synth-ai 0.2.16__py3-none-any.whl → 0.2.19__py3-none-any.whl

synth_ai/cli/commands/eval/core.py

@@ -0,0 +1,1112 @@
+from __future__ import annotations
+
+import asyncio
+import contextlib
+import importlib
+import importlib.util
+import json
+import os
+import sqlite3
+import sys
+import time
+import uuid
+from collections.abc import Sequence
+from functools import lru_cache
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, cast
+
+import click
+from synth_ai.task.config import EvalConfig
+from synth_ai.tracing_v3.session_tracer import SessionTracer
+from synth_ai.utils.task_app_discovery import discover_eval_config_paths
+
+from .errors import (
+    EvalCliError,
+    EvalConfigNotFoundError,
+    EvalConfigParseError,
+    InvalidEvalConfigError,
+    MetadataFilterFormatError,
+    MetadataSQLExecutionError,
+    MetadataSQLResultError,
+    MissingEvalTableError,
+    NoSeedsMatchedError,
+    SeedParseError,
+    TaskInfoUnavailableError,
+    TomlUnavailableError,
+)
+from .validation import validate_eval_options
+
+try:  # Python 3.11+
+    import tomllib as _toml
+except Exception:  # pragma: no cover - fallback
+    _toml = None  # type: ignore[assignment]
+
+__all__ = ["command", "get_command", "format_eval_error"]
+
+if TYPE_CHECKING:
+    from synth_ai.cli.task_apps import AppChoice, TaskAppEntryType
+
+
+@lru_cache(maxsize=1)
+def _task_apps_module():
+    from synth_ai.cli import task_apps as module  # local import to avoid circular deps
+
+    return module
+
+
+@click.command(
+    "eval",
+    help="Run one-off rollouts against a task app and print judge/eval summaries.",
+)
+@click.argument("app_id", type=str, required=False)
+@click.option(
+    "--config",
+    type=click.Path(),
+    default=None,
+    help="Path to eval TOML (short schema). Auto-discovers the first matching file when omitted.",
+)
+@click.option(
+    "--url",
+    "task_app_url",
+    type=str,
+    default=None,
+    help="Base URL of a running task app instead of spawning locally (requires --env-file for secrets).",
+)
+@click.option(
+    "--seeds",
+    default="0,1,2,3,4",
+    help="Comma-separated seeds/indices to evaluate. Use negative numbers to wrap around the dataset.",
+)
+@click.option("--split", default="train", show_default=True, help="Dataset split to use")
+@click.option(
+    "--model",
+    default=None,
+    help="Model identifier. When omitted the CLI will prompt based on task metadata.",
+)
+@click.option(
+    "--env-file",
+    multiple=True,
+    type=click.Path(),
+    help="Env file(s) to load (API keys, etc.). Required when using --url or remote judges.",
+)
+@click.option(
+    "--trace-db",
+    default="traces/v3/synth_ai.db",
+    show_default=True,
+    help="SQLite/Turso URL for storing rollout traces set to 'none' to disable persistence.",
+)
+@click.option(
+    "--metadata",
+    multiple=True,
+    help="Filter tasks by key=value metadata (e.g., --metadata difficulty=easy)",
+)
+@click.option(
+    "--metadata-sql",
+    default=None,
+    help="SQLite query that returns seeds to evaluate (e.g., SELECT seed FROM tasks WHERE difficulty='easy' LIMIT 5)",
+)
+def eval_command(
+    app_id: str | None,
+    config: str | None,
+    task_app_url: str | None,
+    seeds: str,
+    split: str,
+    model: str | None,
+    env_file: Sequence[str],
+    trace_db: str,
+    metadata: Sequence[str],
+    metadata_sql: str | None,
+) -> None:
+    try:
+        return _eval_command_impl(
+            app_id=app_id,
+            config=config,
+            task_app_url=task_app_url,
+            seeds=seeds,
+            split=split,
+            model=model,
+            env_file=env_file,
+            trace_db=trace_db,
+            metadata=metadata,
+            metadata_sql=metadata_sql,
+        )
+    except EvalCliError as exc:
+        raise click.ClickException(format_eval_error(exc)) from exc
+
+
+def _eval_command_impl(
+    app_id: str | None,
+    config: str | None,
+    task_app_url: str | None,
+    seeds: str,
+    split: str,
+    model: str | None,
+    env_file: Sequence[str],
+    trace_db: str,
+    metadata: Sequence[str],
+    metadata_sql: str | None,
+) -> None:
+    """Run rollouts against a task app and report judge statistics.
+
+    By default the command spins up the selected task app in-process, executes the
+    requested seeds, and prints aggregate scores (official and custom judges). When
+    pointing at a remote `--url`, supply matching `--env-file` values so the CLI can
+    forward authentication headers to the running service.
