fix(graph): key checkpoints on graph shape and expose the LLM retry budget

- checkpoint resume keyed only by ticker+date silently continued the old graph
  under a different analyst selection / depth / asset mode; fold a run signature
  into the thread id #1089
- add llm_max_retries + TRADINGAGENTS_LLM_MAX_RETRIES, forwarded to every provider
  when set (int-coerced, rejects negatives/booleans), so a 429 burst can't abort
  a run #1091

tests/test_checkpoint_resume.py

@@ -140,5 +140,79 @@ class TestCheckpointResume(unittest.TestCase):
         self.assertTrue(has_checkpoint(self.tmpdir, self.ticker, self.date))
 
 
+class TestCheckpointSignature(unittest.TestCase):
+    """A different graph shape (analyst selection / depth / asset mode) must not
+    resume the previous run's checkpoint (#1089)."""
+
+    def setUp(self):
+        self.tmpdir = tempfile.mkdtemp()
+        self.ticker = "TEST"
+        self.date = "2026-04-20"
+
+    def test_empty_signature_is_legacy_id(self):
+        self.assertEqual(
+            thread_id(self.ticker, self.date),
+            thread_id(self.ticker, self.date, ""),
+        )
+
+    def test_signature_changes_thread_id(self):
+        legacy = thread_id(self.ticker, self.date)
+        sig_a = thread_id(self.ticker, self.date, "analysts=market,news|asset=stock")
+        sig_b = thread_id(self.ticker, self.date, "analysts=market|asset=stock")
+        self.assertNotEqual(sig_a, sig_b)          # different graph shapes differ
+        self.assertNotEqual(legacy, sig_a)         # signature-keyed differs from legacy
+        self.assertEqual(                          # same inputs are stable
+            sig_a, thread_id(self.ticker, self.date, "analysts=market,news|asset=stock")
+        )
+
+    def test_different_signature_starts_fresh(self):
+        global _should_crash
+        builder = _build_graph()
+        sig1 = "analysts=market,news,fundamentals|asset=stock"
+        sig2 = "analysts=market|asset=stock"       # dropped analysts -> different graph
+
+        _should_crash = True
+        tid1 = thread_id(self.ticker, self.date, sig1)
+        with get_checkpointer(self.tmpdir, self.ticker) as saver:
+            graph = builder.compile(checkpointer=saver)
+            with self.assertRaises(RuntimeError):
+                graph.invoke({"count": 0}, config={"configurable": {"thread_id": tid1}})
+
+        self.assertTrue(has_checkpoint(self.tmpdir, self.ticker, self.date, sig1))
+        # A different graph shape has no checkpoint to resume from.
+        self.assertFalse(has_checkpoint(self.tmpdir, self.ticker, self.date, sig2))
+
+        _should_crash = False
+        tid2 = thread_id(self.ticker, self.date, sig2)
+        self.assertNotEqual(tid1, tid2)
+        with get_checkpointer(self.tmpdir, self.ticker) as saver:
+            graph = builder.compile(checkpointer=saver)
+            result = graph.invoke({"count": 0}, config={"configurable": {"thread_id": tid2}})
+        self.assertEqual(result["count"], 11)
+        # sig1's checkpoint remains untouched.
+        self.assertTrue(has_checkpoint(self.tmpdir, self.ticker, self.date, sig1))
+
+    def test_run_signature_captures_graph_shape(self):
+        from tradingagents.graph.trading_graph import TradingAgentsGraph
+
+        # Build a bare instance to exercise the pure helper without heavy __init__.
+        g = object.__new__(TradingAgentsGraph)
+        g.selected_analysts = ("market", "news")
+        g.config = {"max_debate_rounds": 1, "max_risk_discuss_rounds": 1}
+        base = g._run_signature("stock")
+
+        self.assertNotEqual(base, g._run_signature("crypto"))     # asset mode
+        g.selected_analysts = ("market",)
+        self.assertNotEqual(base, g._run_signature("stock"))      # analyst selection
+        g.selected_analysts = ("market", "news")
+        g.config = {"max_debate_rounds": 3, "max_risk_discuss_rounds": 1}
+        self.assertNotEqual(base, g._run_signature("stock"))      # debate depth
+        g.config = {"max_debate_rounds": 1, "max_risk_discuss_rounds": 5}
+        self.assertNotEqual(base, g._run_signature("stock"))      # risk depth
+        # Stable for identical inputs.
+        g.config = {"max_debate_rounds": 1, "max_risk_discuss_rounds": 1}
+        self.assertEqual(base, g._run_signature("stock"))
+
+
 if __name__ == "__main__":
     unittest.main()

