Implement strict judge feedback: ambiguous task descriptions, server-side payload clues, increased exploration steps

inference.py

@@ -11,7 +11,7 @@ MODEL_NAME = os.getenv("MODEL_NAME", "mistralai/Mistral-7B-Instruct-v0.3")
 HF_TOKEN = os.getenv("HF_TOKEN")
 ENV_URL = os.getenv("ENV_URL", "http://localhost:7860")
 BENCHMARK = "api_debug_env"
-MAX_STEPS = 5
+MAX_STEPS = 10
 SUCCESS_SCORE_THRESHOLD = 0.8
 
 llm = OpenAI(base_url=API_BASE_URL, api_key=HF_TOKEN)

models.py

@@ -35,5 +35,5 @@ class APIObservation(Observation):
 
 class APIState(State):
     task_id: str = Field(default="easy")
-    max_steps: int = Field(default=5)
+    max_steps: int = Field(default=10)
     solved: bool = Field(default=False)

server/api_debug_environment.py

@@ -35,7 +35,7 @@ class APIDebugEnvironment(Environment):
             episode_id=f"ep_{random.randint(10000, 99999)}",
             step_count=0,
             task_id=task_id,
-            max_steps=5,
+            max_steps=10,
             solved=False,
         )
         return APIObservation(

server/mock_api.py

@@ -37,10 +37,10 @@ async def get_token(body: TokenRequest):
 @router.get("/users")
 async def get_users(authorization: Optional[str] = Header(default=None)):
     if not authorization or not authorization.startswith("Bearer "):
-        raise HTTPException(status_code=401, detail="Authorization header missing or malformed")
+        raise HTTPException(status_code=401, detail="Authorization header missing or malformed. Expected format: Bearer <token>. Use 'demo_token_123' for testing.")
     token = authorization.split(" ")[1]
     if token not in _issued_tokens and token != "demo_token_123":
-        raise HTTPException(status_code=401, detail="Invalid token")
+        raise HTTPException(status_code=401, detail="Invalid token. Try 'demo_token_123'.")
     return {"users": [{"id": 1, "name": "Alice"}, {"id": 2, "name": "Bob"}]}
 
 
@@ -104,7 +104,11 @@ async def rate_limited(request: Request, x_retry_after: Optional[str] = Header(d
     window = [t for t in _request_log.get(client_id, []) if now - t < 5]
     if len(window) >= 3 and not x_retry_after:
         _request_log[client_id] = window
-        raise HTTPException(status_code=429, detail="Rate limit exceeded", headers={"Retry-After": "2"})
+        raise HTTPException(
+            status_code=429, 
+            detail={"error": "Rate limit exceeded", "hint": "Please add X-Retry-After header with value 2 to override."},
+            headers={"Retry-After": "2"}
+        )
     window.append(now)
     _request_log[client_id] = window
     return {"data": "rate_limited_resource", "requests_in_window": len(window)}

tasks/easy.py

@@ -1,7 +1,7 @@
 EASY_TASKS = [
     {
         "id": "easy_auth",
-        "description": "The request is missing an Authorization header. The API requires: Authorization: Bearer demo_token_123. Fix the request to GET /mock_api/users successfully.",
+        "description": "Access the protected resource at GET /mock_api/users to get a 200 OK. The current request is failing.",
         "broken_request": {
             "method": "GET",
             "url": "/mock_api/users",
@@ -14,7 +14,7 @@ EASY_TASKS = [
     },
     {
         "id": "easy_content_type",
-        "description": "The request uses Content-Type: text/plain but the API requires application/json. Fix the header and send body with a name field to POST /mock_api/items.",
+        "description": "Create an item with name 'test_item' via POST /mock_api/items to get a 200 OK. The current request payload is rejected by the server.",
         "broken_request": {
             "method": "POST",
             "url": "/mock_api/items",
@@ -27,7 +27,7 @@ EASY_TASKS = [
     },
     {
         "id": "easy_query_param",
-        "description": "The search endpoint requires a query parameter q. The request omits it. Add q=python to the query params for GET /mock_api/search.",
+        "description": "Search for 'python' via GET /mock_api/search. The current request is missing required parameters.",
         "broken_request": {
             "method": "GET",
             "url": "/mock_api/search",

tasks/hard.py

@@ -1,7 +1,7 @@
 HARD_TASKS = [
     {
         "id": "hard_token_exchange",
-        "description": "The protected endpoint needs a fresh token. First POST to /mock_api/auth/token with body client_id abc and client_secret xyz to get an access_token. Then use it as Authorization: Bearer <token> to GET /mock_api/protected. The token in the broken request is expired.",
+        "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.",
         "broken_request": {
             "method": "GET",
             "url": "/mock_api/protected",
@@ -14,7 +14,7 @@ HARD_TASKS = [
     },
     {
         "id": "hard_rate_limit",
-        "description": "The rate_limited endpoint returns 429 if you send more than 3 requests without the X-Retry-After header. After receiving a 429 include header X-Retry-After with value 2.",
+        "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.",
         "broken_request": {
             "method": "GET",
             "url": "/mock_api/rate_limited",
@@ -27,7 +27,7 @@ HARD_TASKS = [
     },
     {
         "id": "hard_pagination",
-        "description": "GET /mock_api/logs returns only the first page with a next_cursor field. Follow the cursor chain by passing cursor as a query param until has_more is false.",
+        "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.",
         "broken_request": {
             "method": "GET",
             "url": "/mock_api/logs",

tasks/medium.py

@@ -1,7 +1,7 @@
 MEDIUM_TASKS = [
     {
         "id": "medium_wrong_method",
-        "description": "Creating an order requires POST not GET. Fix the HTTP method and send body with product_id 5 and qty 2 to /mock_api/orders.",
+        "description": "Create an order for product_id 5 with qty 2 at /mock_api/orders. The given request is failing.",
         "broken_request": {
             "method": "GET",
             "url": "/mock_api/orders",
@@ -14,7 +14,7 @@ MEDIUM_TASKS = [
     },
     {
         "id": "medium_type_mismatch",
-        "description": "The API expects product_id as an integer but the request sends the string five. Fix the type to the integer 5 in the body.",
+        "description": "Submit a new order for product_id 5 and qty 2 at /mock_api/orders. The API is returning a validation error for the current payload.",
         "broken_request": {
             "method": "POST",
             "url": "/mock_api/orders",
@@ -27,7 +27,7 @@ MEDIUM_TASKS = [
     },
     {
         "id": "medium_nested_field",
-        "description": "The profile endpoint expects address to be a dict like street and city keys, not a plain string. Fix the body structure.",
+        "description": "Update the profile with name 'Alice' and address containing street '123 Main St' and city 'Metropolis'. The current payload structure is rejected.",
         "broken_request": {
             "method": "POST",
             "url": "/mock_api/profile",