synth-ai 0.2.9.dev11__py3-none-any.whl → 0.4.1__py3-none-any.whl

synth_ai/sdk/api/train/builders.py

@@ -0,0 +1,895 @@
+from __future__ import annotations
+
+import importlib
+import os
+from collections.abc import Callable
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, cast
+
+import click
+from pydantic import ValidationError
+
+try:
+    _models_module = cast(
+        Any, importlib.import_module("synth_ai.sdk.api.models.supported")
+    )
+    UnsupportedModelError = cast(type[Exception], _models_module.UnsupportedModelError)
+    ensure_allowed_model = cast(
+        Callable[..., None], _models_module.ensure_allowed_model
+    )
+    normalize_model_identifier = cast(
+        Callable[[str], str], _models_module.normalize_model_identifier
+    )
+except Exception as exc:  # pragma: no cover - critical dependency
+    raise RuntimeError("Unable to load supported model helpers") from exc
+
+try:
+    _sft_module = cast(
+        Any, importlib.import_module("synth_ai.sdk.learning.sft.config")
+    )
+    prepare_sft_job_payload = cast(
+        Callable[..., dict[str, Any]], _sft_module.prepare_sft_job_payload
+    )
+except Exception as exc:  # pragma: no cover - critical dependency
+    raise RuntimeError("Unable to load SFT payload helpers") from exc
+
+from synth_ai.core.config.resolver import ConfigResolver
+from synth_ai.core.telemetry import log_info
+
+from .configs import PromptLearningConfig, RLConfig, SFTConfig
+from .supported_algos import (
+    AlgorithmValidationError,
+    ensure_model_supported_for_algorithm,
+    validate_algorithm_config,
+)
+from .utils import TrainError, ensure_api_base
+
+
+@dataclass(slots=True)
+class RLBuildResult:
+    payload: dict[str, Any]
+    task_url: str
+    idempotency: str | None
+
+
+@dataclass(slots=True)
+class SFTBuildResult:
+    payload: dict[str, Any]
+    train_file: Path
+    validation_file: Path | None
+
+
+@dataclass(slots=True)
+class PromptLearningBuildResult:
+    payload: dict[str, Any]
+    task_url: str
+
+
+def _format_validation_error(path: Path, exc: ValidationError) -> str:
+    lines: list[str] = []
+    for error in exc.errors():
+        loc = ".".join(str(part) for part in error.get("loc", ()))
+        msg = error.get("msg", "invalid value")
+        lines.append(f"{loc or '<root>'}: {msg}")
+    details = "\n".join(f"  - {line}" for line in lines) or "  - Invalid configuration"
+    return f"Config validation failed ({path}):\n{details}"
+
+
+def build_rl_payload(
+    *,
+    config_path: Path,
+    task_url: str,
+    overrides: dict[str, Any],
+    idempotency: str | None,
+    allow_experimental: bool | None = None,
+) -> RLBuildResult:
+    ctx: dict[str, Any] = {"config_path": str(config_path), "task_url": task_url}
+    log_info("build_rl_payload invoked", ctx=ctx)
+    # Load and validate config with SDK-level checks
+    from synth_ai.cli.commands.train.validation import validate_rl_config
+    from synth_ai.sdk.api.train.utils import load_toml
+    
+    try:
+        raw_config = load_toml(config_path)
+        validated_config = validate_rl_config(raw_config)  # Adds defaults & validates
+        rl_cfg = RLConfig.from_mapping(validated_config)
+    except ValidationError as exc:
+        raise click.ClickException(_format_validation_error(config_path, exc)) from exc
+
+    data = rl_cfg.to_dict()
+    
+    # Remove smoke section - it's CLI-only and should not be sent to the trainer
+    if "smoke" in data:
+        del data["smoke"]
+    
+    # Ensure required [reference] section for backend validators
+    try:
+        ref_cfg = data.get("reference") if isinstance(data, dict) else None
+        if not isinstance(ref_cfg, dict):
+            data["reference"] = {"placement": "none"}
+        else:
+            ref_cfg.setdefault("placement", "none")
+    except Exception:
+        # Defensive: never fail builder due to optional defaults
+        data["reference"] = {"placement": "none"}
+    try:
+        spec = validate_algorithm_config(
+            rl_cfg.algorithm.model_dump(), expected_family="rl"
+        )
+    except AlgorithmValidationError as exc:
+        raise click.ClickException(str(exc)) from exc
+    services = data.get("services") if isinstance(data.get("services"), dict) else {}
+    model_cfg = rl_cfg.model
+
+    cli_task_url = overrides.get("task_url")
+    env_task_url = task_url or os.environ.get("TASK_APP_URL")
+    config_task_url = services.get("task_url") if isinstance(services, dict) else None
+    final_task_url = ConfigResolver.resolve(
+        "task_app_url",
+        cli_value=cli_task_url,
+        env_value=env_task_url,
+        config_value=config_task_url,
+        required=True,
+    )
+    assert final_task_url is not None  # required=True guarantees non-None
+
+    model_source = (model_cfg.source or "").strip() if model_cfg else ""
+    model_base = (model_cfg.base or "").strip() if model_cfg else ""
+    override_model = (overrides.get("model") or "").strip()
+    if override_model:
+        model_source = override_model
+        model_base = ""
+    if bool(model_source) == bool(model_base):
+        details = (
+            f"Config: {config_path}\n"
+            f"[model].source={model_source!r} | [model].base={model_base!r}"
+        )
+        hint = (
+            "Set exactly one: [model].base for a base model (e.g. 'Qwen/Qwen3-1.7B') "
+            "or [model].source for a fine-tuned model id. Also remove any conflicting "
+            "'[policy].model' entries."
