fix(agent): keep system prompt cache-friendly

docs/ARCHITECTURE.md

@@ -162,8 +162,9 @@ budget), context compression first calls `memory.flush_conversation`, and
 
 ### Memory is pushed, not pulled
 Each turn the system prompt carries SOUL.md (persona), USER.md, MEMORY.md,
-recent daily notes, recalled facts, the skill index, the clock and a running
-summary of long conversations. v2 relied on the model deciding to call a memory
+recent daily notes, recalled facts, the skill index, and a running
+summary of long conversations. The current clock goes with the latest user
+message instead. v2 relied on the model deciding to call a memory
 tool before it knew anything about you. Tools remain for explicit recall,
 remembering, forgetting and history search. Memories learned from outside
 content (an email, a web page), from work nobody asked for, or from an import

sentient/agent/loop.py

@@ -11,7 +11,7 @@
   look-up tools requested together run concurrently; results keep their order.
 - ``Agent.run_turn`` is a chat turn: persists the user message (with
   attachments), builds the system prompt (persona, profile, recalled memory,
-  user model, skills, clock, running conversation summary), runs the loop,
+  user model, skills, running conversation summary), runs the loop,
   persists the transcript (the final reply carries its memory sources: the facts and
   insights in its prompt and the facts memory tools returned), then kicks off background work (fact extraction,
   auto title, context compression). Messages the user sends while a reply runs
@@ -34,6 +34,7 @@
 import weakref
 from collections.abc import AsyncIterator, Awaitable, Callable
 from dataclasses import dataclass, field
+from datetime import datetime
 from typing import Any
 
 from pydantic import ValidationError
@@ -87,6 +88,7 @@
     effective_risk,
     is_unprompted,
 )
+from sentient.tools.builtin.time_tool import resolve_tz
 from sentient.tools.registry import ToolRegistry
 from sentient.tools.rules import (
     SCREEN_SOURCE,
@@ -500,7 +502,6 @@ async def _system_prompt_with_sources(
             skills_index=self.skills.prompt_index(),
             assistant_name=self.config.assistant.name,
             user_name=self.config.assistant.user_name,
-            timezone=self.config.assistant.timezone,
             channel=channel,
             location=self.config.assistant.location,
             user_context=user_context,
@@ -1126,9 +1127,23 @@ def release() -> list[str]:
             system, sources = await self._system_prompt_with_sources(user_text, channel, session)
             history = await self.store.recent_messages(session_id, self.config.chat.history_window)
             convo = history_to_openai(history)
-            # replace the just-stored user message with the rich version (attachments inlined)
+
+            # Replace the just-stored user message with the rich version (attachments inlined).
+            # Keep the stored user message unchanged. Add the current time to the
+            # model-facing user message so it doesn't invalidate the system prompt cache.
             if convo and convo[-1]["role"] == "user":
-                convo[-1] = {"role": "user", "content": await self._user_content(user_text, attachments)}
+                content = await self._user_content(user_text, attachments)
+                tz = resolve_tz(self.config.assistant.timezone)
+                now = datetime.now(tz)
+                time_note = f"[Current time: {now.strftime('%A %Y-%m-%d %H:%M')} ({now.tzinfo})]"
+
+                if isinstance(content, str):
+                    content = f"{time_note}\n\n{content}"
+                else:
+                    content = [{"type": "text", "text": time_note}, *content]
+
+                convo[-1] = {"role": "user", "content": content}
+
             # spoken channels use the voice role: its own model and reasoning effort (default: thinking off)
             role = "voice" if channel in SPOKEN_CHANNELS else "primary"
             if attachments and any(is_image(paths.files_dir() / a) for a in attachments) and self.config.models.roles.vision:

sentient/agent/prompt.py

@@ -2,16 +2,13 @@
 
 Everything the model needs to *be* the user's assistant is injected here every
 turn: persona, who the user is, curated notes, recalled facts, the user model,
-the skill index, and the clock. The legacy system relied on the model calling a
+and the skill index. The legacy system relied on the model calling a
 memory tool before it knew anything; here memory is pushed, and tools are the fallback.
 """
 
 from __future__ import annotations
 
 from collections.abc import Iterable
-from datetime import datetime
-
-from sentient.tools.builtin.time_tool import resolve_tz
 
