app.py

# app.py
import os, json, tempfile, logging
import gradio as gr
import pandas as pd
import numpy as np

# Quiet noisy logs
logging.getLogger("cmdstanpy").setLevel(logging.WARNING)
logging.getLogger("prophet").setLevel(logging.WARNING)

# ==== Tools (your @tool template) ============================================
from smolagents import tool, CodeAgent, OpenAIServerModel

# ---------- Helper (rounding) ----------
def _round_df(df: pd.DataFrame, places: int = 2) -> pd.DataFrame:
    if df is None or df.empty:
        return df
    out = df.copy()
    num_cols = out.select_dtypes(include=["number"]).columns
    out[num_cols] = out[num_cols].astype(float).round(places)
    return out

# ---------- Tool 1: Forecast ----------
from smolagents import tool
import json, pandas as pd, numpy as np

@tool
def forecast_tool(horizon_months: int = 1, use_demo: bool = True, history_csv_path: str = "",
                  use_covariates: bool = False) -> str:
    """
    Forecast monthly demand using a GLOBAL N-HiTS model (fast & accurate).

    Args:
        horizon_months (int): Number of future months to forecast (>=1).
        use_demo (bool): If True, generates synthetic history for FG100/FG200.
        history_csv_path (str): Optional CSV with columns [product_id,date,qty,(optional extra covariates...)].
        use_covariates (bool): If True and extra numeric columns exist, use them as past covariates
                               (for future effects you must provide future values too).

    Returns:
        str: JSON list of {"product_id","period_start","forecast_qty"} for the next horizon_months.
    """
    # --- 1) Load data in long form ---
    if use_demo or not history_csv_path:
        rng = pd.date_range("2023-01-01", periods=24, freq="MS")
        rows = []
        np.random.seed(0)
        for pid, base in [("FG100", 1800), ("FG200", 900)]:
            season = 1 + 0.15 * np.sin(2 * np.pi * (np.arange(len(rng)) / 12.0))
            qty = (base * season).astype(float)
            for d, q in zip(rng, qty):
                rows.append({"product_id": pid, "date": d, "qty": float(q)})
        df = pd.DataFrame(rows)
    else:
        df = pd.read_csv(history_csv_path)
        assert {"product_id","date","qty"} <= set(df.columns), "CSV must have product_id,date,qty"
        df["date"] = pd.to_datetime(df["date"], errors="coerce")
        df = df.dropna(subset=["date"])
        df["qty"] = pd.to_numeric(df["qty"], errors="coerce").fillna(0.0)

    # Ensure proper monthly frequency per SKU
    df = df.copy()
    df["product_id"] = df["product_id"].astype(str)

    # --- 2) Build Darts series (GLOBAL model across SKUs) ---
    from darts import TimeSeries
    series_list = []
    past_cov_list = []  # optional

    extra_cols = [c for c in df.columns if c not in ["product_id","date","qty"]]
    # keep only numeric covariates (categoricals must be pre-encoded)
    num_covs = [c for c in extra_cols if pd.api.types.is_numeric_dtype(df[c])]

    for pid, g in df.groupby("product_id"):
        g = (g.set_index("date")
               .sort_index()
               .resample("MS")
               .agg({**{"qty":"sum"}, **{c:"last" for c in num_covs}})
               .fillna(method="ffill")
               .fillna(0.0))
        y = TimeSeries.from_dataframe(g.reset_index(), time_col="date", value_cols="qty", freq="MS")
        series_list.append(y)

        if use_covariates and num_covs:
            pc = TimeSeries.from_dataframe(g.reset_index(), time_col="date", value_cols=num_covs, freq="MS")
            past_cov_list.append(pc)
        else:
            past_cov_list.append(None)

    # --- 3) Train N-HiTS (fast settings) ---
    from darts.models import NHiTSModel

    H = max(1, int(horizon_months))
    # keep chunk length small for short histories; model is global
    input_chunk = max(6, min(12, min(len(s) for s in series_list) - 1)) if series_list else 12

    model = NHiTSModel(
        input_chunk_length=input_chunk,
        output_chunk_length=min(H, 3),   # can roll to reach H
        n_epochs=60,                     # keep fast; tune up if needed
        batch_size=32,
        random_state=0,
        dropout=0.0,
    )

    if use_covariates and any(pc is not None for pc in past_cov_list):
        model.fit(series=series_list, past_covariates=past_cov_list, verbose=False)
    else:
        model.fit(series=series_list, verbose=False)

    # --- 4) Predict per SKU and return JSON ---
    out = []
    for pid, s, pc in zip(df["product_id"].unique(), series_list, past_cov_list):
        if use_covariates and pc is not None:
            pred = model.predict(n=H, series=s, past_covariates=pc)
        else:
            pred = model.predict(n=H, series=s)
        for ts, val in zip(pred.time_index, pred.values().flatten()):
            out.append({
                "product_id": str(pid),
                "period_start": pd.Timestamp(ts).strftime("%Y-%m-%d"),
                "forecast_qty": float(val)
            })

    return json.dumps(out)

# ---------- Tool 2: Optimize (LP) ----------
@tool
def optimize_supply_tool(forecast_json: str) -> str:
    """
    Optimize a single-month supply plan (LP) using the forecast.

    Args:
        forecast_json (str): JSON string returned by forecast_tool.