+    """
+    module = _task_apps_module()
+    task_app_config_type = module.TaskAppConfig
+    create_task_app = module.create_task_app
+    select_app_choice = module._select_app_choice
+    determine_env_files = module._determine_env_files
+    load_env_files_into_process = module._load_env_files_into_process
+    store_trace = getattr(module, "_store_trace", None)
+    pearson = module._pearson
+    judge_spec_cls = module.JudgeSpec
+    session_tracer_cls = getattr(module, "SessionTracer", None)
+
+    # Parse and validate TOML config
+
+    cfg: dict[str, Any] = {}
+    eval_cfg: EvalConfig | None = None
+    config_path: Path | None = None
+
+    if config:
+        config_path = Path(config)
+    else:
+        auto_configs = discover_eval_config_paths()
+        if auto_configs:
+            config_path = auto_configs[0]
+            click.echo(f"Using eval config: {config_path}")
+
+    if config_path:
+        if _toml is None:
+            raise TomlUnavailableError()
+        if not config_path.exists():
+            raise EvalConfigNotFoundError(str(config_path))
+        try:
+            data = config_path.read_bytes()
+            parsed = _toml.loads(data.decode("utf-8"))
+            if isinstance(parsed, dict):
+                section = parsed.get("eval")
+                if section is None:
+                    cfg = dict(parsed)
+                elif isinstance(section, dict):
+                    cfg = dict(section)
+                else:
+                    raise MissingEvalTableError()
+        except Exception as exc:
+            raise EvalConfigParseError(path=str(config_path), detail=str(exc)) from exc
+
+    if cfg:
+        try:
+            normalized_cfg = validate_eval_options(cfg)
+            normalized_cfg_dict = dict(normalized_cfg)
+            eval_cfg = EvalConfig.from_dict(normalized_cfg_dict)
+            cfg = normalized_cfg_dict
+            click.echo(f"✓ Config validated: {len(eval_cfg.seeds)} seeds, model={eval_cfg.model}")
+        except (ValueError, TypeError) as validation_error:
+            raise InvalidEvalConfigError(detail=str(validation_error)) from validation_error
+    else:
+        cfg = {}
+
+    # CLI args override config
+    if eval_cfg:
+        app_id = app_id or eval_cfg.app_id
+    else:
+        app_id = app_id or (cfg.get("app_id") if isinstance(cfg.get("app_id"), str) else None)  # type: ignore
+
+    metadata_filters: dict[str, str] = {}
+    if eval_cfg:
+        metadata_filters.update(eval_cfg.metadata)
+    else:
+        cfg_metadata = cfg.get("metadata")
+        if isinstance(cfg_metadata, dict):
+            for key, value in cfg_metadata.items():
+                metadata_filters[str(key)] = str(value)
+        elif isinstance(cfg_metadata, list):
+            for item in cfg_metadata:
+                if isinstance(item, str) and "=" in item:
+                    key, value = item.split("=", 1)
+                    metadata_filters[key.strip()] = value.strip()
+
+    for item in metadata or ():
+        if "=" not in item:
+            raise MetadataFilterFormatError(entry=item)
+        key, value = item.split("=", 1)
+        key = key.strip()
+        value = value.strip()
+        if not key or not value:
+            raise MetadataFilterFormatError(entry=item)
+        metadata_filters[key] = value
+
+    metadata_sql_query: str | None = None
+    if eval_cfg and eval_cfg.metadata_sql:
+        metadata_sql_query = eval_cfg.metadata_sql
+    else:
+        cfg_metadata_sql = cfg.get("metadata_sql")
+        if isinstance(cfg_metadata_sql, dict):
+            metadata_sql_query = cfg_metadata_sql.get("query") or cfg_metadata_sql.get("sql")
+        elif isinstance(cfg_metadata_sql, str):
+            metadata_sql_query = cfg_metadata_sql
+
+    if metadata_sql:
+        metadata_sql_query = metadata_sql
+    if metadata_sql_query is not None:
+        metadata_sql_query = str(metadata_sql_query)
+
+    trace_db_url: str | None = None
+    trace_db = (trace_db or "").strip()
+    if trace_db and trace_db.lower() not in {"none", "off", "disable"}:
+        if "://" in trace_db:
+            trace_db_url = trace_db
+        else:
+            trace_path = Path(trace_db).expanduser()
+            trace_path.parent.mkdir(parents=True, exist_ok=True)
+            trace_db_url = f"sqlite+aiosqlite:///{trace_path}"
+    trace_tracer: SessionTracer | None = None
+    if trace_db_url and session_tracer_cls is not None:
+        trace_tracer = cast(SessionTracer, session_tracer_cls(db_url=trace_db_url, auto_save=True))
+
+    # Determine selection params (CLI takes precedence; TOML only fills unset model/seeds/env)
+    if cfg.get("model") and not model:
+        model = str(cfg["model"])  # type: ignore[index]
+    if cfg.get("seeds") and seeds == "0,1,2,3,4":
+        val = cfg["seeds"]
+        if isinstance(val, list):
+            with contextlib.suppress(Exception):
+                seeds = ",".join(str(int(x)) for x in val)
+        elif isinstance(val, str):
+            seeds = val
+        elif isinstance(val, int):
+            seeds = str(val)
+    if cfg.get("env_file") and not env_file:
+        ef = cfg["env_file"]
+        if isinstance(ef, str):
+            env_file = (ef,)  # type: ignore[assignment]
+        elif isinstance(ef, list):
+            env_file = tuple(str(x) for x in ef)  # type: ignore[assignment]
+
+    choice_for_env: AppChoice | None = None
+    entry: TaskAppEntryType | None = None
+    if task_app_url is None:
+        choice_for_env = select_app_choice(app_id, purpose="eval")
+        entry = choice_for_env.ensure_entry()
+
+    env_paths: list[Path] = []
+    if entry is not None:
+        original_env_path = choice_for_env.path if choice_for_env is not None else None
