[codex] Restore Akasha prompt cache boundary

.github/workflows/ci.yml

@@ -54,7 +54,7 @@ jobs:
           PYTHONWARNINGS: "ignore:.*AbstractEventLoopPolicy.*:DeprecationWarning"
         run: |
           .venv/bin/python scripts/check_test_budget.py
-          .venv/bin/pytest -q -W error tests/
+          .venv/bin/pytest -q tests/
 
       - name: Run Web regressions
         run: |

agent/context.py

@@ -273,6 +273,7 @@ def render(
         *,
         system_sections_top: list[PromptSectionRender] | None = None,
         system_sections_bottom: list[PromptSectionRender] | None = None,
+        context_frame_sections: list[PromptSectionRender] | None = None,
     ) -> AssembledTurnInput:
         turn_injection_context = self.build_turn_injection_context(
             turn_injection_prompt=request.turn_injection_prompt
@@ -290,6 +291,7 @@ def render(
             turn_injection_context=turn_injection_context,
             system_sections_top=system_sections_top,
             system_sections_bottom=system_sections_bottom,
+            context_frame_sections=context_frame_sections,
         )
         self._last_render_diagnostics.set(
             (

agent/lifecycle/phases/prompt_render.py

@@ -102,6 +102,7 @@ async def run(self, frame: PromptRenderFrame) -> PromptRenderFrame:
             ),
             system_sections_top=ctx.system_sections_top,
             system_sections_bottom=ctx.system_sections_bottom,
+            context_frame_sections=ctx.context_frame_sections,
         )
         if ctx.extra_hints:
             rendered.messages.append(

agent/lifecycle/types.py

@@ -91,6 +91,9 @@ class PromptRenderCtx:
     system_sections_bottom: list[PromptSectionRender] = field(
         default_factory=list[PromptSectionRender]
     )
+    context_frame_sections: list[PromptSectionRender] = field(
+        default_factory=list[PromptSectionRender]
+    )
 
 
 # 保留旧导入名；运行时结果由 ContextBuilder 的唯一组装结果承载。

agent/prompting/assembler.py

@@ -49,7 +49,6 @@ def set(self, scope: str, section_name: str, signature: str, content: str) -> No
 
 _CONTEXT_FRAME_SECTIONS = {
     "active_skills",
-    "retrieved_memory",
 }
 SYSTEM_CONTEXT_FRAME_MARKER = '<system-reminder data-system-context-frame="true">'
 SYSTEM_CONTEXT_FRAME_END = "</system-reminder>"
@@ -99,6 +98,7 @@ def assemble(
         turn_injection_context: dict[str, str] | None = None,
         system_sections_top: list[PromptSectionRender] | None = None,
         system_sections_bottom: list[PromptSectionRender] | None = None,
+        context_frame_sections: list[PromptSectionRender] | None = None,
     ) -> AssembledTurnInput:
         # assembler 负责把“主 prompt + turn injection + message envelope”
         # 收束成一份统一输入，避免调用方各自手拼消息顺序。
@@ -135,11 +135,23 @@ def assemble(
             for section in all_sections
             if section.name not in _CONTEXT_FRAME_SECTIONS
         ]
-        frame_sections = [
+        built_frame_sections = [
             section
             for section in all_sections
             if section.name in _CONTEXT_FRAME_SECTIONS
         ]
+        contributed_frame_sections = [
+            section
+            for section in (context_frame_sections or [])
+            if section.name not in disabled
+        ]
+        frame_sections = [*built_frame_sections, *contributed_frame_sections]
+        frame_sections.sort(
+            key=lambda section: (
+                section.order is None,
+                section.order if section.order is not None else 0,
+            )
+        )
         for name, content in injection_context.items():
             text = content.strip()
             if text:
@@ -171,6 +183,7 @@ def assemble(
                 *_section_meta(top_sections),
                 *_section_meta(built_sections),
                 *_section_meta(bottom_sections),
+                *_section_meta(contributed_frame_sections),
             ],
         )
 

plugins/akasha/plugin.py

@@ -641,11 +641,12 @@ async def _inject_memory(
     )
     block = result.text_block.strip()
     if block:
-        event.system_sections_bottom.append(
+        event.context_frame_sections.append(
             PromptSectionRender(
-                name="memory",
+                name=RETRIEVED_MEMORY_SECTION,
                 content=block,
                 is_static=False,
+                order=10,
             )
         )
 

