"""FastAPI + SSE backend for the GeoForce dashboard / Streamlit fallback. Endpoints: GET /health → { "ok": true } GET /scenarios → demo/scenarios.yaml entries POST /predict → runs solver and/or surrogate, returns fields POST /query → streams Server-Sent Events from the agent: event: text data: {"text": "..."} event: tool data: {"name": "...", "input": {...}} event: result data: {"final_text": "...", "stop_reason": "..."} event: error data: {"message": "..."} Run with: .venv/bin/uvicorn agent.api:app --host 0.0.0.0 --port 8000 """ from __future__ import annotations import asyncio import json import time from contextlib import AsyncExitStack from dataclasses import dataclass, field from pathlib import Path from typing import Any, AsyncGenerator, Literal from uuid import uuid4 import numpy as np import yaml from fastapi import FastAPI, HTTPException from fastapi.middleware.cors import CORSMiddleware from fastapi.responses import FileResponse from fastapi.staticfiles import StaticFiles from pydantic import BaseModel from sse_starlette.sse import EventSourceResponse from claude_agent_sdk import ( AssistantMessage, ClaudeSDKClient, ResultMessage, TextBlock, ToolUseBlock, ) from agent.runtime import _load_env, build_options from tools.predict_solver import predict as solver_predict from tools.predict_surrogate import predict as surrogate_predict REPO_ROOT = Path(__file__).resolve().parent.parent SCENARIOS_PATH = REPO_ROOT / "demo" / "scenarios.yaml" app = FastAPI(title="GeoForce Agent API", version="0.1") app.add_middleware( CORSMiddleware, allow_origins=["*"], allow_credentials=False, allow_methods=["*"], allow_headers=["*"], ) class QueryRequest(BaseModel): query: str scenario_id: str | None = None @app.get("/health") def health() -> dict[str, Any]: return {"ok": True} @app.get("/scenarios") def scenarios() -> dict[str, Any]: if not SCENARIOS_PATH.exists(): return {"scenarios": []} with SCENARIOS_PATH.open() as f: data = yaml.safe_load(f) or {} return data async def _stream_agent(query: str) -> AsyncGenerator[dict[str, str], None]: _load_env() options = build_options() final_parts: list[str] = [] final_stop: str | None = None try: async with ClaudeSDKClient(options=options) as client: await client.query(query) async for message in client.receive_response(): if isinstance(message, AssistantMessage): for block in message.content: if isinstance(block, TextBlock): final_parts.append(block.text) yield { "event": "text", "data": json.dumps({"text": block.text}), } elif isinstance(block, ToolUseBlock): yield { "event": "tool", "data": json.dumps( {"name": block.name, "input": block.input} ), } elif isinstance(message, ResultMessage): final_stop = getattr(message, "stop_reason", None) if not final_parts and message.result: final_parts.append(message.result) break except Exception as exc: # noqa: BLE001 yield {"event": "error", "data": json.dumps({"message": str(exc)})} return yield { "event": "result", "data": json.dumps( {"final_text": "".join(final_parts), "stop_reason": final_stop} ), } class PredictRequest(BaseModel): scenario_id: str | None = None scenario: dict[str, Any] | None = None engine: Literal["both", "solver", "surrogate"] = "both" def _serialize_field(result: dict[str, Any]) -> dict[str, Any]: t = np.asarray(result["temperature"]) p = np.asarray(result["pressure"]) return { "grid": result["grid"], "temperature": t.tolist(), "pressure": p.tolist(), "t_min": float(t.min()), "t_max": float(t.max()), "p_min_MPa": float(p.min()) / 1.0e6, "p_max_MPa": float(p.max()) / 1.0e6, "elapsed_seconds": float(result.get("elapsed_seconds", 0.0)), } @app.post("/predict") def predict(req: PredictRequest) -> dict[str, Any]: """Run one or both engines on a scenario and return field arrays. Accepts either ``scenario_id`` (pulled from demo/scenarios.yaml) or an inline ``scenario`` dict. Returns solver and/or surrogate results with temperature/pressure arrays suitable for client-side heatmaps. """ scenario: dict[str, Any] | None = req.scenario if scenario is None and req.scenario_id: data = scenarios() match = next( (s for s in data.get("scenarios", []) if s.get("id") == req.scenario_id), None, ) if not match: raise HTTPException(404, f"scenario_id {req.scenario_id!r} not found") scenario = match.get("scenario") if not scenario: raise HTTPException(400, "scenario or scenario_id required") out: dict[str, Any] = {"engine": req.engine} if req.engine in ("both", "solver"): out["solver"] = _serialize_field(solver_predict(scenario)) if req.engine in ("both", "surrogate"): out["surrogate"] = _serialize_field(surrogate_predict(scenario)) return out @app.post("/query") async def query(req: QueryRequest) -> EventSourceResponse: """Stream the agent response as SSE. If ``scenario_id`` is provided, we prepend the matching question from demo/scenarios.yaml so the Streamlit/React