+        env_paths = determine_env_files(entry, env_file, original_path=original_env_path)
+    else:
+        if not env_file:
+            raise click.ClickException("--env-file is required when using --url")
+        for candidate in env_file:
+            p = Path(candidate).expanduser()
+            if not p.exists():
+                raise click.ClickException(f"Env file not found: {p}")
+            env_paths.append(p)
+
+    click.echo("Using env file(s): " + ", ".join(str(p) for p in env_paths))
+    load_env_files_into_process([str(Path(p)) for p in env_paths])
+
+    if task_app_url is None:
+        config = entry.config_factory()  # type: ignore[union-attr]
+        # Help the type checker; runtime check also enforced in server.run_task_app
+        if not isinstance(config, task_app_config_type):
+            raise click.ClickException(
+                "Invalid task app: config_factory did not return TaskAppConfig"
+            )
+        app = create_task_app(config)
+
+    # Determine supported models
+    inference_meta: dict[str, Any] = {}
+    supported: list[str] = []
+    seen_models: set[str] = set()
+
+    def _add_supported_model(candidate: Any) -> None:
+        if not candidate:
+            return
+        text = str(candidate).strip()
+        if not text or text in seen_models:
+            return
+        supported.append(text)
+        seen_models.add(text)
+
+    if task_app_url is None:
+        try:
+            if hasattr(config, "base_task_info") and config.base_task_info:
+                inf_obj = getattr(config.base_task_info, "inference", None)
+                if inf_obj is not None:
+                    if hasattr(inf_obj, "model_dump"):
+                        inference_meta = dict(inf_obj.model_dump(exclude_none=True))  # type: ignore[attr-defined]
+                    elif isinstance(inf_obj, dict):
+                        inference_meta = dict(inf_obj)
+        except Exception:
+            inference_meta = {}
+    else:
+        try:
+            import httpx as _hx
+
+            headers = {}
+            api_key = (os.environ.get("ENVIRONMENT_API_KEY") or "").strip()
+            if api_key:
+                headers["X-API-Key"] = api_key
+            with _hx.Client(base_url=task_app_url, headers=headers, timeout=15.0) as c:
+                info = c.get("/info").json()
+            inf = info.get("inference") if isinstance(info, dict) else None
+            if isinstance(inf, dict):
+                inference_meta = dict(inf)
+        except Exception:
+            inference_meta = {}
+
+    default_model = inference_meta.get("model")
+    if isinstance(default_model, str):
+        _add_supported_model(default_model)
+
+    models_field = inference_meta.get("models")
+    if isinstance(models_field, list):
+        for candidate in models_field:
+            _add_supported_model(candidate)
+
+    supported_models = inference_meta.get("supported_models")
+    if isinstance(supported_models, list):
+        for candidate in supported_models:
+            _add_supported_model(candidate)
+
+    providers = inference_meta.get("providers")
+    if isinstance(providers, list):
+        if "openai" in providers:
+            _add_supported_model("gpt-5")
+        if "groq" in providers:
+            _add_supported_model("groq:llama-3.1-70b-versatile")
+
+    _add_supported_model("synth:qwen-0.6b")
+
+    selected_model = model
+    if not selected_model:
+        if not supported:
+            raise click.ClickException(
+                "No supported models; supply --model or add base_task_info.inference.model"
+            )
+        click.echo("Select model to evaluate:")
+        for idx, m in enumerate(supported, start=1):
+            click.echo(f"  {idx}) {m}")
+        choice_idx = click.prompt("Enter choice", type=click.IntRange(1, len(supported)))
+        selected_model = supported[choice_idx - 1]
+
+    try:
+        seed_values = [int(s.strip()) for s in seeds.split(",") if s.strip()]
+    except Exception as exc:
+        raise SeedParseError(value=seeds) from exc
+
+    import httpx
+
+    headers = {}
+    api_key = (os.environ.get("ENVIRONMENT_API_KEY") or "").strip()
+    if api_key:
+        headers["X-API-Key"] = api_key
+
+    # Precompute optional policy overrides from TOML
+    policy_overrides: dict[str, Any] = {}
+    try:
+        # Accept [eval.policy] table or top-level keys for convenience
+        if isinstance(cfg.get("policy"), dict):
+            policy_overrides.update(dict(cfg["policy"]))
+        # Back-compat: allow temperature/max_tokens at top level
+        for k in (
+            "temperature",
+            "max_tokens",
+            "reasoning_effort",
+            "system_hint",
+            "tool_choice",
+            "inference_url",
+        ):
+            if k in cfg and k not in policy_overrides:
+                policy_overrides[k] = cfg.get(k)
+    except Exception:
+        policy_overrides = {}
+
+    raw_concurrency = cfg.get("concurrency")
+    try:
+        concurrency_limit = int(raw_concurrency) if raw_concurrency is not None else 1
+    except Exception:
+        concurrency_limit = 1
+    if concurrency_limit <= 0:
+        concurrency_limit = 1
+    concurrency_limit = min(concurrency_limit, max(1, len(seed_values)))
+
+    judge_specs: list[Any] = []
+
+    def _register_judge(name_hint: str | None, judge_cfg: dict[str, Any]) -> None:
+        if not judge_cfg:
+            return
+        judge_module = judge_cfg.get("module")
+        judge_path = judge_cfg.get("path")
+        judge_callable_name = judge_cfg.get("callable") or judge_cfg.get("function")
+        if judge_module and judge_path:
+            raise click.ClickException("Judge config cannot set both 'module' and 'path'")
+        if not judge_module and not judge_path:
+            raise click.ClickException("Judge config requires 'module' or 'path'")
+        try:
+            if judge_module:
+                module = importlib.import_module(str(judge_module))
+            else:
+                path = Path(str(judge_path)).expanduser()
+                if not path.exists():
+                    raise click.ClickException(f"Judge module path not found: {path}")
+                spec = importlib.util.spec_from_file_location(
+                    f"_eval_judge_{path.stem}", path
+                )
+                if not spec or not spec.loader:
+                    raise click.ClickException(f"Failed to load judge module from {path}")
+                module = importlib.util.module_from_spec(spec)
+                sys.modules[spec.name] = module
+                spec.loader.exec_module(module)
+        except click.ClickException:
+            raise
+        except Exception as exc:
+            raise click.ClickException(f"Unable to load judge module: {exc}") from exc
+
+        if judge_callable_name:
+            try:
+                judge_fn = getattr(module, str(judge_callable_name))
+            except AttributeError as exc:
+                raise click.ClickException(
+                    f"Judge callable '{judge_callable_name}' not found in module"
+                ) from exc
+        else:
+            if hasattr(module, "judge"):
+                judge_fn = module.judge
+            else:
+                raise click.ClickException("Judge module must expose 'judge' callable")
+
+        if not callable(judge_fn):
+            raise click.ClickException("Judge callable is not callable")
+
+        judge_kwargs = {
+            k: v
+            for k, v in judge_cfg.items()
+            if k not in {"module", "path", "callable", "function", "name"}
+        }
+        display_name = str(
+            judge_cfg.get("name")
+            or name_hint
+            or f"judge{len(judge_specs) + 1}"
+        )
+        judge_specs.append(judge_spec_cls(display_name, judge_fn, judge_kwargs))
+
+    raw_judge_cfg = cfg.get("judge")
+    if isinstance(raw_judge_cfg, dict) and raw_judge_cfg:
+        direct_keys = {"module", "path", "callable", "function", "name"}
+        has_direct_keys = any(key in raw_judge_cfg for key in direct_keys)
+        nested_candidates = [
+            (key, value)
+            for key, value in raw_judge_cfg.items()
+            if isinstance(value, dict)
+        ]
+        if has_direct_keys and not nested_candidates:
+            _register_judge(None, raw_judge_cfg)
+        else:
+            for sub_name, sub_cfg in nested_candidates:
+                _register_judge(sub_name, sub_cfg)
+
+    raw_judges_list = cfg.get("judges")
+    if isinstance(raw_judges_list, list):
+        for _index, entry in enumerate(raw_judges_list, start=1):
+            if isinstance(entry, dict):
+                _register_judge(entry.get("name") or f"judge{len(judge_specs) + 1}", entry)
+
+    records: list[dict[str, Any]] = []
+
+    successes = 0
+    failures = 0
+    # Aggregate outcome stats across successful seeds
+    outcome_sum: float = 0.0
+    outcome_count: int = 0
+    outcome_correct: int = 0
+
+    def _build_task_rows(taskset: Any) -> dict[int, dict[str, Any]]:
+        rows: dict[int, dict[str, Any]] = {}
+        if not isinstance(taskset, dict):
+            return rows
+
+        scenario_ids = taskset.get("scenario_ids") or []
+        loop_ids = taskset.get("loop_ids") or []
+        thread_ids = taskset.get("thread_ids") or []
+        difficulty_map = taskset.get("difficulty_map") or {}
+
+        max_len = max(len(scenario_ids), len(loop_ids), len(thread_ids))
+        for seed in range(max_len):
+            scenario_id = scenario_ids[seed] if seed < len(scenario_ids) else None
+            loop_id = loop_ids[seed] if seed < len(loop_ids) else None
+            thread_id = thread_ids[seed] if seed < len(thread_ids) else None
+            difficulty = None
+            if isinstance(difficulty_map, dict):
+                if scenario_id and scenario_id in difficulty_map:
+                    difficulty = difficulty_map.get(scenario_id)
+                elif str(seed) in difficulty_map:
+                    difficulty = difficulty_map.get(str(seed))
+
+            rows[seed] = {
+                "seed": seed,
+                "scenario_id": scenario_id,
+                "loop_id": loop_id,
+                "thread_id": thread_id,
+                "difficulty": difficulty,
+            }
+        return rows
+
+    def _apply_metadata_filters(
+        rows: dict[int, dict[str, Any]], seeds_list: list[int], filters: dict[str, str]
+    ) -> list[int]:
+        if not filters:
+            return seeds_list
+        filtered: list[int] = []
+        for seed in seeds_list:
+            row = rows.get(seed)
+            if not row:
+                continue
+            include = True
+            for key, expected in filters.items():
+                actual = row.get(key)
+                if actual is None:
+                    include = False
+                    break
+                if str(actual).lower() != expected.lower():
+                    include = False
+                    break
+            if include:
+                filtered.append(seed)
+        return filtered
+
+    def _apply_metadata_sql(
+        rows: dict[int, dict[str, Any]], seeds_list: list[int], query: str
+    ) -> list[int]:
+        """Return seeds that satisfy an arbitrary SQL query.
+
+        The query is executed against an in-memory SQLite table named `tasks`
+        with columns (seed INTEGER, scenario_id TEXT, loop_id TEXT, thread_id TEXT, difficulty TEXT).
+        Any rows whose `seed` value (or first column if `seed` is absent) appear in the result set are retained.
