_only_calls,
-                        "skill_bootstrap_only_rate": turn_skill_bootstrap_only_rate,
-                        "skill_effective_calls": turn_skill_effective_calls,
-                        "skill_effective_rate": turn_skill_effective_rate,
-                        "tool_family_calls": turn_family_calls,
-                        "tool_family_effectiveness": turn_family_effectiveness,
-                        "coverage_contribution_markers": coverage_markers,
-                        "requirement_coverage_contribution_level": contribution_level,
-                    }
-                )
-            if turn_entries:
-                turn_count = len(turn_entries)
-                summary["turn_effectiveness_summary"] = turn_entries
-                summary["turn_effective_rate"] = round(turn_effective_count / turn_count, 3)
-                summary["turn_requirement_coverage_contribution_rate"] = round(coverage_turn_count / turn_count, 3)
-                summary["turn_requirement_coverage_contribution_markers"] = aggregate_markers
-        return summary
-
-    @staticmethod
-    def _extract_tool_names(raw_name: Any) -> list[str]:
-        value = str(raw_name or "").strip()
-        if not value:
-            return []
-        normalized = value.replace("/", " ").replace(",", " ").replace(";", " ")
-        candidates = [part.strip() for part in re.split(r"\s+", normalized) if part.strip()]
-        known = (
-            "list_directory",
-            "read_file",
-            "grep_search",
-            "rg_search",
-            "ast_grep_search",
-            "list_evidence_documents",
-            "search_evidence",
-            "read_evidence_document",
-            "search_evidence_vector",
-            "search_code_vector",
-            "vector_retrieval_status",
-            "get_definition",
-            "find_references",
-            "workspace_symbol",
-            "list_skills",
-            "load_skill",
-            "read_skill_resource",
-            "run_skill_script",
-        )
-        if len(candidates) == 1 and candidates[0] == value:
-            known_sorted = sorted(known, key=len, reverse=True)
-            matched: list[str] = []
-            cursor = 0
-            while cursor < len(value):
-                token = next((name for name in known_sorted if value.startswith(name, cursor)), None)
-                if token is None:
-                    matched = []
-                    break
-                matched.append(token)
-                cursor += len(token)
-            if matched and cursor == len(value):
-                return matched
-        return candidates
-
-    def _log_pydantic_agent_event(
-        self,
-        *,
-        event: Any,
-        run_started_at: float,
-        telemetry: dict[str, Any] | None = None,
-    ) -> None:
-        """Emit structured logs for PydanticAI runtime events."""
-        import time
-
-        event_name = type(event).__name__
-        elapsed = round(max(0.0, time.monotonic() - run_started_at), 2)
-        if isinstance(telemetry, dict):
-            telemetry["event_last_seen_at"] = time.monotonic()
-            telemetry.setdefault("event_tool_effective_calls", 0)
-            telemetry.setdefault("event_skill_execution_effective_tool_calls", 0)
-            effective_count_by_name = telemetry.get("event_tool_effective_count_by_name")
-            if not isinstance(effective_count_by_name, dict):
-                telemetry["event_tool_effective_count_by_name"] = {}
-
-        if event_name == "FunctionToolCallEvent":
-            part = getattr(event, "part", None)
-            raw_tool_name = getattr(part, "tool_name", None) or getattr(part, "name", None) or "unknown"
-            tool_names = self._extract_tool_names(raw_tool_name)
-            primary_tool_name = tool_names[0] if tool_names else str(raw_tool_name)
-            tool_call_id = getattr(part, "tool_call_id", None)
-            raw_args = getattr(part, "args", None)
-            args_repr = ""
-            if raw_args is not None:
-                try:
-                    args_repr = json.dumps(raw_args, sort_keys=True, default=str, ensure_ascii=True)
-                except Exception:
-                    args_repr = str(raw_args)
-            self._log_trace_payload(
-                event_name="pydantic_agent_trace_tool_call",
-                elapsed_sec=elapsed,
-                payload={
-                    "tool_name": primary_tool_name,
-                    "tool_names": tool_names,
-                    "tool_call_id": tool_call_id,
-                    "args": raw_args,
-                },
-            )
-            if isinstance(telemetry, dict):
-                turn_index = int(telemetry.get("event_turn_index", 0) or 0) + 1
-                telemetry["event_turn_index"] = turn_index
-                telemetry["event_tool_calls_started"] = int(telemetry.get("event_tool_calls_started", 0) or 0) + 1
-                telemetry["event_last_tool_activity_at"] = time.monotonic()
-                signatures = telemetry.get("event_tool_signatures")
-                if not isinstance(signatures, list):
-                    signatures = []
-                signatures.append(f"{primary_tool_name}|{args_repr}")
-                telemetry["event_tool_signatures"] = signatures
-
-                if tool_call_id:
-                    started_at_map = telemetry.get("_tool_call_started_at")
-                    if not isinstance(started_at_map, dict):
-                        started_at_map = {}
-                    started_at_map[str(tool_call_id)] = time.monotonic()
-                    telemetry["_tool_call_started_at"] = started_at_map
-                    turn_map = telemetry.get("_tool_call_turn_index")
-                    if not isinstance(turn_map, dict):
-                        turn_map = {}
-                    turn_map[str(tool_call_id)] = turn_index
-                    telemetry["_tool_call_turn_index"] = turn_map
-            logged_turn_index = int(telemetry.get("event_turn_index", 0) or 0) if isinstance(telemetry, dict) else None
-            self.logger.info(
-                "pydantic_agent_tool_call_started",
-                elapsed_sec=elapsed,
-                tool_name=primary_tool_name,
-                tool_names=tool_names if len(tool_names) > 1 else None,
-                tool_call_id=tool_call_id,
-                turn_index=logged_turn_index,
-            )
-            return
-
-        if event_name == "FunctionToolResultEvent":
-            result = getattr(event, "result", None)
-            raw_tool_name = getattr(result, "tool_name", None) or getattr(result, "name", None) or "unknown"