client can fire-and-forget a preset scenario card. """ prompt = req.query if req.scenario_id: data = scenarios() match = next( (s for s in data.get("scenarios", []) if s.get("id") == req.scenario_id), None, ) if match and match.get("question"): prompt = f"{match['question']}\n\nUser override: {req.query}" if req.query else match["question"] return EventSourceResponse(_stream_agent(prompt)) # ---- Multi-turn sessions ---------------------------------------------------- # A "session" holds one long-lived ClaudeSDKClient so the dashboard can have a # real chat with the agent — the model sees prior turns + tool results. Each # session has a per-session asyncio.Lock so two concurrent /query calls don't # interleave on the same transport. Sessions idle > TTL are reaped. SESSION_TTL_SEC = 600 # 10 min idle SESSION_CAP = 32 # evict oldest when full @dataclass class _Session: client: ClaudeSDKClient stack: AsyncExitStack last_used: float lock: asyncio.Lock = field(default_factory=asyncio.Lock) SESSIONS: dict[str, _Session] = {} async def _open_session() -> str: _load_env() stack = AsyncExitStack() try: client = await stack.enter_async_context(ClaudeSDKClient(options=build_options())) except Exception: await stack.aclose() raise sid = uuid4().hex[:16] SESSIONS[sid] = _Session(client=client, stack=stack, last_used=time.time()) return sid async def _close_session(sid: str) -> None: sess = SESSIONS.pop(sid, None) if sess is None: return try: await sess.stack.aclose() except Exception: # noqa: BLE001 — best-effort teardown pass async def _reap_sessions() -> None: while True: try: now = time.time() stale = [sid for sid, s in list(SESSIONS.items()) if now - s.last_used > SESSION_TTL_SEC] for sid in stale: await _close_session(sid) except Exception: # noqa: BLE001 pass await asyncio.sleep(60) @app.on_event("startup") async def _startup() -> None: app.state._reaper = asyncio.create_task(_reap_sessions()) @app.post("/sessions") async def create_session() -> dict[str, str]: """Open a multi-turn session. Returns a session_id the client uses for subsequent /sessions/{id}/query calls.""" if len(SESSIONS) >= SESSION_CAP: oldest = min(SESSIONS, key=lambda k: SESSIONS[k].last_used) await _close_session(oldest) try: sid = await _open_session() except Exception as exc: # noqa: BLE001 raise HTTPException(500, f"failed to open session: {exc}") from exc return {"session_id": sid} @app.delete("/sessions/{sid}") async def drop_session(sid: str) -> dict[str, bool]: await _close_session(sid) return {"ok": True} async def _stream_session( sess: _Session, prompt: str ) -> AsyncGenerator[dict[str, str], None]: # Serialize queries on the same session; SDK transport isn't safe for # concurrent .query() calls. async with sess.lock: sess.last_used = time.time() final_parts: list[str] = [] final_stop: str | None = None try: await sess.client.query(prompt) async for message in sess.client.receive_response(): if isinstance(message, AssistantMessage): for block in message.content: if isinstance(block, TextBlock): final_parts.append(block.text) yield { "event": "text", "data": json.dumps({"text": block.text}), } elif isinstance(block, ToolUseBlock): yield { "event": "tool", "data": json.dumps( {"name": block.name, "input": block.input} ), } elif isinstance(message, ResultMessage): final_stop = getattr(message, "stop_reason", None) if not final_parts and message.result: final_parts.append(message.result) break except Exception as exc: # noqa: BLE001 yield {"event": "error", "data": json.dumps({"message": str(exc)})} return sess.last_used = time.time() yield { "event": "result", "data": json.dumps( {"final_text": "".join(final_parts), "stop_reason": final_stop} ), } @app.post("/sessions/{sid}/query") async def session_query(sid: str, req: QueryRequest) -> EventSourceResponse: """Stream a turn in an existing multi-turn session as SSE.""" sess = SESSIONS.get(sid) if sess is None: raise HTTPException(404, f"session {sid!r} not found or expired") prompt = req.query if req.scenario_id: data = scenarios() match = next( (s for s in data.get("scenarios", []) if s.get("id") == req.scenario_id), None, ) if match and match.get("question"): prompt = ( f"{match['question']}\n\nUser override: {req.query}" if req.query else match["question"] ) return EventSourceResponse(_stream_session(sess, prompt)) # ---- Serve built React dashboard (if present) ------------------------------- # In dev, the dashboard runs on Vite (http://localhost:5173) and proxies /api # to us. In docker/prod we bake `dashboard/dist/` into the image and serve it # at the same origin so there's only one port to expose. _DIST = REPO_ROOT / "dashboard" / "dist" if _DIST.is_dir(): app.mount( "/assets", StaticFiles(directory=_DIST / "assets"), name="assets" ) @app.get("/") def _spa_index() -> FileResponse: return FileResponse(_DIST / "index.html")