+        """
+        if not query:
+            return seeds_list
+        conn = sqlite3.connect(":memory:")
+        try:
+            cur = conn.cursor()
+            cur.execute(
+                "CREATE TABLE tasks (seed INTEGER, scenario_id TEXT, loop_id TEXT, thread_id TEXT, difficulty TEXT)"
+            )
+            insert_stmt = (
+                "INSERT INTO tasks (seed, scenario_id, loop_id, thread_id, difficulty) VALUES (?,?,?,?,?)"
+            )
+            for seed in seeds_list:
+                row = rows.get(seed, {})
+                cur.execute(
+                    insert_stmt,
+                    [
+                        seed,
+                        row.get("scenario_id"),
+                        row.get("loop_id"),
+                        row.get("thread_id"),
+                        row.get("difficulty"),
+                    ],
+                )
+
+            result = cur.execute(query)
+            fetched = result.fetchall()
+            if not fetched:
+                return []
+            description = result.description or []
+            col_names = [col[0] for col in description]
+            seeds_out: list[int] = []
+            for entry in fetched:
+                value = entry[col_names.index("seed")] if "seed" in col_names else entry[0]
+                try:
+                    seeds_out.append(int(value))
+                except Exception as exc:
+                    raise MetadataSQLResultError(
+                        query=query,
+                        detail="non-integer value returned",
+                    ) from exc
+            seeds_set = set(seeds_out)
+            return [seed for seed in seeds_list if seed in seeds_set]
+        except sqlite3.Error as exc:
+            raise MetadataSQLExecutionError(query=query, detail=str(exc)) from exc
+        finally:
+            conn.close()
+
+    async def _run_eval() -> None:
+        nonlocal successes, failures, outcome_sum, outcome_count, outcome_correct, records, seed_values
+
+        if trace_tracer is not None and trace_tracer.db is None:
+            await trace_tracer.initialize()
+
+        if task_app_url is None:
+            transport = httpx.ASGITransport(app=app)  # type: ignore[name-defined]
+            async_client = httpx.AsyncClient(
+                transport=cast(Any, transport),
+                base_url="http://eval.local",
+                timeout=300.0,
+                follow_redirects=True,
+                headers=headers,
+            )
+        else:
+            async_client = httpx.AsyncClient(
+                base_url=task_app_url,
+                timeout=300.0,
+                follow_redirects=True,
+                headers=headers,
+            )
+
+        try:
+            taskset_payload: dict[str, Any] | None = None
+            try:
+                task_info_response = await async_client.get("/task_info")
+            except Exception:
+                task_info_response = None
+            if task_info_response is not None and task_info_response.status_code == 200:
+                with contextlib.suppress(Exception):
+                    payload_json = task_info_response.json()
+                if isinstance(payload_json, dict) and "taskset" in payload_json:
+                    taskset_payload = payload_json.get("taskset")
+                    if not isinstance(taskset_payload, dict):
+                        taskset_payload = None
+                elif isinstance(payload_json, dict):
+                    taskset_payload = payload_json
+
+            available_seeds = list(seed_values)
+            if metadata_sql_query or metadata_filters:
+                if not taskset_payload:
+                    raise TaskInfoUnavailableError()
+                rows = _build_task_rows(taskset_payload)
+                if metadata_sql_query:
+                    available_seeds = _apply_metadata_sql(rows, available_seeds, metadata_sql_query)
+                if metadata_filters:
+                    available_seeds = _apply_metadata_filters(rows, available_seeds, metadata_filters)
+                if not available_seeds:
+                    raise NoSeedsMatchedError()
+                seed_values = available_seeds
+
+            semaphore = asyncio.Semaphore(concurrency_limit)
+
+            async def _run_seed(seed_val: int) -> None:
+                nonlocal successes, failures, outcome_sum, outcome_count, outcome_correct, records
+                # Read env_name and policy_name from config if available
+                env_name = cfg.get("env_name") or (cfg.get("env", {}).get("env_name") if isinstance(cfg.get("env"), dict) else None)
+                policy_name = cfg.get("policy_name") or (cfg.get("policy", {}).get("policy_name") if isinstance(cfg.get("policy"), dict) else None)
+                env_config_overrides = cfg.get("env_config", {}) if isinstance(cfg.get("env_config"), dict) else {}
+                policy_config_overrides = cfg.get("policy_config", {}) if isinstance(cfg.get("policy_config"), dict) else {}
+                
+                # Debug: print config parsing
+                if seed_val == 0:
+                    click.echo(f"[DEBUG] env_name from config: {env_name}")
+                    click.echo(f"[DEBUG] policy_name from config: {policy_name}")
+                
+                # Generate default ops sequence if not provided
+                max_llm_calls = policy_config_overrides.get("max_llm_calls", 10)
+                ops_list = cfg.get("ops", [])
+                if not ops_list:
+                    # Generate default "agent, env" pairs for max_llm_calls
+                    ops_list = ["agent", "env"] * int(max_llm_calls)
+                
+                body = {
+                    "run_id": str(uuid.uuid4()),
+                    "env": {"config": {"split": split, "index": seed_val, **env_config_overrides}, "seed": seed_val},
+                    "policy": {
+                        "policy_name": policy_name or selected_model,
+                        "config": {"model": selected_model, **policy_overrides, **policy_config_overrides},
+                    },
+                    "ops": ops_list,
+                    "record": {
+                        "return_trace": cfg.get("return_trace", True),
+                        "trace_format": cfg.get("trace_format", "structured"),
+                    },
+                    "mode": "eval",  # RolloutMode.EVAL: use inference URLs as-is, no transformations
+                }
+                if env_name:
+                    env_section = body.get("env")
+                    if isinstance(env_section, dict):
+                        env_section["env_name"] = env_name
+                    else:
+                        body["env"] = {"env_name": env_name}
+
+                # Debug: print the body being sent
+                if seed_val == 0:
+                    click.echo(f"[DEBUG] rollout body env: {body['env']}")
+                    click.echo(f"[DEBUG] rollout body policy: {body['policy']}")
+                    click.echo(f"[DEBUG] rollout body mode: {body.get('mode', 'NOT SET')}")
+                rollout_elapsed: float | None = None
+                rollout_start = time.perf_counter()
+                try:
+                    import logging
+                    _log = logging.getLogger(__name__)
+                    _log.info(f"[EVAL_BODY_DEBUG] Sending body with mode={body.get('mode')}")
+                    async with semaphore:
+                        response = await async_client.post("/rollout", json=body)
+                    rollout_elapsed = time.perf_counter() - rollout_start
+                except Exception as exc:
+                    failures += 1
+                    click.echo(f"seed={seed_val} error={exc}")
+                    return
+
+                ok = 200 <= response.status_code < 300
+                if ok:
+                    successes += 1
+                else:
+                    failures += 1
+
+                summary = [f"seed={seed_val}", f"status={response.status_code}"]
+                data: Any
+                try:
+                    data = response.json()
+                except Exception:
+                    data = None
+                
+                # Debug: print validation errors
+                if response.status_code == 422 and data:
+                    click.echo(f"[DEBUG] 422 Validation Error: {data}")
+
+                metrics: dict[str, Any] | None = None
+                completion: str | None = None
+                prompt_index: int | None = None
+                prompt_text: str | None = None
+                task_id: str | None = None
+                task_split: str | None = None
+                task_rubric_id: str | None = None
+
+                trace_namespace: dict[str, Any] | None = None
+                session_trace_dict: dict[str, Any] | None = None
+
+                if isinstance(data, dict):
+                    import logging
+                    _logger = logging.getLogger(__name__)
+                    _logger.info(f"[EVAL_DEBUG] Response data keys: {list(data.keys())}")
+                    if "detail" in data:
+                        _logger.error(f"[EVAL_DEBUG] Task app returned error: {data['detail']}")
+                    trace_namespace = data.get("trace")
+                    _logger.info(f"[EVAL_DEBUG] trace_namespace type: {type(trace_namespace)}, value: {trace_namespace if not isinstance(trace_namespace, dict) else 'dict with keys: ' + str(list(trace_namespace.keys()) if trace_namespace else 'None')}")
+                    if not isinstance(trace_namespace, dict):
+                        raise RuntimeError(
+                            "The 'synth-ai eval' command requires trace payloads in rollout responses. "
+                            "Ensure the rollout request includes 'trace_format': 'structured' and 'return_trace': true, "
+                            "and that task app tracing is enabled (TASKAPP_TRACING_ENABLED=1). "
+                            "Note: This is specific to the eval command - general rollout endpoints don't require traces."
+                        )
+                    # Handle both "compact" and "full" trace formats:
+                    # - compact: trace_namespace contains {session_id, metadata, ...}
+                    # - full: trace_namespace IS the full session_trace dict
+                    session_trace_dict = trace_namespace.get("session_trace")
+                    if not isinstance(session_trace_dict, dict):
+                        # If no session_trace key, assume "full" format where trace itself is the session_trace
+                        if "session_id" in trace_namespace:
+                            session_trace_dict = trace_namespace
+                        else:
+                            raise RuntimeError(
+                                "The 'synth-ai eval' command requires 'session_trace' in the trace payload or a valid full trace format. "
+                                "Ensure the task app is using tracing_v3 and returning structured trace data."
