synth-ai 0.2.9.dev5__tar.gz → 0.2.9.dev7__tar.gz

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/MANIFEST.in

@@ -1,6 +1,5 @@
 include README.md
 include LICENSE
-include synth_ai/install_sqld.sh
 recursive-include synth_ai *.py
 # Only include lightweight config/data files from core packages
 recursive-include synth_ai *.toml
@@ -33,4 +32,4 @@ recursive-exclude synth_ai *.zip
 recursive-exclude synth_ai *.tar
 global-exclude *.pyc
 global-exclude __pycache__
-global-exclude .DS_Store
+global-exclude .DS_Store

{synth_ai-0.2.9.dev5/synth_ai.egg-info → synth_ai-0.2.9.dev7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: synth-ai
-Version: 0.2.9.dev5
+Version: 0.2.9.dev7
 Summary: RL as a service SDK - Core AI functionality and tracing
 Author-email: Synth AI <josh@usesynth.ai>
 License-Expression: MIT

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/common_old/backend.py

@@ -18,4 +18,3 @@ def resolve_backend_url() -> str:
 
 if __name__ == "__main__":
     print(resolve_backend_url())
-

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/crafter_debug_render.py

@@ -10,6 +10,7 @@ Quick local Crafter observation inspector.
 Run:
   uv run python examples/crafter_debug_render.py --base-url http://localhost:8901 --seed 1
 """
+
 import argparse
 import math
 import os
@@ -117,13 +118,17 @@ def format_semantic_map_view(obs: Dict[str, Any], view_size: int = 7) -> str:
                 row_cells.append("void")
         lines.append(" ".join(row_cells))
 
-    legend = f"Visible items: {', '.join(sorted(visible))}" if visible else "No notable items visible"
+    legend = (
+        f"Visible items: {', '.join(sorted(visible))}" if visible else "No notable items visible"
+    )
     return "\n".join(lines) + "\n" + legend
 
 
 async def main():
     parser = argparse.ArgumentParser()
-    parser.add_argument("--base-url", default=os.getenv("CRAFTER_BASE_URL", "http://localhost:8901"))
+    parser.add_argument(
+        "--base-url", default=os.getenv("CRAFTER_BASE_URL", "http://localhost:8901")
+    )
     parser.add_argument("--seed", type=int, default=1)
     args = parser.parse_args()
 
@@ -145,7 +150,11 @@ async def main():
         print(f"Health: {obs.get('health', 10)}/10")
         print(f"Hunger: {obs.get('food', 10)}/10")
         print(f"Energy: {obs.get('energy', 10)}/10")
-        inv_items = ", ".join([f"{k}: {v}" for k, v in inv.items() if v]) if isinstance(inv, dict) else str(inv)
+        inv_items = (
+            ", ".join([f"{k}: {v}" for k, v in inv.items() if v])
+            if isinstance(inv, dict)
+            else str(inv)
+        )
         print(f"Inventory: {inv_items if inv_items else 'empty'}")
         if isinstance(ach, dict):
             unlocked = sum(1 for v in ach.values() if v)
@@ -167,7 +176,9 @@ async def main():
 
         # Cleanup
         try:
-            await client.post(f"{args.base_url}/env/CrafterClassic/terminate", json={"env_id": env_id})
+            await client.post(
+                f"{args.base_url}/env/CrafterClassic/terminate", json={"env_id": env_id}
+            )
         except Exception:
             pass
 
@@ -176,5 +187,3 @@ if __name__ == "__main__":
     import asyncio
 
     asyncio.run(main())
-
-

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/evals_old/compare_models.py

@@ -19,6 +19,7 @@ Analyzes and compares:
 - Performance metrics
 - Cost analysis
 """
+
 import os
 import sys
 from pathlib import Path

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/_backup_synth_qwen/filter_traces_achievements.py

@@ -34,7 +34,8 @@ except Exception:  # pragma: no cover
         )
     except Exception as _import_err:  # pragma: no cover
         raise ImportError(
-            "Could not import FinetuningDataExtractorV3 from synth_ai.") from _import_err
+            "Could not import FinetuningDataExtractorV3 from synth_ai."
+        ) from _import_err
 
 
 def env_list(name: str) -> list[str]:
@@ -133,6 +134,7 @@ async def main() -> None:
     try:
         import numpy as _np
         from collections import Counter as _Counter
+
         async with FinetuningDataExtractorV3(db_url) as _ex:
             _sessions = await _ex.get_all_sessions()
             _ach_counts: _Counter[str] = _Counter()
@@ -155,7 +157,9 @@ async def main() -> None:
             if _rewards:
                 _r = _np.array(_rewards, dtype=float)
                 print("\nReward stats:")
-                print(f"  min={_r.min():.2f} median={_np.median(_r):.2f} mean={_r.mean():.2f} max={_r.max():.2f}")
+                print(
+                    f"  min={_r.min():.2f} median={_np.median(_r):.2f} mean={_r.mean():.2f} max={_r.max():.2f}"
+                )
     except Exception:
         pass
 

