diff --git a/agentdescent/cli.py b/agentdescent/cli.py
index bdd1f24c..109814d0 100644
--- a/agentdescent/cli.py
+++ b/agentdescent/cli.py
@@ -1671,6 +1671,21 @@ def _key_sent_to_openai(block: Any, key_env: str, url_env: str) -> Optional[str]
             f"`default_base_url` (the endpoint URL) or export {url_env}")
 
 
+def _propose_fail_rounds_text(spec: EvolveSpec) -> str:
+    """How many rounds a run with a dead reflector lasts, in words.
+
+    The evolve loop stops after ``PROPOSE_FAIL_ROUNDS`` rounds in which every
+    proposal raised, but a spec asking for fewer rounds ends sooner: the quick
+    trial ``--rounds 1`` was told its run "fails after two rounds".
+    """
+    from .evolution import PROPOSE_FAIL_ROUNDS
+    n = PROPOSE_FAIL_ROUNDS
+    rounds = (spec.evolve or {}).get("rounds") if isinstance(spec.evolve, dict) else None
+    if isinstance(rounds, int) and not isinstance(rounds, bool) and 0 < rounds < n:
+        n = rounds
+    return {1: "one round", 2: "two rounds"}.get(n, f"{n} rounds")
+
+
 def _unusable_refs(spec: EvolveSpec) -> List[str]:
     """Warnings for agents this machine cannot actually run.
 
@@ -1735,9 +1750,9 @@ def _unusable_refs(spec: EvolveSpec) -> List[str]:
                          "for its endpoint" if present else "")
                 out.append(f"{field}: `openai_compatible` needs {key_env} (and "
                            f"{url_env} for a non-OpenAI endpoint){route}; without it "
-                           + ("every proposal raises and the run fails after two rounds "
-                              "with nothing learned" if field == "reflect"
-                              else "every call raises"))
+                           + (f"every proposal raises and the run fails after "
+                              f"{_propose_fail_rounds_text(spec)} with nothing learned"
+                              if field == "reflect" else "every call raises"))
             else:
                 leak = _key_sent_to_openai(block, key_env, url_env)
                 if leak:
diff --git a/changelog.d/872-dead-reflector-rounds.md b/changelog.d/872-dead-reflector-rounds.md
new file mode 100644
index 00000000..79122ac2
--- /dev/null
+++ b/changelog.d/872-dead-reflector-rounds.md
@@ -0,0 +1,4 @@
+Fixed
+The missing-reflector-key warning from `plan` / `algorithms run` now names how
+long the run really lasts: "fails after one round" on a `--rounds 1` quick
+trial instead of always "two rounds" (#872).
diff --git a/tests/test_key_preflight.py b/tests/test_key_preflight.py
index 3cff5371..9fcf0354 100644
--- a/tests/test_key_preflight.py
+++ b/tests/test_key_preflight.py
@@ -155,3 +155,14 @@ def test_claude_without_api_key_env_is_unchanged_and_unset_key_is_clear(monkeypa
     assert seen["init"] == {"max_retries": 0}
     with pytest.raises(RuntimeError, match="set X_KEY"):
         claude("m", api_key_env="X_KEY", retries=1)("p")
+
+
+@pytest.mark.parametrize("rounds,words", [(1, "one round"), (2, "two rounds"),
+                                          (8, "two rounds")])
+def test_dead_reflector_warning_names_the_rounds_the_spec_will_run(tmp_path, rounds, words):
+    # #872: a --rounds 1 quick trial was told it "fails after two rounds".
+    spec = _spec(tmp_path, reflect={"ref": "openai_compatible", "model": "m",
+                                    "api_key_env": "X_UNSET_KEY"})
+    spec.evolve = {"rounds": rounds}
+    [warning] = _unusable_refs(spec)
+    assert warning.endswith(f"the run fails after {words} with nothing learned"), warning