+        )
+        raise click.ClickException(
+            "Invalid model config: exactly one of [model].source or [model].base is required.\n"
+            + details
+            + "\nHint: "
+            + hint
+        )
+
+    try:
+        if model_source:
+            model_source = normalize_model_identifier(model_source)
+        if model_base:
+            model_base = normalize_model_identifier(model_base)
+    except UnsupportedModelError as exc:
+        raise click.ClickException(str(exc)) from exc
+
+    base_model_for_training: str | None = None
+    if model_source:
+        base_model_for_training = ensure_allowed_model(
+            model_source,
+            allow_finetuned_prefixes=True,
+            allow_experimental=allow_experimental,
+        )
+    elif model_base:
+        base_model_for_training = ensure_allowed_model(
+            model_base,
+            allow_finetuned_prefixes=False,
+            allow_experimental=allow_experimental,
+        )
+    if base_model_for_training:
+        try:
+            ensure_model_supported_for_algorithm(base_model_for_training, spec)
+        except AlgorithmValidationError as exc:
+            raise click.ClickException(str(exc)) from exc
+
+    # Force TOML services.task_url to the effective endpoint to avoid split URLs
+    try:
+        if isinstance(data.get("services"), dict):
+            data["services"]["task_url"] = final_task_url
+        else:
+            data["services"] = {"task_url": final_task_url}
+    except Exception:
+        pass
+
+    payload_data: dict[str, Any] = {
+        "endpoint_base_url": final_task_url.rstrip("/"),
+        "config": data,
+    }
+    payload: dict[str, Any] = {
+        "job_type": "rl",
+        "compute": data.get("compute", {}),
+        "data": payload_data,
+        "tags": {"source": "train-cli"},
+    }
+    if model_source:
+        payload_data["model"] = model_source
+    if model_base:
+        payload_data["base_model"] = model_base
+
+    backend = overrides.get("backend")
+    if backend:
+        metadata_default: dict[str, Any] = {}
+        metadata = cast(dict[str, Any], payload.setdefault("metadata", metadata_default))
+        metadata["backend_base_url"] = ensure_api_base(str(backend))
+
+    return RLBuildResult(payload=payload, task_url=final_task_url, idempotency=idempotency)
+
+
+def build_sft_payload(
+    *,
+    config_path: Path,
+    dataset_override: Path | None,
+    allow_experimental: bool | None,
+) -> SFTBuildResult:
+    ctx: dict[str, Any] = {"config_path": str(config_path), "dataset_override": str(dataset_override) if dataset_override else None}
+    log_info("build_sft_payload invoked", ctx=ctx)
+    try:
+        sft_cfg = SFTConfig.from_path(config_path)
+    except ValidationError as exc:
+        raise TrainError(_format_validation_error(config_path, exc)) from exc
+
+    data = sft_cfg.to_dict()
+    try:
+        algo_mapping = sft_cfg.algorithm.model_dump() if sft_cfg.algorithm else None
+        spec = validate_algorithm_config(algo_mapping, expected_family="sft")
+    except AlgorithmValidationError as exc:
+        raise TrainError(str(exc)) from exc
+    data_cfg = data.get("data") if isinstance(data.get("data"), dict) else {}
+    hp_cfg = data.get("hyperparameters") if isinstance(data.get("hyperparameters"), dict) else {}
+    train_cfg = data.get("training") if isinstance(data.get("training"), dict) else {}
+    compute_cfg = data.get("compute") if isinstance(data.get("compute"), dict) else {}
+
+    raw_dataset = dataset_override or sft_cfg.job.data or sft_cfg.job.data_path
+    if not raw_dataset:
+        raise TrainError("Dataset not specified; pass --dataset or set [job].data")
+    dataset_path = Path(raw_dataset)
+    # Resolve relative paths from current working directory, not config directory
+    dataset_path = (
+        dataset_path if dataset_path.is_absolute() else (Path.cwd() / dataset_path)
+    ).resolve()
+    if not dataset_path.exists():
+        raise TrainError(f"Dataset not found: {dataset_path}")
+
+    validation_path = (
+        data_cfg.get("validation_path")
+        if isinstance(data_cfg, dict)
+        else None
+        if isinstance(data_cfg, dict) and isinstance(data_cfg.get("validation_path"), str)
+        else None
+    )
+    validation_file = None
+    if validation_path:
+        vpath = Path(validation_path)
+        # Resolve relative paths from current working directory, not config directory
+        vpath = (vpath if vpath.is_absolute() else (Path.cwd() / vpath)).resolve()
+        if not vpath.exists():
+            click.echo(f"[WARN] Validation dataset {vpath} missing; continuing without validation")
+        else:
+            validation_file = vpath
+
+    hp_block: dict[str, Any] = {
+        "n_epochs": int(hp_cfg.get("n_epochs", 1) if isinstance(hp_cfg, dict) else 1),
+    }
+    for key in (
+        "batch_size",
+        "global_batch",
+        "per_device_batch",
+        "gradient_accumulation_steps",
+        "sequence_length",
+        "learning_rate",
+        "warmup_ratio",
+        "train_kind",
+    ):
+        if isinstance(hp_cfg, dict) and key in hp_cfg:
+            hp_block[key] = hp_cfg[key]
+    if isinstance(hp_cfg, dict) and isinstance(hp_cfg.get("parallelism"), dict):
+        hp_block["parallelism"] = hp_cfg["parallelism"]
+
+    compute_block = {
+        k: compute_cfg[k]
+        for k in ("gpu_type", "gpu_count", "nodes")
+        if isinstance(compute_cfg, dict) and k in compute_cfg
+    }
+
+    effective = {
+        "compute": compute_block,
+        "data": {
+            "topology": data_cfg.get("topology", {})
+            if isinstance(data_cfg, dict) and isinstance(data_cfg.get("topology"), dict)
+            else {}
+        },
+        "training": {
+            k: v
+            for k, v in (train_cfg.items() if isinstance(train_cfg, dict) else [])
+            if k in ("mode", "use_qlora")