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/_backup_synth_qwen/react_agent_lm.py

@@ -5,6 +5,7 @@ This demonstrates using the LM class with Synth models through native integratio
 
 This version uses the new tracing_v3 system with async Turso/SQLite backend.
 """
+
 import argparse
 import asyncio
 import contextlib
@@ -100,13 +101,13 @@ HTTP_TIMEOUT = (
 MAX_RETRIES = 3
 RETRY_DELAY = 1.0
 
+
 # Use the backend
 @asynccontextmanager
 async def _noop_async_context():
     yield
 
 
-
 async def create_experiment_context(
     db_manager: AsyncSQLTraceManager, experiment_name: str, description: str
 ) -> dict[str, Any]:
@@ -1266,9 +1267,8 @@ async def run_episode(
                         info = step_data.get("info", {})
 
                         # Calculate achievement reward if not provided by service
-                        if (
-                            (reward == 0 or reward is None)
-                            and ("achievements_status" in obs and "achievements_status" in prev_obs)
+                        if (reward == 0 or reward is None) and (
+                            "achievements_status" in obs and "achievements_status" in prev_obs
                         ):
                             prev_achievements = prev_obs["achievements_status"]
                             curr_achievements = obs["achievements_status"]

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/_backup_synth_qwen/sft_kickoff.py

@@ -67,9 +67,10 @@ async def create_job(file_id: str) -> str:
         "upload_to_wasabi": bool(scfg.get("upload_to_wasabi", True)),
     }
     headers = {"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json"}
-    async with aiohttp.ClientSession() as session, session.post(
-        f"{API_URL}/fine_tuning/jobs", json=body, headers=headers
-    ) as resp:
+    async with (
+        aiohttp.ClientSession() as session,
+        session.post(f"{API_URL}/fine_tuning/jobs", json=body, headers=headers) as resp,
+    ):
         assert resp.status == 200, await resp.text()
         data = await resp.json()
         return data["id"]

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/synth_qwen_v1/filter_traces_achievements.py

@@ -34,7 +34,8 @@ except Exception:  # pragma: no cover
         )
     except Exception as _import_err:  # pragma: no cover
         raise ImportError(
-            "Could not import FinetuningDataExtractorV3 from synth_ai.") from _import_err
+            "Could not import FinetuningDataExtractorV3 from synth_ai."
+        ) from _import_err
 
 
 def env_list(name: str) -> list[str]:
@@ -133,6 +134,7 @@ async def main() -> None:
     try:
         import numpy as _np
         from collections import Counter as _Counter
+
         async with FinetuningDataExtractorV3(db_url) as _ex:
             _sessions = await _ex.get_all_sessions()
             _ach_counts: _Counter[str] = _Counter()
@@ -155,7 +157,9 @@ async def main() -> None:
             if _rewards:
                 _r = _np.array(_rewards, dtype=float)
                 print("\nReward stats:")
-                print(f"  min={_r.min():.2f} median={_np.median(_r):.2f} mean={_r.mean():.2f} max={_r.max():.2f}")
+                print(
+                    f"  min={_r.min():.2f} median={_np.median(_r):.2f} mean={_r.mean():.2f} max={_r.max():.2f}"
+                )
     except Exception:
         pass
 

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/synth_qwen_v1/hello_ft_model.py

