synth-ai 0.2.17__py3-none-any.whl → 0.2.19__py3-none-any.whl

synth_ai/cli/commands/smoke/core.py

@@ -0,0 +1,1436 @@
+from __future__ import annotations
+
+import asyncio
+import contextlib
+import logging
+import os
+import subprocess
+import sys
+import time
+import uuid
+from pathlib import Path
+from typing import Any
+from urllib.parse import parse_qsl, urlencode, urlparse, urlunparse
+
+import click
+import httpx
+import toml
+from synth_ai.task.client import TaskAppClient
+from synth_ai.task.contracts import (
+    RolloutEnvSpec,
+    RolloutMode,
+    RolloutPolicySpec,
+    RolloutRecordConfig,
+    RolloutRequest,
+    RolloutSafetyConfig,
+)
+from synth_ai.task.validators import (
+    normalize_inference_url,
+    validate_rollout_response_for_rl,
+    validate_task_app_url,
+)
+from synth_ai.tracing_v3.config import resolve_trace_db_settings
+from synth_ai.tracing_v3.turso.daemon import start_sqld
+
+
+def _append_query_param(url: str, key: str, value: str) -> str:
+    parsed = urlparse(url)
+    params = dict(parse_qsl(parsed.query, keep_blank_values=True))
+    params[key] = value
+    new_query = urlencode(params)
+    return urlunparse(parsed._replace(query=new_query))
+
+
+def _ensure_local_libsql() -> None:
+    """Start a local sqld/libSQL instance or abort the smoke test."""
+
+    traces_root = Path(os.getenv("SYNTH_TRACES_DIR", str((Path.cwd() / "traces" / "v3").resolve())))
+    traces_root.mkdir(parents=True, exist_ok=True)
+
+    local_db_path = Path(os.getenv("SQLD_DB_PATH", str(traces_root / "local.db"))).resolve()
+    local_db_path.parent.mkdir(parents=True, exist_ok=True)
+
+    hrana_port = int(os.getenv("SQLD_HTTP_PORT", "8080"))
+    http_port = hrana_port + 1
+    os.environ["SQLD_DB_PATH"] = str(local_db_path)
+    os.environ["SQLD_HTTP_PORT"] = str(hrana_port)
+
+    try:
+        start_sqld(db_path=str(local_db_path), hrana_port=hrana_port, http_port=http_port)
+        started_new = True
+    except Exception as exc:
+        # If address in use, assume an existing sqld instance; verify health below
+        if "Address already in use" in str(exc):
+            started_new = False
+            click.echo(
+                f"[libsql] sqld already running on 127.0.0.1:{hrana_port} (hrana) and 127.0.0.1:{http_port} (http); attempting to reuse", err=True
+            )
+        else:
+            raise click.ClickException(
+                f"Failed to start local sqld on 127.0.0.1:{hrana_port}: {exc}"
+            ) from exc
+
+    health_url = f"http://127.0.0.1:{http_port}/health"
+    deadline = time.time() + 5.0
+    healthy = False
+    while time.time() < deadline:
+        try:
+            resp = httpx.get(health_url, timeout=0.5)
+            if resp.status_code == 200:
+                healthy = True
+                break
+        except Exception:
+            pass
+        time.sleep(0.1)
+
+    if not healthy:
+        msg = (
+            f"Tracing backend not reachable at {health_url}. "
+            "Start sqld manually or disable tracing (TASKAPP_TRACING_ENABLED=0)."
+        )
+        raise click.ClickException(msg)
+
+    click.echo(
+        f"[libsql] sqld ready on libsql://127.0.0.1:{hrana_port} with HTTP API on :{http_port} (started_new={started_new})",
+        err=True,
+    )
+
+    # Python libsql client uses HTTP API port, not Hrana WebSocket port
+    local_dsn = f"http://127.0.0.1:{http_port}"
+    os.environ["LIBSQL_URL"] = local_dsn
+    os.environ["SYNTH_TRACES_DB"] = local_dsn
+    os.environ.pop("LIBSQL_AUTH_TOKEN", None)
+    os.environ.pop("TURSO_AUTH_TOKEN", None)
+
+
+def _refresh_tracing_config() -> None:
+    """Rebuild global tracing configuration so new env vars take effect."""
+
+    from synth_ai.tracing_v3 import config as tracing_config_module
+    from synth_ai.tracing_v3.storage import config as storage_config_module
+
+    tracing_config_module.CONFIG = tracing_config_module.TursoConfig()  # type: ignore[assignment]
+    storage_config_module.STORAGE_CONFIG = storage_config_module.StorageConfig(  # type: ignore[assignment]
+        connection_string=os.environ["SYNTH_TRACES_DB"],
+        backend=storage_config_module.StorageBackend.TURSO_NATIVE,
+    )
+
+
+def _load_smoke_config(config_path: Path | None) -> dict[str, Any]:
+    """Load [smoke] section from TOML config file.
+    
+    Returns an empty dict if no config file or no [smoke] section.
+    """
+    if not config_path:
+        return {}
+    
+    try:
+        with open(config_path) as f:
+            full_config = toml.load(f)
+        
+        smoke_config = full_config.get("smoke", {})
+        
+        if smoke_config:
+            click.echo(f"[smoke] Loaded configuration from {config_path}", err=True)
+            click.echo(f"[smoke] Config keys: {', '.join(smoke_config.keys())}", err=True)
+        
+        return smoke_config
+    except Exception as exc:
+        click.echo(f"[smoke] Warning: Failed to load config from {config_path}: {exc}", err=True)
+        return {}
+
+
+def _kill_process_on_port(port: int) -> None:
+    """Kill any process listening on the given port."""
+    try:
+        # Use lsof to find and kill process on port
+        result = subprocess.run(
+            ["lsof", "-ti", f":{port}"],
+            capture_output=True,
+            text=True,
+            timeout=2,
+        )
+        if result.stdout.strip():
+            pids = result.stdout.strip().split('\n')
+            for pid in pids:
+                try:
+                    subprocess.run(["kill", "-9", pid], timeout=2)
+                    click.echo(f"[smoke] Killed existing process {pid} on port {port}", err=True)
+                except Exception:
+                    pass
+            time.sleep(2.0)  # Give OS time to release port
+    except Exception as exc:
+        click.echo(f"[smoke] Warning: Could not check/kill port {port}: {exc}", err=True)
+
+
+def _start_task_app_server(
+    task_app_name: str,
+    port: int,
+    env_file: str | None,
+    force: bool
+) -> tuple[Any, str]:
+    """Start a task app server in the background using task-app serve.
+    
+    Returns (process, url) tuple.