-            tool_names = self._extract_tool_names(raw_tool_name)
-            primary_tool_name = tool_names[0] if tool_names else str(raw_tool_name)
-            effective = self._is_effective_tool_result_event(result, tool_name=primary_tool_name)
-            tool_call_id = getattr(result, "tool_call_id", None)
-            turn_index = None
-            if isinstance(telemetry, dict):
-                turn_map = telemetry.get("_tool_call_turn_index")
-                if isinstance(turn_map, dict) and tool_call_id:
-                    raw_turn = turn_map.pop(str(tool_call_id), None)
-                    if isinstance(raw_turn, int):
-                        turn_index = raw_turn
-                    telemetry["_tool_call_turn_index"] = turn_map
-            self.logger.info(
-                "pydantic_agent_tool_call_completed",
-                elapsed_sec=elapsed,
-                tool_name=primary_tool_name,
-                tool_names=tool_names if len(tool_names) > 1 else None,
-                tool_call_id=tool_call_id,
-                turn_index=turn_index,
-                effective=effective,
-            )
-            self._log_trace_payload(
-                event_name="pydantic_agent_trace_tool_result",
-                elapsed_sec=elapsed,
-                payload={
-                    "tool_name": primary_tool_name,
-                    "tool_names": tool_names,
-                    "tool_call_id": tool_call_id,
-                    "result": {
-                        "name": getattr(result, "name", None),
-                        "tool_name": getattr(result, "tool_name", None),
-                        "content": getattr(result, "content", None),
-                        "output": getattr(result, "output", None),
-                    },
-                },
-            )
-            if isinstance(telemetry, dict):
-                telemetry["event_tool_calls"] = int(telemetry.get("event_tool_calls", 0) or 0) + 1
-                telemetry["event_tool_calls_completed"] = int(telemetry.get("event_tool_calls_completed", 0) or 0) + 1
-                telemetry["event_last_tool_activity_at"] = time.monotonic()
-                if effective:
-                    telemetry["event_tool_effective_calls"] = (
-                        int(telemetry.get("event_tool_effective_calls", 0) or 0) + 1
-                    )
-
-                seen_tools = telemetry.get("event_tool_names")
-                if not isinstance(seen_tools, list):
-                    seen_tools = []
-                if primary_tool_name and primary_tool_name not in seen_tools:
-                    seen_tools.append(primary_tool_name)
-                telemetry["event_tool_names"] = seen_tools
-
-                if primary_tool_name == "final_result":
-                    result_content = getattr(result, "content", None)
-                    if isinstance(result_content, list):
-                        for item in result_content:
-                            if not isinstance(item, dict):
-                                continue
-                            if str(item.get("type") or "") != "json_invalid":
-                                continue
-                            raw_invalid_input = item.get("input")
-                            if isinstance(raw_invalid_input, str) and raw_invalid_input.strip():
-                                telemetry["event_final_result_invalid_input"] = raw_invalid_input
-                                break
-
-                tool_count_by_name = telemetry.get("event_tool_count_by_name")
-                if not isinstance(tool_count_by_name, dict):
-                    tool_count_by_name = {}
-                tool_count_by_name[primary_tool_name] = int(tool_count_by_name.get(primary_tool_name, 0) or 0) + 1
-                telemetry["event_tool_count_by_name"] = tool_count_by_name
-
-                if effective:
-                    effective_count_by_name = telemetry.get("event_tool_effective_count_by_name")
-                    if not isinstance(effective_count_by_name, dict):
-                        effective_count_by_name = {}
-                    effective_count_by_name[primary_tool_name] = (
-                        int(effective_count_by_name.get(primary_tool_name, 0) or 0) + 1
-                    )
-                    telemetry["event_tool_effective_count_by_name"] = effective_count_by_name
-
-                started_at_map = telemetry.get("_tool_call_started_at")
-                if isinstance(started_at_map, dict) and tool_call_id:
-                    started_at = started_at_map.pop(str(tool_call_id), None)
-                    if isinstance(started_at, (int, float)):
-                        latencies = telemetry.get("event_tool_latencies_ms")
-                        if not isinstance(latencies, list):
-                            latencies = []
-                        latencies.append(int(max(0.0, (time.monotonic() - float(started_at)) * 1000)))
-                        telemetry["event_tool_latencies_ms"] = latencies
-                    telemetry["_tool_call_started_at"] = started_at_map
-
-                skill_discovery_tools = {"list_skills"}
-                skill_bootstrap_tools = {"load_skill"}
-                skill_execution_tools = {"read_skill_resource", "run_skill_script"}
-                skill_tools = skill_discovery_tools | skill_bootstrap_tools | skill_execution_tools
-                if primary_tool_name in skill_tools:
-                    telemetry["event_skill_tool_calls"] = int(telemetry.get("event_skill_tool_calls", 0) or 0) + 1
-                    skill_count_by_name = telemetry.get("event_skill_tool_count_by_name")
-                    if not isinstance(skill_count_by_name, dict):
-                        skill_count_by_name = {}
-                    skill_count_by_name[primary_tool_name] = int(skill_count_by_name.get(primary_tool_name, 0) or 0) + 1
-                    telemetry["event_skill_tool_count_by_name"] = skill_count_by_name
-                if primary_tool_name in skill_discovery_tools:
-                    telemetry["event_skill_discovery_tool_calls"] = (
-                        int(telemetry.get("event_skill_discovery_tool_calls", 0) or 0) + 1
-                    )
-                if primary_tool_name in skill_bootstrap_tools:
-                    telemetry["event_skill_bootstrap_tool_calls"] = (
-                        int(telemetry.get("event_skill_bootstrap_tool_calls", 0) or 0) + 1
-                    )
-                if primary_tool_name in skill_execution_tools:
-                    telemetry["event_skill_execution_tool_calls"] = (
-                        int(telemetry.get("event_skill_execution_tool_calls", 0) or 0) + 1
-                    )
-                    if effective:
-                        telemetry["event_skill_execution_effective_tool_calls"] = (
-                            int(telemetry.get("event_skill_execution_effective_tool_calls", 0) or 0) + 1
-                        )
-                        # Legacy field retained for backward-compatible consumers.
-                        telemetry["event_skill_effective_tool_calls"] = (
-                            int(telemetry.get("event_skill_effective_tool_calls", 0) or 0) + 1
-                        )
-                if isinstance(turn_index, int) and turn_index > 0:
-                    turn_key = str(turn_index)
-                    turn_payloads = telemetry.get("event_turn_effectiveness")
-                    if not isinstance(turn_payloads, dict):
-                        turn_payloads = {}
-                    current_payload = turn_payloads.get(turn_key)
-                    if not isinstance(current_payload, dict):
-                        current_payload = {}
-                    current_payload["tool_calls"] = int(current_payload.get("tool_calls", 0) or 0) + 1
-                    if effective:
-                        current_payload["effective_tool_calls"] = (
-                            int(current_payload.get("effective_tool_calls", 0) or 0) + 1
-                        )
-                    if primary_tool_name in skill_tools:
-                        current_payload["skill_calls"] = int(current_payload.get("skill_calls", 0) or 0) + 1
-                    if primary_tool_name in skill_discovery_tools:
-                        current_payload["skill_discovery_calls"] = (
-                            int(current_payload.get("skill_discovery_calls", 0) or 0) + 1
-                        )
-                    if primary_tool_name in skill_bootstrap_tools:
-                        current_payload["skill_bootstrap_calls"] = (
-                            int(current_payload.get("skill_bootstrap_calls", 0) or 0) + 1
-                        )
-                    if primary_tool_name in skill_execution_tools:
-                        current_payload["skill_execution_calls"] = (
-                            int(current_payload.get("skill_execution_calls", 0) or 0) + 1
-                        )
-                        if effective:
-                            current_payload["skill_execution_effective_calls"] = (
-                                int(current_payload.get("skill_execution_effective_calls", 0) or 0) + 1
-                            )
-                            # Legacy per-turn field kept in sync with effective execution semantics.
-                            current_payload["skill_effective_calls"] = (
-                                int(current_payload.get("skill_effective_calls", 0) or 0) + 1
-                            )
-                    family_name = self._resolve_tool_family(primary_tool_name)
-                    family_calls_raw = current_payload.get("tool_family_calls")
-                    if not isinstance(family_calls_raw, dict):
-                        family_calls_raw = {}
-                    for family_key in self._tool_family_tool_names():
-                        family_calls_raw[family_key] = int(family_calls_raw.get(family_key, 0) or 0)
-                    if family_name:
-                        family_calls_raw[family_name] = int(family_calls_raw.get(family_name, 0) or 0) + 1
-                    current_payload["tool_family_calls"] = family_calls_raw
-                    family_effective_calls_raw = current_payload.get("tool_family_effective_calls")
-                    if not isinstance(family_effective_calls_raw, dict):
-                        family_effective_calls_raw = {}
-                    for family_key in self._tool_family_tool_names():
-                        family_effective_calls_raw[family_key] = int(family_effective_calls_raw.get(family_key, 0) or 0)
-                    if family_name and effective:
-                        family_effective_calls_raw[family_name] = (
-                            int(family_effective_calls_raw.get(family_name, 0) or 0) + 1
-                        )
-                    current_payload["tool_family_effective_calls"] = family_effective_calls_raw
-                    raw_markers = current_payload.get("coverage_contribution_markers")
-                    markers = {
-                        "requirements": bool((raw_markers or {}).get("requirements"))
-                        if isinstance(raw_markers, dict)
-                        else False,
-                        "docs": bool((raw_markers or {}).get("docs")) if isinstance(raw_markers, dict) else False,
-                        "code": bool((raw_markers or {}).get("code")) if isinstance(raw_markers, dict) else False,
-                    }
-                    if family_name and effective:
-                        family_markers = self._family_coverage_markers(family_name)
-                        for marker_key in markers:
-                            markers[marker_key] = bool(markers[marker_key] or family_markers[marker_key])
-                    current_payload["coverage_contribution_markers"] = markers
-                    turn_payloads[turn_key] = current_payload
-                    telemetry["event_turn_effectiveness"] = turn_payloads
-            return
-
-        if event_name == "FinalResultEvent":
-            turn_count = None
-            if isinstance(telemetry, dict):
-                turn_count = int(telemetry.get("event_turn_index", 0) or 0)
-            self.logger.info(
-                "pydantic_agent_final_result_event",
-                elapsed_sec=elapsed,
-                turn_count=turn_count,
-            )
-            return
-
-        if event_name == "PartDeltaEvent" and getattr(self.settings, "agent_event_stream_log_text_deltas", False):
-            delta = getattr(event, "delta", None)
-            delta_text = (
-                getattr(delta, "content_delta", None)
-                or getattr(delta, "text", None)
-                or getattr(delta, "delta", None)
-                or ""
-            )
-            if delta_text:
-                self.logger.info(
-                    "pydantic_agent_text_delta",
-                    elapsed_sec=elapsed,
-                    delta_preview=str(delta_text)[:200],
-                )
-            return
-
-    def _summarize_output(self, output: Any) -> dict[str, Any]:
-        """Summarize agent output for logging.