@@ -44,6 +44,7 @@ async def main() -> None:
         print("===== End Response =====\n")
     except Exception as e:  # always print full failure context
         import traceback
+
         print("\n===== Inference Error =====")
         print(f"Type: {type(e).__name__}")
         print(f"Repr: {repr(e)}")
@@ -51,14 +52,15 @@ async def main() -> None:
         print(traceback.format_exc())
         try:
             from synth_ai.http import HTTPError  # type: ignore
+
             if isinstance(e, HTTPError):
                 print("HTTPError details:")
                 print(f"  status={e.status}")
                 print(f"  url={e.url}")
                 print(f"  message={e.message}")
-                if getattr(e, 'detail', None) is not None:
+                if getattr(e, "detail", None) is not None:
                     print(f"  detail={e.detail}")
-                if getattr(e, 'body_snippet', None):
+                if getattr(e, "body_snippet", None):
                     print(f"  body_snippet={e.body_snippet}")
         except Exception:
             pass
@@ -67,5 +69,3 @@ async def main() -> None:
 
 if __name__ == "__main__":
     asyncio.run(main())
-
-

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/synth_qwen_v1/infer.py

@@ -4,6 +4,7 @@ import asyncio
 
 import sys
 import os
+
 sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(__file__)))))
 
 from synth_ai.inference import InferenceClient
@@ -33,5 +34,3 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
-
-

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/synth_qwen_v1/poll.py

@@ -5,6 +5,7 @@ from typing import Dict
 
 import sys
 import os
+
 sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(__file__)))))
 
 from synth_ai.learning import JobHandle
@@ -27,7 +28,9 @@ async def _run(mode: str | None) -> None:
 
     # Use shared JobHandle poller abstraction (strict=True for FT)
     handle = JobHandle(base, key, job_id, strict=True)
-    final = await handle.poll_until_terminal(interval_seconds=2.0, max_seconds=1800, on_event=_print_event)
+    final = await handle.poll_until_terminal(
+        interval_seconds=2.0, max_seconds=1800, on_event=_print_event
+    )
     print(f"final_status={final.get('status')}")
     ft = final.get("fine_tuned_model")
     if ft:
@@ -41,4 +44,3 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
-

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/synth_qwen_v1/prepare_data.py

@@ -17,12 +17,14 @@ def main() -> None:
 
     # Minimal single-example JSONL
     lines = [
-        json.dumps({
-            "messages": [
-                {"role": "user", "content": "Write a short greeting."},
-                {"role": "assistant", "content": "Hello there!"},
-            ]
-        })
+        json.dumps(
+            {
+                "messages": [
+                    {"role": "user", "content": "Write a short greeting."},
+                    {"role": "assistant", "content": "Hello there!"},
+                ]
+            }
+        )
     ]
     out_path.write_text("\n".join(lines) + "\n")
     validate_jsonl(out_path)
@@ -31,5 +33,3 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
-
-

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/synth_qwen_v1/react_agent_lm.py

@@ -5,6 +5,7 @@ This demonstrates using the LM class with Synth models through native integratio
 
 This version uses the new tracing_v3 system with async Turso/SQLite backend.
 """
+
 import argparse
 import asyncio
 import contextlib
@@ -38,6 +39,7 @@ def _resolve_backend_default() -> str:
     base = base.rstrip("/")
     return base if base.endswith("/api") else f"{base}/api"
 
+
 # Disable httpx logging immediately
 logging.getLogger("httpx").setLevel(logging.ERROR)
 logging.getLogger("httpcore").setLevel(logging.ERROR)
@@ -108,13 +110,13 @@ HTTP_TIMEOUT = (
 MAX_RETRIES = 3
 RETRY_DELAY = 1.0
 
+
 # Use the backend
 @asynccontextmanager
 async def _noop_async_context():
     yield
 
 
-
 async def create_experiment_context(
     db_manager: AsyncSQLTraceManager, experiment_name: str, description: str
 ) -> dict[str, Any]:
@@ -1274,9 +1276,8 @@ async def run_episode(
                         info = step_data.get("info", {})
 
                         # Calculate achievement reward if not provided by service
-                        if (
-                            (reward == 0 or reward is None)
-                            and ("achievements_status" in obs and "achievements_status" in prev_obs)
+                        if (reward == 0 or reward is None) and (
+                            "achievements_status" in obs and "achievements_status" in prev_obs
                         ):
                             prev_achievements = prev_obs["achievements_status"]
                             curr_achievements = obs["achievements_status"]

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/synth_qwen_v1/run_crafter_sft_job.py