+                            )
+                    metrics = data.get("metrics") if isinstance(data.get("metrics"), dict) else None
+                    if metrics:
+                        mean_return = metrics.get("mean_return") or metrics.get("total_reward")
+                        outcome = metrics.get("outcome_score")
+                        if mean_return is not None:
+                            summary.append(f"mean_return={mean_return}")
+                        if outcome is not None:
+                            summary.append(f"outcome={outcome}")
+                            try:
+                                val = float(outcome)
+                                outcome_sum += val
+                                outcome_count += 1
+                                if val >= 0.5:
+                                    outcome_correct += 1
+                            except Exception:
+                                pass
+                    trajs = (
+                        data.get("trajectories")
+                        if isinstance(data.get("trajectories"), list)
+                        else None
+                    )
+                    if trajs:
+                        first = trajs[0] if trajs else None
+                        steps = first.get("steps") if isinstance(first, dict) else None
+                        if isinstance(steps, list) and steps:
+                            step0 = steps[0]
+                            tool_calls = step0.get("tool_calls") or step0.get("tools") or []
+                            if isinstance(tool_calls, list):
+                                summary.append(f"tool_calls={len(tool_calls)}")
+                            obs = step0.get("obs") if isinstance(step0, dict) else None
+                            if isinstance(obs, dict):
+                                idx_val = obs.get("prompt_index")
+                                if isinstance(idx_val, int):
+                                    prompt_index = idx_val
+                                prompt_raw = obs.get("prompt")
+                                if isinstance(prompt_raw, str):
+                                    prompt_text = prompt_raw
+                                if task_id is None:
+                                    candidate_id = obs.get("task_id")
+                                    if isinstance(candidate_id, str) and candidate_id:
+                                        task_id = candidate_id
+                                if task_split is None:
+                                    candidate_split = obs.get("task_split")
+                                    if isinstance(candidate_split, str) and candidate_split:
+                                        task_split = candidate_split
+                                if task_rubric_id is None:
+                                    candidate_rid = obs.get("task_rubric_id")
+                                    if isinstance(candidate_rid, str) and candidate_rid:
+                                        task_rubric_id = candidate_rid
+                        final = first.get("final") if isinstance(first, dict) else None
+                        if isinstance(final, dict):
+                            final_obs = final.get("observation")
+                            if isinstance(final_obs, dict):
+                                comp_val = final_obs.get("completion")
+                                if isinstance(comp_val, str):
+                                    completion = comp_val
+                                if task_id is None:
+                                    candidate_id = final_obs.get("task_id")
+                                    if isinstance(candidate_id, str) and candidate_id:
+                                        task_id = candidate_id
+                                if task_split is None:
+                                    candidate_split = final_obs.get("task_split")
+                                    if isinstance(candidate_split, str) and candidate_split:
+                                        task_split = candidate_split
+                                if task_rubric_id is None:
+                                    candidate_rid = final_obs.get("task_rubric_id")
+                                    if isinstance(candidate_rid, str) and candidate_rid:
+                                        task_rubric_id = candidate_rid
+                            final_info = final.get("info")
+                            if isinstance(final_info, dict):
+                                if task_id is None:
+                                    candidate_id = final_info.get("task_id")
+                                    if isinstance(candidate_id, str) and candidate_id:
+                                        task_id = candidate_id
+                                if task_split is None:
+                                    candidate_split = final_info.get("task_split")
+                                    if isinstance(candidate_split, str) and candidate_split:
+                                        task_split = candidate_split
+                                if task_rubric_id is None:
+                                    candidate_rid = final_info.get("task_rubric_id")
+                                    if isinstance(candidate_rid, str) and candidate_rid:
+                                        task_rubric_id = candidate_rid
+                    if task_id:
+                        summary.append(f"task_id={task_id}")
+                    click.echo(" ".join(summary))
+                    with contextlib.suppress(Exception):
+                        click.echo(json.dumps(data, indent=2))
+                else:
+                    click.echo(" ".join(summary))
+
+                official_score = None
+                if isinstance(metrics, dict):
+                    for key in ("mean_return", "total_reward", "outcome_score"):
+                        val = metrics.get(key)
+                        if isinstance(val, int | float):
+                            official_score = float(val)
+                            break
+                if official_score is None and isinstance(data, dict):
+                    try:
+                        reward_val = data["trajectories"][0]["steps"][0].get("reward")
+                        if isinstance(reward_val, int | float):
+                            official_score = float(reward_val)
+                    except Exception:
+                        pass
+
+                if official_score is not None:
+                    if official_score < 0.0:
+                        official_score = 0.0
+                    elif official_score > 1.0:
+                        official_score = min(1.0, official_score)
+
+                judge_scores: dict[str, float | None] = {}
+                judges_timings: dict[str, float | None] = {}
+                timings: dict[str, Any] = {
+                    "rollout_s": rollout_elapsed,
+                    "judges": judges_timings,
+                }
+                if judge_specs:
+                    for spec in judge_specs:
+                        score_value: float | None = None
+                        judge_elapsed: float | None = None
+                        # Run judges for all tasks (text-based and trajectory-based)
+                        # Text-based tasks have completion, trajectory-based tasks use response
+                        judge_payload = {
+                            "seed": seed_val,
+                            "prompt_index": prompt_index,
+                            "prompt": prompt_text,
+                            "completion": completion,
+                            "metrics": metrics,