+    """
+    import subprocess
+    import time as time_module
+    
+    # Build command using task-app serve (for TaskAppConfig-based apps)
+    cmd = [
+        "nohup",
+        "uvx", "synth-ai",
+        "task-app", "serve", task_app_name,
+        "--port", str(port),
+    ]
+    
+    if env_file:
+        cmd.extend(["--env-file", env_file])
+    
+    if force:
+        cmd.append("--force")
+    
+    # Resolve the synth-ai root directory
+    import synth_ai
+    synth_ai_root = Path(synth_ai.__file__).resolve().parent.parent
+    
+    click.echo(f"[smoke] Starting task app '{task_app_name}' on port {port}...", err=True)
+    click.echo(f"[smoke] Command: {' '.join(cmd)}", err=True)
+    click.echo(f"[smoke] Working directory: {synth_ai_root}", err=True)
+    
+    # nohup requires output redirection to a file
+    # Open file, start process, then close file handle so process is fully detached
+    # Run from synth-ai root so task app discovery works
+    nohup_log = Path(synth_ai_root) / "nohup_task_app.out"
+    
+    # Inherit SYNTH_QUIET environment variable to suppress patch messages
+    env = os.environ.copy()
+    if os.getenv("SYNTH_QUIET"):
+        env["SYNTH_QUIET"] = "1"
+    
+    with open(nohup_log, "w") as log_file:
+        proc = subprocess.Popen(
+            cmd,
+            stdout=log_file,
+            stderr=subprocess.STDOUT,
+            text=True,
+            cwd=str(synth_ai_root),
+            env=env,
+        )
+    # File is closed immediately so process is detached
+    
+    # Wait for server to be ready
+    url = f"http://localhost:{port}"
+    click.echo(f"[smoke] Waiting for task app to be ready at {url}...", err=True)
+    
+    import httpx
+    deadline = time.time() + 120.0  # Give it 2 minutes for initial setup
+    attempt = 0
+    last_log_line = None
+    while time.time() < deadline:
+        attempt += 1
+        try:
+            resp = httpx.get(f"{url}/health", timeout=1.0)
+            # Accept both 200 and 400 - 400 means server is up but auth is failing (which is fine for smoke test)
+            if resp.status_code in (200, 400):
+                click.echo(f"[smoke] Task app ready at {url} (status={resp.status_code})", err=True)
+                return proc, url
+        except Exception:
+            pass
+        
+        # Show polling progress every 5 seconds with last log line
+        if attempt % 10 == 0:
+            elapsed = int(time.time() - (deadline - 120.0))
+            # Try to read last line from nohup log
+            try:
+                if nohup_log.exists():
+                    with open(nohup_log) as f:
+                        lines = f.readlines()
+                        if lines:
+                            # Get last non-empty line
+                            for line in reversed(lines[-10:]):
+                                stripped = line.strip()
+                                if stripped and stripped != last_log_line:
+                                    last_log_line = stripped
+                                    # Truncate if too long
+                                    if len(stripped) > 80:
+                                        stripped = stripped[:77] + "..."
+                                    click.echo(f"[smoke] Waiting ({elapsed}s): {stripped}", err=True)
+                                    break
+                            else:
+                                click.echo(f"[smoke] Still waiting for task app... ({elapsed}s elapsed)", err=True)
+                        else:
+                            click.echo(f"[smoke] Still waiting for task app... ({elapsed}s elapsed)", err=True)
+            except Exception:
+                click.echo(f"[smoke] Still waiting for task app... ({elapsed}s elapsed)", err=True)
+        
+        # Check if process died
+        if proc.poll() is not None:
+            # Build a manual command that the user can copy-paste
+            manual_cmd_parts = ["uvx", "synth-ai", "task-app", "serve", task_app_name, "--port", str(port)]
+            if env_file:
+                manual_cmd_parts.extend(["--env-file", env_file])
+            if force:
+                manual_cmd_parts.append("--force")
+            
+            raise click.ClickException(
+                f"Task app '{task_app_name}' process exited unexpectedly (code={proc.returncode}). "
+                f"Check that the task app name is correct and .env has required keys. "
+                f"Try running manually: {' '.join(manual_cmd_parts)}"
+            )
+        
+        time_module.sleep(0.5)
+    
+    proc.kill()
+    raise click.ClickException("Task app failed to start within 120 seconds")
+
+
+def _start_sqld_server(
+    db_path: str,
+    hrana_port: int,
+    http_port: int
+) -> Any:
+    """Start sqld server in the background.
+    
+    Returns the process handle.
+    """
+    import shutil
+    import subprocess
+    
+    # Check if sqld is available
+    sqld_bin = shutil.which("sqld")
+    if not sqld_bin:
+        click.echo("[smoke] Warning: sqld not found in PATH, skipping auto-start", err=True)
+        click.echo("[smoke] Install sqld: brew install sqld", err=True)
+        return None
+    
+    # Ensure db directory exists
+    db_path_obj = Path(db_path).expanduser().resolve()
+    db_path_obj.parent.mkdir(parents=True, exist_ok=True)
+    
+    # Kill any existing processes on these ports
+    for port in [hrana_port, http_port]:
+        _kill_process_on_port(port)
+    
+    cmd = [
+        sqld_bin,
+        "--db-path", str(db_path_obj),
+        "--hrana-listen-addr", f"127.0.0.1:{hrana_port}",
+        "--http-listen-addr", f"127.0.0.1:{http_port}",
+    ]
+    
+    click.echo("[smoke] Starting sqld server...", err=True)
+    click.echo(f"[smoke] DB path: {db_path_obj}", err=True)
+    click.echo(f"[smoke] Hrana port: {hrana_port}, HTTP port: {http_port}", err=True)
+    click.echo(f"[smoke] Command: {' '.join(cmd)}", err=True)
+    
+    # Redirect to devnull to avoid process dying from pipe buffer issues
+    proc = subprocess.Popen(
+        cmd,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+        text=True,
+    )
+    
+    # Wait for server to be ready
+    health_url = f"http://127.0.0.1:{http_port}/health"
+    click.echo(f"[smoke] Waiting for sqld to be ready at {health_url}...", err=True)
+    
+    deadline = time.time() + 10.0
+    while time.time() < deadline:
+        try:
+            resp = httpx.get(health_url, timeout=0.5)
+            if resp.status_code == 200:
+                click.echo("[smoke] sqld ready", err=True)
+                # Set environment variables for tracing
+                os.environ["SQLD_DB_PATH"] = str(db_path_obj)
+                os.environ["SQLD_HTTP_PORT"] = str(hrana_port)
+                os.environ["LIBSQL_URL"] = f"http://127.0.0.1:{http_port}"
+                os.environ["SYNTH_TRACES_DB"] = f"http://127.0.0.1:{http_port}"
+                return proc
+        except Exception:
+            pass
+        
+        # Check if process died
+        if proc.poll() is not None:
+            click.echo(f"[smoke] Warning: sqld process exited with code {proc.returncode}", err=True)
+            return None
+        
+        time.sleep(0.2)
+    
+    click.echo("[smoke] Warning: sqld health check timed out, continuing anyway...", err=True)
+    return proc
+
+class MockRLTrainer:
+    """Minimal trainer emulator with a local FastAPI mock for GPT-5-Nano.