+        },
+    }
+
+    validation_cfg = (
+        train_cfg.get("validation")
+        if isinstance(train_cfg, dict) and isinstance(train_cfg.get("validation"), dict)
+        else None
+    )
+    if isinstance(validation_cfg, dict):
+        hp_block.update(
+            {
+                "evaluation_strategy": validation_cfg.get("evaluation_strategy", "steps"),
+                "eval_steps": int(validation_cfg.get("eval_steps", 0) or 0),
+                "save_best_model_at_end": bool(validation_cfg.get("save_best_model_at_end", True)),
+                "metric_for_best_model": validation_cfg.get("metric_for_best_model", "val.loss"),
+                "greater_is_better": bool(validation_cfg.get("greater_is_better", False)),
+            }
+        )
+        effective.setdefault("training", {})["validation"] = {
+            "enabled": bool(validation_cfg.get("enabled", True))
+        }
+
+    raw_model = (sft_cfg.job.model or "").strip()
+    if not raw_model:
+        model_block = data.get("model")
+        if isinstance(model_block, str):
+            raw_model = model_block.strip()
+    if not raw_model:
+        raise TrainError("Model not specified; set [job].model or [model].base in the config")
+
+    try:
+        base_model = ensure_allowed_model(
+            raw_model,
+            allow_finetuned_prefixes=False,
+            allow_experimental=allow_experimental,
+        )
+    except UnsupportedModelError as exc:
+        raise TrainError(str(exc)) from exc
+    
+    if base_model:
+        try:
+            ensure_model_supported_for_algorithm(base_model, spec)
+        except AlgorithmValidationError as exc:
+            raise TrainError(str(exc)) from exc
+
+    try:
+        payload = prepare_sft_job_payload(
+            model=raw_model,
+            training_file=None,
+            hyperparameters=hp_block,
+            metadata={"effective_config": effective},
+            training_type="sft_offline",
+            training_file_field="training_file_id",
+            require_training_file=False,
+            include_training_file_when_none=True,
+            allow_finetuned_prefixes=False,
+        )
+    except UnsupportedModelError as exc:
+        raise TrainError(str(exc)) from exc
+    except ValueError as exc:
+        raise TrainError(str(exc)) from exc
+
+    return SFTBuildResult(payload=payload, train_file=dataset_path, validation_file=validation_file)
+
+
+def build_prompt_learning_payload(
+    *,
+    config_path: Path,
+    task_url: str | None,
+    overrides: dict[str, Any],
+    allow_experimental: bool | None = None,
+) -> PromptLearningBuildResult:
+    """Build payload for prompt learning job (MIPRO or GEPA)."""
+    ctx: dict[str, Any] = {"config_path": str(config_path), "task_url": task_url}
+    log_info("build_prompt_learning_payload invoked", ctx=ctx)
+    from pydantic import ValidationError
+
+    from .configs.prompt_learning import load_toml
+
+    # SDK-SIDE VALIDATION: Catch errors BEFORE sending to backend
+    from .validators import validate_prompt_learning_config
+    
+    raw_config = load_toml(config_path)
+    validate_prompt_learning_config(raw_config, config_path)
+    
+    try:
+        pl_cfg = PromptLearningConfig.from_path(config_path)
+    except ValidationError as exc:
+        raise click.ClickException(_format_validation_error(config_path, exc)) from exc
+    
+    # Early validation: Check required fields for GEPA
+    if pl_cfg.algorithm == "gepa":
+        if not pl_cfg.gepa:
+            raise click.ClickException(
+                "GEPA config missing: [prompt_learning.gepa] section is required"
+            )
+        if not pl_cfg.gepa.evaluation:
+            raise click.ClickException(
+                "GEPA config missing: [prompt_learning.gepa.evaluation] section is required"
+            )
+        train_seeds = getattr(pl_cfg.gepa.evaluation, "train_seeds", None) or getattr(pl_cfg.gepa.evaluation, "seeds", None)
+        if not train_seeds:
+            raise click.ClickException(
+                "GEPA config missing train_seeds: [prompt_learning.gepa.evaluation] must have 'train_seeds' or 'seeds' field"
+            )
+        val_seeds = getattr(pl_cfg.gepa.evaluation, "val_seeds", None) or getattr(pl_cfg.gepa.evaluation, "validation_seeds", None)
+        if not val_seeds:
+            raise click.ClickException(
+                "GEPA config missing val_seeds: [prompt_learning.gepa.evaluation] must have 'val_seeds' or 'validation_seeds' field"
+            )
+    
+    cli_task_url = overrides.get("task_url") or task_url
+    env_task_url = os.environ.get("TASK_APP_URL")
+    config_task_url = (pl_cfg.task_app_url or "").strip() or None
+    
+    # For prompt learning, prefer config value over env if config is explicitly set
+    # This allows TOML files to specify task_app_url without env var interference
+    # But CLI override always wins
+    if cli_task_url:
+        # CLI override takes precedence
+        final_task_url = ConfigResolver.resolve(
+            "task_app_url",
+            cli_value=cli_task_url,
+            env_value=None,  # Don't check env when CLI is set
+            config_value=config_task_url,
+            required=True,
+        )
+    elif config_task_url:
+        # Config explicitly set - use it (ignore env var to avoid conflicts)
+        final_task_url = config_task_url
+    else:
+        # No config, fall back to env or error
+        final_task_url = ConfigResolver.resolve(
+            "task_app_url",
+            cli_value=None,
+            env_value=env_task_url,
+            config_value=None,
+            required=True,
+        )
+    assert final_task_url is not None  # required=True guarantees non-None
+    
+    # Get task_app_api_key from config or environment
+    config_api_key = (pl_cfg.task_app_api_key or "").strip() or None
+    cli_api_key = overrides.get("task_app_api_key")
