feat(runtime): añade AgentState y nodos LangGraph (validate, llm, propose, approve, finalize)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-10 10:56:25 +02:00
parent 006d781df0
commit 683cc69edc
3 changed files with 284 additions and 0 deletions
@@ -0,0 +1,193 @@
+"""Factory de nodos LangGraph parametrizados por engine, policy, llm, agent_def."""
+
+from __future__ import annotations
+
+import json
+import time
+from collections.abc import Awaitable, Callable
+from datetime import UTC, datetime
+from typing import Any
+from uuid import UUID
+
+import structlog
+from langgraph.types import interrupt
+
+from agentforge_core.domain.agent import AgentDefinition
+from agentforge_core.domain.policy import PolicyDefinition
+from agentforge_core.guardrails.base import GuardrailEngine
+from agentforge_core.llm.base import LLMProvider, Message
+from agentforge_core.runtime.state import AgentState
+
+log = structlog.get_logger(__name__)
+
+NodeFn = Callable[[AgentState], Awaitable[dict[str, Any]]]
+
+
+def _now() -> datetime:
+    return datetime.now(UTC)
+
+
+def _step(name: str, started: float, **detail: Any) -> dict[str, Any]:
+    return {
+        "step": name,
+        "timestamp": _now().isoformat(),
+        "duration_ms": int((time.perf_counter() - started) * 1000),
+        "detail": detail,
+    }
+
+
+def build_node_validate_input(engine: GuardrailEngine, policy: PolicyDefinition) -> NodeFn:
+    async def validate_input(state: AgentState) -> dict[str, Any]:
+        started = time.perf_counter()
+        trace_id = UUID(state["trace_id"])
+        violations = await engine.validate_input(state["user_input"], policy, trace_id)
+        any_blocked = any(v.blocked for v in violations)
+        new_violations = state.get("violations", []) + [
+            v.model_dump(mode="json") for v in violations
+        ]
+        update: dict[str, Any] = {
+            "violations": new_violations,
+            "decision_path": [_step("validate_input", started, n_violations=len(violations))],
+        }
+        if any_blocked:
+            update["status"] = "blocked_by_guardrail"
+        return update
+
+    return validate_input
+
+
+def build_node_llm_reason(provider: LLMProvider, agent_def: AgentDefinition) -> NodeFn:
+    async def llm_reason(state: AgentState) -> dict[str, Any]:
+        started = time.perf_counter()
+        messages = [
+            Message(role="system", content=agent_def.system_prompt),
+            Message(role="user", content=state["user_input"]),
+        ]
+        try:
+            result = await provider.complete(
+                messages=messages,
+                temperature=agent_def.llm.temperature,
+                max_tokens=agent_def.llm.max_tokens,
+            )
+        except Exception as exc:
+            log.error("llm_failed", trace_id=state["trace_id"], error=str(exc))
+            return {
+                "status": "failed",
+                "error": "llm_unavailable",
+                "decision_path": [_step("llm_reason", started, ok=False, error=str(exc))],
+            }
+        return {
+            "raw_llm_output": result.content,
+            "messages": [m.model_dump() for m in messages],
+            "decision_path": [
+                _step(
+                    "llm_reason",
+                    started,
+                    model=result.model,
+                    tokens_in=result.tokens_in,
+                    tokens_out=result.tokens_out,
+                    latency_ms=result.latency_ms,
+                )
+            ],
+        }
+
+    return llm_reason
+
+
+def build_node_validate_output(engine: GuardrailEngine, policy: PolicyDefinition) -> NodeFn:
+    async def validate_output(state: AgentState) -> dict[str, Any]:
+        started = time.perf_counter()
+        if state.get("status") in {"failed", "blocked_by_guardrail"}:
+            return {}
+        try:
+            parsed = json.loads(state.get("raw_llm_output") or "{}")
+        except json.JSONDecodeError as exc:
+            return {
+                "status": "failed",
+                "error": "output_schema_mismatch",
+                "decision_path": [_step("validate_output", started, ok=False, error=str(exc))],