-
-        Extracts key metrics from the output without logging full content.
-        """
-        summary: dict[str, Any] = {}
-
-        # Common fields across output models
-        if hasattr(output, "confidence"):
-            summary["confidence"] = round(output.confidence, 2)
-        if hasattr(output, "score"):
-            summary["score"] = output.score
-        if hasattr(output, "score_missing"):
-            summary["score_missing"] = bool(getattr(output, "score_missing", False))
-        if hasattr(output, "posture_score"):
-            summary["posture_score"] = output.posture_score
-
-        # Count findings/vulnerabilities
-        if hasattr(output, "issues") and output.issues:
-            summary["issues_count"] = len(output.issues)
-        if hasattr(output, "vulnerabilities") and output.vulnerabilities:
-            summary["vulnerabilities_count"] = len(output.vulnerabilities)
-        if hasattr(output, "assessments") and output.assessments:
-            summary["assessments_count"] = len(output.assessments)
-
-        # Tech stack if available
-        if hasattr(output, "tech_stack") and output.tech_stack:
-            summary["tech_stack"] = output.tech_stack[:3]  # First 3 items
-
-        return summary
-
-    def _normalize_agent_output(self, output: Any, result_type: type[ResultType]) -> Any:
-        """Normalize stream/text payloads into expected result type shape."""
-        if result_type is str:
-            if isinstance(output, str):
-                if self.settings.protocol == LLMProtocolType.ANTHROPIC:
-                    return self._normalize_anthropic_text_payload_with_diagnostics(
-                        output,
-                        source="normalize_agent_output:str",
-                    )
-                return output
-            if output is None:
-                return ""
-            if isinstance(output, bytes):
-                decoded = output.decode("utf-8", errors="replace")
-                if self.settings.protocol == LLMProtocolType.ANTHROPIC:
-                    return self._normalize_anthropic_text_payload_with_diagnostics(
-                        decoded,
-                        source="normalize_agent_output:bytes",
-                    )
-                return decoded
-            if isinstance(output, (dict, list)):
-                return json.dumps(output, ensure_ascii=False)
-            return str(output)
-
-        if isinstance(output, result_type):
-            return output
-
-        validator_json = getattr(result_type, "model_validate_json", None)
-        validator_obj = getattr(result_type, "model_validate", None)
-
-        if isinstance(output, str):
-            text_payload = output
-            if self.settings.protocol == LLMProtocolType.ANTHROPIC:
-                text_payload = self._normalize_anthropic_text_payload_with_diagnostics(
-                    text_payload,
-                    source="normalize_agent_output:model",
-                )
-            normalized_payload = self._strip_json_fence(text_payload)
-
-            if callable(validator_json):
-                try:
-                    return validator_json(normalized_payload)
-                except Exception:
-                    for candidate in reversed(self._extract_json_object_candidates(normalized_payload)):
-                        try:
-                            return validator_json(candidate)
-                        except Exception:
-                            continue
-
-            if callable(validator_obj):
-                try:
-                    decoded = json.loads(normalized_payload)
-                except Exception:
-                    decoded = None
-                if decoded is not None:
-                    try:
-                        return validator_obj(decoded)
-                    except Exception:
-                        pass
-            return output
-
-        if callable(validator_obj) and isinstance(output, (dict, list)):
-            try:
-                return validator_obj(output)
-            except Exception:
-                return output
-
-        return output
-
-    def _build_response_quality_summary(self, output: Any, telemetry: dict[str, Any] | None = None) -> dict[str, Any]:
-        """Build response-quality diagnostics from output + observed tool activity."""
