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(agents): brief spawned agents to do the work, not a read-only audit

rubenluetke10-beepec068fc99206
diff --git a/jarvis/brain/workspace_tool.py b/jarvis/brain/workspace_tool.py
index 66e6683df1..88e3282b71 100644
--- a/jarvis/brain/workspace_tool.py
+++ b/jarvis/brain/workspace_tool.py
@@ -2,6 +2,7 @@
from __future__ import annotations
+from jarvis.core.agent_brief import AGENT_BRIEF_RULE
from jarvis.core.protocols import ExecutionContext, ToolResult, WorkspaceOrchestrationGateway
@@ -38,7 +39,9 @@ class WorkspaceOrchestrationTool:
"Accepted means delivered, not completed; uncertain delivery must not be retried. "
"After a proven pre-write refusal, resolve again for a fresh request_id "
"before a new attempt. "
- "Use context with the same IDs to inspect recorded results. No prompt rewriting is needed."
+ "Use context with the same IDs to inspect recorded results. No prompt rewriting is needed. "
+ "A prompt for create, open_workspace or send is a work order the agent carries "
+ "out, never a read-only request unless the user asked for one (see prompt)."
)
schema = {
"type": "object",
@@ -71,9 +74,15 @@ class WorkspaceOrchestrationTool:
"project_id",
"workspace_id",
"terminal_id",
- "prompt",
)
},
+ "prompt": {
+ "type": "string",
+ "description": (
+ "create/open_workspace/send: the full, self-contained task brief; "
+ "respond: the answer to type. " + AGENT_BRIEF_RULE
+ ),
+ },
# No pattern: a schema rejection would discard the whole call over
# one mistyped character. The orchestrator repairs the id instead.
"request_id": {
diff --git a/jarvis/core/agent_brief.py b/jarvis/core/agent_brief.py
new file mode 100644
index 0000000000..5c1fb5516e
--- /dev/null
+++ b/jarvis/core/agent_brief.py
@@ -0,0 +1,29 @@
+"""How Jarvis words a task it hands to another agent.
+
+Every hand-off surface (workspace coding panes, the agent society, the live
+voice backends) writes its own brief, and none of them said what that brief
+should ask for. Left alone, the model read "check", "look into", "deep dive"
+or "audit" as a request for a report and wrote "READ-ONLY" or "no changes"
+into the brief on its own, so the agent found the problem and stopped (live
+2026-10-02: "have it look at what is wrong with our plugins" became
+"Autonomously perform a careful, evidence-based, READ-ONLY audit").
+One rule, shared by every surface, so they cannot drift apart.
+"""
+
+from __future__ import annotations
+
+from typing import Final
+
+#: Appended to every tool description / instruction that makes Jarvis write a
+#: brief for another agent. Kept short: it rides in the voice tool catalog,
+#: which has a byte budget (``jarvis.live.tools._CATALOG_BYTE_BUDGET``).
+AGENT_BRIEF_RULE: Final[str] = (
+ "Write the brief as a work order the agent completes on its own: find the cause, "
+ "make the change, verify it, then report what changed. 'Check', 'look into', "
+ "'deep dive' or 'audit' include fixing what is found. Never add read-only, "
+ "no-changes or report-only limits yourself; use them only when the user "
+ "explicitly asked just for analysis, a plan or an answer. Let the agent decide "
+ "steps without asking back, except before destructive or outward actions."
+)
+
+__all__ = ["AGENT_BRIEF_RULE"]
diff --git a/jarvis/live/config.py b/jarvis/live/config.py
index 835ce871a6..dea3db6fb4 100644
--- a/jarvis/live/config.py
+++ b/jarvis/live/config.py
@@ -4,6 +4,7 @@ from __future__ import annotations
from pydantic import BaseModel, ConfigDict, Field
+from jarvis.core.agent_brief import AGENT_BRIEF_RULE
from jarvis.live.product import PRODUCT_BRIEF
@@ -51,6 +52,8 @@ class LiveConfig(BaseModel):
"A request for a NEW coding agent (or several: 'two Claude Code agents') is "
"workspace-orchestrate create in the named or visible workspace, with cli, "
"count and the task as prompt; never an existing agent and never spawn_worker. "
+ + AGENT_BRIEF_RULE
+ + " "
"Computer-use tasks use the selected thinking model and the same credential. "
+ self.backend_instructions
),
diff --git a/jarvis/live/native.py b/jarvis/live/native.py
index b04cffcf4b..a058e6ad73 100644
--- a/jarvis/live/native.py
+++ b/jarvis/live/native.py
@@ -12,6 +12,7 @@ from types import SimpleNamespace
from typing import Any, Literal
from uuid import uuid4
+from jarvis.core.agent_brief import AGENT_BRIEF_RULE
from jarvis.core.paths import user_data_dir
from jarvis.core.runtime_refs import get_supervisor_tool_gateway
from jarvis.core.tool_budget import VOICE_TOOL_BUDGET_S
@@ -153,7 +154,8 @@ class NativeLiveVoiceSession(LiveVoiceSession):
"prompt) in the named or visible workspace; never spawn_worker. "
"Explicit references override the visible workspace; ask on ambiguity. "
"Do not switch the UI to address another workspace. Reuse request_id on "
- "retries and never replay uncertain delivery."
