Skip to content

Personal AI, On Personal Devices

Python

fix(server): load SOUL.md / USER.md context in streaming chat (#449)

Arush Wadhawan09b19193fe54
diff --git a/src/openjarvis/server/agent_manager_routes.py b/src/openjarvis/server/agent_manager_routes.py
index 91459ba7..9c2da535 100644
--- a/src/openjarvis/server/agent_manager_routes.py
+++ b/src/openjarvis/server/agent_manager_routes.py
@@ -439,6 +439,26 @@ _SAMPLER_PARAM_KEYS = (
)
+def _build_managed_system_prompt(system_prompt: str, app_config: Any) -> str:
+ """Build the streaming managed-agent system prompt via SystemPromptBuilder.
+
+ Routes the agent's own ``system_prompt`` through the same builder the
+ CLI/ask path uses, so SOUL.md / MEMORY.md / USER.md persona files are
+ injected for streaming chat too (#431). Returns the assembled prompt
+ (caller decides whether to append a SYSTEM message); an agent with
+ neither persona nor template yields an empty string, preserving the
+ prior no-SYSTEM-message behavior.
+ """
+ from openjarvis.prompt.builder import SystemPromptBuilder
+
+ builder = SystemPromptBuilder(
+ agent_template=system_prompt or "",
+ memory_files_config=getattr(app_config, "memory_files", None),
+ system_prompt_config=getattr(app_config, "system_prompt", None),
+ )
+ return builder.build()
+
+
def _sampler_kwargs(config: Dict[str, Any]) -> Dict[str, Any]:
"""Extract per-agent sampler params from a managed agent's config (#386)."""
out: Dict[str, Any] = {}
@@ -811,8 +831,23 @@ async def _stream_managed_agent(
# Build conversation messages from history + current input
llm_messages: List[Message] = []
- if system_prompt:
- llm_messages.append(Message(role=Role.SYSTEM, content=system_prompt))
+
+ # Wire the SystemPromptBuilder to inject SOUL.md / MEMORY.md / USER.md
+ # persona files (parity with the CLI/ask path) — see #431.
+ app_config = getattr(app_state, "config", None)
+ if app_config is None:
+ from openjarvis.core.config import load_config
+
+ app_config = load_config()
+
+ final_system_prompt = _build_managed_system_prompt(
+ system_prompt or "", app_config
+ )
+
+ if final_system_prompt and final_system_prompt.strip():
+ llm_messages.append(
+ Message(role=Role.SYSTEM, content=final_system_prompt.strip())
+ )
# Resolve agent type and class for DeepResearch tool wiring
agent_type = agent_record.get("agent_type", "")
diff --git a/tests/server/test_managed_agent_streaming.py b/tests/server/test_managed_agent_streaming.py
index cf57047b..143c7d3e 100644
--- a/tests/server/test_managed_agent_streaming.py
+++ b/tests/server/test_managed_agent_streaming.py
@@ -15,6 +15,7 @@ pytest.importorskip("fastapi") # agent_manager_routes imports FastAPI at module
