Fix Issue 3: Add 60s timeout to API calls in prod_test.py

tests/prod_test.py

@@ -10,6 +10,7 @@ import purpose_agent as pa
 from purpose_agent.llm_backend import resolve_backend, ChatMessage
 from purpose_agent.orchestrator import Environment
 from purpose_agent.types import State
+from purpose_agent.hardening import with_timeout
 
 b = resolve_backend("openrouter:meta-llama/llama-3.3-70b-instruct", api_key=OR)
 P = F = 0
@@ -19,6 +20,10 @@ def ok(n, c, d=""):
     icon = "PASS" if c else "FAIL"
     print(f"  [{icon}] {n}" + (f" — {d}" if d else ""))
 
+def timed_run(fn, timeout_s=60, default=None, label="api_call"):
+    """Run fn with a timeout to prevent hanging on slow API calls."""
+    return with_timeout(fn, timeout_s=timeout_s, default=default, label=label)()
+
 print(f"Purpose Agent v{pa.__version__} — 0-Day Production Test")
 print(f"Model: Llama-3.3-70B via OpenRouter\n")
 
@@ -27,7 +32,7 @@ print("LEVEL 1: purpose()")
 team = pa.purpose("Write code", model=b)
 ok("Auto-detect coding team", [a.name for a in team._agents] == ["architect","coder","tester"], str([a.name for a in team._agents]))
 t0 = time.time()
-result = team.run("Check if number is prime", verbose=False)
+result = timed_run(lambda: team.run("Check if number is prime", verbose=False), timeout_s=90, default="", label="team_run")
 ok("Run real task", len(result) > 5, f"{len(result)} chars {time.time()-t0:.0f}s")
 team.teach("Add types")
 ok("Teach works", True)
@@ -35,10 +40,10 @@ time.sleep(0.3)
 
 # ═══ LEVEL 2 ═══
 print("\nLEVEL 2: Backend + Knowledge")
-resp = b.generate([ChatMessage(role="user", content="Say ok")], temperature=0, max_tokens=5)
+resp = timed_run(lambda: b.generate([ChatMessage(role="user", content="Say ok")], temperature=0, max_tokens=5), timeout_s=30, default="", label="simple_api")
 ok("API call", len(resp) > 0, repr(resp[:20]))
 kt = pa.purpose("Answer Qs", model=b, knowledge=["Python created by Guido 1991."])
-ans = kt.ask("Who created Python?")
+ans = timed_run(lambda: kt.ask("Who created Python?"), timeout_s=60, default="", label="knowledge_ask")
 ok("Knowledge team", len(ans) > 3, ans[:40])
 time.sleep(0.3)
 
@@ -94,7 +99,11 @@ for name, purpose, tests in [
         available_actions={"submit_code": "Submit code in params.code", "DONE": "Done"},
         optimize_every_n_tasks=99)
     t0 = time.time()
-    r = orch.run_task(purpose=purpose, initial_state=env.reset(), max_steps=2)
+    r = timed_run(
+        lambda p=purpose, e=env: orch.run_task(purpose=p, initial_state=e.reset(), max_steps=2),
+        timeout_s=60, default=type('R', (), {'final_state': State(data={}), 'trajectory': type('T', (), {'steps': [], 'final_phi': None, 'success_rate': 0})()})(),
+        label=f"coding_{name}",
+    )
     ok(name, r.final_state.data.get("all_passed", False),
        f'{r.final_state.data.get("pass_rate",0):.0%} {time.time()-t0:.0f}s')
     time.sleep(0.3)
@@ -107,7 +116,9 @@ orch2 = pa.Orchestrator(llm=b, environment=env2,
 orch2.optimizer.min_reward_threshold = 0.01
 counts = []
 for _ in range(3):
-    try: orch2.run_task(purpose="fib(5)=5", initial_state=env2.reset(), max_steps=2)
+    try:
+        timed_run(lambda: orch2.run_task(purpose="fib(5)=5", initial_state=env2.reset(), max_steps=2),
+                  timeout_s=60, label="self_improve")
     except: pass
     counts.append(len(orch2.optimizer.heuristic_library))
     time.sleep(0.3)