-        summary = self._summarize_output(output)
-        quality: dict[str, Any] = {}
-        quality["score"] = summary.get("score")
-        quality["score_missing"] = bool(summary.get("score_missing", False))
-        quality["confidence"] = summary.get("confidence")
-        quality["issues_count"] = summary.get("issues_count", 0)
-        quality["vulnerabilities_count"] = summary.get("vulnerabilities_count", 0)
-        quality["assessments_count"] = summary.get("assessments_count", 0)
-        has_reasoning_field = hasattr(output, "reasoning")
-        quality["reasoning_chars"] = len(str(getattr(output, "reasoning", "") or "")) if has_reasoning_field else 0
-        quality["files_analyzed"] = summary.get("files_analyzed", 0)
-        tech_stack = summary.get("tech_stack")
-        quality["tech_stack"] = tech_stack
-        tool_count_by_name = telemetry.get("event_tool_count_by_name") if isinstance(telemetry, dict) else {}
-        if not isinstance(tool_count_by_name, dict):
-            tool_count_by_name = {}
-        observed_tool_calls = int(
-            (telemetry or {}).get("event_tool_calls_completed", (telemetry or {}).get("event_tool_calls", 0)) or 0
-        )
-        quality["observed_tool_calls"] = observed_tool_calls
-        has_effective_counter = isinstance(telemetry, dict) and "event_tool_effective_calls" in telemetry
-        observed_tool_effective_calls = int(
-            (telemetry or {}).get("event_tool_effective_calls", observed_tool_calls) or 0
-        )
-        quality["observed_tool_effective_calls"] = observed_tool_effective_calls
-        quality["observed_read_file_calls"] = int(tool_count_by_name.get("read_file", 0) or 0)
-        observed_skill_execution_calls = int(
-            (telemetry or {}).get(
-                "event_skill_execution_tool_calls", (telemetry or {}).get("event_skill_effective_tool_calls", 0)
-            )
-            or 0
-        )
-        observed_skill_effective_calls = int(
-            (telemetry or {}).get(
-                "event_skill_execution_effective_tool_calls",
-                (telemetry or {}).get("event_skill_effective_tool_calls", 0),
-            )
-            or 0
-        )
-        quality["observed_skill_execution_calls"] = observed_skill_execution_calls
-        quality["observed_skill_effective_calls"] = observed_skill_effective_calls
-        efficiency_summary = self._build_tool_efficiency_summary(telemetry if isinstance(telemetry, dict) else None)
-        turn_effectiveness_summary = efficiency_summary.get("turn_effectiveness_summary")
-        if isinstance(turn_effectiveness_summary, list):
-            quality["turn_effectiveness_summary"] = turn_effectiveness_summary
-            quality["observed_turn_count"] = len(turn_effectiveness_summary)
-        requirement_markers = efficiency_summary.get("turn_requirement_coverage_contribution_markers")
-        if isinstance(requirement_markers, dict):
-            quality["requirement_coverage_contribution_markers"] = {
-                "requirements": bool(requirement_markers.get("requirements")),
-                "docs": bool(requirement_markers.get("docs")),
-                "code": bool(requirement_markers.get("code")),
-            }
-        requirement_rate = efficiency_summary.get("turn_requirement_coverage_contribution_rate")
-        if isinstance(requirement_rate, (int, float)):
-            quality["requirement_coverage_contribution_rate"] = round(float(requirement_rate), 3)
-        turn_effective_rate = efficiency_summary.get("turn_effective_rate")
-        if isinstance(turn_effective_rate, (int, float)):
-            quality["turn_effective_rate"] = round(float(turn_effective_rate), 3)
-
-        flags: list[str] = []
-        if has_reasoning_field and quality["reasoning_chars"] < 80:
-            flags.append("reasoning_too_short")
-        if quality["observed_read_file_calls"] >= 3 and isinstance(tech_stack, list):
-            normalized_stack = [str(item).strip().lower() for item in tech_stack]
-            if normalized_stack and all(("unknown" in item) for item in normalized_stack):
-                flags.append("unknown_tech_stack_with_code_reads")
-        score = quality.get("score")
-        score_present = not bool(quality.get("score_missing", False))
-        quality["score_present"] = score_present
-        high_activity_calls = observed_tool_effective_calls if has_effective_counter else observed_tool_calls
-        if score_present and isinstance(score, (int, float)) and float(score) <= 2.5 and high_activity_calls >= 8:
-            flags.append("very_low_score_after_high_tool_activity")
-        if hasattr(output, "notes") and hasattr(output, "recommendations"):
-            doc_signal_count = (
-                len(getattr(output, "notes", []) or [])
-                + len(getattr(output, "recommendations", []) or [])
-                + len(getattr(output, "missing_sections", []) or [])
-                + len(getattr(output, "strengths", []) or [])
-            )
-            quality["doc_signal_count"] = doc_signal_count
-            if quality["observed_tool_calls"] >= 1 and doc_signal_count == 0:
-                flags.append("docs_signal_too_thin")
-        if hasattr(output, "vulnerabilities") and hasattr(output, "recommendations"):
-            security_signal_count = (
-                len(getattr(output, "vulnerabilities", []) or [])
-                + len(getattr(output, "recommendations", []) or [])
-                + len(getattr(output, "security_controls", []) or [])
-            )
-            quality["security_signal_count"] = security_signal_count
-            if quality["observed_tool_calls"] >= 8 and security_signal_count == 0:
-                flags.append("security_signal_too_thin")
-        quality["quality_flags"] = flags
-        quality["quality_ok"] = len(flags) == 0
-        return quality
-
-    @staticmethod
-    def _strip_json_fence(text: str) -> str:
-        trimmed = text.strip()
-        if trimmed.startswith("```"):
-            lines = trimmed.splitlines()
-            if len(lines) >= 2 and lines[-1].strip() == "```":
-                return "\n".join(lines[1:-1]).strip()
-            return "\n".join(lines[1:]).strip()
-        return trimmed
-
-    def _build_model_settings(self) -> ModelSettings:
-        settings: dict[str, Any] = {
-            "temperature": 0.0,
-            "max_tokens": self.settings.max_tokens_per_request,
-        }
-        from vds_audit_orchestrator.config import inject_reasoning_effort
-
-        inject_reasoning_effort(settings, self.settings.reasoning_effort, self.settings.protocol)
-        if self._should_disable_parallel_tool_calls():
-            # Local OpenAI-compatible gateways often perform better with serialized
-            # tool invocation; this reduces bursty tool-call fan-out and improves
-            # bounded-budget stability.