 TOOL_RULES = """## Working rules
 - If a request can be done with a tool, do it and report the outcome briefly. Do not narrate what you are about to do.
@@ -100,15 +97,12 @@ def build_system_prompt(
     skills_index: str,
     assistant_name: str,
     user_name: str,
-    timezone: str,
     channel: str,
     location: str = "",
     user_context: str = "",
     tool_names: Iterable[str] | None = None,
     read_only_apps: Iterable[str] | None = None,
 ) -> str:
-    tz = resolve_tz(timezone)
-    now = datetime.now(tz)
     parts: list[str] = []
 
     soul = snapshot.get("soul") or f"You are {assistant_name}, a personal assistant."
@@ -137,7 +131,6 @@ def build_system_prompt(
 
     parts.append(
         "## Context\n"
-        f"- Now: {now.strftime('%A %Y-%m-%d %H:%M')} ({now.tzinfo})\n"
         f"- User: {user_name or 'unknown (ask once, then remember)'}\n"
         f"- User location: {location or 'unknown'}\n"
         f"- Channel: {channel}\n"

tests/channels/test_telegram.py

@@ -288,7 +288,9 @@ async def test_message_while_running_is_queued_without_steer(tg, gated, monkeypa
     gated.gate.set()
     await tg.ch.wait_idle()
     assert gated.stream_calls == 2
-    assert gated.calls[-1]["messages"][-1]["content"] == "two"
+    content = gated.calls[-1]["messages"][-1]["content"]
+    assert "[Current time:" in content
+    assert content.endswith("two")
     assert tg.api.screen()[-1] == "second reply"
 
 
@@ -364,7 +366,9 @@ async def spy(method, **kwargs):
     await tg.say(42, None, **VOICE)
     assert voice.received == [(b"OggS-voice", "telegram-voice-u1.ogg")]
     assert any("<i>Heard:</i> what&#x27;s the weather" in t or "<i>Heard:</i> what's the weather" in t for t in tg.api.screen())
-    assert llm.calls[-1]["messages"][-1]["content"] == "what's the weather in Pune"
+    content = llm.calls[-1]["messages"][-1]["content"]
+    assert "[Current time:" in content
+    assert content.endswith("what's the weather in Pune")
     assert tg.api.screen()[-1] == "It's sunny."
     voices = [kwargs for method, kwargs in sent if method == "sendVoice"]
     assert len(voices) == 1
@@ -424,7 +428,13 @@ async def test_photo_and_document_become_attachments(tg, llm):
     saved = paths.files_dir() / "uploads" / "telegram-photo-big.jpg"
     assert saved.read_bytes() == b"\xff\xd8\xff\xe0fakejpeg"
     content = llm.calls[-1]["messages"][-1]["content"]
-    assert isinstance(content, list) and "what is this" in content[0]["text"]
+    assert isinstance(content, list)
+    assert "[Current time:" in content[0]["text"]
+    assert any(
+        block.get("type") == "text" and "what is this" in block.get("text", "")
+        for block in content
+    )
+    assert any(block.get("type") == "image_url" for block in content)
     doc = {"document": {"file_id": "d1", "file_unique_id": "doc", "file_name": "notes.txt", "mime_type": "text/plain", "file_size": 22}}
     await tg.say(42, None, **doc)
     assert "hello from a text file" in llm.calls[-1]["messages"][-1]["content"]