+
+    In ``synthetic`` mode it emits deterministic tool calls so the rollout can
+    progress without relying on external inference. In ``openai`` mode it acts
+    as a thin proxy around the real OpenAI chat completions endpoint (useful to
+    reproduce production behaviour locally).
+    """
+
+    def __init__(self, *, port: int = 0, backend: str = "synthetic") -> None:
+        self.port = port
+        self.backend = backend.lower().strip() or "synthetic"
+        self._server = None
+        self._task: asyncio.Task | None = None
+        self._openai_endpoint = os.getenv(
+            "SMOKE_OPENAI_ENDPOINT", "https://api.openai.com/v1/chat/completions"
+        )
+        self._openai_api_key = (
+            os.getenv("SMOKE_OPENAI_API_KEY") or os.getenv("OPENAI_API_KEY") or ""
+        )
+
+    def _build_app(self):
+        import json
+
+        from fastapi import Body, FastAPI
+        from fastapi.responses import JSONResponse
+
+        try:
+            logger = logging.getLogger(__name__)
+        except Exception:  # pragma: no cover - logging failures should not crash
+            logger = None
+
+        app = FastAPI()
+        backend = self.backend
+
+        @app.post("/v1/chat/completions")
+        async def chat_completions(body: dict = Body(...), cid: str | None = None):
+            log = logger or logging.getLogger("MockRLTrainer")
+            try:
+                msg_count = len(body.get("messages") or [])
+            except Exception:
+                msg_count = -1
+            click.echo(
+                f"[mock-rl] ← request backend={backend} model={body.get('model')} messages={msg_count} cid={cid}",
+                err=True,
+            )
+
+            # Explicit Body(...) avoids FastAPI interpreting parameters as query args
+            model = (body.get("model") or "gpt-5-nano")
+            messages = body.get("messages") or []
+            tools = body.get("tools") or []
+
+            # Decide whether to emit a tool call (to drive env steps) or plain text
+            emit_tool = False
+            tool_name = ""
+            for t in tools:
+                try:
+                    if (t or {}).get("type") == "function":
+                        fn = (t or {}).get("function") or {}
+                        name = (fn or {}).get("name") or ""
+                        if name:
+                            tool_name = name
+                            emit_tool = True
+                            break
+                except Exception:
+                    continue
+
+            # Simple heuristic actions to move/explore then interact
+            actions = ["move_right", "move_right", "move_down", "move_left", "do"]
+
+            correlation = cid
+
+            if backend == "openai":
+                if not self._openai_api_key:
+                    return JSONResponse(
+                        {
+                            "error": "OPENAI_API_KEY (or SMOKE_OPENAI_API_KEY) is required for mock backend 'openai'"
+                        },
+                        status_code=500,
+                    )
+                try:
+                    from examples.task_apps.crafter.task_app.synth_envs_hosted.inference.openai_client import (
+                        OpenAIClient as _HostedOpenAIClient,
+                    )
+
+                    hosted_client = _HostedOpenAIClient(
+                        base_url=self._openai_endpoint,
+                        api_key=self._openai_api_key,
+                    )
+                except Exception as exc:
+                    if logger is not None:
+                        logger.error("MockRLTrainer failed to import HostedOpenAIClient: %s", exc)
+                    return JSONResponse(
+                        {"error": f"OpenAI proxy unavailable: {exc}"},
+                        status_code=500,
+                    )
+
+                try:
+                    result = await hosted_client.generate_with_retries(  # type: ignore[attr-defined]
+                        request=body,
+                        base_url=self._openai_endpoint,
+                        max_retries=0,
+                    )
+                except Exception as exc:
+                    if logger is not None:
+                        logger.error("MockRLTrainer OpenAI generate failed: %s", exc)
+                    return JSONResponse(
+                        {"error": f"OpenAI proxy request failed: {exc}"},
+                        status_code=502,
+                    )
+
+                if isinstance(result, dict):
+                    data_typed = dict(result)
+                    synth_meta = data_typed.get("synth")
+                    if not isinstance(synth_meta, dict):
+                        synth_meta = {}
+                        data_typed["synth"] = synth_meta
+                    if correlation:
+                        synth_meta.setdefault("cid", correlation)
+
+                    # Fallback: if the upstream response failed to emit tool calls,
+                    # synthesize a deterministic action plan so the rollout can proceed.
+                    try:
+                        choices = data_typed.get("choices") or []
+                        first = choices[0] if choices else {}
+                        message = first.get("message") if isinstance(first, dict) else {}
+                        tc = message.get("tool_calls") if isinstance(message, dict) else None
+                        if not tc:
+                            if logger is not None:
+                                logger.warning(
+                                    "MockRLTrainer fallback: OpenAI returned no tool calls; injecting deterministic actions."
+                                )
+                            fallback_message = dict(message or {})
+                            fallback_message.setdefault("role", "assistant")
+                            fallback_message["content"] = ""
+                            fallback_message["tool_calls"] = [
+                                {
+                                    "id": f"call_{uuid.uuid4().hex[:8]}",
+                                    "type": "function",
+                                    "function": {
+                                        "name": tool_name or "interact_many",
+                                        "arguments": json.dumps({"actions": actions}),
+                                    },
+                                }
+                            ]
+                            fallback_message["function_call"] = {
+                                "name": tool_name or "interact_many",
+                                "arguments": json.dumps({"actions": actions}),
+                            }
+                            if choices:
+                                choices[0]["message"] = fallback_message
+                            else:
+                                data_typed["choices"] = [
+                                    {
+                                        "index": 0,
+                                        "message": fallback_message,
+                                        "finish_reason": "tool_calls",
+                                    }
+                                ]
+                    except Exception as exc:
+                        if logger is not None:
+                            logger.debug("MockRLTrainer fallback injection failed: %s", exc)
+
+                    tool_call_count = 0
+                    try:
+                        choices = data_typed.get("choices") or []
+                        first = choices[0] if choices else {}
+                        message = first.get("message") if isinstance(first, dict) else {}
+                        if isinstance(message, dict):
+                            tool_call_count = len(message.get("tool_calls") or [])
+                    except Exception:
+                        tool_call_count = 0
+
+                    log.info(
+                        "MockRLTrainer proxy returning response with %s tool calls (cid=%s)",
+                        tool_call_count,
+                        cid,
+                    )
+                    if tool_call_count == 0:
+                        log.error(
+                            "MockRLTrainer proxy still missing tool_calls after fallback injection (cid=%s)",
+                            cid,
+                        )
+                        click.echo(
+                            "[mock-rl] ✗ proxy response missing tool_calls; failing request", err=True
+                        )
+                    return JSONResponse(data_typed)
+                return JSONResponse(result)
+
+            if emit_tool:
+                # Emit BOTH legacy function_call and modern tool_calls for broad compatibility
+                message_payload = {
+                    "role": "assistant",
+                    "content": "",
+                    "function_call": {
+                        "name": tool_name,
+                        "arguments": json.dumps({"actions": actions}),
+                    },
+                    "tool_calls": [
+                        {
+                            "id": f"call_{uuid.uuid4().hex[:8]}",
+                            "type": "function",
+                            "function": {
+                                "name": tool_name,
+                                "arguments": json.dumps({"actions": actions}),
+                            },
+                        }
+                    ],
+                }
+                finish_reason = "tool_calls"
+            else:
+                # Fallback: echo last user content as plain text
+                click.echo(
+                    f"[mock-rl] ! no tool schema supplied; returning text response (cid={cid})",
+                    err=True,
+                )
+                log.warning(
+                    "MockRLTrainer received request without tool schema; responding with text content (cid=%s)",
+                    cid,
+                )
+                last_user = next((m.get("content", "") for m in reversed(messages) if m.get("role") == "user"), "")
+                text = (last_user or "").strip()
+                if len(text) > 160:
+                    text = text[:160] + "..."