+    env_api_key = os.environ.get("ENVIRONMENT_API_KEY")
+    task_app_api_key = ConfigResolver.resolve(
+        "task_app_api_key",
+        cli_value=cli_api_key,
+        env_value=env_api_key,
+        config_value=config_api_key,
+        required=True,
+    )
+    
+    # Build config dict for backend
+    config_dict = pl_cfg.to_dict()
+    
+    # ASSERT: MIPRO fields exist in Pydantic model
+    if pl_cfg.algorithm == "mipro":
+        assert pl_cfg.mipro is not None, "pl_cfg.mipro is None for MIPRO algorithm"
+        bootstrap_in_model = getattr(pl_cfg.mipro, "bootstrap_train_seeds", None)
+        online_in_model = getattr(pl_cfg.mipro, "online_pool", None)
+        assert bootstrap_in_model is not None, f"pl_cfg.mipro.bootstrap_train_seeds is None! pl_cfg.mipro keys: {dir(pl_cfg.mipro) if pl_cfg.mipro else 'N/A'}"
+        assert online_in_model is not None, f"pl_cfg.mipro.online_pool is None! pl_cfg.mipro keys: {dir(pl_cfg.mipro) if pl_cfg.mipro else 'N/A'}"
+    
+    # Ensure task_app_url and task_app_api_key are set
+    pl_section = config_dict.get("prompt_learning", {})
+    if isinstance(pl_section, dict):
+        pl_section["task_app_url"] = final_task_url
+        pl_section["task_app_api_key"] = task_app_api_key
+        
+        # GEPA: Extract train_seeds from nested structure for backwards compatibility
+        # Backend checks for train_seeds at top level before parsing nested structure
+        if pl_cfg.algorithm == "gepa" and pl_cfg.gepa:
+            # Try to get train_seeds directly from the gepa config object first
+            train_seeds = None
+            if pl_cfg.gepa.evaluation:
+                train_seeds = getattr(pl_cfg.gepa.evaluation, "train_seeds", None) or getattr(pl_cfg.gepa.evaluation, "seeds", None)
+            
+            # If not found, try from serialized dict
+            if not train_seeds:
+                gepa_section = pl_section.get("gepa", {})
+                # Handle case where gepa_section might still be a Pydantic model
+                if hasattr(gepa_section, "model_dump"):
+                    gepa_section = gepa_section.model_dump(mode="python")
+                elif hasattr(gepa_section, "dict"):
+                    gepa_section = gepa_section.dict()
+                
+                if isinstance(gepa_section, dict):
+                    eval_section = gepa_section.get("evaluation", {})
+                    # Handle case where eval_section might still be a Pydantic model
+                    if hasattr(eval_section, "model_dump"):
+                        eval_section = eval_section.model_dump(mode="python")
+                    elif hasattr(eval_section, "dict"):
+                        eval_section = eval_section.dict()
+                    
+                    if isinstance(eval_section, dict):
+                        train_seeds = eval_section.get("train_seeds") or eval_section.get("seeds")
+                    
+                    # Update gepa_section back to pl_section in case we converted it
+                    pl_section["gepa"] = gepa_section
+            
+            # Add train_seeds to top level for backwards compatibility
+            if train_seeds and not pl_section.get("train_seeds"):
+                pl_section["train_seeds"] = train_seeds
+            if train_seeds and not pl_section.get("evaluation_seeds"):
+                pl_section["evaluation_seeds"] = train_seeds
+        
+        # MIPRO: CRITICAL - Ensure bootstrap_train_seeds and online_pool are ALWAYS in mipro section
+        # Handle Pydantic model serialization - mipro might be a model object, not a dict
+        if pl_cfg.algorithm == "mipro":
+            mipro_section = pl_section.get("mipro", {})
+            
+            # ASSERT: Check what we got from to_dict()
+            assert pl_cfg.mipro is not None, "pl_cfg.mipro is None"
+            bootstrap_before_convert = getattr(pl_cfg.mipro, "bootstrap_train_seeds", None)
+            online_before_convert = getattr(pl_cfg.mipro, "online_pool", None)
+            assert bootstrap_before_convert is not None, f"bootstrap_train_seeds is None in model! mipro_section type: {type(mipro_section)}, keys: {list(mipro_section.keys()) if isinstance(mipro_section, dict) else 'N/A'}"
+            assert online_before_convert is not None, f"online_pool is None in model! mipro_section type: {type(mipro_section)}, keys: {list(mipro_section.keys()) if isinstance(mipro_section, dict) else 'N/A'}"
+            
+            # Convert Pydantic model to dict if needed
+            if hasattr(mipro_section, "model_dump"):
+                mipro_section = mipro_section.model_dump(mode="python")
+            elif hasattr(mipro_section, "dict"):
+                mipro_section = mipro_section.dict()
+            
+            if not isinstance(mipro_section, dict):
+                mipro_section = {}
+            
+            # ASSERT: After conversion, check if fields are present
+            bootstrap_after_convert = mipro_section.get("bootstrap_train_seeds")
+            online_after_convert = mipro_section.get("online_pool")
+            if bootstrap_after_convert is None:
+                raise AssertionError(f"bootstrap_train_seeds missing after conversion! mipro_section keys: {list(mipro_section.keys())}, bootstrap_before_convert: {bootstrap_before_convert}")
+            if online_after_convert is None:
+                raise AssertionError(f"online_pool missing after conversion! mipro_section keys: {list(mipro_section.keys())}, online_before_convert: {online_before_convert}")
+            
+            # CRITICAL: Get fields from Pydantic model FIRST (most reliable)
+            # These fields MUST be present - get them from the source of truth
+            if not pl_cfg.mipro:
+                raise ValueError(
+                    "MIPRO config missing: pl_cfg.mipro is None. "
+                    "Ensure [prompt_learning.mipro] section exists in TOML."