+                            "response": data,
+                            "trace": trace_namespace,
+                        }
+                        try:
+                            judge_start = time.perf_counter()
+                            result = spec.fn(judge_payload, **spec.kwargs)
+                            judge_elapsed = time.perf_counter() - judge_start
+                            if isinstance(result, int | float):
+                                score_value = float(result)
+                        except Exception as exc:
+                            if judge_elapsed is None:
+                                judge_elapsed = time.perf_counter() - judge_start
+                            click.echo(f"seed={seed_val} judge[{spec.name}]_error={exc}")
+                        judges_timings[spec.name] = judge_elapsed
+                        judge_scores[spec.name] = score_value
+
+                if trace_tracer is not None and trace_namespace:
+                    storage_metadata = {
+                        "eval_seed": seed_val,
+                        "prompt_index": prompt_index,
+                        "task_id": task_id,
+                        "task_split": task_split,
+                        "task_rubric_id": task_rubric_id,
+                        "official_score": official_score,
+                        "judge_scores": judge_scores,
+                        "model": selected_model,
+                        "prompt": prompt_text,
+                        "completion": completion,
+                    }
+                    if store_trace is not None:
+                        await store_trace(trace_tracer, trace_namespace, storage_metadata)
+
+                records.append(
+                    {
+                        "seed": seed_val,
+                        "prompt_index": prompt_index,
+                        "task_id": task_id,
+                        "task_split": task_split,
+                        "task_rubric_id": task_rubric_id,
+                        "official_score": official_score,
+                        "judge_scores": judge_scores,
+                        "timings": timings,
+                    }
+                )
+
+            await asyncio.gather(*[_run_seed(seed_val) for seed_val in seed_values])
+        finally:
+            await async_client.aclose()
+
+    try:
+        asyncio.run(_run_eval())
+    finally:
+        if trace_tracer is not None and trace_tracer.db is not None:
+            asyncio.run(trace_tracer.db.close())
+
+    click.echo(
+        f"Eval complete: {successes} ok, {failures} failed; model={selected_model}, split={split}"
+    )
+
+    if outcome_count > 0:
+        mean_outcome = outcome_sum / float(outcome_count)
+        frac_right = outcome_correct / float(outcome_count)
+        click.echo(
+            f"Outcome summary: correct={outcome_correct}/{outcome_count} ({frac_right:.2%}), mean_outcome={mean_outcome:.3f}"
+        )
+
+    if records:
+        judge_specs = judge_specs or []  # ensure iterable
+        official_scores = [
+            r["official_score"] for r in records if r["official_score"] is not None
+        ]
+        if official_scores:
+            click.echo(f"  Official mean: {sum(official_scores) / len(official_scores):.3f}")
+        else:
+            click.echo("  Official mean: n/a")
+
+        for spec in judge_specs:
+            spec_scores = [
+                record["judge_scores"].get(spec.name)
+                for record in records
+                if record["judge_scores"].get(spec.name) is not None
+            ]
+            if spec_scores:
+                mean_spec = sum(spec_scores) / len(spec_scores)
+                click.echo(f"  [{spec.name}] mean: {mean_spec:.3f}")
+            else:
+                click.echo(f"  [{spec.name}] mean: n/a")
+
+            paired = [
+                (
+                    record["official_score"],
+                    record["judge_scores"].get(spec.name),
+                )
+                for record in records
+                if record["official_score"] is not None
+                and record["judge_scores"].get(spec.name) is not None
+            ]
+            if len(paired) >= 2:
+                corr = pearson(
+                    [p[0] for p in paired if p[0] is not None],
+                    [p[1] for p in paired if p[1] is not None],
+                )
+                if corr is not None:
+                    click.echo(f"    Pearson r: {corr:.3f}")
+                else:
+                    click.echo("    Pearson r: undefined (zero variance)")
+            else:
+                click.echo("    Pearson r: n/a (need ≥2 paired scores)")
+
+        header = ["Seed", "Prompt", "Official"]
+        header.extend(spec.name for spec in judge_specs)
+        rows: list[list[str]] = []
+        for record in sorted(records, key=lambda r: (r["seed"], r.get("prompt_index") or -1)):
+            seed_val = str(record["seed"])
+            prompt_idx = (
+                str(record["prompt_index"])
+                if record["prompt_index"] is not None
+                else "-"
+            )
+            official_val = (
+                f"{record['official_score']:.3f}"
+                if record["official_score"] is not None
+                else "-"
+            )
+            row = [seed_val, prompt_idx, official_val]
+            for spec in judge_specs:
+                score_val = record["judge_scores"].get(spec.name)
+                row.append(f"{score_val:.3f}" if isinstance(score_val, int | float) else "-")
+            rows.append(row)
+
+        widths = [len(col) for col in header]
+        for row in rows:
+            for idx, cell in enumerate(row):
+                widths[idx] = max(widths[idx], len(cell))
+
+        click.echo("")
+        click.echo("  ".join(h.ljust(widths[idx]) for idx, h in enumerate(header)))
+        click.echo("  ".join("-" * widths[idx] for idx in range(len(header))))
+        for row in rows:
+            click.echo("  ".join(cell.ljust(widths[idx]) for idx, cell in enumerate(row)))
+
+
+
+command = eval_command
+
+
+def get_command() -> click.Command:
+    """Return the Click command implementing task-app evaluation."""
+    return command
+
+
+def format_eval_error(err: EvalCliError) -> str:
+    if isinstance(err, TomlUnavailableError):
+        hint = err.hint or "Install tomli or use Python 3.11+."
+        return f"TOML parser not available. {hint}"
+    if isinstance(err, EvalConfigNotFoundError):
+        return f"Eval config not found: {err.path}"
+    if isinstance(err, EvalConfigParseError):
+        return f"Failed to parse TOML '{err.path}': {err.detail}"
+    if isinstance(err, MissingEvalTableError):
+        return "Config must contain an [eval] table."
+    if isinstance(err, InvalidEvalConfigError):
+        return f"Invalid eval config: {err.detail}"
+    if isinstance(err, SeedParseError):
+        return f"Unable to parse seeds from '{err.value}'. Provide comma-separated integers."
+    if isinstance(err, MetadataFilterFormatError):
+        return f"Metadata filter '{err.entry}' must be key=value."
+    if isinstance(err, TaskInfoUnavailableError):
+        return "Task metadata filters require the task app to expose /task_info metadata."
+    if isinstance(err, NoSeedsMatchedError):
+        hint = err.hint or "Adjust the metadata filters or seed list."
+        return f"No seeds match the provided metadata filters. {hint}"
+    if isinstance(err, MetadataSQLExecutionError):
+        return f"Failed to execute metadata SQL query '{err.query}': {err.detail}"
+    if isinstance(err, MetadataSQLResultError):
+        return f"metadata SQL query '{err.query}' must return integer seed values ({err.detail})"
+    return str(err)