from openjarvis.core.types import Role # noqa: E402
from openjarvis.server.agent_manager_routes import ( # noqa: E402
+ _build_managed_system_prompt,
_instantiate_managed_tool,
_replay_history_messages,
_sampler_kwargs,
@@ -26,12 +27,26 @@ class TestReplayHistoryToolCalls:
def test_assistant_tool_calls_are_replayed_with_results(self):
history = [ # DESC order (newest first), as list_messages returns
- {"id": "m2", "direction": "agent_to_user", "content": "", "tool_calls": [
- {"tool": "shell_exec", "arguments": '{"command":"pwd"}',
- "result": "/home/u", "success": True, "latency": 1.0},
- ]},
- {"id": "m1", "direction": "user_to_agent", "content": "run pwd",
- "tool_calls": None},
+ {
+ "id": "m2",
+ "direction": "agent_to_user",
+ "content": "",
+ "tool_calls": [
+ {
+ "tool": "shell_exec",
+ "arguments": '{"command":"pwd"}',
+ "result": "/home/u",
+ "success": True,
+ "latency": 1.0,
+ },
+ ],
+ },
+ {
+ "id": "m1",
+ "direction": "user_to_agent",
+ "content": "run pwd",
+ "tool_calls": None,
+ },
]
msgs = _replay_history_messages(history, exclude_id="current")
# Chronological: user, assistant(tool_calls), tool(result)
@@ -46,8 +61,12 @@ class TestReplayHistoryToolCalls:
def test_plain_assistant_without_tool_calls(self):
history = [
- {"id": "m1", "direction": "agent_to_user", "content": "hi",
- "tool_calls": None},
+ {
+ "id": "m1",
+ "direction": "agent_to_user",
+ "content": "hi",
+ "tool_calls": None,
+ },
]
msgs = _replay_history_messages(history, exclude_id="current")
assert len(msgs) == 1
@@ -56,10 +75,18 @@ class TestReplayHistoryToolCalls:
def test_excludes_current_message(self):
history = [
- {"id": "cur", "direction": "user_to_agent", "content": "now",
- "tool_calls": None},
- {"id": "old", "direction": "user_to_agent", "content": "before",
- "tool_calls": None},
+ {
+ "id": "cur",
+ "direction": "user_to_agent",
+ "content": "now",
+ "tool_calls": None,
+ },
+ {
+ "id": "old",
+ "direction": "user_to_agent",
+ "content": "before",
+ "tool_calls": None,
+ },
]
msgs = _replay_history_messages(history, exclude_id="cur")
assert [m.content for m in msgs] == ["before"]
@@ -69,13 +96,24 @@ class TestSamplerKwargs:
"""#386 — sampler params forwarded only when set."""
def test_reads_present_keys(self):
- cfg = {"temperature": 0.7, "repetition_penalty": 1.1, "top_p": 0.9,
- "top_k": 40, "min_p": 0.05, "frequency_penalty": 0.2,
- "presence_penalty": 0.1}
+ cfg = {
+ "temperature": 0.7,
+ "repetition_penalty": 1.1,
+ "top_p": 0.9,
+ "top_k": 40,
+ "min_p": 0.05,
+ "frequency_penalty": 0.2,
+ "presence_penalty": 0.1,
+ }
out = _sampler_kwargs(cfg)