    Returns:
        str: JSON with summary + readable tables (not raw solver output).
    """
    from scipy.optimize import linprog

    rows = json.loads(forecast_json)
    # Use first month per product
    demand = {}
    for r in rows:
        p = r["product_id"]
        if p not in demand:
            demand[p] = float(r["forecast_qty"])

    P = sorted(demand.keys()) or ["FG100", "FG200"]
    price = {"FG100": 98.0, "FG200": 120.0}
    conv  = {"FG100": 12.5, "FG200": 15.0}
    r1    = {"FG100": 0.03, "FG200": 0.05}
    r2    = {"FG100": 0.02, "FG200": 0.01}

    RMs = ["RM_A", "RM_B"]
    rm_cost = {"RM_A": 20.0, "RM_B": 30.0}
    rm_start = {"RM_A": 1000.0, "RM_B": 100.0}
    rm_cap   = {"RM_A": 5000.0, "RM_B": 5000.0}
    bom = {
        "FG100": {"RM_A": 0.8, "RM_B": 0.2 * 1.02},
        "FG200": {"RM_A": 1.0, "RM_B": 0.1},
    }
    r1_cap, r2_cap = 320.0, 480.0

    nP, nR = len(P), len(RMs)
    pidx = {p:i for i,p in enumerate(P)}
    ridx = {r:i for i,r in enumerate(RMs)}

    def i_prod(p): return pidx[p]
    def i_sell(p): return nP + pidx[p]
    def i_einv(p): return 2*nP + pidx[p]
    def i_pur(r):  return 3*nP + ridx[r]
    def i_einr(r): return 3*nP + nR + ridx[r]

    n_vars = 3*nP + 2*nR
    c = np.zeros(n_vars)
    bounds = [None]*n_vars

    for p in P:
        c[i_prod(p)] += conv[p]
        c[i_sell(p)] -= price[p]
        c[i_einv(p)] += 0.0
        bounds[i_prod(p)] = (0, None)
        bounds[i_sell(p)] = (0, demand[p])  # demand as upper bound (no backorders)
        bounds[i_einv(p)] = (0, None)
    for r in RMs:
        c[i_pur(r)]  += rm_cost[r]
        c[i_einr(r)] += 0.0
        bounds[i_pur(r)]  = (0, rm_cap[r])
        bounds[i_einr(r)] = (0, None)

    # Equalities
    Aeq, beq = [], []
    for p in P:
        row = np.zeros(n_vars); row[i_prod(p)]=1; row[i_sell(p)]=-1; row[i_einv(p)]=-1
        Aeq.append(row); beq.append(0.0)  # start_inv=0 in this demo
    for r in RMs:
        row = np.zeros(n_vars); row[i_pur(r)]=1; row[i_einr(r)]=-1
        for p in P: row[i_prod(p)] -= bom[p].get(r,0.0)
        Aeq.append(row); beq.append(-rm_start[r])

    # Inequalities (resources)
    Aub, bub = [], []
    row = np.zeros(n_vars);  [row.__setitem__(i_prod(p), r1[p]) for p in P];  Aub.append(row); bub.append(r1_cap)
    row = np.zeros(n_vars);  [row.__setitem__(i_prod(p), r2[p]) for p in P];  Aub.append(row); bub.append(r2_cap)

    from scipy.optimize import linprog
    res = linprog(c, A_ub=np.array(Aub), b_ub=np.array(bub), A_eq=np.array(Aeq), b_eq=np.array(beq),
                  bounds=bounds, method="highs")
    if not res.success:
        return json.dumps({"status": "FAILED", "message": res.message})

    x = res.x
    def v(idx): return float(x[idx])

    # Compose human-friendly tables
    prod_tbl = []
    revenue = 0.0; conv_cost = 0.0
    for p in P:
        produce = v(i_prod(p)); sell = v(i_sell(p))
        prod_tbl.append({"Product": p, "Produce": produce, "Sell": sell, "Unit Price": price[p], "Conv. Cost/u": conv[p]})
        revenue += sell*price[p]; conv_cost += produce*conv[p]

    raw_tbl = []
    rm_purch_cost = 0.0
    for r in RMs:
        purchase = v(i_pur(r))
        cons = float(sum(bom[p].get(r,0.0)*v(i_prod(p)) for p in P))
        cost = purchase*rm_cost[r]; rm_purch_cost += cost
        raw_tbl.append({"Raw": r, "Purchase": purchase, "Consume": cons, "Cost/u": rm_cost[r], "Total Cost": cost})

    r1_used = float(sum(r1[p]*v(i_prod(p)) for p in P))
    r2_used = float(sum(r2[p]*v(i_prod(p)) for p in P))
    res_tbl = [
        {"Resource": "R1", "Used": r1_used, "Cap": r1_cap, "Slack": r1_cap - r1_used},
        {"Resource": "R2", "Used": r2_used, "Cap": r2_cap, "Slack": r2_cap - r2_used},
    ]
    profit = revenue - conv_cost - rm_purch_cost

    out = {
        "status": "OPTIMAL",
        "kpis": {"Profit": profit, "Revenue": revenue, "Conv. Cost": conv_cost, "RM Purchase Cost": rm_purch_cost},
        "products": prod_tbl,
        "raw_materials": raw_tbl,
        "resources": res_tbl
    }
    return json.dumps(out)

# ---------- Tool 3: MD61 (simulate) ----------
@tool
def update_sap_md61_tool(forecast_json: str, plant: str = "PLANT01", uom: str = "EA", mrp_area: str = "") -> str:
    """
    Prepare an MD61-style demand upload (SIMULATION ONLY).

    Args:
        forecast_json (str): JSON string returned by forecast_tool.
        plant (str): SAP plant (WERKS). Defaults to 'PLANT01'.
        uom (str): Unit of measure. Defaults to 'EA'.
        mrp_area (str): Optional MRP area.