+                )
+            
+            bootstrap_from_model = getattr(pl_cfg.mipro, "bootstrap_train_seeds", None)
+            online_from_model = getattr(pl_cfg.mipro, "online_pool", None)
+            test_from_model = getattr(pl_cfg.mipro, "test_pool", None)
+            reference_from_model = getattr(pl_cfg.mipro, "reference_pool", None)
+            
+            # FORCE these fields into mipro_section (model is source of truth)
+            # Use model values if present, otherwise keep existing dict values
+            assert bootstrap_from_model is not None, f"bootstrap_from_model is None! pl_cfg.mipro: {pl_cfg.mipro}"
+            assert online_from_model is not None, f"online_from_model is None! pl_cfg.mipro: {pl_cfg.mipro}"
+            
+            mipro_section["bootstrap_train_seeds"] = bootstrap_from_model
+            mipro_section["online_pool"] = online_from_model
+            
+            if test_from_model is not None:
+                mipro_section["test_pool"] = test_from_model
+            elif not mipro_section.get("test_pool") and pl_section.get("test_pool"):
+                mipro_section["test_pool"] = pl_section["test_pool"]
+            
+            if reference_from_model is not None:
+                mipro_section["reference_pool"] = reference_from_model
+            elif not mipro_section.get("reference_pool") and pl_section.get("reference_pool"):
+                mipro_section["reference_pool"] = pl_section["reference_pool"]
+            
+            # ASSERT: Fields are now in mipro_section
+            assert mipro_section.get("bootstrap_train_seeds") is not None, f"bootstrap_train_seeds STILL missing after forcing! mipro_section keys: {list(mipro_section.keys())}"
+            assert mipro_section.get("online_pool") is not None, f"online_pool STILL missing after forcing! mipro_section keys: {list(mipro_section.keys())}"
+            
+            # CRITICAL: Validate fields are present BEFORE override merge
+            # If they're missing here, we'll check overrides after they're extracted
+            # For now, just ensure they're in mipro_section from TOML
+            if not mipro_section.get("bootstrap_train_seeds") and bootstrap_from_model is None:
+                raise ValueError(
+                    f"MIPRO config missing bootstrap_train_seeds in TOML. "
+                    f"pl_cfg.mipro.bootstrap_train_seeds={bootstrap_from_model}, "
+                    f"mipro_section keys={list(mipro_section.keys())}, "
+                    f"pl_section keys={list(pl_section.keys())[:10]}. "
+                    f"Ensure [prompt_learning.mipro] has bootstrap_train_seeds or provide override."
+                )
+            if not mipro_section.get("online_pool") and online_from_model is None:
+                raise ValueError(
+                    f"MIPRO config missing online_pool in TOML. "
+                    f"pl_cfg.mipro.online_pool={online_from_model}, "
+                    f"mipro_section keys={list(mipro_section.keys())}, "
+                    f"pl_section keys={list(pl_section.keys())[:10]}. "
+                    f"Ensure [prompt_learning.mipro] has online_pool or provide override."
+                )
+            
+            # Extract env_name from mipro section to top-level (backend expects it there)
+            mipro_env_name = mipro_section.get("env_name")
+            if mipro_env_name and not pl_section.get("env_name") and not pl_section.get("task_app_id"):
+                pl_section["env_name"] = mipro_env_name
+            
+            # CRITICAL: Update mipro section back to config_dict IMMEDIATELY
+            # This ensures fields are present before override merge
+            pl_section["mipro"] = mipro_section
+            config_dict["prompt_learning"] = pl_section
+            
+            # ASSERT: Fields are in config_dict before override merge
+            assert config_dict.get("prompt_learning", {}).get("mipro", {}).get("bootstrap_train_seeds") is not None, \
+                f"bootstrap_train_seeds missing from config_dict before override merge! config_dict keys: {list(config_dict.keys())}"
+            assert config_dict.get("prompt_learning", {}).get("mipro", {}).get("online_pool") is not None, \
+                f"online_pool missing from config_dict before override merge! config_dict keys: {list(config_dict.keys())}"
+    else:
+        config_dict["prompt_learning"] = {
+            "task_app_url": final_task_url,
+            "task_app_api_key": task_app_api_key,
+        }
+    
+    # Build payload matching backend API format
+    # Extract nested overrides if present, otherwise use flat overrides directly
+    # The experiment queue passes flat overrides like {"prompt_learning.policy.model": "..."}
+    # But some SDK code passes nested like {"overrides": {"prompt_learning.policy.model": "..."}}
