Merge pull request #679 from gilbert-barajas/fix/serve-persona-prompt-builder

fix: served/SDK agents silently lose their persona (SOUL.md never reaches the model over HTTP)

.github/clone-stats/badge.json

@@ -1,7 +1,7 @@
 {
   "schemaVersion": 1,
   "label": "Git Clones",
-  "message": "189,482",
+  "message": "190,252",
   "color": "green",
   "namedLogo": "git"
 }

.github/clone-stats/clone-data.json

@@ -1,6 +1,6 @@
 {
-  "total_clones": 189482,
-  "last_updated": "2026-08-13T07:22:13Z",
+  "total_clones": 190252,
+  "last_updated": "2026-08-14T07:19:51Z",
   "daily": {
     "2026-03-27": 2189,
     "2026-03-28": 1874,
@@ -140,6 +140,7 @@
     "2026-08-09": 1076,
     "2026-08-10": 1060,
     "2026-08-11": 2182,
-    "2026-08-12": 641
+    "2026-08-12": 641,
+    "2026-08-13": 770
   }
 }

frontend/src-tauri/Info.plist

@@ -0,0 +1,14 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+  <key>NSAppTransportSecurity</key>
+  <dict>
+    <!-- The desktop API URL is user-configurable, so it cannot be represented
+         by Tauri's single, build-time exceptionDomain setting. Keep this
+         exception scoped to WKWebView; native URLSession traffic retains ATS. -->
+    <key>NSAllowsArbitraryLoadsInWebContent</key>
+    <true/>
+  </dict>
+</dict>
+</plist>

frontend/src-tauri/tauri.conf.json

@@ -24,7 +24,7 @@
       }
     ],
     "security": {
-      "csp": "default-src 'self'; script-src 'self' 'unsafe-inline'; style-src 'self' 'unsafe-inline'; connect-src 'self' http://localhost:* http://127.0.0.1:* ws://localhost:* ws://127.0.0.1:*; img-src 'self' data: blob:"
+      "csp": "default-src 'self'; script-src 'self' 'unsafe-inline'; style-src 'self' 'unsafe-inline'; connect-src 'self' http: https: ws: wss:; img-src 'self' data: blob:"
     }
   },
   "bundle": {
@@ -45,7 +45,6 @@
     "macOS": {
       "entitlements": "Entitlements.plist",
       "minimumSystemVersion": "10.15",
-      "exceptionDomain": "",
       "frameworks": [],
       "providerShortName": null,
       "signingIdentity": "-"

frontend/src/components/Chat/MessageBubble.tsx

@@ -103,6 +103,24 @@ function CopyMessageButton({ content }: { content: string }) {
 export function MessageBubble({ message, isLive = false }: Props) {
   const isUser = message.role === 'user';
 
+  const cleanContent = useMemo(() => stripThinkTags(message.content), [message.content]);
+
+  // Build a ref→source lookup once per render. Memoized so the rehype plugin
+  // identity stays stable until the source list actually changes.
+  const sourcesMap = useMemo(() => {
+    const m = new Map<number, NonNullable<ChatMessage['researchSources']>[number]>();
+    for (const s of message.researchSources ?? []) {
+      if (typeof s.ref === 'number') m.set(s.ref, s);
+    }
+    return m;
+  }, [message.researchSources]);
+
+  const rehypePlugins = useMemo(() => {
+    const base: any[] = [[rehypeHighlight, { detect: true }], rehypeKatex];
+    if (sourcesMap.size > 0) base.push([rehypeCitations, { sources: sourcesMap }]);
+    return base;
+  }, [sourcesMap]);
+
   if (isUser) {
     return (
       <div className="flex justify-end mb-4">
@@ -122,24 +140,6 @@ export function MessageBubble({ message, isLive = false }: Props) {
     );
   }
 
-  const cleanContent = useMemo(() => stripThinkTags(message.content), [message.content]);
-
-  // Build a ref→source lookup once per render. Memoized so the rehype plugin
-  // identity stays stable until the source list actually changes.
-  const sourcesMap = useMemo(() => {
-    const m = new Map<number, NonNullable<ChatMessage['researchSources']>[number]>();
-    for (const s of message.researchSources ?? []) {
-      if (typeof s.ref === 'number') m.set(s.ref, s);
-    }
-    return m;
-  }, [message.researchSources]);
-
-  const rehypePlugins = useMemo(() => {
-    const base: any[] = [[rehypeHighlight, { detect: true }], rehypeKatex];
-    if (sourcesMap.size > 0) base.push([rehypeCitations, { sources: sourcesMap }]);
-    return base;
-  }, [sourcesMap]);
-
   return (
     <div className="group mb-6">
       {/* Deep Research timeline (steps + status) */}