-            settings["parallel_tool_calls"] = False
-            self.logger.info(
-                "openai_parallel_tool_calls_disabled",
-                base_url=self.settings.base_url,
-                protocol=self.settings.protocol.value,
-            )
-        if self._should_suppress_anthropic_tool_choice():
-            # Some Anthropic-compatible proxy gateways reject tool_choice=auto after
-            # translation to OpenAI-compatible payloads; explicit null suppresses this
-            # field while preserving tool definitions/calls.
-            settings["extra_body"] = build_structured_output_extra_body(
-                self.settings.base_url,
-                self.settings.protocol,
-            )
-            self.logger.info(
-                "anthropic_proxy_tool_choice_suppressed",
-                base_url=self.settings.base_url,
-                protocol=self.settings.protocol.value,
-            )
-        elif self._should_suppress_openai_tool_choice():
-            settings["extra_body"] = build_structured_output_extra_body(
-                self.settings.base_url,
-                self.settings.protocol,
-            )
-            self.logger.info(
-                "openai_proxy_tool_choice_suppressed",
-                base_url=self.settings.base_url,
-                protocol=self.settings.protocol.value,
-            )
-        if self._is_openai_codex_protocol(self.settings.protocol):
-            settings["openai_store"] = False
-        return ModelSettings(
-            **settings,
-        )
-
-    def _build_run_model_settings(self, max_tokens: int) -> dict[str, Any]:
-        """Build per-run model settings kwargs used by Agent(...)."""
-        settings: dict[str, Any] = {
-            "temperature": 0.0,
-            "max_tokens": max_tokens,
-        }
-        from vds_audit_orchestrator.config import inject_reasoning_effort
-
-        inject_reasoning_effort(settings, self.settings.reasoning_effort, self.settings.protocol)
-        if self._should_disable_parallel_tool_calls():
-            settings["parallel_tool_calls"] = False
-        if self._should_suppress_anthropic_tool_choice() or self._should_suppress_openai_tool_choice():
-            settings["extra_body"] = build_structured_output_extra_body(
-                self.settings.base_url,
-                self.settings.protocol,
-            )
-        if self._is_openai_codex_protocol(self.settings.protocol):
-            settings["openai_store"] = False
-        return settings
-
-    def _should_disable_parallel_tool_calls(self) -> bool:
-        """Disable parallel tool calls for local OpenAI-compatible endpoints."""
-        if self.settings.protocol != LLMProtocolType.OPENAI:
-            return False
-        endpoint = str(self.settings.base_url or "").strip()
-        if not endpoint:
-            endpoint = "http://127.0.0.1:11434"
-        parsed = urlparse(endpoint)
-        host = (parsed.hostname or "").lower()
-        return host in {"localhost", "127.0.0.1", "0.0.0.0", "::1"}
-
-    def _should_suppress_anthropic_tool_choice(self) -> bool:
-        """Return True for non-official Anthropic endpoints with tool_choice incompatibility."""
-        return should_suppress_anthropic_tool_choice(self.settings.base_url, self.settings.protocol)
-
-    def _should_suppress_openai_tool_choice(self) -> bool:
-        """Return True for OpenAI-compatible endpoints with explicit tool_choice incompatibility."""
-        return should_suppress_openai_tool_choice(self.settings.base_url, self.settings.protocol)
-
-    # ------------------------------------------------------------------
-    # FR-70 / TSK-641.2: Event stream telemetry resilience
-    # ------------------------------------------------------------------
-
-    def _populate_telemetry_from_result(
-        self,
-        result: Any,
-        run_telemetry: dict[str, Any],
-    ) -> None:
-        """Populate *run_telemetry* from PydanticAI ``RunResult`` when the event
-        stream handler was removed (fallback path).
-
-        This prevents post-run guards from seeing zeroed-out counters and
-        raising ``RuntimeError("agentic usage guard unsatisfied")`` even though
-        tools actually ran successfully.
-
-        Three fallback data sources are used:
-        1. ``result.usage()`` — request count, token counts, tool_calls count.