    Returns:
        str: JSON with {"status":"SIMULATED","csv_path":"...","preview":[...]}.
    """
    rows = json.loads(forecast_json)
    md61 = [{
        "Material": r["product_id"], "Plant": plant, "MRP_Area": mrp_area,
        "Req_Date": r["period_start"], "Req_Qty": float(r["forecast_qty"]),
        "UoM": uom, "Version": "00"
    } for r in rows]
    df = pd.DataFrame(md61)
    tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".csv")
    df.to_csv(tmp.name, index=False)
    return json.dumps({"status": "SIMULATED", "csv_path": tmp.name, "preview": md61[:5]})

# ---------- Upgrade A: Data Quality Guard ----------
@tool
def data_quality_guard_tool(use_demo: bool = True, history_csv_path: str = "", z: float = 3.5) -> str:
    """
    Scan monthly history for outliers and likely stockout zeros. Demo-friendly.

    Args:
        use_demo (bool): If True, use synthetic FG100/FG200 history.
        history_csv_path (str): Optional CSV path with columns [product_id,date,qty].
        z (float): Robust z-threshold (MAD-based) for outlier flags. Default 3.5.

    Returns:
        str: JSON {"status":"OK","issues":[{product_id,period_start,qty,flag,suggested_action,suggested_value?}]}
    """
    # Load history
    if use_demo or not history_csv_path:
        rng = pd.date_range("2023-01-01", periods=24, freq="MS")
        rows = []
        np.random.seed(0)
        for pid, base in [("FG100", 1800), ("FG200", 900)]:
            season = 1 + 0.15 * np.sin(2 * np.pi * (np.arange(len(rng)) / 12.0))
            qty = (base * season).astype(float)
            for d, q in zip(rng, qty):
                rows.append({"product_id": pid, "date": d, "qty": float(q)})
        df = pd.DataFrame(rows)
    else:
        df = pd.read_csv(history_csv_path)
        assert {"product_id","date","qty"} <= set(df.columns), "CSV must have product_id,date,qty"
        df["date"] = pd.to_datetime(df["date"], errors="coerce")
        df = df.dropna(subset=["date"])
        df["qty"] = pd.to_numeric(df["qty"], errors="coerce").fillna(0.0)

    issues = []
    for pid, g in df.groupby("product_id"):
        s = g.set_index("date")["qty"].resample("MS").sum().asfreq("MS").fillna(0.0)
        if len(s) < 6:
            continue
        med = s.rolling(6, min_periods=3).median()
        mad = (s - med).abs().rolling(6, min_periods=3).median()
        robust_z = (0.6745 * (s - med) / mad.replace(0, np.nan)).abs()

        for ts, y in s.items():
            flag = None; action = None; val = None
            rz = robust_z.get(ts, np.nan)
            # Outlier
            if not np.isnan(rz) and rz > z:
                flag = "OUTLIER"
                action = "cap_to_rolling_median"
                val = float(med.get(ts, np.nan)) if not np.isnan(med.get(ts, np.nan)) else None
            # Likely stockout zero in-between non-zeros
            prev_nonzero = s[:ts].iloc[:-1].ne(0).any() if len(s[:ts])>1 else False
            next_nonzero = s[ts:].iloc[1:].ne(0).any() if len(s[ts:])>1 else False
            if y == 0 and prev_nonzero and next_nonzero:
                flag = flag or "ZERO_BETWEEN_NONZERO"
                action = action or "impute_neighbor_avg"
                idx = s.index.get_loc(ts)
                neighbors = []
                if idx>0: neighbors.append(s.iloc[idx-1])
                if idx<len(s)-1: neighbors.append(s.iloc[idx+1])
                val = float(np.mean(neighbors)) if neighbors else None
            if flag:
                issues.append({
                    "product_id": pid,
                    "period_start": ts.strftime("%Y-%m-%d"),
                    "qty": float(y),
                    "flag": flag,
                    "suggested_action": action,
                    "suggested_value": None if val is None or np.isnan(val) else float(val)
                })
    return json.dumps({"status":"OK","issues":issues})

# ---------- Upgrade B: Scenario Explorer ----------
@tool
def scenario_explorer_tool(forecast_json: str) -> str:
    """
    Run a small set of what-if scenarios and summarize results.

    Args:
        forecast_json (str): JSON string from forecast_tool (uses first month per SKU).

    Returns:
        str: JSON {"status":"OK","scenarios":[{name,profit,r1_used,r1_slack,r2_used,r2_slack}]}
    """
    from scipy.optimize import linprog

    base_rows = json.loads(forecast_json)
    demand = {}
    for r in base_rows:
        p = r["product_id"]
        if p not in demand:
            demand[p] = float(r["forecast_qty"])
    P = sorted(demand.keys()) or ["FG100","FG200"]

    # Base constants
    price = {"FG100": 98.0, "FG200": 120.0}
    conv  = {"FG100": 12.5, "FG200": 15.0}
    r1    = {"FG100": 0.03, "FG200": 0.05}
    r2    = {"FG100": 0.02, "FG200": 0.01}
    RMs = ["RM_A","RM_B"]
    rm_cost = {"RM_A": 20.0, "RM_B": 30.0}
    rm_start = {"RM_A": 1000.0, "RM_B": 100.0}
    rm_cap = {"RM_A": 5000.0, "RM_B": 5000.0}
    bom = {
        "FG100": {"RM_A": 0.8, "RM_B": 0.2*1.02},
        "FG200": {"RM_A": 1.0, "RM_B": 0.1},
    }
    r1_cap0, r2_cap0 = 320.0, 480.0

