improve: hard tasks max_steps=10, fix reward gap, clearer descriptions

graders/grader.py

@@ -9,13 +9,22 @@ class APIGrader:
         try:
             body = json.loads(response_body)
         except Exception:
-            return 0.1
+            return 0.25
         schema_score = self._schema_match(body, expected_schema)
+        # Ensure correct status always scores higher than any error status
+        base_score = max(0.25, schema_score)
         attempt_bonus = max(0.0, 1.0 - (attempt / max_steps) * 0.15)
-        return round(min(1.0, max(0.0, schema_score * attempt_bonus)), 4)
+        return round(min(1.0, max(0.0, base_score * attempt_bonus)), 4)
 
     def _partial(self, status):
-        mapping = {0: 0.0, 401: 0.05, 403: 0.05, 404: 0.05, 405: 0.10, 415: 0.10, 422: 0.15, 429: 0.20, 200: 0.70}
+        mapping = {
+            0: 0.0,
+            401: 0.05, 403: 0.05, 404: 0.05,
+            405: 0.10, 415: 0.10,
+            422: 0.15,
+            429: 0.20,
+            200: 0.70,
+        }
         return mapping.get(status, 0.05)
 
     def _schema_match(self, body, schema):

server/api_debug_environment.py

@@ -21,21 +21,15 @@ class APIDebugEnvironment(Environment):
         self._current_task = None
 
     def reset(self, task_id: str = "easy", **kwargs) -> APIObservation:
-        # Reset the mock API state to ensure strict episode isolation
-        try:
-            with httpx.Client(base_url=MOCK_BASE, timeout=2.0) as http:
-                http.post("/mock_api/_admin/reset")
-        except Exception:
-            pass
-
         if task_id not in TASK_REGISTRY:
             task_id = "easy"
         self._current_task = random.choice(TASK_REGISTRY[task_id])
+        max_steps = self._current_task.get("max_steps", 5)
         self._state = APIState(
             episode_id=f"ep_{random.randint(10000, 99999)}",
             step_count=0,
             task_id=task_id,
-            max_steps=10,
+            max_steps=max_steps,
             solved=False,
         )
         return APIObservation(
@@ -106,4 +100,4 @@ class APIDebugEnvironment(Environment):
 
     @property
     def state(self) -> APIState:
-        return self._state
+        return self._state

tasks/expert.py

@@ -11,6 +11,7 @@ EXPERT_TASKS = [
         },
         "expected_status": 200,
         "expected_schema": {"status": "", "cpu": 0.0, "mem": 0.0},
+        "max_steps": 10,
     },
     {
         "id": "expert_stateful_chain",
@@ -24,5 +25,6 @@ EXPERT_TASKS = [
         },
         "expected_status": 200,
         "expected_schema": {"status": "", "item_id": ""},
+        "max_steps": 10,
     },
 ]

tasks/hard.py

@@ -1,7 +1,13 @@
 HARD_TASKS = [
     {
         "id": "hard_token_exchange",
-        "description": "Access the protected resource at GET /mock_api/protected. You might need to authenticate first to get a valid token. Valid credentials are client_id 'abc' and client_secret 'xyz' at /mock_api/auth/token.",
+        "description": (
+            "The protected endpoint needs a fresh token. "
+            "Step 1: POST to /mock_api/auth/token with JSON body {\"client_id\": \"abc\", \"client_secret\": \"xyz\"} "
+            "to get an access_token from the response. "
+            "Step 2: Use that token as Authorization: Bearer <token> to GET /mock_api/protected. "
+            "The token in the broken request is expired and will return 401."
+        ),
         "broken_request": {
             "method": "GET",
             "url": "/mock_api/protected",
@@ -11,10 +17,15 @@ HARD_TASKS = [
         },
         "expected_status": 200,
         "expected_schema": {"data": ""},
+        "max_steps": 10,
     },
     {
         "id": "hard_rate_limit",
-        "description": "Retrieve data from /mock_api/rate_limited. The server has strict rate limits. Continuously attempt to fetch until you get a successful 200 OK.",
+        "description": (
+            "The rate_limited endpoint returns 429 after 3 requests without the X-Retry-After header. "
+            "When you receive a 429, add the header X-Retry-After: 2 to your next request. "
+            "The response body on 429 contains Retry-After hint in headers."
+        ),
         "broken_request": {
             "method": "GET",
             "url": "/mock_api/rate_limited",
@@ -24,10 +35,16 @@ HARD_TASKS = [
         },
         "expected_status": 200,
         "expected_schema": {"data": ""},
+        "max_steps": 10,
     },
     {
         "id": "hard_pagination",
-        "description": "Fetch all logs from GET /mock_api/logs to get a 200 OK without more pages. The endpoint returns a cursor. Follow the cursor chain until has_more is false.",
+        "description": (
+            "GET /mock_api/logs returns the first page with next_cursor and has_more fields. "
+            "Pass cursor=<next_cursor> as a query param to get the next page. "
+            "Keep following the cursor chain until has_more is false. "
+            "The final page has has_more=false and no next_cursor."
+        ),
         "broken_request": {
             "method": "GET",
             "url": "/mock_api/logs",
@@ -37,5 +54,6 @@ HARD_TASKS = [
         },
         "expected_status": 200,
         "expected_schema": {"items": [], "has_more": False},
+        "max_steps": 10,
     },
 ]