+    config_overrides = overrides.get("overrides", {}) if "overrides" in overrides else overrides
+    # Remove non-override keys (backend, task_url, metadata, auto_start)
+    config_overrides = {
+        k: v for k, v in config_overrides.items()
+        if k not in ("backend", "task_url", "metadata", "auto_start")
+    }
+    
+    # ASSERT: Check MIPRO fields BEFORE override merge
+    if pl_cfg.algorithm == "mipro":
+        pre_merge_mipro = config_dict.get("prompt_learning", {}).get("mipro", {})
+        assert pre_merge_mipro.get("bootstrap_train_seeds") is not None, \
+            f"bootstrap_train_seeds missing BEFORE override merge! pre_merge_mipro keys: {list(pre_merge_mipro.keys()) if isinstance(pre_merge_mipro, dict) else 'NOT DICT'}"
+        assert pre_merge_mipro.get("online_pool") is not None, \
+            f"online_pool missing BEFORE override merge! pre_merge_mipro keys: {list(pre_merge_mipro.keys()) if isinstance(pre_merge_mipro, dict) else 'NOT DICT'}"
+    
+    # CRITICAL: Merge overrides into config_dict BEFORE sending to backend
+    # This ensures early validation in backend sees merged values
+    # Use the same _deep_update logic from experiment_queue/config_utils.py
+    if config_overrides:
+        from synth_ai.cli.local.experiment_queue.config_utils import _deep_update
+        _deep_update(config_dict, config_overrides)
+        
+        # ASSERT: Check MIPRO fields AFTER override merge
+        if pl_cfg.algorithm == "mipro":
+            post_merge_mipro = config_dict.get("prompt_learning", {}).get("mipro", {})
+            assert post_merge_mipro.get("bootstrap_train_seeds") is not None, \
+                f"bootstrap_train_seeds missing AFTER override merge! post_merge_mipro keys: {list(post_merge_mipro.keys()) if isinstance(post_merge_mipro, dict) else 'NOT DICT'}, overrides: {list(config_overrides.keys())[:5]}"
+            assert post_merge_mipro.get("online_pool") is not None, \
+                f"online_pool missing AFTER override merge! post_merge_mipro keys: {list(post_merge_mipro.keys()) if isinstance(post_merge_mipro, dict) else 'NOT DICT'}, overrides: {list(config_overrides.keys())[:5]}"
+    
+    # After merging overrides (or if no overrides), re-run MIPRO reorganization to ensure
+    # fields like bootstrap_train_seeds and online_pool are in the right place
+    # This must run AFTER override merge so we catch fields from overrides
+    pl_section_in_dict = config_dict.get("prompt_learning", {})
+    if pl_cfg.algorithm == "mipro" and isinstance(pl_section_in_dict, dict):
+        mipro_section = pl_section_in_dict.get("mipro", {})
+        if not isinstance(mipro_section, dict):
+            mipro_section = {}
+        
+        # CRITICAL: After override merge, ensure fields are in mipro section
+        # Check both top-level and nested locations (overrides may have added them anywhere)
+        if not mipro_section.get("bootstrap_train_seeds") and pl_section_in_dict.get("bootstrap_train_seeds"):
+            mipro_section["bootstrap_train_seeds"] = pl_section_in_dict["bootstrap_train_seeds"]
+        
+        if not mipro_section.get("online_pool") and pl_section_in_dict.get("online_pool"):
+            mipro_section["online_pool"] = pl_section_in_dict["online_pool"]
+        
+        if not mipro_section.get("test_pool") and pl_section_in_dict.get("test_pool"):
+            mipro_section["test_pool"] = pl_section_in_dict["test_pool"]
+        
+        if not mipro_section.get("reference_pool") and pl_section_in_dict.get("reference_pool"):
+            mipro_section["reference_pool"] = pl_section_in_dict["reference_pool"]
+        
+        # ASSERT: Fields should be in mipro_section after reorganization
+        assert mipro_section.get("bootstrap_train_seeds") is not None, \
+            f"bootstrap_train_seeds missing after reorganization! mipro_section keys: {list(mipro_section.keys())}, pl_section keys: {list(pl_section_in_dict.keys())[:10]}"
+        assert mipro_section.get("online_pool") is not None, \
+            f"online_pool missing after reorganization! mipro_section keys: {list(mipro_section.keys())}, pl_section keys: {list(pl_section_in_dict.keys())[:10]}"
+        
+        # Update mipro section back to config_dict
+        pl_section_in_dict["mipro"] = mipro_section
+        config_dict["prompt_learning"] = pl_section_in_dict
+        
+        # ASSERT: Fields are in config_dict after reorganization
+        assert config_dict.get("prompt_learning", {}).get("mipro", {}).get("bootstrap_train_seeds") is not None, \
+            "bootstrap_train_seeds missing from config_dict after reorganization!"
+        assert config_dict.get("prompt_learning", {}).get("mipro", {}).get("online_pool") is not None, \
+            "online_pool missing from config_dict after reorganization!"