    scenarios = [
        {"name":"Base", "mult":1.0, "r1_cap":r1_cap0, "r2_cap":r2_cap0, "rm_cost_B_add":0.0, "conv_fg100_mult":1.0, "rmA_start_mult":1.0},
        {"name":"+10% Demand", "demand_mult":1.10, "r1_cap":r1_cap0, "r2_cap":r2_cap0, "rm_cost_B_add":0.0, "conv_fg100_mult":1.0, "rmA_start_mult":1.0},
        {"name":"-10% R1 Cap", "mult":1.0, "r1_cap":0.9*r1_cap0, "r2_cap":r2_cap0, "rm_cost_B_add":0.0, "conv_fg100_mult":1.0, "rmA_start_mult":1.0},
        {"name":"+₹5 RM_B", "mult":1.0, "r1_cap":r1_cap0, "r2_cap":r2_cap0, "rm_cost_B_add":5.0, "conv_fg100_mult":1.0, "rmA_start_mult":1.0},
        {"name":"+10% FG100 ConvCost", "mult":1.0, "r1_cap":r1_cap0, "r2_cap":r2_cap0, "rm_cost_B_add":0.0, "conv_fg100_mult":1.10, "rmA_start_mult":1.0},
        {"name":"-20% RM_A Start", "mult":1.0, "r1_cap":r1_cap0, "r2_cap":r2_cap0, "rm_cost_B_add":0.0, "conv_fg100_mult":1.0, "rmA_start_mult":0.80},
    ]

    results = []
    for sc in scenarios:
        dem_mult = sc.get("demand_mult", sc.get("mult",1.0))
        d = {p: demand[p]*dem_mult for p in P}
        r1_cap = sc["r1_cap"]; r2_cap = sc["r2_cap"]
        rm_cost_local = rm_cost.copy(); rm_cost_local["RM_B"] += sc["rm_cost_B_add"]
        conv_local = conv.copy(); conv_local["FG100"] = conv_local["FG100"] * sc["conv_fg100_mult"]
        rm_start_local = rm_start.copy(); rm_start_local["RM_A"] = rm_start_local["RM_A"] * sc["rmA_start_mult"]

        # Build and solve LP
        nP, nR = len(P), len(RMs)
        pidx = {p:i for i,p in enumerate(P)}; ridx = {r:i for i,r in enumerate(RMs)}
        def i_prod(p): return pidx[p]
        def i_sell(p): return nP + pidx[p]
        def i_einv(p): return 2*nP + pidx[p]
        def i_pur(r):  return 3*nP + ridx[r]
        def i_einr(r): return 3*nP + nR + ridx[r]
        n_vars = 3*nP + 2*nR
        c = np.zeros(n_vars); bounds = [None]*n_vars

        for p in P:
            c[i_prod(p)] += conv_local[p]; c[i_sell(p)] -= price[p]; c[i_einv(p)] += 0.0
            bounds[i_prod(p)] = (0, None); bounds[i_sell(p)] = (0, d[p]); bounds[i_einv(p)] = (0, None)
        for r in RMs:
            c[i_pur(r)] += rm_cost_local[r]; c[i_einr(r)] += 0.0
            bounds[i_pur(r)] = (0, rm_cap[r]); bounds[i_einr(r)] = (0, None)

        Aeq, beq = [], []
        for p in P:
            row = np.zeros(n_vars); row[i_prod(p)]=1; row[i_sell(p)]=-1; row[i_einv(p)]=-1
            Aeq.append(row); beq.append(0.0)
        for r in RMs:
            row = np.zeros(n_vars); row[i_pur(r)]=1; row[i_einr(r)]=-1
            for p in P: row[i_prod(p)] -= bom[p].get(r,0.0)
            Aeq.append(row); beq.append(-rm_start_local[r])

        Aub, bub = [], []
        row = np.zeros(n_vars); [row.__setitem__(i_prod(p), r1[p]) for p in P]; Aub.append(row); bub.append(r1_cap)
        row = np.zeros(n_vars); [row.__setitem__(i_prod(p), r2[p]) for p in P]; Aub.append(row); bub.append(r2_cap)

        from scipy.optimize import linprog
        res = linprog(c, A_ub=np.array(Aub), b_ub=np.array(bub), A_eq=np.array(Aeq), b_eq=np.array(beq),
                      bounds=bounds, method="highs")
        if not res.success:
            results.append({"name": sc["name"], "status":"FAILED", "profit": None})
            continue

        x = res.x
        def v(idx): return float(x[idx])

        r1_used = float(sum(r1[p]*v(i_prod(p)) for p in P))
        r2_used = float(sum(r2[p]*v(i_prod(p)) for p in P))
        revenue = float(sum(price[p]*v(i_sell(p)) for p in P))
        conv_cost = float(sum(conv_local[p]*v(i_prod(p)) for p in P))
        rm_purch_cost = float(sum(v(i_pur(r))*rm_cost_local[r] for r in RMs))
        profit = revenue - conv_cost - rm_purch_cost

        results.append({
            "name": sc["name"],
            "status":"OPTIMAL",
            "profit": profit,
            "r1_used": r1_used, "r1_slack": r1_cap - r1_used,
            "r2_used": r2_used, "r2_slack": r2_cap - r2_used
        })

    return json.dumps({"status":"OK","scenarios":results})

# ---------- Upgrade C: Plan Explainer ----------
@tool
def plan_explainer_tool(plan_json: str) -> str:
    """
    Generate a plain-language explanation of the plan (demo assumptions).

    Args:
        plan_json (str): JSON from optimize_supply_tool.