+ "retries and never replay uncertain delivery. "
+ + AGENT_BRIEF_RULE
),
history=tuple(
{"role": item["role"], "text": item["delta"]} for item in self._initial_seed
diff --git a/jarvis/plugins/tool/delegate_to_agent.py b/jarvis/plugins/tool/delegate_to_agent.py
index fe9606f810..9fe5d381c6 100644
--- a/jarvis/plugins/tool/delegate_to_agent.py
+++ b/jarvis/plugins/tool/delegate_to_agent.py
@@ -23,6 +23,7 @@ import logging
from collections.abc import Callable
from typing import Any, Final
+from jarvis.core.agent_brief import AGENT_BRIEF_RULE
from jarvis.core.protocols import ExecutionContext, ToolResult
from jarvis.society.communication import REPLY_POLICY_SCHEMA, select_reply_policy
@@ -156,7 +157,7 @@ class DelegateToAgentTool:
"type": "string",
"description": (
"An actionable objective and scope faithful to the user's intent, "
- "not just a paraphrase."
+ "not just a paraphrase. " + AGENT_BRIEF_RULE
),
},
"reply_policy": REPLY_POLICY_SCHEMA,
diff --git a/tests/unit/core/test_agent_brief.py b/tests/unit/core/test_agent_brief.py
new file mode 100644
index 0000000000..71e4ba7b85
--- /dev/null
+++ b/tests/unit/core/test_agent_brief.py
@@ -0,0 +1,47 @@
+"""Every surface that briefs another agent carries the same work-order rule.
+
+Without it the model turned "check", "look into" or "deep dive" into a
+"READ-ONLY audit" on its own and the agent stopped at a report
+(live 2026-10-02). See ``jarvis/core/agent_brief.py``.
+"""
+
+from __future__ import annotations
+
+import inspect
+
+from jarvis.brain.workspace_tool import WorkspaceOrchestrationTool
+from jarvis.core.agent_brief import AGENT_BRIEF_RULE
+from jarvis.live import native
+from jarvis.live.config import LiveConfig
+from jarvis.plugins.tool.delegate_to_agent import DelegateToAgentTool
+
+
+def test_rule_asks_for_the_work_and_reserves_read_only_for_an_explicit_ask():
+ rule = AGENT_BRIEF_RULE.lower()
+ assert "make the change" in rule
+ assert "never add read-only" in rule
+ assert "explicitly asked" in rule
+ assert "deep dive" in rule
+
+
+def test_workspace_prompt_field_carries_the_rule():
+ prompt = WorkspaceOrchestrationTool.schema["properties"]["prompt"]
+ assert prompt["type"] == "string"
+ assert AGENT_BRIEF_RULE in prompt["description"]
+ assert "read-only" in WorkspaceOrchestrationTool.description
+
+
+def test_society_task_field_carries_the_rule():
+ task = DelegateToAgentTool.schema["properties"]["task"]["description"]
+ assert AGENT_BRIEF_RULE in task
+
+
+def test_live_delegation_backend_is_instructed_with_the_rule():
+ wire = LiveConfig(configured=True, backend_model="chosen-model").session_config(
+ language="en", tools=[]
+ )
+ assert AGENT_BRIEF_RULE in wire["delegation"]["responses"]["instructions"]
+
+
+def test_native_live_session_instructs_the_rule():
+ assert "AGENT_BRIEF_RULE" in inspect.getsource(native.NativeLiveVoiceSession)