perf: preserve conversation prefixes across agent turns

tests/unit/test_prefix_cache.py

@@ -0,0 +1,63 @@
+import json
+from types import SimpleNamespace
+
+from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
+
+from artemis.llm.prefix_cache import PromptPrefixTracker
+from artemis.llm.router import ModelEndpoint
+from artemis.services.llm import RobustChatModelWrapper
+
+
+def test_prefix_diagnostics_distinguish_append_rewrite_and_tool_changes_without_payloads():
+    tracker = PromptPrefixTracker()
+    messages = [SystemMessage(content="private rules"), HumanMessage(content="private screenshot")]
+    options = {"tools": [{"name": "click"}, {"name": "observe"}]}
+    assert tracker.observe("task:flash", messages, options)["cold_start"]
+    messages.extend(
+        [
+            AIMessage(content="", tool_calls=[{"name": "click", "args": {"x": 5}, "id": "c1"}]),
+            ToolMessage(content="clicked", tool_call_id="c1"),
+        ]
+    )
+    extended = tracker.observe("task:flash", messages, options)
+    assert extended["append_only"] and extended["shared_prefix_messages"] == 2
+    messages[1] = HumanMessage(content="compressed summary")
+    rewritten = tracker.observe("task:flash", messages, options)
+    assert rewritten["first_changed_message"] == 2 and not rewritten["append_only"]
+    reordered = tracker.observe("task:flash", messages, {"tools": list(reversed(options["tools"]))})
+    assert reordered["tools_or_format_changed"] and not reordered["append_only"]
+    assert tracker.observe("task:checker", messages, options)["cold_start"]
+    assert "private" not in json.dumps(tracker._previous)
+
+
+def test_bound_wrappers_reuse_task_prefix_tracker_after_message_adaptation():
+    class Model:
+        kwargs = {"tools": [{"type": "function", "name": "observe"}]}
+
+        def bind_tools(self, *_args, **_kwargs):
+            return self
+
+    ctx = SimpleNamespace(data_engine=SimpleNamespace(current_session_id="session"))
+    wrapper = RobustChatModelWrapper(Model(), ctx=ctx, endpoint=ModelEndpoint(), role="flash")
+    messages = [SystemMessage(content="rules"), HumanMessage(content="step 1")]
+
+    def diagnostics(items):
+        bound = wrapper.bind_tools([])  # Flash binds again for each call
+        args, kwargs = bound._prepare_model_call((items,), {})
+        return bound._diagnostics(args, kwargs)
+
+    first = diagnostics(messages)
+    messages.append(HumanMessage(content="step 2"))
+    second = diagnostics(messages)
+    assert first["prefix_cache"]["cold_start"]
+    assert second["prefix_cache"]["append_only"]
+    assert first["prompt_cache_key"] == second["prompt_cache_key"]
+    assert not any("rules" in str(value) for value in ctx.prompt_prefix_tracker._previous.values())
+
+    ctx.transcript_ledger = SimpleNamespace(
+        context_phase=1, last_compaction={"reason": "context_pressure"}
+    )
+    third = diagnostics(messages[:1] + [HumanMessage(content="history summary")])
+    assert third["prefix_cache"]["context_phase"] == 1
+    assert third["prefix_cache"]["last_compaction"]["reason"] == "context_pressure"
+    assert not third["prefix_cache"]["append_only"]