@@ -37,13 +37,12 @@ def parse_args() -> argparse.Namespace:
     p.add_argument("--mode", choices=["local", "dev", "prod"], default=None)
     p.add_argument(
         "--db",
-        default=str(
-            Path(__file__).resolve().parents[3]
-            / "traces/v3/synth_ai.db/dbs/default/data"
-        ),
+        default=str(Path(__file__).resolve().parents[3] / "traces/v3/synth_ai.db/dbs/default/data"),
         help="Path to sqld internal data file or sqlite+aiosqlite URL",
     )
-    p.add_argument("--output", default=str(Path(__file__).parent / "data" / "training_crafter.jsonl"))
+    p.add_argument(
+        "--output", default=str(Path(__file__).parent / "data" / "training_crafter.jsonl")
+    )
     p.add_argument("--min-achievements", type=int, default=2)
     p.add_argument("--max-cost", type=float, default=10.0)
     p.add_argument("--max-tokens", type=int, default=100000)
@@ -72,12 +71,14 @@ async def extract_jsonl_from_traces(db_url: str, output_path: str, cfg: dict[str
         from synth_ai.environments.examples.crafter_classic.agent_demos.crafter_modal_ft.filter_traces_sft_turso import (  # type: ignore
             FinetuningDataExtractorV3 as _Ex,
         )
+
         Extractor = _Ex
     except Exception:
         try:
             from synth_ai.environments.examples.crafter_classic.agent_demos.crafter_openai_ft.filter_traces_sft_turso import (  # type: ignore
                 FinetuningDataExtractorV3 as _Ex,
             )
+
             Extractor = _Ex
         except Exception as e:
             raise ImportError("FinetuningDataExtractorV3 not available in current build") from e
@@ -110,7 +111,11 @@ async def extract_jsonl_from_traces(db_url: str, output_path: str, cfg: dict[str
                     """,
                     {"session_id": sid},
                 )
-                session_models = model_df["model_name"].tolist() if model_df is not None and not model_df.empty else []
+                session_models = (
+                    model_df["model_name"].tolist()
+                    if model_df is not None and not model_df.empty
+                    else []
+                )
                 if not any(m in session_models for m in models):
                     continue
             ach = await ex.get_session_achievements(sid) or []
@@ -203,5 +208,3 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
-
-

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/synth_qwen_v1/run_ft_job.py

@@ -65,12 +65,14 @@ def ensure_training_jsonl(path: Path) -> Path:
     if not path.exists():
         # Minimal JSONL with a single example
         lines: list[str] = [
-            json.dumps({
-                "messages": [
-                    {"role": "user", "content": "Write a short greeting."},
-                    {"role": "assistant", "content": "Hello there!"},
-                ]
-            })
+            json.dumps(
+                {
+                    "messages": [
+                        {"role": "user", "content": "Write a short greeting."},
+                        {"role": "assistant", "content": "Hello there!"},
+                    ]
+                }
+            )
         ]
         path.write_text("\n".join(lines) + "\n")
     # Validate using shared SDK validator
@@ -83,7 +85,9 @@ async def run(args: argparse.Namespace) -> None:
     base_url, api_key = load_env(args.mode)
     # Force canonical prod base when prod mode (or override) is selected
     try:
-        if (args.mode == "prod") or (os.getenv("SYNTH_BACKEND_URL_OVERRIDE", "").strip().lower() == "prod"):
+        if (args.mode == "prod") or (
+            os.getenv("SYNTH_BACKEND_URL_OVERRIDE", "").strip().lower() == "prod"
+        ):
             base_url = _resolve_backend_default()
             # Also export for any downstream helpers that read env
             os.environ["PROD_BACKEND_URL"] = base_url
@@ -123,8 +127,8 @@ async def run(args: argparse.Namespace) -> None:
                     "gpu_type": "A10G",
                     "container_count": 1,
                 }
-            }
-        }
+            },
+        },
     }
 
     create_resp = await ft.create_sft_job(
@@ -194,6 +198,7 @@ async def run(args: argparse.Namespace) -> None:
         except Exception as e:
             # Always print full error details and traceback
             import traceback
+
             try:
                 from synth_ai.http import HTTPError  # type: ignore
             except Exception:  # pragma: no cover - fallback if import shape changes
@@ -206,7 +211,7 @@ async def run(args: argparse.Namespace) -> None:
                 print("Traceback:")
                 print(tb)
             # If HTTP error from backend, surface structured fields
-            if 'HTTPError' in str(type(e)) or (isinstance((), tuple) and False):
+            if "HTTPError" in str(type(e)) or (isinstance((), tuple) and False):
                 pass
             try:
                 if HTTPError and isinstance(e, HTTPError):  # type: ignore[arg-type]
@@ -214,9 +219,9 @@ async def run(args: argparse.Namespace) -> None:
                     print(f"  status={e.status}")
                     print(f"  url={e.url}")
                     print(f"  message={e.message}")
-                    if getattr(e, 'detail', None) is not None:
+                    if getattr(e, "detail", None) is not None:
                         print(f"  detail={e.detail}")
-                    if getattr(e, 'body_snippet', None):
+                    if getattr(e, "body_snippet", None):
                         print(f"  body_snippet={e.body_snippet}")
             except Exception:
                 pass

