Ready for first real test

agents/graph.py

@@ -41,11 +41,22 @@ RESEARCH_KICKOFF = (
     "call get_option_chain(underlying, 'call' or 'put') and copy an exact contract_symbol."
 )
 
-DEEPEN_NUDGE = (
-    "You're not finished. Keep researching before finalizing: evaluate additional candidate "
-    "tickers you haven't examined yet, pull bars/news/fundamentals for them, verify with "
-    "web_search, and stress-test your leading ideas against the alternatives. Continue using tools."
-)
+# Rotating directive nudges — cycled so a long (~2hr) research session stays PRODUCTIVE
+# (does new, varied work) instead of looping "am I done?". Keep using tools.
+DEEPEN_NUDGES = [
+    "Not done yet. Screen 5-10 MORE candidate tickers in your style you haven't examined; pull "
+    "get_bars / get_news / get_fundamentals on them.",
+    "Build the BEAR case for your current top pick: what would make it wrong? Find data that "
+    "challenges your thesis (web_search + web_fetch recent coverage).",
+    "Use run_analysis to write a real script: fetch prices with yfinance, compute momentum / "
+    "moving averages / volatility / correlations, and save a chart to back your best idea.",
+    "Zoom out: check the macro + sector context (rates, the sector ETF, peers) via web_search and "
+    "get_bars. Does it support or undercut your picks?",
+    "Compare your leading picks head-to-head against 2-3 sector peers on valuation and momentum "
+    "(get_fundamentals + run_analysis). Are you picking the best expression of the thesis?",
+    "Re-examine catalysts and timing for each candidate (earnings dates, events via get_news / "
+    "web_search) and refine entry, stop, and take-profit levels with run_python.",
+]
 
 FINALIZE_PROMPT = (
     "Research complete. Return a JSON object of the form {{\"picks\": [ ... ]}} containing about "
@@ -176,6 +187,7 @@ def run_agent(
     deadline_ts: float,
     step_budget: int,
     min_steps: int = 0,
+    min_seconds: float = 0.0,
     checkpointer=None,
     model_name: str = "",
 ) -> AgentOutput:
@@ -195,6 +207,7 @@ def run_agent(
     running_summary = ""
     reasoning_trace: list[str] = []
     steps_used = 0
+    nudge_count = 0
     ended_reason = "natural"
 
     def snapshot(status: AgentStatus, picks: list[Pick] | None = None, errors=None) -> AgentOutput:
@@ -229,11 +242,16 @@ def run_agent(
 
         tool_calls = getattr(ai, "tool_calls", None)
         if not tool_calls:
-            # Force depth: if the model wants to stop early, nudge it to keep researching
-            # until it has done at least min_steps (still bounded by the wall-clock deadline).
-            if steps_used < min_steps and time.time() < deadline_ts:
-                messages.append(HumanMessage(content=DEEPEN_NUDGE))
-                reasoning_trace.append(f"[nudge] deepen research ({steps_used}/{min_steps} steps)")
+            # Force depth: if the model tries to finalize before the TIME floor (or step floor),
+            # inject a rotating directive nudge so it keeps doing NEW research. Bounded by deadline.
+            elapsed = time.time() - start
+            if (elapsed < min_seconds or steps_used < min_steps) and time.time() < deadline_ts:
+                nudge = DEEPEN_NUDGES[nudge_count % len(DEEPEN_NUDGES)]
+                nudge_count += 1
+                messages.append(HumanMessage(content=nudge))
+                reasoning_trace.append(
+                    f"[nudge #{nudge_count}] {elapsed/60:.0f}/{min_seconds/60:.0f}min elapsed, step {steps_used}"
+                )
                 continue
             if ai.content:
                 reasoning_trace.append(f"[note] {str(ai.content)[:200]}")
@@ -241,7 +259,9 @@ def run_agent(
             break
 
         for tc in tool_calls:
-            name, args = tc["name"], tc.get("args", {}) or {}
+            # gpt-oss can leak harmony channel tokens into the tool name (e.g. "get_bars<|channel|>")
+            name = tc["name"].split("<|")[0].strip()
+            args = tc.get("args", {}) or {}
             tool = tools_by_name.get(name)
             try:
                 result = tool.invoke(args) if tool else f"ERROR: unknown tool {name}"