+        
+        # CRITICAL: Verify MIPRO required fields are present after merge
+        # This is the final check before sending to backend (runs whether or not overrides were applied)
+        final_mipro_section = pl_section_in_dict.get("mipro", {})
+        
+        # DEBUG: Log what we found for troubleshooting
+        bootstrap_seeds = final_mipro_section.get("bootstrap_train_seeds")
+        online_pool_val = final_mipro_section.get("online_pool")
+        
+        if not bootstrap_seeds or not online_pool_val:
+            # Log debug info before raising error
+            import logging
+            logger = logging.getLogger(__name__)
+            logger.error(
+                "MIPRO config validation failed:\n"
+                f"  bootstrap_train_seeds in mipro section: {bootstrap_seeds is not None}\n"
+                f"  online_pool in mipro section: {online_pool_val is not None}\n"
+                f"  mipro_section keys: {list(final_mipro_section.keys()) if isinstance(final_mipro_section, dict) else 'NOT A DICT'}\n"
+                f"  pl_section keys: {list(pl_section_in_dict.keys()) if isinstance(pl_section_in_dict, dict) else 'NOT A DICT'}\n"
+                f"  config_overrides keys: {list(config_overrides.keys()) if config_overrides else 'NO OVERRIDES'}"
+            )
+        
+        if not bootstrap_seeds:
+            raise ValueError(
+                "MIPRO config missing bootstrap_train_seeds after applying overrides. "
+                "Ensure it's set in TOML at [prompt_learning.mipro] or [prompt_learning] level, "
+                "or provided via override 'prompt_learning.mipro.bootstrap_train_seeds'. "
+                f"Current mipro_section keys: {list(final_mipro_section.keys()) if isinstance(final_mipro_section, dict) else 'NOT A DICT'}"
+            )
+        if not online_pool_val:
+            raise ValueError(
+                "MIPRO config missing online_pool after applying overrides. "
+                "Ensure it's set in TOML at [prompt_learning.mipro] or [prompt_learning] level, "
+                "or provided via override 'prompt_learning.mipro.online_pool'. "
+                f"Current mipro_section keys: {list(final_mipro_section.keys()) if isinstance(final_mipro_section, dict) else 'NOT A DICT'}"
+            )
+    
+    # CRITICAL: Final validation - ensure MIPRO fields are in config_body before sending
+    pl_section_final = config_dict.get("prompt_learning", {})
+    if pl_cfg.algorithm == "mipro":
+        mipro_final = pl_section_final.get("mipro", {})
+        if not isinstance(mipro_final, dict):
+            mipro_final = {}
+        
+        # If fields are still missing, FORCE them from the Pydantic model one last time
+        if pl_cfg.mipro:
+            if not mipro_final.get("bootstrap_train_seeds"):
+                bootstrap_val = getattr(pl_cfg.mipro, "bootstrap_train_seeds", None)
+                if bootstrap_val is not None:
+                    mipro_final["bootstrap_train_seeds"] = bootstrap_val
+            
+            if not mipro_final.get("online_pool"):
+                online_val = getattr(pl_cfg.mipro, "online_pool", None)
+                if online_val is not None:
+                    mipro_final["online_pool"] = online_val
+            
+            if not mipro_final.get("test_pool"):
+                test_val = getattr(pl_cfg.mipro, "test_pool", None)
+                if test_val is not None:
+                    mipro_final["test_pool"] = test_val
+            
+            if not mipro_final.get("reference_pool"):
+                ref_val = getattr(pl_cfg.mipro, "reference_pool", None)
+                if ref_val is not None:
+                    mipro_final["reference_pool"] = ref_val
+        
+        # Update back to config_dict
+        pl_section_final["mipro"] = mipro_final
+        config_dict["prompt_learning"] = pl_section_final
+        
+        # ASSERT: Fields are in config_dict after final check
+        assert config_dict.get("prompt_learning", {}).get("mipro", {}).get("bootstrap_train_seeds") is not None, \
+            f"bootstrap_train_seeds missing from config_dict after final check! mipro_final keys: {list(mipro_final.keys())}"
+        assert config_dict.get("prompt_learning", {}).get("mipro", {}).get("online_pool") is not None, \
+            f"online_pool missing from config_dict after final check! mipro_final keys: {list(mipro_final.keys())}"
+        
+        # FINAL ASSERTION: These fields MUST be present
+        assert mipro_final.get("bootstrap_train_seeds") is not None, \
+            f"CRITICAL: bootstrap_train_seeds missing from config_body! mipro_final keys: {list(mipro_final.keys())}, pl_cfg.mipro.bootstrap_train_seeds: {getattr(pl_cfg.mipro, 'bootstrap_train_seeds', None) if pl_cfg.mipro else 'N/A'}"
+        assert mipro_final.get("online_pool") is not None, \
+            f"CRITICAL: online_pool missing from config_body! mipro_final keys: {list(mipro_final.keys())}, pl_cfg.mipro.online_pool: {getattr(pl_cfg.mipro, 'online_pool', None) if pl_cfg.mipro else 'N/A'}"
+    
+    # ASSERT: Verify critical overrides are reflected in config_body
+    pl_section_in_dict = config_dict.get("prompt_learning", {})
+    if config_overrides:
+        # Check rollout budget override
+        rollout_budget_key = "prompt_learning.gepa.rollout.budget"
+        if rollout_budget_key in config_overrides:
+            expected_budget = config_overrides[rollout_budget_key]
+            gepa_section = pl_section_in_dict.get("gepa", {})
+            actual_budget = gepa_section.get("rollout", {}).get("budget") if isinstance(gepa_section, dict) else None
+            if actual_budget is not None:
+                assert actual_budget == expected_budget, (
+                    f"Rollout budget mismatch: config_body has {actual_budget} but override specifies {expected_budget}. "
+                    f"This indicates the override wasn't applied correctly."
+                )
+        
+        # Check model override
+        model_key = "prompt_learning.policy.model"
+        if model_key in config_overrides:
+            expected_model = config_overrides[model_key]
+            policy_section = pl_section_in_dict.get("policy", {})
+            actual_model = policy_section.get("model") if isinstance(policy_section, dict) else None
+            if actual_model is not None:
+                assert actual_model == expected_model, (
+                    f"Model mismatch: config_body has {actual_model} but override specifies {expected_model}. "
+                    f"This indicates the override wasn't applied correctly."