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/finetuning_old/synth_qwen_v1/util.py

@@ -11,6 +11,7 @@ from synth_ai.config.base_url import get_backend_from_env
 try:
     from dotenv import load_dotenv  # type: ignore[reportMissingImports]
 except Exception:  # pragma: no cover
+
     def load_dotenv(*args, **kwargs):  # type: ignore[no-redef]
         return False
 
@@ -71,7 +72,9 @@ def load_env(mode: str | None = None) -> tuple[str, str]:
             or os.getenv("TESTING_LOCAL_SYNTH_API_KEY", "").strip()
         )
         if not base_url or not api_key:
-            raise RuntimeError("Missing LOCAL_BACKEND_URL or DEV_SYNTH_API_KEY/TESTING_LOCAL_SYNTH_API_KEY in environment/.env")
+            raise RuntimeError(
+                "Missing LOCAL_BACKEND_URL or DEV_SYNTH_API_KEY/TESTING_LOCAL_SYNTH_API_KEY in environment/.env"
+            )
     elif mode == "dev":
         base_url = os.getenv("DEV_BACKEND_URL", "").strip()
         api_key = os.getenv("DEV_SYNTH_API_KEY", "").strip()
@@ -85,7 +88,9 @@ def load_env(mode: str | None = None) -> tuple[str, str]:
             or os.getenv("SYNTH_API_KEY", "").strip()
         )
         if not api_key:
-            raise RuntimeError("Missing PROD_SYNTH_API_KEY/TESTING_PROD_SYNTH_API_KEY/SYNTH_API_KEY in environment/.env")
+            raise RuntimeError(
+                "Missing PROD_SYNTH_API_KEY/TESTING_PROD_SYNTH_API_KEY/SYNTH_API_KEY in environment/.env"
+            )
     base_url = base_url.rstrip("/")
     print(f"SYNTH backend: {base_url} (mode={mode})")
     # Also print masked API key and source

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/rl/configs/eval_base_qwen.toml

@@ -6,7 +6,7 @@ num_episodes = 50
 seed_start = 0
 
 [policy]
-inference_url = "http://localhost:8000/api/inference"
+inference_url = "https://agent-learning.onrender.com/api/inference"
 max_tokens = 128
 temperature = 0.0
 

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/rl/configs/rl_from_base_qwen17.toml

@@ -11,7 +11,7 @@ base = "Qwen/Qwen3-1.7B"
 
 [policy]
 model = "Qwen/Qwen3-1.7B"
-inference_url = "http://localhost:8000/api/inference"
+inference_url = "https://agent-learning.onrender.com/api/inference"
 max_tokens = 1028
 temperature = 0.2
 

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/rl/download_dataset.py

@@ -20,10 +20,12 @@ def extract_examples(dataset: Any, *, limit: int | None) -> list[dict[str, str]]
         solution = item.get("solution") or ""
         if isinstance(solution, list):
             solution = "\n".join(str(part) for part in solution)
-        examples.append({
-            "problem": problem,
-            "solution": solution,
-        })
+        examples.append(
+            {
+                "problem": problem,
+                "solution": solution,
+            }
+        )
     return examples
 
 
@@ -35,12 +37,26 @@ def write_jsonl(path: Path, rows: list[dict[str, str]]) -> None:
 
 
 def main() -> None:
-    parser = argparse.ArgumentParser(description="Download MATH dataset splits to JSONL for offline use")
-    parser.add_argument("--output-dir", default="examples/rl/data", help="Directory to write <split>.jsonl files")
-    parser.add_argument("--dataset", default="nlile/hendrycks-MATH-benchmark", help="Hugging Face dataset identifier")
-    parser.add_argument("--config", default="algebra", help="Hugging Face dataset config (if required)")
-    parser.add_argument("--splits", nargs="*", default=["train", "validation", "test"], help="Splits to download")
-    parser.add_argument("--limit", type=int, default=None, help="Optional cap on examples per split")
+    parser = argparse.ArgumentParser(
+        description="Download MATH dataset splits to JSONL for offline use"
+    )
+    parser.add_argument(
+        "--output-dir", default="examples/rl/data", help="Directory to write <split>.jsonl files"
+    )
+    parser.add_argument(
+        "--dataset",
+        default="nlile/hendrycks-MATH-benchmark",
+        help="Hugging Face dataset identifier",
+    )
+    parser.add_argument(
+        "--config", default="algebra", help="Hugging Face dataset config (if required)"
+    )
+    parser.add_argument(
+        "--splits", nargs="*", default=["train", "validation", "test"], help="Splits to download"
+    )
+    parser.add_argument(
+        "--limit", type=int, default=None, help="Optional cap on examples per split"
+    )
     args = parser.parse_args()
 
     output_dir = Path(args.output_dir).expanduser()