    Returns:
        str: JSON {"status":"OK","summary": "... human-readable text ..."}
    """
    d = json.loads(plan_json)
    k = d.get("kpis", {})
    products = d.get("products", [])
    resources = d.get("resources", [])

    # Contribution by product
    contribs = []
    for row in products:
        rev = row["Sell"] * row["Unit Price"]
        conv = row["Produce"] * row["Conv. Cost/u"]
        contribs.append((row["Product"], rev - conv))
    contribs.sort(key=lambda x: x[1], reverse=True)
    top_prod = contribs[0][0] if contribs else "N/A"

    # Binding resource (min slack)
    bind_res = None
    if resources:
        r_sorted = sorted(resources, key=lambda r: r.get("Slack", 0.0))
        bind_res = r_sorted[0]["Resource"]

    summary = (
        f"Profit: ₹{k.get('Profit',0):,.2f} (Revenue ₹{k.get('Revenue',0):,.2f} − "
        f"Conv. Cost ₹{k.get('Conv. Cost',0):,.2f} − RM Cost ₹{k.get('RM Purchase Cost',0):,.2f}). "
        f"Top driver: {top_prod}. "
        f"Bottleneck check: {bind_res} has the least slack."
    )

    return json.dumps({"status":"OK","summary": summary})

# ---------- NEW: Bottleneck Search (finite-difference + policy) ----------
@tool
def bottleneck_search_tool(forecast_json: str, policy_json: str = "") -> str:
    """
    Find the best practical lever to relax (within ~1 month) via small scenario probes.
    Supports resource overtime, RM expedite (cap), and demand_lift (promo ROI-checked).

    Args:
        forecast_json (str): JSON from forecast_tool (first month per SKU used).
        policy_json (str): Optional JSON with levers and costs.

    Returns:
        str: JSON {"status":"OK","diagnostics":{...},"recommendations":[...]}
    """
    import json, numpy as np
    from scipy.optimize import linprog

    # --- Defaults (edit to your reality) ---
    policy = {
        "R1_overtime":   {"type":"resource","target":"R1","step":10.0,"max_delta":80.0,"unit_cost":600.0},
        "R2_overtime":   {"type":"resource","target":"R2","step":10.0,"max_delta":40.0,"unit_cost":800.0},
        "RM_A_expedite": {"type":"rm_expedite","target":"RM_A","step":200.0,"max_delta":1500.0,"unit_premium":8.0},
        "RM_B_expedite": {"type":"rm_expedite","target":"RM_B","step":100.0,"max_delta":800.0,"unit_premium":12.0},
        "Factory_expansion": {"type":"blocked","reason":"Not relaxable within 1 month"}
        # You can add demand levers via policy_json:
        # "promo_FG100": {"type":"demand_lift","target":"FG100","step":100,"max_delta":600,"unit_cost":10}
    }
    if policy_json:
        try: policy.update(json.loads(policy_json))
        except Exception: pass

    # --- Problem primitives (same as your optimizer) ---
    rows = json.loads(forecast_json)
    demand = {}
    for r in rows:
        p = r["product_id"]
        if p not in demand:
            demand[p] = float(r["forecast_qty"])
    P = sorted(demand.keys()) or ["FG100","FG200"]

    price = {"FG100":98.0,"FG200":120.0}
    conv  = {"FG100":12.5,"FG200":15.0}
    r1    = {"FG100":0.03,"FG200":0.05}
    r2    = {"FG100":0.02,"FG200":0.01}
    RMs = ["RM_A","RM_B"]
    rm_cost = {"RM_A":20.0,"RM_B":30.0}
    rm_start = {"RM_A":1000.0,"RM_B":100.0}
    rm_cap   = {"RM_A":5000.0,"RM_B":5000.0}
    bom = {"FG100":{"RM_A":0.8,"RM_B":0.204},"FG200":{"RM_A":1.0,"RM_B":0.1}}
    r1_cap0, r2_cap0 = 320.0, 480.0

    # Helpful unit margin for demand_lift (approx variable cost)
    unit_var_cost = {p: conv[p] + sum(bom[p].get(rm,0.0)*rm_cost[rm] for rm in RMs) for p in P}
    unit_margin   = {p: price[p] - unit_var_cost[p] for p in P}

    # --- LP builder (inner solve) ---
    def solve(dem, r1_cap, r2_cap, rm_cap_adj):
        nP, nR = len(P), len(RMs)
        pidx = {p:i for i,p in enumerate(P)}; ridx = {r:i for i,r in enumerate(RMs)}
        def i_prod(p): return pidx[p]
        def i_sell(p): return nP + pidx[p]
        def i_einv(p): return 2*nP + pidx[p]
        def i_pur(r):  return 3*nP + ridx[r]
        def i_einr(r): return 3*nP + nR + ridx[r]

        n_vars = 3*nP + 2*nR
        c = np.zeros(n_vars); bounds=[None]*n_vars

        for p in P:
            c[i_prod(p)] += conv[p]
            c[i_sell(p)] -= price[p]
            bounds[i_prod(p)] = (0,None)
            bounds[i_sell(p)] = (0, dem[p])
            bounds[i_einv(p)] = (0,None)
        for r in RMs:
            c[i_pur(r)]  += rm_cost[r]
            bounds[i_pur(r)]  = (0, rm_cap_adj[r])
            bounds[i_einr(r)] = (0,None)

        Aeq, beq = [], []
        for p in P:
            row = np.zeros(n_vars); row[i_prod(p)]=1; row[i_sell(p)]=-1; row[i_einv(p)]=-1
            Aeq.append(row); beq.append(0.0)
        for r in RMs:
            row = np.zeros(n_vars); row[i_pur(r)]=1; row[i_einr(r)]=-1
            for p in P: row[i_prod(p)] -= bom[p].get(r,0.0)
            Aeq.append(row); beq.append(-rm_start[r])

