Skip to content

Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.

fix(brain): send the society reply policy once and cap core memory honestly

rubenluetke10-beepfbac18fa3e62
diff --git a/jarvis/brain/manager.py b/jarvis/brain/manager.py
index baf9c4fd7a..4762c32c64 100644
--- a/jarvis/brain/manager.py
+++ b/jarvis/brain/manager.py
@@ -2397,9 +2397,16 @@ _WRITTEN_CHAT_STYLE = (
"with plain-text headers and no emoji column."
)
+# The written-only rules, for a prompt that already carries the shared policy
+# (a society agent's briefing has it under "How to reply to the person").
+_WRITTEN_CHAT_RULES = _WRITTEN_CHAT_STYLE
+
# Keep the API chat and the main brain's written delivery equally conversational.
_WRITTEN_CHAT_STYLE += "\n" + CONVERSATIONAL_RESPONSE_STYLE
+#: Upper bound for the legacy core-memory block in the system prompt.
+_CORE_MEMORY_MAX_CHARS = 20_000
+
def _is_written_turn() -> bool:
"""True when the current turn is a typed (written) turn, not voice.
@@ -4225,7 +4232,11 @@ class BrainManager:
_TOOL_ROUTING_RULES,
self._render_live_tool_block(),
getattr(self, "_evidence_directive", ""),
- _WRITTEN_CHAT_STYLE,
+ # The briefing already carries the shared reply policy; sending
+ # it twice costs ~2.3k characters on every society API turn.
+ _WRITTEN_CHAT_RULES
+ if CONVERSATIONAL_RESPONSE_STYLE in (private.system_extra or "")
+ else _WRITTEN_CHAT_STYLE,
]
identity = getattr(self, "_active_turn_identity", None)
if identity:
@@ -4374,11 +4385,10 @@ class BrainManager:
except Exception: # noqa: BLE001
pass
cm = self._core_memory.render_system_prompt_block()
- # Cap substantially larger than the old 400 characters — otherwise
- # even 5-10 facts get cut off mid-block and the LLM claims it knows
- # nothing. 2500 corresponds to ~600 tokens, stays prompt-cache-friendly.
- if len(cm) > 2500:
- cm = cm[:20_000] + "…"
+ # A safety cap only (the 2500-character cost cap was retired): cut,
+ # and mark the cut, only when the block really is longer.
+ if len(cm) > _CORE_MEMORY_MAX_CHARS:
+ cm = cm[:_CORE_MEMORY_MAX_CHARS] + "…"
parts.append(cm)
# Skills-Brain-Integration (Track B): surface the installed, active
diff --git a/tests/unit/agent_chat/test_response_style.py b/tests/unit/agent_chat/test_response_style.py
index ddea242c7b..06d45a351f 100644
--- a/tests/unit/agent_chat/test_response_style.py
+++ b/tests/unit/agent_chat/test_response_style.py
@@ -2,6 +2,8 @@
from pathlib import Path
+import pytest
+
from jarvis.agent_chat import jarvis_harness
from jarvis.agent_chat.runner_api import system_prompt
from jarvis.brain.manager import _WRITTEN_CHAT_STYLE
@@ -48,3 +50,30 @@ def test_both_voice_personas_include_reporting_without_losing_spoken_rules() ->
assert "REPORTING RESULTS" in prompt
assert "SPOKEN" in prompt
assert "spell every number" in prompt
+
+
[email protected]("briefed", [True, False])
+def test_society_api_turn_carries_the_reply_policy_exactly_once(briefed: bool) -> None:
+ from types import SimpleNamespace
+
+ from jarvis.brain.manager import _TURN_OVERRIDE, BrainManager
+ from jarvis.brain.turn_override import TurnOverride
+
+ briefing = "## How to reply to the person\n" + CONVERSATIONAL_RESPONSE_STYLE
+ manager = BrainManager.__new__(BrainManager)
+ manager._config = SimpleNamespace(performance=SimpleNamespace(cache_optimized_prompt=True))
+ manager._render_live_tool_block = lambda: "TOOLS"
+ manager._reply_language_directive = lambda: "LANGUAGE"
+ token = _TURN_OVERRIDE.set(
+ TurnOverride(
+ provider="test",
+ system_extra=briefing if briefed else "## Agent\nScout",
+ tool_context={"tool_origin": "society"},
+ )
+ )
+ try:
+ prompt = manager._build_system_prompt()
+ finally:
+ _TURN_OVERRIDE.reset(token)
+ assert prompt.count(CONVERSATIONAL_RESPONSE_STYLE) == 1
+ assert "WRITTEN CHAT STYLE" in prompt