+                message_payload = {"role": "assistant", "content": f"MOCK(gpt-5-nano): {text or 'ack'}"}
+                finish_reason = "stop"
+
+            response = {
+                "id": f"cmpl_{uuid.uuid4().hex[:12]}",
+                "object": "chat.completion",
+                "created": int(asyncio.get_event_loop().time()),
+                "model": model,
+                "choices": [{"index": 0, "message": message_payload, "finish_reason": finish_reason}],
+                "usage": {"prompt_tokens": 32, "completion_tokens": 16, "total_tokens": 48},
+                "synth": {"cid": correlation},
+            }
+            if finish_reason == "tool_calls":
+                # Type-safe extraction of tool call count
+                tc = 0
+                try:
+                    choices = response.get("choices")
+                    if isinstance(choices, list) and choices:
+                        first_choice = choices[0]
+                        if isinstance(first_choice, dict):
+                            msg = first_choice.get("message")
+                            if isinstance(msg, dict):
+                                tool_calls = msg.get("tool_calls")
+                                if isinstance(tool_calls, list):
+                                    tc = len(tool_calls)
+                except Exception:
+                    pass
+                log.debug(
+                    "MockRLTrainer synthetic response emitting %s tool calls (cid=%s)",
+                    tc,
+                    cid,
+                )
+                assert tc > 0, "MockRLTrainer synthetic response missing tool_calls"
+                click.echo(
+                    f"[mock-rl] → response tool_calls={tc} backend={backend} cid={cid}",
+                    err=True,
+                )
+            else:
+                click.echo(
+                    f"[mock-rl] → response finish_reason={finish_reason} backend={backend} cid={cid}",
+                    err=True,
+                )
+            return JSONResponse(response)
+
+        return app
+
+    async def start(self) -> None:
+        import socket
+
+        import uvicorn
+
+        def _allocate_port() -> int:
+            nonlocal socket
+            if self.port:
+                with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as probe:
+                    probe.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+                    try:
+                        probe.bind(("127.0.0.1", self.port))
+                        return self.port
+                    except OSError:
+                        pass
+            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as probe:
+                probe.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+                probe.bind(("127.0.0.1", 0))
+                self.port = probe.getsockname()[1]
+                return self.port
+
+        retries = 0
+        while True:
+            selected_port = _allocate_port()
+            config = uvicorn.Config(
+                self._build_app(),
+                host="127.0.0.1",
+                port=selected_port,
+                log_level="warning",
+            )
+            self._server = uvicorn.Server(config)
+            self._task = asyncio.create_task(self._server.serve())
+
+            for _ in range(100):
+                if getattr(self._server, "started", False):
+                    break
+                if self._task.done():
+                    break
+                await asyncio.sleep(0.05)
+
+            if getattr(self._server, "started", False):
+                try:
+                    logging.getLogger(__name__).info(
+                        "MockRLTrainer started on http://127.0.0.1:%s (backend=%s)",
+                        self.port,
+                        self.backend,
+                    )
+                    click.echo(
+                        f"[mock-rl] server ready http://127.0.0.1:{self.port} backend={self.backend}",
+                        err=True,
+                    )
+                except Exception:
+                    pass
+                return
+
+            # Startup failed; stop server and retry on a new port if possible
+            await self.stop()
+            if retries >= 5:
+                raise RuntimeError("MockRLTrainer failed to start after multiple attempts")
+            self.port = 0
+            retries += 1
+
+    async def stop(self) -> None:
+        if self._server is not None:
+            self._server.should_exit = True
+        if self._task is not None:
+            with contextlib.suppress(Exception):
+                await asyncio.wait_for(self._task, timeout=2.0)
+        self._task = None
+        self._server = None
+        click.echo("[mock-rl] server stopped", err=True)
+
+async def _run_smoke_async(
+    *,
+    task_app_url: str,
+    api_key: str | None,
+    env_name_opt: str | None,
+    policy_name: str,
+    model: str,
+    inference_url_opt: str | None,
+    inference_policy: str | None,
+    max_steps: int,
+    return_trace: bool,
+    use_mock: bool,
+    mock_port: int,
+    mock_backend: str,
+    config_path: Path | None,
+    rollouts: int = 1,
+    group_size: int = 1,
+    batch_size: int | None = None,
+) -> int:
+    # If config is provided, derive defaults (URL/env/model)
+    cfg: Any | None = None
+    if config_path is not None:
+        try:
+            from synth_ai.api.train.configs.rl import (
+                RLConfig as _RLConfig,  # lazy import to avoid heavy deps when unused
+            )
+            cfg = _RLConfig.from_path(config_path)
+        except Exception as exc:
+            click.echo(f"Failed to load RL config {config_path}: {exc}", err=True)
+            return 2
+
+        # Prefer explicit CLI --url; only use config services.task_url if URL not provided
+        try:
+            if not task_app_url and cfg.services and getattr(cfg.services, "task_url", None):
+                task_app_url = cfg.services.task_url
+        except Exception:
+            pass
+        # Fill env and model if not explicitly set
+        try:
+            if not env_name_opt and cfg.rollout and getattr(cfg.rollout, "env_name", None):
+                env_name_opt = cfg.rollout.env_name
+        except Exception:
+            pass
+        try:
+            if model == "gpt-5-nano":
+                # Prefer smoke config model over policy model for smoke tests
+                smoke_cfg = getattr(cfg, "smoke", None)
+                smoke_model = None
+                if smoke_cfg and hasattr(smoke_cfg, "model"):
+                    smoke_model = smoke_cfg.model
+                if smoke_model:
+                    model = str(smoke_model).strip()
+                elif cfg.policy:
+                    if getattr(cfg.policy, "model_name", None):
+                        model = str(cfg.policy.model_name).strip()
+                    elif getattr(cfg.policy, "source", None):
+                        model = str(cfg.policy.source).strip()
+                elif cfg.model and getattr(cfg.model, "source", None):
+                    model = str(cfg.model.source).strip()
+                elif cfg.model and getattr(cfg.model, "base", None):
+                    model = str(cfg.model.base).strip()