        Aub, bub = [], []
        row = np.zeros(n_vars); [row.__setitem__(i_prod(p), r1[p]) for p in P]; Aub.append(row); bub.append(r1_cap)
        row = np.zeros(n_vars); [row.__setitem__(i_prod(p), r2[p]) for p in P]; Aub.append(row); bub.append(r2_cap)

        res = linprog(c, A_ub=np.array(Aub), b_ub=np.array(bub), A_eq=np.array(Aeq), b_eq=np.array(beq),
                      bounds=bounds, method="highs")
        if not res.success:
            return None
        x = res.x
        def v(idx): return float(x[idx])

        revenue = sum(price[p]*v(i_sell(p)) for p in P)
        conv_cost = sum(conv[p]*v(i_prod(p)) for p in P)
        rm_purch_cost = sum(v(i_pur(r))*rm_cost[r] for r in RMs)
        r1_used = sum(r1[p]*v(i_prod(p)) for p in P)
        r2_used = sum(r2[p]*v(i_prod(p)) for p in P)
        return {
            "profit": revenue - conv_cost - rm_purch_cost,
            "r1_used": r1_used, "r2_used": r2_used,
        }

    # Baseline
    base = solve(demand, r1_cap0, r2_cap0, rm_cap.copy())
    if base is None:
        return json.dumps({"status":"FAILED","message":"Baseline infeasible."})
    base_profit = float(base["profit"])
    diag = {
        "demand_met": True,  # sells hit demand ceiling in this model
        "r1_slack": round(r1_cap0 - base["r1_used"], 2),
        "r2_slack": round(r2_cap0 - base["r2_used"], 2),
        "rm_caps_hit": {rm: False for rm in RMs}
    }
    # If everything is slack, tell the user plainly
    all_slack = (diag["r1_slack"] > 1e-3 and diag["r2_slack"] > 1e-3)

    # --- Probe levers ---
    recs = []
    for name, cfg in policy.items():
        if cfg.get("type") == "blocked":
            recs.append({"lever":name,"recommended":0,"est_profit_lift":0,"est_net_gain":0,
                         "rationale":cfg.get("reason","Not feasible in horizon")})
            continue

        best_gain = 0.0; best_delta = 0.0; best_lift = 0.0
        step = float(cfg.get("step",0)); maxd = float(cfg.get("max_delta",0))
        d = 0.0
        while d <= maxd + 1e-6:
            dem = demand.copy()
            r1_cap = r1_cap0; r2_cap = r2_cap0
            rm_cap_adj = rm_cap.copy()
            relax_cost = 0.0

            t = cfg["type"]
            if t == "resource" and cfg["target"] == "R1":
                r1_cap += d; relax_cost = d * float(cfg["unit_cost"])
            elif t == "resource" and cfg["target"] == "R2":
                r2_cap += d; relax_cost = d * float(cfg["unit_cost"])
            elif t == "rm_expedite":
                key = "RM_A" if cfg["target"]=="RM_A" else "RM_B"
                rm_cap_adj[key] = rm_cap[key] + d
                relax_cost = d * float(cfg["unit_premium"])
            elif t == "demand_lift":
                sku = cfg["target"]
                if sku in dem:
                    dem[sku] = dem[sku] + d
                    relax_cost = d * float(cfg["unit_cost"])
                else:
                    d += step; continue
            else:
                d += step; continue

            res = solve(dem, r1_cap, r2_cap, rm_cap_adj)
            if res is None:
                d += step; continue
            lift = float(res["profit"] - base_profit)
            net = lift - relax_cost
            if net > best_gain:
                best_gain, best_delta, best_lift = net, d, lift
            d += step

        rationale = ("Beneficial at small Δ" if best_gain>0 else
                     ("Capacity not binding; try demand/cost levers" if all_slack else
                      "No positive net gain within feasible range"))
        recs.append({"lever":name,"recommended":round(best_delta,2),
                     "est_profit_lift":round(best_lift,2),
                     "est_net_gain":round(best_gain,2),
                     "rationale":rationale})

    recs.sort(key=lambda r: r["est_net_gain"], reverse=True)
    return json.dumps({"status":"OK","diagnostics":diag,"recommendations":recs})

# ==== Agent (end-to-end) ======================================================
def make_agent():
    api_key = os.environ.get("OPENAI_API_KEY", "")
    if not api_key:
        raise RuntimeError("OPENAI_API_KEY not set. Add it as a Space secret.")
    model = OpenAIServerModel(model_id="gpt-4o-mini", api_key=api_key, temperature=0)
    return CodeAgent(tools=[
        forecast_tool, optimize_supply_tool, update_sap_md61_tool,
        data_quality_guard_tool, scenario_explorer_tool, plan_explainer_tool, bottleneck_search_tool
    ], model=model, add_base_tools=False, stream_outputs=False)

SYSTEM_PLAN = (
    "Run the pipeline and return one JSON:\n"
    "1) forecast_tool(...)\n"
    "2) optimize_supply_tool(forecast_json)\n"
    "3) update_sap_md61_tool(forecast_json, ...)\n"
    "Return: {'forecast': <json>, 'plan': <json>, 'md61': <json>}"
)

def run_agentic(h, plant, demo_flag, file_obj):
    agent = make_agent()
    if file_obj is not None:
        path = file_obj.name
        prompt = (f"{SYSTEM_PLAN}\n"
                  f"Use forecast_tool(horizon_months={int(h)}, use_demo=False, history_csv_path='{path}'). "
                  f"Then run the other two steps as specified. Return only the final JSON.")
    else:
        prompt = (f"{SYSTEM_PLAN}\n"
                  f"Use forecast_tool(horizon_months={int(h)}, use_demo={bool(demo_flag)}). "
                  f"Then run the other two steps as specified. Return only the final JSON.")
    return agent.run(prompt)

