"""`pred_label=None` server crash repro for the Lexsi platform team.

Reproduces the bug documented in `docs/sdk_issues.md` §2:
`ProjectConfig.pred_label` is typed `Optional[str]` in the SDK, but the
server-side training pipeline does direct subscript access on the key
and crashes on `None`. Activity Log surfaces a bare `'pred_label'` line
followed by `#05-013` (XAI step) or `#05-001` (model build step)
depending on which handler runs first.

Failure mode:
  Activity Log →  Started building TabICL_v1 model
                  Building Model
                  'pred_label'
                  Failed while building model. #05-001
  Client-side  →  Exception("") (empty message) from `upload_data`

Workaround (what the agent does):
  Set `pred_label="Prediction"` unconditionally in the ProjectConfig.

Dataset: `data/context_df.csv` — same 682-row PKDD binary
classification training set used by the TabICL repro.

How to run:

    export SDK_ACCESS_TOKEN=<your token>
    export LEXSI_ORG_NAME=<org, default "personal">
    export LEXSI_WORKSPACE_NAME=<workspace>
    uv run python scripts/repro_tabicl/repro_pred_label_none.py
"""

from __future__ import annotations

import os
import sys
import time
from datetime import datetime
from pathlib import Path

import pandas as pd

HERE = Path(__file__).resolve().parent
CONTEXT_CSV = HERE / "data" / "context_df.csv"


def _require_env() -> tuple[str, str, str]:
    token = os.environ.get("SDK_ACCESS_TOKEN")
    org = os.environ.get("LEXSI_ORG_NAME", "personal")
    ws = os.environ.get("LEXSI_WORKSPACE_NAME")
    if not token or not ws:
        print(
            "error: set SDK_ACCESS_TOKEN and LEXSI_WORKSPACE_NAME (and "
            "LEXSI_ORG_NAME if not 'personal').",
            file=sys.stderr,
        )
        sys.exit(2)
    return token, org, ws


def main() -> None:
    token, org_name, ws_name = _require_env()

    from lexsi_sdk import xai as lexsi  # type: ignore[import-not-found]

    print(f"# Lexsi pred_label=None repro — {datetime.now().isoformat(timespec='seconds')}")
    print(f"# org={org_name!r} workspace={ws_name!r}")

    lexsi.login(sdk_access_token=token)
    org = lexsi.organization(org_name)
    ws = org.workspace(ws_name)

    proj_name = f"repronone{int(time.time()) % 1_000_000}"
    print(f"\n# creating fresh tabular project: {proj_name}")
    project = None
    last_err: Exception | None = None
    for attempt in range(1, 4):
        try:
            project = ws.create_project(
                project_name=proj_name,
                modality="tabular",
                project_type="classification",
            )
            break
        except Exception as e:
            last_err = e
            print(f"  create_project attempt {attempt}/3 failed: {e}")
            time.sleep(2)
    if project is None:
        raise RuntimeError(f"create_project failed after 3 attempts: {last_err}")
    print(f"# project ready: {proj_name}")

    context_df = pd.read_csv(CONTEXT_CSV)
    train_tag = "reprotrain"
    print(f"\n# training: {context_df.shape} target=y_default tag={train_tag!r}")

    
    
    config = {
        "unique_identifier": "loan_id",
        "true_label": "y_default",
        "tag": train_tag,
        "model_name": "TabICL",
        "pred_label": None,
        "feature_exclude": [],
        "feature_encodings": {},
        "drop_duplicate_uid": True,
        "handle_errors": True,
        "handle_data_imbalance": False,
        "sample_percentage": None,
        "xai_method": [],
    }

    print("\n# uploading training data with pred_label=None — expect server "
          "crash visible in Activity Log as `'pred_label'` + #05-001 or #05-013")
    try:
        project.upload_data(
            data=context_df,
            tag=train_tag,
            config=config,
            compute_type="T4.small",
        )
        print("# (unexpected) upload_data returned without raising — "
              "check Activity Log; the model build may still fail asynchronously.")
    except Exception as e:
        print(f"# upload_data raised: {type(e).__name__}: {e!r}")

    print(f"\n# done. project left behind for inspection: {proj_name}")
    print("# delete with: project.delete_project()")


if __name__ == "__main__":
    main()