+        except Exception:
+            pass
+
+    base = validate_task_app_url(task_app_url)
+    mock_backend = (mock_backend or "synthetic").strip().lower()
+
+    # Discover environment if not provided
+    async with TaskAppClient(base_url=base, api_key=api_key) as client:
+        # Probe basic info quickly
+        try:
+            _ = await client.health()
+        except Exception:
+            click.echo("Auth or connectivity check failed on /health. If this endpoint requires a key, pass --api-key or set ENVIRONMENT_API_KEY.", err=True)
+            # Continue; rollout may still clarify the error
+
+        # Fetch a sample task instance to infer environment name if not provided
+        env_name = env_name_opt
+        if not env_name:
+            try:
+                ti = await client.task_info(seeds=[0])
+                # task_info returns TaskInfo or list[TaskInfo]; normalize
+                info: Any = ti[0] if isinstance(ti, list) else ti
+                env_name = getattr(info, "environment", None) or getattr(info, "task", {}).get("name")  # type: ignore[attr-defined]
+            except Exception:
+                env_name = None
+        if not env_name:
+            click.echo("Could not infer environment name; pass --env-name.", err=True)
+            return 2
+
+        # Build ops: alternating agent/env for max_steps
+        ops: list[str] = []
+        for _ in range(max_steps):
+            ops.append("agent")
+            ops.append("env")
+
+        # Inference URL: user override > preset > local mock > Synth API default
+        synth_base = (os.getenv("SYNTH_API_BASE") or os.getenv("SYNTH_BASE_URL") or "https://api.synth.run").rstrip("/")
+        # Avoid double '/api' if base already includes it
+        if synth_base.endswith("/api"):
+            default_infer = f"{synth_base}/inference/v1/chat/completions"
+        else:
+            default_infer = f"{synth_base}/api/inference/v1/chat/completions"
+
+        # Helper to execute one or more rollouts and return exit code
+        async def __do_rollouts(inference_url_raw: str) -> int:
+            successes = 0
+            total_steps = 0
+            nonzero_returns = 0
+            v3_traces = 0
+
+            # Derive sampling params from config if present
+            sampling: dict[str, Any] = {}
+            try:
+                if cfg and cfg.policy:
+                    if getattr(cfg.policy, "temperature", None) is not None:
+                        sampling["temperature"] = cfg.policy.temperature
+                    if getattr(cfg.policy, "top_p", None) is not None:
+                        sampling["top_p"] = cfg.policy.top_p
+                    if getattr(cfg.policy, "max_tokens", None) is not None:
+                        sampling["max_tokens"] = cfg.policy.max_tokens
+            except Exception:
+                pass
+
+            num_outer = batch_size if (batch_size is not None and batch_size > 0) else max(1, int(rollouts))
+            for i in range(num_outer):
+                for g in range(max(1, int(group_size))):
+                    if inference_url_raw.startswith("/"):
+                        inference_url_abs = f"{base}{inference_url_raw}"
+                    else:
+                        inference_url_abs = inference_url_raw
+                    inference_url_norm = normalize_inference_url(inference_url_abs)
+                    correlation_id = f"smoke-{uuid.uuid4()}"
+                    inference_url_with_cid = _append_query_param(inference_url_norm, "cid", correlation_id)
+
+                    run_id = correlation_id
+                    policy_cfg: dict[str, Any] = {
+                        "model": model,
+                        "inference_url": inference_url_with_cid,
+                    }
+                    if sampling:
+                        policy_cfg.update(sampling)
+
+                    request = RolloutRequest(
+                        run_id=run_id,
+                        env=RolloutEnvSpec(env_name=env_name, config={}, seed=i),
+                        policy=RolloutPolicySpec(policy_name=policy_name, config=policy_cfg),
+                        ops=ops,
+                        record=RolloutRecordConfig(
+                            trajectories=True,
+                            logprobs=False,
+                            value=False,
+                            return_trace=return_trace,
+                            trace_format=("structured" if return_trace else "compact"),
+                        ),
+                        on_done="reset",
+                        safety=RolloutSafetyConfig(max_ops=max_steps * 4, max_time_s=900.0),
+                        training_session_id=None,
+                        synth_base_url=synth_base,
+                        mode=RolloutMode.RL,
+                    )
+
+                    try:
+                        click.echo(f">> POST /rollout run_id={run_id} env={env_name} policy={policy_name} url={inference_url_with_cid}")
+                        click.echo(f"   ops={ops[:10]}{'...' if len(ops) > 10 else ''}")
+                        response = await client.rollout(request)
+                    except Exception as exc:
+                        click.echo(f"Rollout[{i}:{g}] failed: {type(exc).__name__}: {exc}", err=True)
+                        import traceback
+                        click.echo(f"Traceback: {traceback.format_exc()}", err=True)
+                        continue
+
+                    successes += 1
+                    try:
+                        validate_rollout_response_for_rl(response.model_dump())
+                    except Exception as vexc:
+                        click.echo(f"  ⚠ RL response validation warning: {vexc}", err=True)
+
+                    pm = response.pipeline_metadata or {}
+                    inferred_url = pm.get("inference_url") if isinstance(pm, dict) else None
+                    metrics = response.metrics
+                    if inferred_url:
+                        click.echo(f"  rollout[{i}:{g}] inference_url: {inferred_url}")
+                    click.echo(f"  rollout[{i}:{g}] episodes={metrics.num_episodes} steps={metrics.num_steps} mean_return={metrics.mean_return:.4f}")
+
+                    total_steps += int(metrics.num_steps)
+                    if (metrics.mean_return or 0.0) != 0.0:
+                        nonzero_returns += 1
+                    if response.trace is not None and isinstance(response.trace, dict):
+                        v3_traces += 1
+
+                    if i == 0 and g == 0:
+                        try:
+                            traj0 = response.trajectories[0]
+                            step_meta_url = None
+                            for step in traj0.steps:
+                                info = getattr(step, "info", None) or {}
+                                meta = info.get("meta") if isinstance(info, dict) else None
+                                if isinstance(meta, dict) and meta.get("inference_url"):
+                                    step_meta_url = meta.get("inference_url")
+                                    break
+                            if step_meta_url:
+                                click.echo(f"  step.meta.inference_url: {str(step_meta_url)[:120]}...")