# ==== UI Helpers (pretty) =====================================================
def parse_forecast(json_str):
    df = pd.DataFrame(json.loads(json_str))
    df = df[["product_id","period_start","forecast_qty"]].rename(columns={
        "product_id":"Product","period_start":"Period Start","forecast_qty":"Forecast Qty"
    })
    return _round_df(df)

def parse_plan(json_str):
    d = json.loads(json_str)
    kpis = pd.DataFrame([d["kpis"]]).rename(columns={
        "Conv. Cost":"Conversion Cost", "RM Purchase Cost":"RM Purchase Cost"
    })
    prod = pd.DataFrame(d["products"])
    raw  = pd.DataFrame(d["raw_materials"])
    res  = pd.DataFrame(d["resources"])
    return d["status"], _round_df(kpis), _round_df(prod), _round_df(raw), _round_df(res)

def parse_md61(json_str):
    d = json.loads(json_str)
    prev = _round_df(pd.DataFrame(d.get("preview", [])))
    path = d.get("csv_path", "")
    return d.get("status",""), prev, path

def parse_data_quality(json_str):
    d = json.loads(json_str)
    df = _round_df(pd.DataFrame(d.get("issues", [])))
    return df

def parse_scenarios(json_str):
    d = json.loads(json_str)
    df = _round_df(pd.DataFrame(d.get("scenarios", [])))
    if not df.empty:
        cols = ["name","status","profit","r1_used","r1_slack","r2_used","r2_slack"]
        df = df[cols]
        df = df.rename(columns={"name":"Scenario","status":"Status","profit":"Profit",
                                "r1_used":"R1 Used","r1_slack":"R1 Slack","r2_used":"R2 Used","r2_slack":"R2 Slack"})
    return df

def parse_recs(json_str):
    d = json.loads(json_str)
    df = _round_df(pd.DataFrame(d.get("recommendations", [])))
    if not df.empty:
        df = df.rename(columns={
            "lever":"Lever","recommended":"Recommended Δ","est_profit_lift":"Est. Profit Lift",
            "est_net_gain":"Est. Net Gain","rationale":"Rationale"
        })
    return df

# ==== Gradio UI ==============================================================
with gr.Blocks(title="Forecast → Optimize → SAP MD61") as demo:
    gr.Markdown("## 🧭 Workflow\n"
                "### 1) **Forecast** → 2) **Optimize Supply** → 3) **Prepare MD61**\n"
                "Run them **manually** below, or use the **agent** to do end-to-end in one click.")

    with gr.Tab("Manual (Step-by-step)"):
        with gr.Row():
            horizon = gr.Number(label="Horizon (months)", value=1, precision=0)
            plant   = gr.Textbox(label="SAP Plant (WERKS)", value="PLANT01")
        with gr.Row():
            use_demo = gr.Checkbox(label="Use demo synthetic history", value=True)
            file = gr.File(label="Or upload history.csv (product_id,date,qty)", file_types=[".csv"])

        # States
        forecast_state = gr.State("")
        plan_state = gr.State("")
        md61_state = gr.State("")

        gr.Markdown("### ➤ Step 1: Forecast")
        run_f = gr.Button("Run Step 1 — Forecast")
        forecast_tbl = gr.Dataframe(label="Forecast (first horizon month per SKU)", interactive=False)
        forecast_note = gr.Markdown("")

        gr.Markdown("### ➤ Step 2: Optimize Supply")
        run_o = gr.Button("Run Step 2 — Optimize")
        plan_status = gr.Markdown("")
        plan_kpis   = gr.Dataframe(label="KPIs", interactive=False)
        plan_prod   = gr.Dataframe(label="Products Plan", interactive=False)
        plan_raw    = gr.Dataframe(label="Raw Materials", interactive=False)
        plan_res    = gr.Dataframe(label="Resources", interactive=False)

        gr.Markdown("### ➤ Step 3: Prepare MD61 (Simulated)")
        run_m = gr.Button("Run Step 3 — MD61")
        md61_status = gr.Markdown("")
        md61_prev   = gr.Dataframe(label="MD61 Preview", interactive=False)
        md61_file   = gr.File(label="Download CSV", interactive=False)

        # Handlers
        def do_forecast(h, demo_flag, f):
            hist_path = "" if (f is None) else f.name
            fj = forecast_tool(horizon_months=int(h), use_demo=(f is None) and bool(demo_flag),
                               history_csv_path=hist_path)
            df = parse_forecast(fj)
            return fj, df, f"Forecast generated for {df['Product'].nunique()} product(s)."

        run_f.click(do_forecast, inputs=[horizon, use_demo, file], outputs=[forecast_state, forecast_tbl, forecast_note])

        def do_optimize(fj):
            if not fj:
                return "", pd.DataFrame(), pd.DataFrame(), pd.DataFrame(), pd.DataFrame(), "⚠️ Run Step 1 first."
            pj = optimize_supply_tool(fj)
            status, kpis, prod, raw, res = parse_plan(pj)
            return pj, kpis, prod, raw, res, f"Optimization status: **{status}**"

        run_o.click(do_optimize, inputs=[forecast_state],
                    outputs=[plan_state, plan_kpis, plan_prod, plan_raw, plan_res, plan_status])

        def do_md61(fj, plant):
            if not fj:
                return "", pd.DataFrame(), None, "⚠️ Run Step 1 first."
            mj = update_sap_md61_tool(fj, plant=plant, uom="EA")
            status, prev, path = parse_md61(mj)
            return mj, prev, path, f"MD61 status: **{status}**"

        run_m.click(do_md61, inputs=[forecast_state, plant], outputs=[md61_state, md61_prev, md61_file, md61_status])

