Revert max_steps to 7, stop wasting steps on Wikipedia direct visits

max_steps=12 blew Cerebras' daily token quota partway through the run, auto-failing the back half of questions (25% -> 10%). Reverted to 7. Also added an explicit instruction against visit_webpage on wikipedia.org URLs, which always 403s - use wikipedia_search instead.

app.py

@@ -40,6 +40,7 @@ If you are asked for a number, don't use commas to write it, and don't use units
 If you are asked for a string, don't use articles or abbreviations (e.g. for cities), and write digits in plain text unless specified otherwise.
 If you are asked for a comma separated list, apply the above rules to each element depending on whether it's a number or a string.
 Work efficiently: if a search or lookup doesn't find what you need after 1-2 tries, try a meaningfully different approach rather than repeating similar queries, and give your best-guess final_answer rather than exhausting all steps.
+Never call visit_webpage on a wikipedia.org URL - it always returns 403 Forbidden. Use the wikipedia_search tool for Wikipedia content instead.
 """
 
 class TokenPacer:
@@ -137,12 +138,15 @@ class BasicAgent:
                     "collections", "statistics", "datetime", "io", "openpyxl", "PIL",
                 ]),
             ],
-            # 12 (was 7): the 25% run showed several questions still making
-            # real progress (finding the right source) when cut off at max
-            # steps, forcing a garbled non-tool-call final answer instead of
-            # a real one. TokenPacer sleeps stayed short (6-8s) at 7 steps,
-            # so there's TPM budget headroom to let runs go longer.
-            max_steps=12,
+            # Reverted 12 -> 7: raising it caused several questions to run
+            # 12-13 steps each with growing context, and the cumulative
+            # token usage blew Cerebras' *daily* quota partway through the
+            # run (confirmed: "Tokens per day limit exceeded" starting
+            # around question 13) - every question after that auto-failed,
+            # dropping the score from 25% to 10%. 7 steps was empirically
+            # better: fewer wasted tokens per question, more questions
+            # actually get a real shot before the daily budget runs out.
+            max_steps=7,
             step_callbacks=[
                 MemoryTrimmer(),
                 TokenPacer(tokens_per_minute_budget=int(os.getenv("RATE_LIMIT_TOKENS_PER_MINUTE", "25000"))),