+                )
+        
+        # Check provider override
+        provider_key = "prompt_learning.policy.provider"
+        if provider_key in config_overrides:
+            expected_provider = config_overrides[provider_key]
+            policy_section = pl_section_in_dict.get("policy", {})
+            actual_provider = policy_section.get("provider") if isinstance(policy_section, dict) else None
+            if actual_provider is not None:
+                assert actual_provider == expected_provider, (
+                    f"Provider mismatch: config_body has {actual_provider} but override specifies {expected_provider}. "
+                    f"This indicates the override wasn't applied correctly."
+                )
+    
+    # FINAL CHECK: Ensure config_body has correct structure for backend
+    # Backend expects: {"prompt_learning": {...}} (full TOML structure)
+    if "prompt_learning" not in config_dict:
+        raise ValueError(
+            "config_dict must have 'prompt_learning' key. "
+            f"Found keys: {list(config_dict.keys())}"
+        )
+    
+    # CRITICAL: Final validation - check MIPRO fields are in config_body before sending
+    if pl_cfg.algorithm == "mipro":
+        pl_section_final = config_dict.get("prompt_learning", {})
+        mipro_section_final = pl_section_final.get("mipro", {}) if isinstance(pl_section_final, dict) else {}
+        
+        # ASSERT: Fields MUST be present before building payload
+        assert mipro_section_final.get("bootstrap_train_seeds") is not None, \
+            f"CRITICAL ASSERTION FAILED: bootstrap_train_seeds missing from config_body before sending to backend! " \
+            f"mipro_section keys: {list(mipro_section_final.keys()) if isinstance(mipro_section_final, dict) else 'NOT A DICT'}, " \
+            f"pl_section keys: {list(pl_section_final.keys()) if isinstance(pl_section_final, dict) else 'NOT A DICT'}, " \
+            f"pl_cfg.mipro.bootstrap_train_seeds: {getattr(pl_cfg.mipro, 'bootstrap_train_seeds', None) if pl_cfg.mipro else 'N/A'}"
+        assert mipro_section_final.get("online_pool") is not None, \
+            f"CRITICAL ASSERTION FAILED: online_pool missing from config_body before sending to backend! " \
+            f"mipro_section keys: {list(mipro_section_final.keys()) if isinstance(mipro_section_final, dict) else 'NOT A DICT'}, " \
+            f"pl_section keys: {list(pl_section_final.keys()) if isinstance(pl_section_final, dict) else 'NOT A DICT'}, " \
+            f"pl_cfg.mipro.online_pool: {getattr(pl_cfg.mipro, 'online_pool', None) if pl_cfg.mipro else 'N/A'}"
+        
+        if not mipro_section_final.get("bootstrap_train_seeds"):
+            import json
+            raise ValueError(
+                "CRITICAL: bootstrap_train_seeds missing from config_body before sending to backend. "
+                f"mipro_section keys: {list(mipro_section_final.keys()) if isinstance(mipro_section_final, dict) else 'NOT A DICT'}, "
+                f"pl_section keys: {list(pl_section_final.keys()) if isinstance(pl_section_final, dict) else 'NOT A DICT'}, "
+                f"config_overrides: {json.dumps(list(config_overrides.keys()) if config_overrides else [])}"
+            )
+        if not mipro_section_final.get("online_pool"):
+            import json
+            raise ValueError(
+                "CRITICAL: online_pool missing from config_body before sending to backend. "
+                f"mipro_section keys: {list(mipro_section_final.keys()) if isinstance(mipro_section_final, dict) else 'NOT A DICT'}, "
+                f"pl_section keys: {list(pl_section_final.keys()) if isinstance(pl_section_final, dict) else 'NOT A DICT'}, "
+                f"config_overrides: {json.dumps(list(config_overrides.keys()) if config_overrides else [])}"
+            )
+    
+    payload: dict[str, Any] = {
+        "algorithm": pl_cfg.algorithm,
+        "config_body": config_dict,
+        "overrides": config_overrides,
+        "metadata": overrides.get("metadata", {}),
+        "auto_start": overrides.get("auto_start", True),
+    }
+    
+    # CRITICAL DEBUG: Print MIPRO section structure before sending
+    if pl_cfg.algorithm == "mipro":
+        import json
+        mipro_debug = config_dict.get("prompt_learning", {}).get("mipro", {})
+        print("\n🔍 DEBUG: MIPRO section in config_body before sending:")
+        print(f"  Type: {type(mipro_debug)}")
+        print(f"  Keys: {list(mipro_debug.keys()) if isinstance(mipro_debug, dict) else 'NOT A DICT'}")
+        print(f"  bootstrap_train_seeds present: {mipro_debug.get('bootstrap_train_seeds') is not None}")
+        print(f"  online_pool present: {mipro_debug.get('online_pool') is not None}")
+        if isinstance(mipro_debug, dict):
+            print(f"  bootstrap_train_seeds value: {mipro_debug.get('bootstrap_train_seeds')}")
+            print(f"  online_pool value: {mipro_debug.get('online_pool')}")
+        # Print full mipro section (truncated)
+        mipro_json = json.dumps(mipro_debug, indent=2, default=str)
+        print(f"  Full mipro section (first 500 chars):\n{mipro_json[:500]}")
+        print("🔍 END DEBUG\n")
+    
+    backend = overrides.get("backend")
+    if backend:
+        metadata_default: dict[str, Any] = {}
+        metadata = cast(dict[str, Any], payload.setdefault("metadata", metadata_default))
+        metadata["backend_base_url"] = ensure_api_base(str(backend))
+    
+    return PromptLearningBuildResult(payload=payload, task_url=final_task_url)
+
+
+__all__ = [
+    "PromptLearningBuildResult",
+    "RLBuildResult",
+    "SFTBuildResult",
+    "build_prompt_learning_payload",
+    "build_rl_payload",
+    "build_sft_payload",
+]