    with gr.Tab("Agentic (End-to-end)"):
        gr.Markdown("One click: the agent runs all three steps with OpenAI.")
        with gr.Row():
            a_horizon = gr.Number(label="Horizon (months)", value=1, precision=0)
            a_plant   = gr.Textbox(label="SAP Plant (WERKS)", value="PLANT01")
        with gr.Row():
            a_demo = gr.Checkbox(label="Use demo synthetic history", value=True)
            a_file = gr.File(label="Or upload history.csv", file_types=[".csv"])
        run_all = gr.Button("Run End-to-end (Agent)")
        out_json = gr.Textbox(label="Agent Raw JSON (for inspection)", lines=6)
        with gr.Accordion("Pretty Outputs", open=True):
            a_forecast_tbl = gr.Dataframe(label="Forecast", interactive=False)
            a_plan_kpis = gr.Dataframe(label="KPIs", interactive=False)
            a_plan_prod = gr.Dataframe(label="Products Plan", interactive=False)
            a_plan_raw  = gr.Dataframe(label="Raw Materials", interactive=False)
            a_plan_res  = gr.Dataframe(label="Resources", interactive=False)
            a_md61_prev = gr.Dataframe(label="MD61 Preview", interactive=False)
            a_md61_file = gr.File(label="Download MD61 CSV", interactive=False)

        def do_agent(h, p, demo_flag, f):
            def to_obj(x):
                return x if isinstance(x, (dict, list)) else json.loads(x)
            def to_str(x):
                return x if isinstance(x, str) else json.dumps(x)

            try:
                res = run_agentic(h, p, demo_flag, f)   # may be dict or str
                out = to_obj(res)

                forecast_json = to_str(out["forecast"])
                plan_json     = to_str(out["plan"])
                md61_json     = to_str(out["md61"])

                f_df = parse_forecast(forecast_json)
                _, kpis, prod, raw, res_tbl = parse_plan(plan_json)
                _, prev, csv_path = parse_md61(md61_json)

                pretty = {
                    "forecast": json.loads(forecast_json),
                    "plan": json.loads(plan_json),
                    "md61": json.loads(md61_json),
                }
                return (json.dumps(pretty, indent=2), f_df, kpis, prod, raw, res_tbl, prev, csv_path)
            except Exception as e:
                return (f"Agent error: {e}", pd.DataFrame(), pd.DataFrame(), pd.DataFrame(), pd.DataFrame(),
                        pd.DataFrame(), pd.DataFrame(), None)

        run_all.click(do_agent, inputs=[a_horizon, a_plant, a_demo, a_file],
                      outputs=[out_json, a_forecast_tbl, a_plan_kpis, a_plan_prod, a_plan_raw, a_plan_res, a_md61_prev, a_md61_file])

    # --------- Upgrades ---------
    with gr.Tab("Upgrades"):
        gr.Markdown("### 🧹 Data Quality Guard")
        with gr.Row():
            dq_use_demo = gr.Checkbox(label="Use demo synthetic history", value=True)
            dq_file = gr.File(label="Or upload history.csv", file_types=[".csv"])
            dq_z = gr.Number(label="Robust z-threshold", value=3.5)
        run_dq = gr.Button("Run Data Quality Guard")
        dq_tbl = gr.Dataframe(label="Issues (flags & suggestions)", interactive=False)

        def do_dq(demo_flag, f, zval):
            hist_path = "" if (f is None) else f.name
            out = data_quality_guard_tool(use_demo=(f is None) and bool(demo_flag),
                                          history_csv_path=hist_path, z=float(zval))
            return parse_data_quality(out)

        run_dq.click(do_dq, inputs=[dq_use_demo, dq_file, dq_z], outputs=[dq_tbl])

        gr.Markdown("### 🔀 Scenario Explorer")
        run_sc = gr.Button("Run Scenarios on current Forecast (Step 1)")
        sc_tbl = gr.Dataframe(label="Scenario Outcomes", interactive=False)

        def do_sc(fj):
            if not fj:
                return pd.DataFrame()
            out = scenario_explorer_tool(fj)
            return parse_scenarios(out)

        run_sc.click(do_sc, inputs=[forecast_state], outputs=[sc_tbl])

        gr.Markdown("### 📝 Plan Explainer")
        run_ex = gr.Button("Explain Current Plan (Step 2)")
        expl_md = gr.Markdown("")

        def do_explain(pj):
            if not pj:
                return "⚠️ Run Step 2 first."
            out = plan_explainer_tool(pj)
            return json.loads(out)["summary"]

        run_ex.click(do_explain, inputs=[plan_state], outputs=[expl_md])

        gr.Markdown("### 🎯 Bottleneck Finder (Practical Levers)")
        policy_box = gr.Textbox(
            label="Optional policy JSON (overtime/expedite limits & costs). Leave blank for sensible defaults.",
            lines=4, value=""
        )
        run_bn = gr.Button("Find Best Bottleneck (uses current Forecast)")
        bn_tbl = gr.Dataframe(label="Ranked Recommendations", interactive=False)

        def do_bn(fj, policy):
            if not fj:
                return pd.DataFrame()
            out = bottleneck_search_tool(fj, policy_json=policy or "")
            return parse_recs(out)

        run_bn.click(do_bn, inputs=[forecast_state, policy_box], outputs=[bn_tbl])

if __name__ == "__main__":
    # Needs OPENAI_API_KEY in env for agent tab; manual tabs work without it.
    if not os.environ.get("OPENAI_API_KEY"):
        print("⚠️  Set OPENAI_API_KEY (Space secret) to use Agentic tab.")
    demo.launch()