+                        except Exception:
+                            pass
+
+                    try:
+                        try:
+                            metrics_dump = response.metrics.model_dump()
+                        except Exception:
+                            metrics_dump = {
+                                "episode_returns": getattr(response.metrics, "episode_returns", None),
+                                "mean_return": getattr(response.metrics, "mean_return", None),
+                                "num_steps": getattr(response.metrics, "num_steps", None),
+                                "num_episodes": getattr(response.metrics, "num_episodes", None),
+                                "outcome_score": getattr(response.metrics, "outcome_score", None),
+                                "events_score": getattr(response.metrics, "events_score", None),
+                            }
+                        click.echo("  reward.info (metrics): " + str(metrics_dump))
+
+                        try:
+                            traj = response.trajectories[0]
+                            step_rewards = []
+                            all_achievements = set()
+                            for st in getattr(traj, "steps", []) or []:
+                                try:
+                                    step_rewards.append(getattr(st, "reward", None))
+                                except Exception:
+                                    step_rewards.append(None)
+                                # Extract achievements from step info
+                                try:
+                                    step_info = getattr(st, "info", None)
+                                    if isinstance(step_info, dict):
+                                        achievements_status = step_info.get("achievements_status")
+                                        if isinstance(achievements_status, dict):
+                                            for ach_name, ach_val in achievements_status.items():
+                                                if ach_val:
+                                                    all_achievements.add(str(ach_name))
+                                except Exception:
+                                    pass
+                            click.echo("  reward.per_step: " + str(step_rewards))
+                            if all_achievements:
+                                click.echo(f"  achievements: {sorted(all_achievements)}")
+                            else:
+                                click.echo("  achievements: none")
+                        except Exception:
+                            pass
+
+                        # Extract and display tool calls from v3 trace
+                        # 
+                        # IMPORTANT: Tool calls are extracted from the structured v3 trace format.
+                        # The trace must be requested with return_trace=True for this to work.
+                        # 
+                        # Trace structure:
+                        #   trace.event_history[] - list of events (policy calls, env steps)
+                        #     ├─ event.call_records[] - LLM calls made during this event
+                        #        ├─ call_record.output_tool_calls[] - tool calls from LLM response
+                        #           ├─ tool_call.name - function name (e.g., "interact_many")
+                        #           └─ tool_call.arguments_json - JSON string of arguments
+                        #
+                        # This provides visibility into what actions the policy is taking,
+                        # which is critical for debugging RL training issues.
+                        tr = response.trace if isinstance(response.trace, dict) else None
+                        if tr:
+                            event_history = tr.get("event_history", [])
+                            tool_call_count = 0
+                            
+                            # Extract tool calls from event_history call_records
+                            if event_history and isinstance(event_history, list):
+                                for event in event_history:
+                                    if not isinstance(event, dict):
+                                        continue
+                                    # Policy events contain call_records with LLM interactions
+                                    call_records = event.get("call_records")
+                                    if call_records and isinstance(call_records, list):
+                                        for call_record in call_records:
+                                            if isinstance(call_record, dict):
+                                                # Extract tool calls from this LLM call
+                                                output_tool_calls = call_record.get("output_tool_calls", [])
+                                                if output_tool_calls and isinstance(output_tool_calls, list):
+                                                    for tc in output_tool_calls:
+                                                        if isinstance(tc, dict):
+                                                            fn_name = tc.get("name", "unknown")
+                                                            fn_args = tc.get("arguments_json", "{}")
+                                                            # Display tool call with truncated args for readability
+                                                            click.echo(f"  TOOL_CALL[{tool_call_count}]: {fn_name}({fn_args[:100]}{'...' if len(fn_args) > 100 else ''})")
+                                                            tool_call_count += 1
+                            
+                            if tool_call_count > 0:
+                                click.echo(f"  ✓ {tool_call_count} tool calls executed")
+                            else:
+                                # No tool calls found - might indicate:
+                                # 1. return_trace=False (trace not requested)
+                                # 2. Policy didn't make tool calls (unlikely for most RL tasks)
+                                # 3. Trace format mismatch (structure changed)
+                                click.echo("  ⚠ No tool calls found in trace")
+                        else:
+                            click.echo("  ⚠ Trace not available")
+                    except Exception as e:
+                        click.echo(f"  trace error: {e}", err=True)
+
+            click.echo("✓ Smoke rollouts complete")
+            denom = num_outer * max(1, int(group_size))
+            click.echo(f"  successes={successes}/{denom} total_steps={total_steps} v3_traces={v3_traces}/{denom} nonzero_returns={nonzero_returns}/{denom}")
+
+            if successes == 0:
+                click.echo("  ⚠ All rollouts failed", err=True)
+                return 3
+            if v3_traces < successes:
+                click.echo("  ⚠ Some rollouts missing v3 traces (trace field)", err=True)
+            if total_steps == 0:
+                click.echo("  ⚠ No steps executed; check ops/policy config", err=True)
+
+            return 0
+
+        # Initialize to default; policy/flags may override below
+        inference_url_raw = inference_url_opt or default_infer
+        mock: MockRLTrainer | None = None
+        preset = (inference_policy or "").strip().lower()
+
+        # Respect explicit preset overrides
+        if preset == "mock":
+            use_mock = True
+        elif preset == "gpt-5-nano":
+            if not inference_url_opt:
+                inference_url_raw = default_infer
+            if not model:
+                model = "gpt-5-nano"
+        elif preset == "openai":
+            inference_url_raw = "https://api.openai.com/v1/chat/completions"
+        elif preset == "groq":
+            inference_url_raw = "https://api.groq.com/openai/v1/chat/completions"
+
+        # Start mock proxy only when explicitly requested
+        if use_mock:
+            backend_choice = mock_backend
+            if backend_choice == "openai" and not (
+                os.getenv("SMOKE_OPENAI_API_KEY") or os.getenv("OPENAI_API_KEY")
+            ):
+                click.echo(
+                    "  ⚠ OPENAI_API_KEY not configured; falling back to synthetic mock.",
+                    err=True,
+                )
+                backend_choice = "synthetic"
+            mock = MockRLTrainer(port=mock_port, backend=backend_choice)
+            await mock.start()
+            inference_url_raw = f"http://127.0.0.1:{mock.port}"
+
+        try:
+            result = await __do_rollouts(inference_url_raw)
+        finally:
+            if mock is not None:
+                with contextlib.suppress(Exception):
+                    await mock.stop()
+        return result
+async def _run_train_step(
+    *,
+    task_app_url: str,
+    api_key: str | None,
+    env_name_opt: str | None,
+    policy_name: str,
+    model: str,
+    inference_policy: str | None,
+    inference_url_opt: str | None,
+    max_steps: int,
+    return_trace: bool,
+    use_mock: bool,
+    mock_backend: str,
+    mock_port: int,
+    config_path: Path | None,
+    parallel: int,
+) -> int:
+    import time
+    start = time.perf_counter()
+
+    async def one(seed_idx: int) -> dict[str, Any]:
+        t0 = time.perf_counter()
+        try:
+            code = await _run_smoke_async(
+                task_app_url=task_app_url,
+                api_key=api_key,
+                env_name_opt=env_name_opt,
+                policy_name=policy_name,
+                model=model,
+                inference_policy=inference_policy,
+                inference_url_opt=inference_url_opt,
+                max_steps=max_steps,
+                return_trace=return_trace,
+                use_mock=use_mock,
+                mock_backend=mock_backend,
+                mock_port=mock_port,
+                config_path=config_path,
+                rollouts=1,
+                group_size=1,
+                batch_size=None,
+            )
+            wall_ms = (time.perf_counter() - t0) * 1000.0
+            return {"exit": int(code), "wall_ms": wall_ms}
+        except Exception as e:
+            wall_ms = (time.perf_counter() - t0) * 1000.0
+            return {"exit": 99, "wall_ms": wall_ms, "error": f"{type(e).__name__}: {e}"}
+
+    # Launch N rollouts concurrently
+    tasks = [one(i) for i in range(max(1, int(parallel)))]
+    results = await asyncio.gather(*tasks, return_exceptions=False)
+    total_wall_ms = (time.perf_counter() - start) * 1000.0
+
+    # Print summary
+    def _exit_code(result: dict[str, Any]) -> int:
+        value = result.get("exit")
+        if isinstance(value, (int, float)):
+            return int(value)
+        if isinstance(value, str) and value.strip():
+            try:
+                return int(value.strip())
+            except ValueError:
+                return 1
+        return 1
+
+    successes = sum(1 for r in results if _exit_code(r) == 0)
+    avg_wall = sum(float(r.get("wall_ms", 0.0)) for r in results) / max(len(results), 1)
+    click.echo("✓ Train-step emulation complete")
+    click.echo(f"  parallel={parallel} successes={successes}/{len(results)} total_wall_ms={total_wall_ms:.1f} avg_rollout_wall_ms={avg_wall:.1f}")
+
+    # Show brief failure codes to aid diagnosis
+    if successes < len(results):
+        codes: dict[int, int] = {}
+        for r in results:
+            if not isinstance(r, dict):
+                continue
+            c = _exit_code(r)
+            codes[c] = codes.get(c, 0) + 1
+        click.echo(f"  failure_codes={codes}")
+
+    return 0 if successes == len(results) else 3
+
+
+@click.command("smoke")
+@click.option("--url", "task_app_url", type=str, default=lambda: os.getenv("TASK_APP_URL", "http://localhost:8765"), help="Task app base URL.")