tests/channels/test_whatsapp.py

@@ -314,7 +314,12 @@ async def test_photo_and_document_become_attachments(wa, llm):
     await wa.say("what is this", media=[WAMedia(kind="image", name="whatsapp-image-1.jpg", mime="image/jpeg", size=12, ref="img")])
     assert (paths.files_dir() / "uploads" / "whatsapp-image-1.jpg").read_bytes() == b"\xff\xd8\xff\xe0fakejpeg"
     content = llm.calls[-1]["messages"][-1]["content"]
-    assert isinstance(content, list) and "what is this" in content[0]["text"]
+    assert isinstance(content, list)
+    assert any(
+        block.get("type") == "text"
+        and "what is this" in block.get("text", "")
+        for block in content
+    )
     await wa.say(media=[WAMedia(kind="document", name="notes.txt", mime="text/plain", size=22, ref="doc")])
     assert "hello from a text file" in llm.calls[-1]["messages"][-1]["content"]
     await wa.say(media=[WAMedia(kind="document", name="huge.zip", size=50 * 1024 * 1024, ref="big")])
@@ -426,9 +431,15 @@ async def test_voice_note_in_and_voice_note_out(wa, llm, monkeypatch):
     await wa.say(media=[WAMedia(kind="voice", name="whatsapp-voice-ABC.ogg", mime="audio/ogg; codecs=opus", size=10, ref="v1")])
     assert voice.received == [(b"OggS-voice", "whatsapp-voice-ABC.ogg")]
     assert "_Heard:_ what's the weather in Pune" in wa.hub.screen()
-    assert llm.calls[-1]["messages"][-1]["content"] == "what's the weather in Pune"
+    content = llm.calls[-1]["messages"][-1]["content"]
+    assert isinstance(content, str)
+    assert content.endswith("what's the weather in Pune")
     assert wa.hub.last() == "It's sunny."
     notes = [s for s in wa.hub.sent if s["op"] == "voice"]
+    print("Voice replies enabled:", wa.app.config.channels.whatsapp.voice_replies)
+    print("Voice service:", repr(wa.app.voice))
+    print("Sent operations:", [s["op"] for s in wa.hub.sent])
+    print("All sent messages:", repr(wa.hub.sent))
     assert len(notes) == 1 and notes[0]["data"].startswith(b"OggS") and notes[0]["seconds"] >= 1
 
 

tests/test_agent_upgrades.py

@@ -422,6 +422,58 @@ def broken(res):
 
 
 # ---------------------------------------------------------------------------- prompt caching
+async def test_current_time_is_in_user_message_not_system_prompt(config, isolated_home):
+    llm = FakeProvider(replies=["First answer.", "Second answer."])
+    s = await start(config, isolated_home, llm, "current-time")
+    try:
+        sid = await s.store.create_session(channel="cli")
+
+        async for _ in s.agent.run_turn(sid, "What time is it?", channel="cli"):
+            pass
+
+        first_messages = llm.calls[0]["messages"]
+        first_system = first_messages[0]["content"]
+        first_user = first_messages[-1]["content"]
+
+        assert "[Current time:" not in first_system
+        assert "[Current time:" in first_user
+        assert "What time is it?" in first_user
+
+        async for _ in s.agent.run_turn(sid, "And now?", channel="cli"):
+            pass
+
+        second_messages = llm.calls[1]["messages"]
+        second_system = second_messages[0]["content"]
+        second_user = second_messages[-1]["content"]
+
+        assert second_system == first_system
+        assert "[Current time:" in second_user
+        assert "And now?" in second_user
+
+        rows = await s.store.recent_messages(sid, 20)
+        assert rows[0]["content"] == "What time is it?"
+        assert rows[2]["content"] == "And now?"
+    finally:
+        await s.stop()
+
+
+def test_system_prompt_does_not_depend_on_current_time():
+    from sentient.agent.prompt import build_system_prompt
+
+    kwargs = {
+        "snapshot": {},
+        "facts": [],
+        "skills_index": "",
+        "assistant_name": "Sentient",
+        "user_name": "Maya",
+        "channel": "desktop",
+    }
+
+    prompt = build_system_prompt(**kwargs)
+
+    assert "- Now:" not in prompt
+    assert "[Current time:" not in prompt
+
 def test_prompt_cache_only_for_anthropic():
     messages = [{"role": "system", "content": "persona"}, {"role": "user", "content": "hi"}]
     tools = [{"type": "function", "function": {"name": "a"}}, {"type": "function", "function": {"name": "b"}}]
@@ -603,7 +655,7 @@ async def test_selection_triggers_surface_new_capabilities():
 