{synth_ai-0.2.9.dev5 → synth_ai-0.2.9.dev7}/examples/rl/run_eval.py

@@ -25,7 +25,10 @@ class TaskAppClient:
     async def __aenter__(self) -> "TaskAppClient":
         headers = {"X-API-Key": self.api_key} if self.api_key else {}
         self._client = httpx.AsyncClient(
-            base_url=self.base_url, headers=headers, timeout=httpx.Timeout(120.0), follow_redirects=True
+            base_url=self.base_url,
+            headers=headers,
+            timeout=httpx.Timeout(120.0),
+            follow_redirects=True,
         )
         return self
 
@@ -39,7 +42,10 @@ class TaskAppClient:
         if self._client is None:
             headers = {"X-API-Key": self.api_key} if self.api_key else {}
             self._client = httpx.AsyncClient(
-                base_url=self.base_url, headers=headers, timeout=httpx.Timeout(120.0), follow_redirects=True
+                base_url=self.base_url,
+                headers=headers,
+                timeout=httpx.Timeout(120.0),
+                follow_redirects=True,
             )
         return self._client
 
@@ -103,8 +109,7 @@ def _math_tool_schema() -> List[Dict[str, Any]]:
                         "answer": {
                             "type": "string",
                             "description": "Final answer in simplest form",
-                        }
-                        ,
+                        },
                         "explanation": {
                             "type": "string",
                             "description": "Optional explanation of reasoning",
@@ -203,9 +208,7 @@ async def _choose_actions(
 
     if provider == "groq":
         # Task app proxies Groq requests; reuse existing headers on the client
-        response = await client.client.post(
-            "/proxy/groq/v1/chat/completions", json=payload
-        )
+        response = await client.client.post("/proxy/groq/v1/chat/completions", json=payload)
         response.raise_for_status()
         body = response.json()
     else:
@@ -223,18 +226,15 @@ async def _choose_actions(
                 headers=headers or None,
             )
         except httpx.ReadTimeout as exc:
-            raise RuntimeError(
-                "Inference request timed out. Check the inference service." ) from exc
+            raise RuntimeError("Inference request timed out. Check the inference service.") from exc
         try:
             body = response.json()
         except Exception:
             body = {"raw": response.text[:800]}
         if response.status_code >= 500:
-            raise RuntimeError(
-                f"Inference server error {response.status_code}: {body}")
+            raise RuntimeError(f"Inference server error {response.status_code}: {body}")
         if response.status_code >= 400:
-            raise RuntimeError(
-                f"Inference request invalid ({response.status_code}): {body}")
+            raise RuntimeError(f"Inference request invalid ({response.status_code}): {body}")
     tool_calls = _parse_tool_calls(body)
     return tool_calls, body
 
@@ -371,7 +371,9 @@ async def main() -> None:
     seed_start = int(cfg.get("seed_start") or 0)
 
     policy_cfg = _default_policy_cfg(cfg)
-    provider_hint = cfg.get("provider") or cfg.get("policy", {}).get("provider") or policy_cfg.get("provider")
+    provider_hint = (
+        cfg.get("provider") or cfg.get("policy", {}).get("provider") or policy_cfg.get("provider")
+    )
     provider = _detect_provider(model, provider_hint)
     policy_cfg.pop("provider", None)
 
@@ -412,7 +414,7 @@ async def main() -> None:
             problem = data.get("problem")
             tool_calls = data.get("tool_calls") or []
             print(
-                f"Episode {episode+1}/{episodes} seed={seed} status={status} reward={data.get('reward')}\n"
+                f"Episode {episode + 1}/{episodes} seed={seed} status={status} reward={data.get('reward')}\n"
                 f"  problem: {problem!r}\n"
                 f"  tool   : {tool_calls!r}\n"
                 f"  answer : {answer!r}\n  expected: {expected!r}",