+@click.option(
+    "--api-key",
+    type=str,
+    default=lambda: os.getenv("ENVIRONMENT_API_KEY", ""),
+    envvar="ENVIRONMENT_API_KEY",
+    help="Environment API key (X-API-Key).",
+)
+@click.option("--env-name", type=str, default=None, help="Environment name to roll out (auto-detected if possible).")
+@click.option("--policy-name", type=str, default="react", help="Policy name to pass to task app.")
+@click.option("--model", type=str, default="gpt-5-nano", help="Model id to route in inference payload.")
+@click.option(
+    "--policy",
+    "inference_policy",
+    type=click.Choice(["mock", "gpt-5-nano", "openai", "groq"], case_sensitive=False),
+    default=None,
+    help="Inference route preset (mock, gpt-5-nano via Synth, OpenAI or Groq).",
+)
+@click.option("--inference-url", type=str, default=None, help="Override inference URL (default: Synth API chat completions).")
+@click.option("--max-steps", type=int, default=3, show_default=True, help="Number of agent/env step pairs.")
+@click.option("--return-trace", is_flag=True, help="Request v3 trace in response if supported.")
+@click.option("--use-mock/--no-mock", default=True, show_default=True, help="Use local mock inference server (GPT-5-Nano emulation).")
+@click.option(
+    "--mock-backend",
+    type=click.Choice(["synthetic", "openai"], case_sensitive=False),
+    default="synthetic",
+    show_default=True,
+    help="Mock inference backend: synthetic deterministic tooling or OpenAI passthrough.",
+)
+@click.option("--mock-port", type=int, default=0, show_default=True, help="Port for local mock inference server (0 = auto).")
+@click.option("--config", type=click.Path(exists=True, dir_okay=False, path_type=Path), default=None, help="RL TOML config to derive URL/env/model.")
+@click.option("--env-file", type=click.Path(exists=True, dir_okay=False, path_type=Path), default=None, help="Path to .env to load before running.")
+@click.option("--rollouts", type=int, default=1, show_default=True, help="Number of rollouts (seeds 0..N-1).")
+@click.option("--group-size", type=int, default=1, show_default=True, help="Completions per seed to emulate GRPO grouping.")
+@click.option("--batch-size", type=int, default=None, help="Alias for rollouts; when set, overrides --rollouts.")
+@click.option(
+    "--parallel",
+    type=int,
+    default=0,
+    show_default=True,
+    help="Emulate a train step by running this many rollouts concurrently (0 = sequential).",
+)
+def command(
+    task_app_url: str,
+    api_key: str,
+    env_name: str | None,
+    policy_name: str,
+    model: str,
+    inference_policy: str | None,
+    inference_url: str | None,
+    max_steps: int,
+    return_trace: bool,
+    use_mock: bool,
+    mock_backend: str,
+    mock_port: int,
+    config: Path | None,
+    env_file: Path | None,
+    rollouts: int,
+    group_size: int,
+    batch_size: int | None,
+    parallel: int,
+) -> None:
+    """Smoke-test a Task App by emulating a trainer rollout using GPT-5-Nano.
+
+    This command posts a minimal RL rollout to the task app, with a valid
+    OpenAI-compatible inference URL including a trace correlation id, and
+    validates that the response contains the fields required by the RL trainer
+    (e.g. pipeline_metadata.inference_url and per-step info.meta.inference_url).
+    
+    If --config is provided, loads settings from the [smoke] section in the TOML file.
+    CLI arguments override TOML values.
+    """
+    
+    # Load [smoke] section from TOML if config is provided
+    smoke_config = _load_smoke_config(config)
+    
+    # Track background processes for cleanup
+    background_procs: list[Any] = []
+    
+    try:
+        # Auto-start sqld if configured
+        if smoke_config.get("sqld_auto_start"):
+            sqld_db_path = smoke_config.get("sqld_db_path", "./traces/local.db")
+            sqld_hrana_port = smoke_config.get("sqld_hrana_port", 8080)
+            sqld_http_port = smoke_config.get("sqld_http_port", 8081)
+            
+            sqld_proc = _start_sqld_server(
+                db_path=sqld_db_path,
+                hrana_port=sqld_hrana_port,
+                http_port=sqld_http_port,
+            )
+            if sqld_proc:
+                background_procs.append(("sqld", sqld_proc))
+        
+        # Auto-start task app if configured
+        task_app_override_url = None
+        if smoke_config.get("task_app_name"):
+            task_app_name = smoke_config["task_app_name"]
+            task_app_port = smoke_config.get("task_app_port", 8765)
+            task_app_env_file = smoke_config.get("task_app_env_file")
+            task_app_force = smoke_config.get("task_app_force", True)
+            
+            task_app_proc, task_app_url = _start_task_app_server(
+                task_app_name=task_app_name,
+                port=task_app_port,
+                env_file=task_app_env_file,
+                force=task_app_force,
+            )
+            background_procs.append(("task_app", task_app_proc))
+            task_app_override_url = task_app_url
+            click.echo(f"[smoke] Task app started, will use URL: {task_app_url}", err=True)
+    except Exception as exc:
+        # Cleanup any processes that did start
+        for proc_name, proc in background_procs:
+            if proc and proc.poll() is None:
+                click.echo(f"[smoke] Cleaning up {proc_name}...", err=True)
+                proc.terminate()
+                try:
+                    proc.wait(timeout=3)
+                except Exception:
+                    proc.kill()
+        
+        click.echo(f"[smoke] ERROR: Auto-start failed: {exc}", err=True)
+        raise click.ClickException(f"Auto-start failed: {exc}") from exc
+    
+    # Apply TOML defaults (CLI args take precedence)
+    # Override task_url with auto-started task app URL if applicable
+    if task_app_override_url:
+        task_app_url = task_app_override_url
+    # For string/int args: use TOML value if CLI value matches the default
+    ctx = click.get_current_context()
+    
+    # Helper to check if a CLI param was explicitly provided or is using default
+    def use_toml_default(param_name: str, cli_value: Any, toml_key: str) -> Any:
+        """Use TOML value if CLI param is at its default, otherwise use CLI value."""
