fix(agent_init): judge the served window only for local endpoints; clamp after the floor as before

Review findings on the first cut: * A hosted provider with a stale model.ollama_num_ctx passed the floor for a 40K model although nothing ever raises a hosted window. The served window now counts only when the endpoint is local (is_local_endpoint), the same gate the num_ctx probe uses. * Moving the whole num_ctx phase ahead of the floor also moved tek's compressor clamp ahead of it, which flipped two observables: a model.context_length above a sub-64K num_ctx was rejected instead of constructed-and-clamped, and the floor's message reported the clamped value with advice (set model.context_length) that could not help. The phase is split: resolution runs before the floor, the clamp (_clamp_compressor_to_ollama_num_ctx) stays at its original position, so every case main constructed still constructs with identical compressor numbers and the floor's message is unchanged. Tests: the harness is a module-level helper so the new class no longer re-collects the parent's tests (13 -> 11 collected); the negative pins the local-endpoint gate (red when the gate is dropped) instead of a case main already rejected.
2026-09-19 11:38:54 +05:30
parent 3de7140bcd
commit afaa53e5fe
2 changed files with 45 additions and 39 deletions
--- a/agent/agent_init.py
+++ b/agent/agent_init.py
@@ -1912,11 +1912,11 @@ def _enforce_minimum_context(agent):
    # Reject windows below the 64K floor needed for reliable tool-calling; an explicit
    # positive model.context_length on LM Studio is allowed below the floor.
    _ctx = getattr(agent.context_compressor, "context_length", 0)
-    # An Ollama server serves num_ctx, not the GGUF's advertised window: a Modelfile or
+    # A local Ollama server serves num_ctx, not the GGUF's advertised window: a Modelfile or
    # model.ollama_num_ctx at 64K+ is a usable window even when the metadata says 40K (#100437).
-    _served = getattr(agent, "_ollama_num_ctx", None)
-    if isinstance(_served, int) and not isinstance(_served, bool) and _served > 0:
-        _ctx = max(_ctx or 0, _served)
+    # Only a local endpoint can honour num_ctx, so a stale override never admits a hosted model.
+    if agent._ollama_num_ctx and agent.base_url and is_local_endpoint(agent.base_url):
+        _ctx = max(_ctx or 0, agent._ollama_num_ctx)
    _allow_lmstudio_explicit_below_floor = (
        str(agent.provider or "").strip().lower() == "lmstudio"
        and isinstance(agent._config_context_length, int)
@@ -2043,6 +2043,9 @@ def _configure_ollama_num_ctx(agent, _model_cfg, _config_context_length):
            "Ollama num_ctx: will request %d tokens (model max from /api/show)",
            agent._ollama_num_ctx,
        )
+
+
+def _clamp_compressor_to_ollama_num_ctx(agent):
    # Recalibrate the compressor to the served window: every request runs at num_ctx, so a
    # trigger derived from the probed model window could sit above it and never fire.
    # A config that sets only model.ollama_num_ctx (without model.context_length) previously left the
@@ -2331,12 +2334,12 @@ def init_agent(
        agent, _agent_cfg, base_url
    )
    _build_context_engine(agent, _agent_cfg, cs, _custom_providers, _effective_context_length, session_db)
-    # num_ctx before the floor: the served Ollama window is part of what the floor judges.
    _configure_ollama_num_ctx(agent, _model_cfg, _config_context_length)
    _enforce_minimum_context(agent)
    _warn_nonagentic_hermes_model(agent)
    _inject_context_engine_tools(agent)
    _init_usage_state(agent)
+    _clamp_compressor_to_ollama_num_ctx(agent)
    _emit_compression_summary(agent, cs)
    _snapshot_primary_runtime(agent)

--- a/tests/agent/test_ollama_num_ctx.py
+++ b/tests/agent/test_ollama_num_ctx.py
@@ -140,39 +140,40 @@ class TestQueryOllamaSupportsVision:
 # ═══════════════════════════════════════════════════════════════════════