pytest.ini

@@ -2,7 +2,7 @@
 asyncio_mode = auto
 asyncio_default_fixture_loop_scope = function
 asyncio_default_test_loop_scope = session
-addopts = -W error
+addopts = -W error -W ignore::DeprecationWarning:starlette.testclient
 testpaths = tests
 markers =
     scenario_mvp: agent loop 基础场景测试

tests/test_plugin_composition_lifecycle.py

@@ -1,15 +1,18 @@
 from __future__ import annotations
 
 import asyncio
-from contextlib import asynccontextmanager
+from contextlib import asynccontextmanager, nullcontext
 from datetime import datetime
 import subprocess
 import sys
+from types import SimpleNamespace
 from typing import Any, AsyncIterator, cast
+from unittest.mock import AsyncMock
 
 import pytest
 
 from agent.core.response_parser import ResponseMetadata
+from agent.context import MessageEnvelopeBuilder
 from agent.lifecycle.composition import (
     AFTER_REASONING_CLEANUP_EVENT,
     AFTER_REASONING_PREPROCESS_EVENT,
@@ -61,6 +64,12 @@
 from core.memory.events import MemoryWritten, RetrievalCompleted
 from agent.retrieval.events import build_retrieval_completed
 from agent.retrieval.protocol import RetrievalRequest
+from agent.prompting import PromptAssembler, PromptSectionRender
+from plugins.akasha.plugin import _inject_memory
+from plugins.openai_compatible.driver import (
+    _merge_leading_system_messages,
+    _normalize_messages,
+)
 
 
 @asynccontextmanager
@@ -93,6 +102,76 @@ def _prompt_ctx() -> PromptRenderCtx:
     )
 
 
+async def _assert_akasha_inserts_first_user_context_frame_block() -> None:
+    ctx = _prompt_ctx()
+    runtime = SimpleNamespace(
+        query=AsyncMock(return_value=MemoryQueryResult(text_block="fresh recall"))
+    )
+    diagnostics = SimpleNamespace(
+        operation=lambda _name: nullcontext(),
+        measure=lambda _name, _value: None,
+    )
+
+    await _inject_memory(ctx, cast(Any, runtime), cast(Any, diagnostics))
+
+    assert ctx.system_sections_bottom == []
+    assert [
+        (section.name, section.content, section.order)
+        for section in ctx.context_frame_sections
+    ] == [("memory", "fresh recall", 10)]
+
+
+def _assert_context_frame_keeps_dynamic_memory_after_stable_history() -> None:
+    history = [
+        {"role": "user", "content": "old question"},
+        {"role": "assistant", "content": "old answer"},
+    ]
+
+    class _ContextStub:
+        _envelope_builder = MessageEnvelopeBuilder()
+
+        @staticmethod
+        def _build_system_prompt_sections(**_kwargs: object) -> list[PromptSectionRender]:
+            return [
+                PromptSectionRender("stable", "stable system", True, order=20),
+                PromptSectionRender("active_skills", "active skill", False, order=50),
+            ]
+
+    assembler = PromptAssembler(cast(Any, _ContextStub()))
+
+    def assemble(memory: str):
+        return assembler.assemble(
+            history=history,
+            current_message="current question",
+            multimodal=False,
+            context_frame_sections=[
+                PromptSectionRender("memory", memory, False, order=10)
+            ],
+        )
+
+    first = assemble("recall one")
+    second = assemble("recall two")
+    provider_messages = _merge_leading_system_messages(
+        _normalize_messages(first.messages)
+    )
+
+    assert first.system_prompt == second.system_prompt == "stable system"
+    assert first.messages[:3] == second.messages[:3]
+    assert [message["role"] for message in provider_messages] == [
+        "system",
+        "user",
+        "assistant",
+        "user",
+        "user",
+    ]
+    reminder = provider_messages[-2]
+    assert reminder["role"] == "user"
+    assert str(reminder["content"]).startswith("<system-reminder")
+    assert str(reminder["content"]).index("## memory") < str(
+        reminder["content"]
+    ).index("## active_skills")
+
+
 def _before_turn_ctx() -> BeforeTurnCtx:
     return BeforeTurnCtx(
         session_key="session",
@@ -489,6 +568,9 @@ async def enqueue() -> None:
 
 @pytest.mark.asyncio
 async def test_retrieval_completed_event_payload() -> None:
+    await _assert_akasha_inserts_first_user_context_frame_block()
+    _assert_context_frame_keeps_dynamic_memory_after_stable_history()
+
     observed: list[RetrievalCompleted] = []
     root = CompositionRoot("retrieval-completed")