-        2. ``result.all_messages()`` — walk message parts to count tool-call
-           and tool-return parts and extract tool names.
-        3. If neither works the dict is left unchanged (guards will still see
-           whatever partial data the event stream captured before failing).
-        """
-        # -- Fallback 1: result.usage() ---------------------------------
-        try:
-            usage = result.usage()
-            if usage is not None:
-                usage_tool_calls = int(getattr(usage, "tool_calls", 0) or 0)
-                usage_requests = int(getattr(usage, "requests", 0) or 0)
-                # Only backfill when event-stream counters are still at zero.
-                if int(run_telemetry.get("event_tool_calls_completed", 0) or 0) < 1 and usage_tool_calls > 0:
-                    run_telemetry["event_tool_calls_completed"] = usage_tool_calls
-                    run_telemetry["event_tool_calls_started"] = usage_tool_calls
-                    run_telemetry["event_tool_calls"] = usage_tool_calls
-                if int(run_telemetry.get("event_tool_effective_calls", 0) or 0) < 1 and usage_tool_calls > 0:
-                    # Usage payload has no per-tool usefulness signal; preserve monotonic counters.
-                    run_telemetry["event_tool_effective_calls"] = usage_tool_calls
-                if not run_telemetry.get("usage"):
-                    run_telemetry["usage"] = {
-                        "requests": usage_requests,
-                        "input_tokens": int(getattr(usage, "input_tokens", 0) or 0),
-                        "output_tokens": int(getattr(usage, "output_tokens", 0) or 0),
-                        "tool_calls": usage_tool_calls,
-                    }
-        except Exception:
-            pass  # usage() may not be available on all result types
-
-        # -- Fallback 2: result.all_messages() ---------------------------
-        try:
-            messages = result.all_messages() if callable(getattr(result, "all_messages", None)) else []
-            if messages and not run_telemetry.get("event_tool_names"):
-                tool_names: list[str] = []
-                tool_call_count = 0
-                for msg in messages:
-                    parts = getattr(msg, "parts", None)
-                    if not parts:
-                        continue
-                    for part in parts:
-                        part_kind = getattr(part, "part_kind", "")
-                        if part_kind == "tool-call":
-                            tool_call_count += 1
-                            tool_name = getattr(part, "tool_name", None)
-                            if tool_name:
-                                tool_names.append(tool_name)
-                if tool_names:
-                    run_telemetry["event_tool_names"] = list(set(tool_names))
-                    # Build count-by-name
-                    count_by_name: dict[str, int] = {}
-                    for tn in tool_names:
-                        count_by_name[tn] = count_by_name.get(tn, 0) + 1
-                    run_telemetry["event_tool_count_by_name"] = count_by_name
-                    # Populate skill-class telemetry from generic tool counters when
-                    # event-stream skill counters are missing.
-                    derived_skill_counts = {
-                        name: int(count_by_name.get(name, 0) or 0)
-                        for name in ("list_skills", "load_skill", "read_skill_resource", "run_skill_script")
-                        if int(count_by_name.get(name, 0) or 0) > 0
-                    }
-                    if derived_skill_counts:
-                        existing_skill_counts = run_telemetry.get("event_skill_tool_count_by_name")
-                        merged_skill_counts: dict[str, int] = (
-                            {str(k): int(v or 0) for k, v in existing_skill_counts.items() if str(k).strip()}
-                            if isinstance(existing_skill_counts, dict)
-                            else {}
-                        )
-                        for skill_tool, count in derived_skill_counts.items():
-                            merged_skill_counts[skill_tool] = max(
-                                int(merged_skill_counts.get(skill_tool, 0) or 0), count
-                            )
-                        run_telemetry["event_skill_tool_count_by_name"] = merged_skill_counts
-                        derived_discovery = int(merged_skill_counts.get("list_skills", 0) or 0)
-                        derived_bootstrap = int(merged_skill_counts.get("load_skill", 0) or 0)
-                        derived_execution = int(merged_skill_counts.get("read_skill_resource", 0) or 0) + int(
-                            merged_skill_counts.get("run_skill_script", 0) or 0
-                        )
-                        effective_count_by_name = run_telemetry.get("event_tool_effective_count_by_name")
-                        derived_execution_effective = 0
-                        if isinstance(effective_count_by_name, dict):
-                            derived_execution_effective = int(
-                                effective_count_by_name.get("read_skill_resource", 0) or 0
-                            ) + int(effective_count_by_name.get("run_skill_script", 0) or 0)
-                        if (
-                            int(run_telemetry.get("event_skill_discovery_tool_calls", 0) or 0) < 1
-                            and derived_discovery > 0
-                        ):
-                            run_telemetry["event_skill_discovery_tool_calls"] = derived_discovery
-                        if (
-                            int(run_telemetry.get("event_skill_bootstrap_tool_calls", 0) or 0) < 1
-                            and derived_bootstrap > 0
-                        ):
-                            run_telemetry["event_skill_bootstrap_tool_calls"] = derived_bootstrap
-                        if (
-                            int(run_telemetry.get("event_skill_execution_tool_calls", 0) or 0) < 1