+        if not smoke_config or toml_key not in smoke_config:
+            return cli_value
+        
+        param = next((p for p in ctx.command.params if p.name == param_name), None)
+        if not param:
+            return cli_value
+        
+        # Check if value was explicitly provided (not default)
+        # If it matches the default, use TOML value
+        param_default = param.default() if callable(param.default) else param.default
+        if cli_value == param_default:
+            toml_value = smoke_config[toml_key]
+            click.echo(f"[smoke] Using {toml_key}={toml_value} from config", err=True)
+            return toml_value
+        
+        return cli_value
+    
+    # Apply TOML defaults
+    task_app_url = use_toml_default("task_app_url", task_app_url, "task_url")
+    env_name = use_toml_default("env_name", env_name, "env_name")
+    policy_name = use_toml_default("policy_name", policy_name, "policy_name")
+    model = use_toml_default("model", model, "model")
+    inference_policy = use_toml_default("inference_policy", inference_policy, "policy")
+    inference_url = use_toml_default("inference_url", inference_url, "inference_url")
+    max_steps = use_toml_default("max_steps", max_steps, "max_steps")
+    return_trace = use_toml_default("return_trace", return_trace, "return_trace")
+    use_mock = use_toml_default("use_mock", use_mock, "use_mock")
+    mock_backend = use_toml_default("mock_backend", mock_backend, "mock_backend")
+    mock_port = use_toml_default("mock_port", mock_port, "mock_port")
+    api_key = use_toml_default("api_key", api_key, "api_key")
+
+    # Auto-configure tracing to avoid interactive prompts
+    try:
+        os.environ.setdefault("CI", "true")
+        os.environ.setdefault("SYNTH_TRACING_AUTO_YES", "1")
+        # Derive a default traces directory relative to CWD
+        traces_dir = os.environ.get("SYNTH_TRACES_DIR")
+        if not traces_dir:
+            traces_dir = str((Path.cwd() / "traces" / "v3").resolve())
+            os.environ["SYNTH_TRACES_DIR"] = traces_dir
+        with contextlib.suppress(Exception):
+            Path(traces_dir).mkdir(parents=True, exist_ok=True)
+        _ensure_local_libsql()
+        # Prefer a libsql/turso/sqld URL when provided to enable concurrent writes
+        libsql_url = (
+            os.getenv("TRACING_DB_URL")
+            or os.getenv("LIBSQL_URL")
+            or os.getenv("TURSO_DATABASE_URL")
+            or os.getenv("LIBSQL_HTTP_URL")
+        )
+        if libsql_url:
+            os.environ.setdefault("LIBSQL_URL", libsql_url)
+
+        auth_hint = (
+            os.getenv("TRACING_DB_AUTH_TOKEN")
+            or os.getenv("LIBSQL_AUTH_TOKEN")
+            or os.getenv("TURSO_AUTH_TOKEN")
+        )
+        if auth_hint:
+            os.environ.setdefault("LIBSQL_AUTH_TOKEN", auth_hint)
+
+        resolved_url, resolved_token = resolve_trace_db_settings()
+        os.environ.setdefault("SYNTH_TRACES_DB", resolved_url)
+        if resolved_token and not (
+            os.getenv("LIBSQL_AUTH_TOKEN") or os.getenv("TURSO_AUTH_TOKEN")
+        ):
+            os.environ["LIBSQL_AUTH_TOKEN"] = resolved_token
+
+        _refresh_tracing_config()
+    except Exception:
+        pass
+
+    # Load env file(s) before resolving API key
+    try:
+        # Explicit --env-file takes precedence
+        if env_file is not None:
+            try:
+                from dotenv import load_dotenv as _ld
+                _ld(env_file, override=False)
+            except Exception:
+                pass
+        else:
+            # Best-effort auto-discovery from CWD
+            try:
+                from dotenv import find_dotenv as _fd
+                from dotenv import load_dotenv as _ld
+                _ld(_fd(usecwd=True), override=False)
+            except Exception:
+                pass
+
+        # If api_key not passed, try to read from env now
+        if not api_key:
+            api_key = os.getenv("ENVIRONMENT_API_KEY", "")
+    except Exception:
+        pass
+
+    try:
+        if parallel and parallel > 0:
+            exit_code = asyncio.run(
+                _run_train_step(
+                    task_app_url=task_app_url,
+                    api_key=(api_key or None),
+                    env_name_opt=env_name,
+                    policy_name=policy_name,
+                    model=model,
+                    inference_policy=inference_policy,
+                    inference_url_opt=inference_url,
+                    max_steps=max_steps,
+                    return_trace=return_trace,
+                    use_mock=use_mock,
+                    mock_backend=mock_backend,
+                    mock_port=mock_port,
+                    config_path=config,
+                    parallel=parallel,
+                )
+            )
+        else:
+            exit_code = asyncio.run(
+                _run_smoke_async(
+                    task_app_url=task_app_url,
+                    api_key=(api_key or None),
+                    env_name_opt=env_name,
+                    policy_name=policy_name,
+                    model=model,
+                    inference_policy=inference_policy,
+                    inference_url_opt=inference_url,
+                    max_steps=max_steps,
+                    return_trace=return_trace,
+                    use_mock=use_mock,
+                    mock_backend=mock_backend,
+                    mock_port=mock_port,
+                    config_path=config,
+                    rollouts=rollouts,
+                    group_size=group_size,
+                    batch_size=batch_size,
+                )
+            )
+    except KeyboardInterrupt:
+        click.echo("Interrupted", err=True)
+        sys.exit(130)
+    finally:
+        # Cleanup background processes
+        for proc_name, proc in background_procs:
+            if proc and proc.poll() is None:
+                click.echo(f"[smoke] Stopping {proc_name}...", err=True)
+                proc.terminate()
+                try:
+                    proc.wait(timeout=5)
+                except Exception:
+                    proc.kill()
+        if background_procs:
+            click.echo("[smoke] Background services stopped", err=True)
+    
+    sys.exit(exit_code)
+
+
+def register(cli: click.Group) -> None:
+    cli.add_command(command)