 def test_capability_rules_follow_available_tools():
     base = dict(snapshot={}, facts=[], skills_index="", assistant_name="Sentient", user_name="Sarthak",
-                timezone="UTC", channel="desktop")
+                channel="desktop")
     plain = build_system_prompt(**base, tool_names=["memory_recall"])
     assert "Open browser button" not in plain and "execute_code" not in plain
     rich = build_system_prompt(
@@ -678,7 +730,7 @@ async def test_steer_messages_are_marked_and_package_schemas_listed(config, isol
 
 
 async def test_messaging_and_phone_channels(config, isolated_home):
-    base = dict(snapshot={}, facts=[], skills_index="", assistant_name="Sentient", user_name="Sarthak", timezone="UTC")
+    base = dict(snapshot={}, facts=[], skills_index="", assistant_name="Sentient", user_name="Sarthak")
     assert "chatting in Telegram" in build_system_prompt(**base, channel="telegram")
     assert "No wide tables" in build_system_prompt(**base, channel="discord")
     llm = FakeProvider(replies=["Hi."])

tests/test_claude_code.py

@@ -179,7 +179,12 @@ async def test_chat_reply_streams_through_claude_code(make_app, fake_claude, iso
     assert not run["has_gateway_token"] and run["tool_search"] == "false"
     assert "Sentient" in run["system"]  # Sentient's own system prompt replaces Claude Code's
     [message] = run["stdin"]
-    assert message["type"] == "user" and message["message"]["content"][0]["text"] == "say hello"
+    content = message["message"]["content"]
+    assert isinstance(content, list)
+    assert any(
+        block.get("type") == "text" and block.get("text", "").endswith("say hello")
+        for block in content
+    )
     # it ran in a scratch folder under Sentient's home, removed afterwards; nothing is left running
     assert Path(run["cwd"]).is_relative_to(isolated_home) and not Path(run["cwd"]).exists()
     assert not claude_code._live

tests/test_screen_share.py

@@ -90,8 +90,17 @@ def test_a_screen_capture_goes_to_the_vision_role_and_marks_the_chat(config, iso
         role = "vision" if vision else "primary"
         content = _user_message(llm, role)["content"]
         assert isinstance(content, list)
-        assert content[1]["image_url"]["url"].startswith("data:image/png;base64,")
-        assert "Screen capture attached: screens/Inbox - Mail.png" in content[0]["text"]
+        image_block = next(
+            block for block in content
+            if block.get("type") == "image_url"
+        )
+        image_url = image_block["image_url"]["url"]
+        assert image_url.startswith("data:image/png;base64,")
+        assert any(
+            block.get("type") == "text"
+            and "Screen capture attached: screens/Inbox - Mail.png" in block.get("text", "")
+            for block in content
+        )
         listed = next(s for s in client.get("/api/sessions").json() if s["id"] == sid)
         assert listed["untrusted"] == "your screen"
 
@@ -104,7 +113,13 @@ def test_a_plain_image_upload_does_not_mark_the_chat(config, isolated_home, monk
         r = client.post("/api/chat", json={"text": "what is this?", "attachments": [up["name"]]})
         sid = json.loads(r.text.splitlines()[0])["session_id"]
         assert llm.calls[-1]["role"] == "vision"
-        assert "Image attached: uploads/cat.png" in _user_message(llm, "vision")["content"][0]["text"]
+        content = _user_message(llm, "vision")["content"]
+        assert isinstance(content, list)
+        assert any(
+            block.get("type") == "text"
+            and "Image attached: uploads/cat.png" in block.get("text", "")
+            for block in content
+        )
         listed = next(s for s in client.get("/api/sessions").json() if s["id"] == sid)
         assert listed["untrusted"] == ""