- assert out == {"repetition_penalty": 1.1, "top_p": 0.9, "top_k": 40,
- "min_p": 0.05, "frequency_penalty": 0.2,
- "presence_penalty": 0.1}
+ assert out == {
+ "repetition_penalty": 1.1,
+ "top_p": 0.9,
+ "top_k": 40,
+ "min_p": 0.05,
+ "frequency_penalty": 0.2,
+ "presence_penalty": 0.1,
+ }
# temperature/max_tokens are handled separately, not here.
assert "temperature" not in out
@@ -98,34 +136,107 @@ class TestInstantiateManagedTool:
def test_memory_tool_gets_backend(self):
backend = object()
- app_state = SimpleNamespace(memory_backend=backend, config=None,
- channel_bridge=None)
- tool = _instantiate_managed_tool(_FakeTool, "memory_store",
- engine=object(), model="m",
- app_state=app_state)
+ app_state = SimpleNamespace(
+ memory_backend=backend, config=None, channel_bridge=None
+ )
+ tool = _instantiate_managed_tool(
+ _FakeTool, "memory_store", engine=object(), model="m", app_state=app_state
+ )
assert tool.kwargs == {"backend": backend}
def test_llm_tool_gets_engine_and_model(self):
engine = object()
- app_state = SimpleNamespace(memory_backend=None, config=None,
- channel_bridge=None)
- tool = _instantiate_managed_tool(_FakeTool, "llm", engine=engine,
- model="qwen", app_state=app_state)
+ app_state = SimpleNamespace(
+ memory_backend=None, config=None, channel_bridge=None
+ )
+ tool = _instantiate_managed_tool(
+ _FakeTool, "llm", engine=engine, model="qwen", app_state=app_state
+ )
assert tool.kwargs == {"engine": engine, "model": "qwen"}
def test_channel_tool_gets_channel(self):
bridge = object()
- app_state = SimpleNamespace(memory_backend=None, config=None,
- channel_bridge=bridge)
- tool = _instantiate_managed_tool(_FakeTool, "channel_send",
- engine=object(), model="m",
- app_state=app_state)
+ app_state = SimpleNamespace(
+ memory_backend=None, config=None, channel_bridge=bridge
+ )
+ tool = _instantiate_managed_tool(
+ _FakeTool, "channel_send", engine=object(), model="m", app_state=app_state
+ )
assert tool.kwargs == {"channel": bridge}
def test_plain_tool_gets_no_injection(self):
- app_state = SimpleNamespace(memory_backend=None, config=None,
- channel_bridge=None)
- tool = _instantiate_managed_tool(_FakeTool, "calculator",
- engine=object(), model="m",
- app_state=app_state)
+ app_state = SimpleNamespace(
+ memory_backend=None, config=None, channel_bridge=None
+ )
+ tool = _instantiate_managed_tool(
+ _FakeTool, "calculator", engine=object(), model="m", app_state=app_state
+ )
assert tool.kwargs == {}
+
+
+class TestBuildManagedSystemPrompt:
+ """#431 — the streaming managed-agent path must run the agent's
+ system_prompt through SystemPromptBuilder so SOUL.md / MEMORY.md /
+ USER.md persona files are injected (parity with the CLI/ask path),
+ instead of using the raw config system_prompt verbatim.
+ """
+
+ def _config_with_soul(self, tmp_path):
+ from openjarvis.core.config import (
+ MemoryFilesConfig,
+ SystemPromptConfig,
+ )
+
+ soul = tmp_path / "SOUL.md"
+ soul.write_text("You are Jarvis, a meticulous local-first assistant.")
+ # Other persona files point at non-existent paths — only SOUL is set.
+ return SimpleNamespace(
+ memory_files=MemoryFilesConfig(soul_path=str(soul)),
+ system_prompt=SystemPromptConfig(),
+ )
+
+ def test_injects_soul_persona(self, tmp_path):
+ app_config = self._config_with_soul(tmp_path)
+ result = _build_managed_system_prompt(
+ system_prompt="You are a helpful assistant.",
+ app_config=app_config,
+ )
+ # The persona file content is injected (the #431 fix)...
+ assert "meticulous local-first assistant" in result
+ # ...alongside the agent's own template.
+ assert "You are a helpful assistant." in result
+
+ def test_agent_template_is_preserved(self, tmp_path):
+ from openjarvis.core.config import MemoryFilesConfig, SystemPromptConfig
+
+ # No persona files configured (all default paths).
+ app_config = SimpleNamespace(
+ memory_files=MemoryFilesConfig(),
+ system_prompt=SystemPromptConfig(),
+ )
+ result = _build_managed_system_prompt(
+ system_prompt="Plain agent.",
+ app_config=app_config,
+ )
+ # The agent's own template is carried into the assembled prompt.
+ assert "Plain agent." in result
+
+ def test_matches_cli_builder_output(self, tmp_path):
+ """Parity check: the helper produces exactly what a directly-
+ constructed SystemPromptBuilder produces (same path the CLI uses),
+ so streaming chat and `jarvis ask` assemble the prompt identically.
+ """
+ from openjarvis.core.config import MemoryFilesConfig, SystemPromptConfig
+ from openjarvis.prompt.builder import SystemPromptBuilder
+
+ app_config = SimpleNamespace(
+ memory_files=MemoryFilesConfig(),
+ system_prompt=SystemPromptConfig(),
+ )
+ helper_out = _build_managed_system_prompt("Agent X.", app_config)
+ direct_out = SystemPromptBuilder(
+ agent_template="Agent X.",
+ memory_files_config=app_config.memory_files,
+ system_prompt_config=app_config.system_prompt,
+ ).build()
+ assert helper_out == direct_out