fix(tui-gateway): scope the hot-reload context pin to the configured default route

The Desktop config-save path re-applied model.context_length to any session, undoing the switch path's scoping; both live paths now share config_context_length_for_runtime.

tests/tui_gateway/test_compression_config_hot_reload.py

@@ -163,6 +163,7 @@ def _neutral_session(**compression_ctor):
     agent = SimpleNamespace(
         model="unset-test-model",
         provider="",
+        base_url="",
         context_compressor=compressor,
         compression_enabled=True,
         compression_idle_compact_after_seconds=0,

tests/tui_gateway/test_live_config_context_pin_mirror.py

@@ -24,6 +24,7 @@ def _session(config_context_length=PIN):
     agent = SimpleNamespace(
         model="model-a",
         provider="custom:acme",
+        base_url="http://127.0.0.1:8123/v1",
         _config_context_length=config_context_length,
         context_compressor=compressor,
         compression_enabled=True,
@@ -55,8 +56,28 @@ def test_changed_context_length_refreshes_the_agent_pin_too(monkeypatch):
     """A new ``model.context_length`` must land on both cached copies."""
     session, compressor = _session()
 
-    _sync(monkeypatch, session, {"model": {"context_length": 400_000}, "compression": {}})
+    _sync(monkeypatch, session, {"model": {"default": "model-a", "provider": "custom:acme",
+                                           "context_length": 400_000}, "compression": {}})
 
     assert compressor._config_context_length == 400_000
     assert compressor.context_length == 400_000
     assert session["agent"]._config_context_length == 400_000
+
+
+def test_hot_reload_does_not_pin_a_session_on_another_route(monkeypatch):
+    """The pin describes the configured default route; a session that switched elsewhere must not
+    inherit it on the next config save (same scoping as the switch path)."""
+    session, compressor = _session(config_context_length=None)
+    session["agent"].base_url = "https://openrouter.ai/api/v1"
+    session["agent"].provider = "openrouter"
+    monkeypatch.setattr(cc_mod, "get_model_context_length", lambda *a, **k: 256_000)
+
+    _sync(monkeypatch, session, {
+        "model": {"default": "model-b", "provider": "custom:acme", "base_url": "http://127.0.0.1:8123/v1",
+                  "context_length": PIN},
+        "custom_providers": [{"name": "acme", "base_url": "http://127.0.0.1:8123/v1", "models": {}}],
+        "compression": {},
+    })
+
+    assert compressor._config_context_length is None
+    assert session["agent"]._config_context_length is None

tui_gateway/session_compression.py

@@ -85,7 +85,7 @@ def _apply_live_compression_config(agent: Any, cfg: dict | None) -> None:
     cfg = cfg if isinstance(cfg, dict) else {}
     compression = cfg.get("compression") if isinstance(cfg.get("compression"), dict) else {}
     model_cfg = cfg.get("model") if isinstance(cfg.get("model"), dict) else {}
-    from agent.agent_init import set_config_context_length
+    from agent.agent_init import config_context_length_for_runtime, set_config_context_length
     enabled_raw = compression.get("enabled", True)
     agent.compression_enabled = enabled_raw if isinstance(enabled_raw, bool) else str(enabled_raw).lower() in {"true", "1", "yes"}
     agent.codex_responses_native_compaction = is_truthy_value(compression.get("codex_responses_native", False))
@@ -137,16 +137,17 @@ def _apply_live_compression_config(agent: Any, cfg: dict | None) -> None:
         cc.threshold_percent = cc._effective_threshold_percent(cc.context_length, base)
     except Exception:
         cc.threshold_percent = pct
-    raw_ctx = model_cfg.get("context_length")
-    if raw_ctx is not None:
-        with contextlib.suppress(TypeError, ValueError):
-            if (new_ctx := int(raw_ctx)) > 0:
-                # Both cached copies: the compressor's (its own re-resolution) and the agent's
-                # (switch/fallback + every display surface). Writing one left the other stale, so the
-                # session showed a pinned ceiling while compressing against a different window (#116467).
-                set_config_context_length(agent, new_ctx)
-                with contextlib.suppress(Exception):
-                    cc.context_length = new_ctx
+    # Same scoping rule as construction and the switch path: the pin describes the configured default
+    # route, so a session that /model-switched elsewhere must not have it re-applied on a config save
+    # (None = absent, invalid, or scoped out).
+    new_ctx = config_context_length_for_runtime(agent, cfg)
+    if new_ctx is not None:
+        # Both cached copies: the compressor's (its own re-resolution) and the agent's
+        # (switch/fallback + every display surface). Writing one left the other stale, so the
+        # session showed a pinned ceiling while compressing against a different window (#116467).
+        set_config_context_length(agent, new_ctx)
+        with contextlib.suppress(Exception):
+            cc.context_length = new_ctx
     elif getattr(cc, "_config_context_length", None) is not None:
         # model.context_length removed: drop the override and force re-inference from model metadata on
         # next access (construction's deferred resolution); re-applies the small-context floor too.