diff --git a/agent/model_metadata.py b/agent/model_metadata.py
index 2ce157f11e9ac..b4bf55d529580 100644
--- a/agent/model_metadata.py
+++ b/agent/model_metadata.py
@@ -2659,7 +2659,17 @@ def estimate_messages_tokens_rough(messages: List[Dict[str, Any]]) -> int:
 # request, just a truncated answer the provider happily returns and the user
 # cannot use — the max_tokens=1 defect.  ``None`` (leave the cap alone, let the
 # provider's own budget decide) is the honest answer there instead.
-_OUTPUT_FIT_MIN_USABLE = 512
+#
+# Deliberately the same value as ``conversation_loop.
+# _LENGTH_CONTINUE_MIN_HEADROOM_TOKENS``: that guard refuses to scaffold another
+# length-continuation round below it, on the grounds that there is no room left
+# for a useful completion.  Same question, so the same answer — and the two must
+# not disagree.  A clamp below that threshold would fight it: the truncated
+# response starts a continuation round, the boost there sizes the next cap to
+# the real context headroom, and this clamp shrinks it straight back to a value
+# the continuation guard itself considers too small to write into.  (No import —
+# conversation_loop imports this module, not the other way round.)
+_OUTPUT_FIT_MIN_USABLE = 2048
 
 
 def output_tokens_that_fit(
@@ -2697,15 +2707,16 @@ def output_tokens_that_fit(
     this local estimate makes it converge to a fitting value in one step.
 
     The reservation degrades in two tiers as the window fills (see the constants
-    above).  The preferred 1.2x cushion is multiplicative, so it exhausts the
-    window on its own once the estimate passes ``(context_length - 2048) / 1.2``
-    — ~82.5% fill.  That is where the *cushion* becomes unaffordable, not where
-    the request stops fitting, so past that point the reservation drops to the
-    1.15x safety contract and keeps reporting real caps (~9,000 tokens at 83%
-    fill of a 200,000 window, tapering to ``min_output``).  ``None`` arrives
-    only when the contract itself admits nothing usable, at ~87% fill — the
-    ``1/1.15`` ceiling — and there the proactive clamp stands down and lets the
-    provider report its own authoritative budget on the reactive path.
+    above).  The preferred 1.2x cushion is multiplicative, so it stops leaving a
+    usable cap once the estimate passes ``(context_length - 3584) / 1.2`` —
+    ~81.8% fill of a 200,000-token window.  That is where the *cushion* becomes
+    unaffordable, not where the request stops fitting, so past that point the
+    reservation drops to the 1.15x safety contract and keeps reporting real caps
+    (~10,900 tokens at 82% fill, tapering to ``min_output``).  ``None`` arrives
+    only when the contract itself admits nothing usable, at ~85.8% fill — where
+    the ``1/1.15`` ceiling meets the usable floor — and there the proactive
+    clamp stands down and lets the provider report its own authoritative budget
+    on the reactive path.
     """
     if not isinstance(context_length, int) or context_length <= 0:
         return None
diff --git a/tests/test_output_fit_preflight_clamp.py b/tests/test_output_fit_preflight_clamp.py
index 249bc75391824..1ccb8500720f0 100644
--- a/tests/test_output_fit_preflight_clamp.py
+++ b/tests/test_output_fit_preflight_clamp.py
@@ -9,6 +9,7 @@
 from types import SimpleNamespace
 
 from agent.chat_completion_helpers import _preflight_clamp_output_tokens
+from agent.conversation_loop import _LENGTH_CONTINUE_MIN_HEADROOM_TOKENS
 from agent.model_metadata import (
     _OUTPUT_FIT_MIN_USABLE,
     _OUTPUT_FIT_WINDOW_MARGIN,
@@ -191,10 +192,23 @@ class TestHighFillBandDegradesGracefully:
     round-trip per turn.
 
     So the reservation now degrades to the 1.15x safety contract instead of
-    vanishing.  ``None`` is still the answer past the ``1/1.15`` ceiling, where
-    nothing usable genuinely fits.
+    vanishing.  ``None`` is still the answer past ~85.8% fill, where the
+    contract can no longer clear the usable floor.
     """
 
+    def test_no_reported_cap_is_below_the_continuation_guard(self):
+        # ``_OUTPUT_FIT_MIN_USABLE`` is deliberately the same value as the
+        # length-continuation head-room guard, which refuses to scaffold another
+        # round below it because there is no room left for a useful completion.
+        # A clamp under that threshold would fight it: the truncated response
+        # starts a continuation round, the boost there sizes the next cap to the
+        # real context head-room, and the pre-flight clamp shrinks it straight
+        # back to a value the guard itself calls too small to write into.
+        assert _OUTPUT_FIT_MIN_USABLE >= _LENGTH_CONTINUE_MIN_HEADROOM_TOKENS
+        for tokens in range(1_000, _CTX_200K, 500):
+            fit = output_tokens_that_fit(_CTX_200K, _msgs(tokens))
+            assert fit is None or fit >= _LENGTH_CONTINUE_MIN_HEADROOM_TOKENS
+
     def test_reported_fit_where_the_preferred_cushion_admits_none(self):
         # The case from the report: at est=166,008 in a 200,000 window the 1.2x
         # reservation overruns the window outright, while the 1.15x contract
@@ -207,10 +221,11 @@ def test_reported_fit_where_the_preferred_cushion_admits_none(self):
         assert _fits_for_a_denser_server(msgs, fit, _CTX_200K)
 
     def test_band_is_covered_without_a_hole(self):
-        # Sweep the whole band the old formula gave up on.  Every fill level
-        # from the 1.2x limit (~82.7%) up to the 1.15x ceiling (~86.9%) must
-        # report a real cap, and the caps must taper rather than jump to None.
-        for pct in range(83, 87):
+        # Sweep the band the old formula gave up on.  Every fill level from the
+        # 1.2x limit (~82.7%) up to where the contract stops clearing the usable
+        # floor (~85.8%) must report a real cap, and the caps must taper rather
+        # than jump to None.
+        for pct in range(83, 86):
             msgs = _msgs(int(_CTX_200K * pct / 100))
             fit = output_tokens_that_fit(_CTX_200K, msgs)
             assert fit is not None, f"no cap reported at {pct}% fill"