-                            and derived_execution > 0
-                        ):
-                            run_telemetry["event_skill_execution_tool_calls"] = derived_execution
-                        if (
-                            int(run_telemetry.get("event_skill_execution_effective_tool_calls", 0) or 0) < 1
-                            and derived_execution_effective > 0
-                        ):
-                            run_telemetry["event_skill_execution_effective_tool_calls"] = derived_execution_effective
-                        if int(run_telemetry.get("event_skill_tool_calls", 0) or 0) < 1:
-                            run_telemetry["event_skill_tool_calls"] = sum(merged_skill_counts.values())
-                        if (
-                            int(run_telemetry.get("event_skill_effective_tool_calls", 0) or 0) < 1
-                            and derived_execution_effective > 0
-                        ):
-                            run_telemetry["event_skill_effective_tool_calls"] = derived_execution_effective
-                if tool_call_count > 0 and int(run_telemetry.get("event_tool_calls_completed", 0) or 0) < 1:
-                    run_telemetry["event_tool_calls_completed"] = tool_call_count
-                    run_telemetry["event_tool_calls_started"] = tool_call_count
-                    run_telemetry["event_tool_calls"] = tool_call_count
-                if tool_call_count > 0 and int(run_telemetry.get("event_tool_effective_calls", 0) or 0) < 1:
-                    run_telemetry["event_tool_effective_calls"] = tool_call_count
-        except Exception:
-            pass  # all_messages() may not be available
-
-        fallback_tool_calls = int(run_telemetry.get("event_tool_calls_completed", 0) or 0)
-        self.logger.info(
-            "event_stream_telemetry_fallback_result",
-            fallback_tool_calls=fallback_tool_calls,
-            fallback_tool_names=run_telemetry.get("event_tool_names"),
-            fallback_source="result.usage+all_messages",
-        )
-
-    def _record_pydantic_usage(self, result: Any, prompt_type: str) -> dict[str, int] | None:
-        """Record PydanticAI RunResult usage with the global tracker."""
-        try:
-            usage = result.usage()
-        except Exception:
-            return None
-
-        if usage is None:
-            return None
-
-        input_tokens = int(getattr(usage, "input_tokens", 0) or 0)
-        output_tokens = int(getattr(usage, "output_tokens", 0) or 0)
-        request_count = int(getattr(usage, "requests", 0) or 0)
-        tool_calls = int(getattr(usage, "tool_calls", 0) or 0)
-        model_name = self._get_model_name()
-
-        record = global_tracker.record_usage(
-            model=model_name,
-            prompt_type=prompt_type,
-            input_tokens=input_tokens,
-            output_tokens=output_tokens,
-        )
-        self.logger.info(
-            "pydantic_agent_usage_summary",
-            prompt_type=prompt_type,
-            model_name=model_name,
-            input_tokens=input_tokens,
-            output_tokens=output_tokens,
-            request_count=request_count,
-            tool_calls=tool_calls,
-        )
-        record_llm_usage(
-            model=model_name,
-            prompt_tokens=input_tokens,
-            completion_tokens=output_tokens,
-            cost=record.cost,
-        )
-        return {
-            "input_tokens": input_tokens,
-            "output_tokens": output_tokens,
-            "request_count": request_count,
-            "tool_calls": tool_calls,
-        }
-
-    def _get_model_name(self) -> str:
-        if self._pydantic_model is not None and hasattr(self._pydantic_model, "model_name"):
-            return str(self._pydantic_model.model_name)
-        return self.settings.model_standard
-
-    def _result(
-        self,
-        agent_name: str,
-        success: bool,
-        findings: list[dict[str, Any]] | None = None,
-        metadata: dict[str, Any] | None = None,
-        error: str | AuditError | None = None,
-        spawned_tasks: list[dict[str, Any]] | None = None,
-    ) -> AgentResult:
-        """Create an AgentResult with consistent pattern.
-
-        Args:
-            agent_name: Name identifier for this agent
-            success: Whether the analysis succeeded
-            findings: List of finding dictionaries
-            metadata: Additional metadata about the analysis
-            error: Error message if analysis failed
-            spawned_tasks: List of task specs to be spawned by the dispatcher
-
-        Returns:
-            AgentResult instance
-        """
-        normalized_error: str | None
-        normalized_error = str(error) if isinstance(error, AuditError) else error
-        return AgentResult(
-            agent_name=agent_name,
-            success=success,
-            findings=findings or [],
-            metadata=metadata or {},
-            error=normalized_error,
-            spawned_tasks=spawned_tasks or [],
-        )
-
-    def _spawn_task(
-        self,
-        task_type: str,
-        description: str,
-        assignee: str,
-        input_data: dict[str, Any] | None = None,
-        priority: int = 100,
-    ) -> dict[str, Any]:
-        """Create a task spec to be spawned by the dispatcher.
-
-        Args:
-            task_type: Type of task (e.g., 'api_analysis', 'security_scan')
-            description: Human-readable description of the task
-            assignee: Agent name that should handle this task
-            input_data: Optional input data for the spawned task
-            priority: Task priority (lower = higher priority, default 100)
-
-        Returns:
-            Task specification dictionary for the dispatcher
-        """
-        return {
-            "type": task_type,
-            "description": description,
-            "assignee": assignee,
-            "input_data": input_data or {},
-            "priority": priority,
-            "created_by": self.__class__.__name__,
-        }