+def _build_agent(cfg, probed_ctx, base_url="http://localhost:11434/v1"):
+    import agent.context_compressor as cc_mod
+    with (
+        patch("model_tools.get_tool_definitions", return_value=[]),
+        patch("model_tools.check_toolset_requirements", return_value={}),
+        patch("agent.process_bootstrap.OpenAI"),
+        patch("hermes_cli.config.load_config", return_value=cfg),
+        patch("hermes_cli.config.load_config_readonly", return_value=cfg),
+        patch(
+            "agent.model_metadata.get_model_context_length",
+            return_value=probed_ctx,
+        ),
+        patch.object(
+            cc_mod, "get_model_context_length", return_value=probed_ctx,
+        ),
+    ):
+        from run_agent import AIAgent
+        return AIAgent(
+            model="gemma3:27b",
+            api_key="ollama",
+            base_url=base_url,
+            quiet_mode=True,
+            skip_context_files=True,
+            skip_memory=True,
+        )
+
+
 class TestCompressorClampsToNumCtx:
    """A config setting ONLY model.ollama_num_ctx (no model.context_length)
    must not leave the compressor targeting the probed model window while
    requests run at the smaller served num_ctx."""

-    def _build_agent(self, cfg, probed_ctx):
-        import agent.context_compressor as cc_mod
-        with (
-            patch("model_tools.get_tool_definitions", return_value=[]),
-            patch("model_tools.check_toolset_requirements", return_value={}),
-            patch("agent.process_bootstrap.OpenAI"),
-            patch("hermes_cli.config.load_config", return_value=cfg),
-            patch("hermes_cli.config.load_config_readonly", return_value=cfg),
-            patch(
-                "agent.model_metadata.get_model_context_length",
-                return_value=probed_ctx,
-            ),
-            patch.object(
-                cc_mod, "get_model_context_length", return_value=probed_ctx,
-            ),
-        ):
-            from run_agent import AIAgent
-            return AIAgent(
-                model="gemma3:27b",
-                api_key="ollama",
-                base_url="http://localhost:11434/v1",
-                quiet_mode=True,
-                skip_context_files=True,
-                skip_memory=True,
-            )
-
    def test_num_ctx_only_config_clamps_compressor_window(self):
-        agent = self._build_agent(
+        agent = _build_agent(
            {"agent": {}, "model": {"ollama_num_ctx": 65536}}, probed_ctx=262144
        )
        assert agent._ollama_num_ctx == 65536
@@ -183,7 +184,7 @@ class TestCompressorClampsToNumCtx:
        assert agent.context_compressor.threshold_tokens < 65536

    def test_larger_num_ctx_does_not_inflate_compressor_window(self):
-        agent = self._build_agent(
+        agent = _build_agent(
            {"agent": {}, "model": {"ollama_num_ctx": 131072}}, probed_ctx=65536
        )
        # num_ctx above the resolved window must not RAISE the compressor
@@ -191,16 +192,18 @@ class TestCompressorClampsToNumCtx:
        assert agent.context_compressor.context_length == 65536


-class TestServedNumCtxSatisfiesTheFloor(TestCompressorClampsToNumCtx):
-    """#100437: the 64K floor judges the window Ollama actually serves. A Modelfile or
-    model.ollama_num_ctx at 64K+ is usable even when the GGUF metadata advertises 40K, so
+class TestServedNumCtxSatisfiesTheFloor:
+    """#100437: the 64K floor judges the window a local Ollama server actually serves. A Modelfile
+    or model.ollama_num_ctx at 64K+ is usable even when the GGUF metadata advertises 40K, so
    construction must succeed; the compressor still targets the smaller probed window."""

    def test_explicit_num_ctx_above_the_floor_admits_a_small_metadata_window(self):
-        agent = self._build_agent({"agent": {}, "model": {"ollama_num_ctx": 65536}}, probed_ctx=40960)
+        agent = _build_agent({"agent": {}, "model": {"ollama_num_ctx": 65536}}, probed_ctx=40960)
        assert agent._ollama_num_ctx == 65536
        assert agent.context_compressor.context_length == 40960  # one-directional clamp unchanged

-    def test_served_window_below_the_floor_is_still_rejected(self):
+    def test_served_window_counts_only_for_a_local_endpoint(self):
+        """Only a local server honours num_ctx; a stale override must not admit a hosted 40K model."""
        with pytest.raises(ValueError, match="below the minimum"):
-            self._build_agent({"agent": {}, "model": {"ollama_num_ctx": 32768}}, probed_ctx=40960)
+            _build_agent({"agent": {}, "model": {"ollama_num_ctx": 65536}}, probed_ctx=40960,
+                         base_url="https://openrouter.ai/api/v1")