tests/test_llm_max_retries.py

@@ -0,0 +1,100 @@
+"""Configurable LLM SDK retry budget (#1090/#1091).
+
+A single transient 429 burst used to kill an otherwise-healthy multi-agent run
+because each provider SDK's max_retries (default 2) was not exposed. This adds an
+opt-in llm_max_retries knob forwarded to every provider chat client.
+"""
+from __future__ import annotations
+
+import importlib
+
+import pytest
+
+import tradingagents.default_config as default_config_module
+from tradingagents.graph.trading_graph import TradingAgentsGraph, _coerce_max_retries
+
+# --- coercion / validation -------------------------------------------------
+
+@pytest.mark.unit
+@pytest.mark.parametrize("value,expected", [(0, 0), (2, 2), (10, 10), ("6", 6)])
+def test_coerce_accepts_non_negative_ints_and_numeric_strings(value, expected):
+    assert _coerce_max_retries(value) == expected
+
+
+@pytest.mark.unit
+@pytest.mark.parametrize("bad", [-1, "-3"])
+def test_coerce_rejects_negative(bad):
+    with pytest.raises(ValueError, match=">= 0"):
+        _coerce_max_retries(bad)
+
+
+@pytest.mark.unit
+@pytest.mark.parametrize("bad", [True, False])
+def test_coerce_rejects_booleans(bad):
+    with pytest.raises(ValueError, match="boolean"):
+        _coerce_max_retries(bad)
+
+
+@pytest.mark.unit
+@pytest.mark.parametrize("bad", ["abc", "1.5", None])
+def test_coerce_rejects_non_integers(bad):
+    with pytest.raises(ValueError, match="integer"):
+        _coerce_max_retries(bad)
+
+
+# --- forwarding into provider kwargs --------------------------------------
+
+def _bare_graph(config):
+    g = object.__new__(TradingAgentsGraph)
+    g.config = config
+    return g
+
+
+@pytest.mark.unit
+def test_not_forwarded_when_unset():
+    kwargs = _bare_graph({"llm_provider": "openai", "llm_max_retries": None})._get_provider_kwargs()
+    assert "max_retries" not in kwargs
+
+
+@pytest.mark.unit
+@pytest.mark.parametrize("provider", ["openai", "anthropic", "google"])
+def test_forwarded_across_providers(provider):
+    kwargs = _bare_graph({"llm_provider": provider, "llm_max_retries": 6})._get_provider_kwargs()
+    assert kwargs["max_retries"] == 6
+
+
+@pytest.mark.unit
+def test_forwarded_env_string_is_coerced():
+    # env vars arrive as strings; the consumer coerces (like temperature)
+    kwargs = _bare_graph({"llm_provider": "openai", "llm_max_retries": "4"})._get_provider_kwargs()
+    assert kwargs["max_retries"] == 4
+
+
+@pytest.mark.unit
+def test_invalid_config_value_fails_loudly():
+    with pytest.raises(ValueError):
+        _bare_graph({"llm_provider": "openai", "llm_max_retries": -1})._get_provider_kwargs()
+
+
+# --- env overlay -----------------------------------------------------------
+
+def _reload_with_env(monkeypatch, **overrides):
+    for key in list(default_config_module._ENV_OVERRIDES):
+        monkeypatch.delenv(key, raising=False)
+    for key, val in overrides.items():
+        monkeypatch.setenv(key, val)
+    return importlib.reload(default_config_module)
+
+
+@pytest.mark.unit
+def test_default_is_none(monkeypatch):
+    dc = _reload_with_env(monkeypatch)
+    assert dc.DEFAULT_CONFIG["llm_max_retries"] is None
+
+
+@pytest.mark.unit
+def test_env_override_sets_config(monkeypatch):
+    dc = _reload_with_env(monkeypatch, TRADINGAGENTS_LLM_MAX_RETRIES="8")
+    # None-default key: env value arrives as a string and is coerced downstream.
+    assert dc.DEFAULT_CONFIG["llm_max_retries"] == "8"
+    assert _coerce_max_retries(dc.DEFAULT_CONFIG["llm_max_retries"]) == 8