+            }
+        trace_id = UUID(state["trace_id"])
+        violations = await engine.validate_output(parsed, policy, trace_id)
+        any_blocked = any(v.blocked for v in violations)
+        new_violations = state.get("violations", []) + [
+            v.model_dump(mode="json") for v in violations
+        ]
+        update: dict[str, Any] = {
+            "parsed_output": parsed,
+            "violations": new_violations,
+            "decision_path": [_step("validate_output", started, n_violations=len(violations))],
+        }
+        if any_blocked:
+            update["status"] = "blocked_by_guardrail"
+        return update
+
+    return validate_output
+
+
+def build_node_propose_actions() -> NodeFn:
+    async def propose_actions(state: AgentState) -> dict[str, Any]:
+        started = time.perf_counter()
+        if state.get("status") in {"failed", "blocked_by_guardrail"}:
+            return {}
+        parsed = state.get("parsed_output") or {}
+        actions = parsed.get("proposed_actions", [])
+        return {
+            "proposed_actions": actions,
+            "decision_path": [_step("propose_actions", started, n_actions=len(actions))],
+        }
+
+    return propose_actions
+
+
+def build_node_approve_gate(agent_def: AgentDefinition) -> NodeFn:
+    async def approve_gate(state: AgentState) -> dict[str, Any]:
+        started = time.perf_counter()
+        if state.get("status") in {"failed", "blocked_by_guardrail"}:
+            return {}
+        actions = state.get("proposed_actions", [])
+        risky = [
+            a
+            for a in actions
+            if int(a.get("risk_score", 1)) >= agent_def.risk_threshold_for_hitl
+            or bool(a.get("requires_approval"))
+        ]
+        if not risky:
+            return {"decision_path": [_step("approve_gate", started, hitl=False)]}
+        # Pausa la ejecución hasta que llegue resume(human_decision={...})
+        decision = interrupt({"awaiting_actions": risky})
+        return {
+            "human_decision": decision,
+            "decision_path": [_step("approve_gate", started, hitl=True, resumed=True)],
+        }
+
+    return approve_gate
+
+
+def build_node_finalize() -> NodeFn:
+    async def finalize(state: AgentState) -> dict[str, Any]:
+        started = time.perf_counter()
+        if state.get("status") in {"failed", "blocked_by_guardrail"}:
+            return {"decision_path": [_step("finalize", started, skipped=True)]}
+        parsed = state.get("parsed_output") or {}
+        decision = state.get("human_decision") or {}
+        approved_ids = set(decision.get("approved_action_ids", [])) if decision else None
+        actions = state.get("proposed_actions", [])
+        if approved_ids is not None:
+            final_actions = [a for a in actions if a.get("id") in approved_ids]
+            if decision.get("rejected"):
+                return {
+                    "status": "failed",
+                    "error": "rejected_by_human",
+                    "decision_path": [_step("finalize", started, rejected=True)],
+                }
+        else:
+            final_actions = actions
+        final_output = {**parsed, "approved_actions": final_actions}
+        return {
+            "final_output": final_output,
+            "status": "completed",
+            "decision_path": [_step("finalize", started, n_approved=len(final_actions))],
+        }
+
+    return finalize
@@ -0,0 +1,25 @@
+"""Estado tipado del grafo LangGraph."""
+
+from __future__ import annotations
+
+import operator
+from typing import Annotated, Any, TypedDict
+
+
+class AgentState(TypedDict, total=False):
+    """Estado mutable que fluye por el grafo; `decision_path` se acumula."""
+
+    trace_id: str
+    agent_name: str
+    agent_version: str
+    user_input: str
+    messages: list[dict[str, Any]]
+    raw_llm_output: str | None
+    parsed_output: dict[str, Any] | None
+    proposed_actions: list[dict[str, Any]]
+    violations: list[dict[str, Any]]
+    decision_path: Annotated[list[dict[str, Any]], operator.add]
+    status: str
+    error: str | None
+    human_decision: dict[str, Any] | None
+    final_output: dict[str, Any] | None