diff --git a/python/freetoken/server/api_server.py b/python/freetoken/server/api_server.py
index 5592a2b0d..7267b1a9c 100644
--- a/python/freetoken/server/api_server.py
+++ b/python/freetoken/server/api_server.py
@@ -710,6 +710,30 @@ def _reasoning_geometry(state: Any) -> dict | None:
     return {"gears": list(gears), "default": default, "kwargs": kwargs}
 
 
+def kv_pool_geometry(state: Any) -> tuple[int, int]:
+    """The allocated KV pool as ``(num_pages, page_size)``, most-recent-truth first: the last
+    rebuild's result, else the running UserReply snapshot, else the load-time ("meta", …) ack.
+    ``num_pages`` is 0 until one of those lands (still loading / older engine build).
+
+    Split out of cache_geometry so /v1/models can clamp the advertised context to the pool
+    without recomputing the whole panel payload -- both callers must agree on the resolution
+    order or they would report different capacities for the same server.
+    """
+    stats = getattr(state, "stats", None)
+    last = getattr(state, "last_rebuild", None) or {}
+    pools = getattr(state, "cache_pools", None) or {}
+    num_pages = int(
+        last.get("num_pages")
+        or getattr(stats, "kv_total_pages", 0)
+        or pools.get("num_pages", 0)
+        or 0
+    )
+    page_size = int(
+        pools.get("page_size", 0) or getattr(getattr(state, "config", None), "page_size", 1) or 1
+    )
+    return num_pages, page_size
+
+
 def cache_geometry(state: Any) -> dict:
     """Current cache geometry for the desktop cache panel. Each pool size resolves
     most-recent-truth first: the last rebuild's result, else the running UserReply snapshot
@@ -725,11 +749,10 @@ def cache_geometry(state: Any) -> dict:
     config = state.config
     last = getattr(state, "last_rebuild", None) or {}
     pools = getattr(state, "cache_pools", None) or {}
-    num_pages = int(last.get("num_pages") or tr.kv_total_pages or pools.get("num_pages", 0) or 0)
+    num_pages, page_size = kv_pool_geometry(state)
     num_mamba_slots = int(
         last.get("mamba_slots") or tr.mamba_total_slots or pools.get("num_mamba_slots", 0) or 0
     )
-    page_size = int(pools.get("page_size", 0) or getattr(config, "page_size", 1) or 1)
     moe_cache_size = last.get("moe_cache_size")
     if moe_cache_size is None:
         moe_cache_size = int(pools.get("moe_cache_size", 0) or 0) or configured_moe_cache_size(
diff --git a/python/freetoken/server/openai_api.py b/python/freetoken/server/openai_api.py
index dc2f73a97..087b2a4e4 100644
--- a/python/freetoken/server/openai_api.py
+++ b/python/freetoken/server/openai_api.py
@@ -701,10 +701,31 @@ def _served_model_name(state: Any) -> str:
 
 
 def _model_context_length(state: Any) -> int | None:
-    """The model ceiling, not `min(ceiling, KV budget)`: a rebuild moves the latter, and agents
-    read this once at startup."""
+    """The usable context: ``min(model ceiling, allocated KV tokens)``.
+
+    This is the clamp the engine already applies to its own ``max_seq_len`` (Engine.__init__,
+    and again in _refresh_seq_state after a rebuild), and therefore the limit the scheduler
+    admits against. The frontend process holds an unclamped ServerArgs copy, so publishing
+    ``config.max_seq_len`` here advertised the model's positional ceiling instead -- e.g. 262144
+    against a 178176-token pool, so the engine and this route disagreed about one quantity.
+
+    `ft launch` reads this to size each agent's context window (opencode's ``limit.context``,
+    codex's ``context_window``, ``CLAUDE_CODE_MAX_CONTEXT_TOKENS``), so overstating it stops
+    those agents compacting before the KV pool runs out.
+    """
     try:  # never 500 a metadata route: max_seq_len walks into the HF config on some builds
         value = int(state.config.max_seq_len)
     except Exception:  # noqa: BLE001
         return None
-    return value if value > 0 else None
+    if value <= 0:
+        return None
+    try:
+        # Local import: api_server imports this module, so a module-level one would cycle.
+        from .api_server import kv_pool_geometry
+
+        num_pages, page_size = kv_pool_geometry(state)
+        kv_tokens = num_pages * page_size
+    except Exception:  # noqa: BLE001
+        kv_tokens = 0
+    # 0 before the ("meta", …) ack lands; the ceiling is the best available answer until then.
+    return min(value, kv_tokens) if kv_tokens > 0 else value
diff --git a/tests/server/test_openai_api.py b/tests/server/test_openai_api.py
index e33018e19..d9d565ff6 100644
--- a/tests/server/test_openai_api.py
+++ b/tests/server/test_openai_api.py
@@ -442,19 +442,67 @@ def test_models_route_returns_served_model_name():
     assert card["max_model_len"] is None and card["context_length"] is None
 
 
+def _models_card(state):
+    app = FastAPI()
+    register_openai_routes(app, lambda: state, lambda: {})
+    return TestClient(app).get("/v1/models").json()["data"][0]
+
+
 def test_models_route_publishes_the_model_context_length():
     """`ft launch` reads this to size each agent's context window."""
     state = FakeState([])
     state.config.max_seq_len = 262144
-    app = FastAPI()
-    register_openai_routes(app, lambda: state, lambda: {})
 
-    card = TestClient(app).get("/v1/models").json()["data"][0]
+    card = _models_card(state)
 
+    # No pool geometry resolved yet (still loading, or pre-("meta", …) ack): the model
+    # ceiling is the only answer available.
     assert card["max_model_len"] == 262144
     assert card["context_length"] == 262144
 
 
+def test_models_route_clamps_the_context_length_to_the_kv_pool():
+    """The engine clamps its own max_seq_len to the allocated pool and the scheduler admits
+    against that, so publishing the unclamped ceiling here made `ft launch` size each agent's
+    compaction window past what the server can actually hold (#448)."""
+    state = FakeState([])
+    state.config.max_seq_len = 262144
+    state.config.page_size = 1
+    state.cache_pools = {"num_pages": 178176, "page_size": 1}
+
+    card = _models_card(state)
+
+    assert card["max_model_len"] == 178176
+    assert card["context_length"] == 178176
+
+
+def test_models_route_keeps_the_ceiling_when_the_pool_exceeds_it():
+    """A pool larger than the model's positional ceiling does not extend the context."""
+    state = FakeState([])
+    state.config.max_seq_len = 32768
+    state.config.page_size = 1
+    state.cache_pools = {"num_pages": 178176, "page_size": 1}
+
+    card = _models_card(state)
+
+    assert card["max_model_len"] == 32768
+    assert card["context_length"] == 32768
+
+
+def test_models_route_prefers_the_last_rebuild_over_the_load_time_pool():
+    """A rebuild moves the pool; /v1/models must follow it rather than freeze the load-time
+    allocation (same most-recent-truth order /v1/cache/status reports)."""
+    state = FakeState([])
+    state.config.max_seq_len = 262144
+    state.config.page_size = 1
+    state.cache_pools = {"num_pages": 178176, "page_size": 1}
+    state.last_rebuild = {"num_pages": 40000}
+
+    card = _models_card(state)
+
+    assert card["max_model_len"] == 40000
+
+
 async def _collect(generator):
     return [chunk async for chunk in generator]