frontend/src/lib/useAgentEvents.test.ts

@@ -0,0 +1,66 @@
+import { afterEach, beforeEach, describe, expect, it } from 'vitest';
+
+import { buildWsUrl } from './useAgentEvents';
+
+const SETTINGS_KEY = 'openjarvis-settings';
+
+class MemoryStorage {
+  private store = new Map<string, string>();
+
+  getItem(key: string): string | null {
+    return this.store.get(key) ?? null;
+  }
+
+  setItem(key: string, value: string): void {
+    this.store.set(key, String(value));
+  }
+}
+
+beforeEach(() => {
+  (globalThis as unknown as { localStorage: MemoryStorage }).localStorage =
+    new MemoryStorage();
+});
+
+afterEach(() => {
+  (globalThis as unknown as { localStorage?: MemoryStorage }).localStorage =
+    undefined;
+});
+
+describe('buildWsUrl', () => {
+  it('authenticates agent events with the configured API key', () => {
+    localStorage.setItem(
+      SETTINGS_KEY,
+      JSON.stringify({
+        apiUrl: 'https://jarvis.example.com:8443',
+        apiKey: 'secret+/=',
+      }),
+    );
+
+    const url = new URL(buildWsUrl('agent/one'));
+
+    expect(url.origin).toBe('wss://jarvis.example.com:8443');
+    expect(url.pathname).toBe('/v1/agents/events');
+    expect(url.searchParams.get('agent_id')).toBe('agent/one');
+    expect(url.searchParams.get('token')).toBe('secret+/=');
+  });
+
+  it('normalizes a versioned API base without duplicating /v1', () => {
+    localStorage.setItem(
+      SETTINGS_KEY,
+      JSON.stringify({ apiUrl: 'http://192.0.2.10:8000/v1/' }),
+    );
+
+    expect(buildWsUrl()).toBe('ws://192.0.2.10:8000/v1/agents/events');
+  });
+
+  it('omits the token for a keyless server', () => {
+    localStorage.setItem(
+      SETTINGS_KEY,
+      JSON.stringify({ apiUrl: 'http://localhost:8000' }),
+    );
+
+    const url = new URL(buildWsUrl('agent-one'));
+
+    expect(url.searchParams.has('token')).toBe(false);
+  });
+});

frontend/src/lib/useAgentEvents.ts

@@ -1,5 +1,5 @@
 import { useEffect, useRef } from 'react';
-import { getBase } from './api';
+import { getApiKey, getBase } from './api';
 
 export interface AgentEvent {
   type: string;
@@ -7,19 +7,16 @@ export interface AgentEvent {
   data: Record<string, unknown>;
 }
 
-function buildWsUrl(agentId?: string): string {
+export function buildWsUrl(agentId?: string): string {
   const base = getBase();
-  let origin: string;
-  if (base) {
-    origin = base.replace(/^http/, 'ws');
-  } else {
-    const loc = window.location;
-    origin = `${loc.protocol === 'https:' ? 'wss:' : 'ws:'}//${loc.host}`;
-  }
-  const path = '/v1/agents/events';
-  return agentId
-    ? `${origin}${path}?agent_id=${encodeURIComponent(agentId)}`
-    : `${origin}${path}`;
+  const url = new URL('/v1/agents/events', base || window.location.origin);
+  url.protocol = url.protocol === 'https:' ? 'wss:' : 'ws:';
+
+  if (agentId) url.searchParams.set('agent_id', agentId);
+  const apiKey = getApiKey();
+  if (apiKey) url.searchParams.set('token', apiKey);
+
+  return url.toString();
 }
 
 /**

src/openjarvis/agents/monitor_operative.py

@@ -127,6 +127,7 @@ class MonitorOperativeAgent(ToolUsingAgent):
         memory_backend: Optional[Any] = None,
         interactive: bool = False,
         confirm_callback=None,
+        prompt_builder: Optional[Any] = None,
         **kwargs: Any,
     ) -> None:
         super().__init__(
@@ -139,7 +140,7 @@ class MonitorOperativeAgent(ToolUsingAgent):
             max_tokens=max_tokens,
             interactive=interactive,
             confirm_callback=confirm_callback,
-            prompt_builder=kwargs.get("prompt_builder"),
+            prompt_builder=prompt_builder,
         )
         # Validate strategies
         if memory_extraction not in VALID_MEMORY_EXTRACTION:

src/openjarvis/agents/operative.py

@@ -58,6 +58,7 @@ class OperativeAgent(ToolUsingAgent):
         memory_backend: Optional[Any] = None,
         interactive: bool = False,
         confirm_callback=None,
+        prompt_builder: Optional[Any] = None,
         **kwargs: Any,
     ) -> None:
         super().__init__(
@@ -70,7 +71,7 @@ class OperativeAgent(ToolUsingAgent):
             max_tokens=max_tokens,
             interactive=interactive,
             confirm_callback=confirm_callback,
-            prompt_builder=kwargs.get("prompt_builder"),
+            prompt_builder=prompt_builder,
         )
         self._system_prompt = system_prompt or ""
         self._operator_id = operator_id

src/openjarvis/agents/orchestrator.py

@@ -13,6 +13,7 @@ Supports two modes:
 from __future__ import annotations
 
 import concurrent.futures
+import json
 import re
 from typing import Any, List, Optional
 
@@ -142,12 +143,21 @@ class OrchestratorAgent(ToolUsingAgent):
                 tool_call = ToolCall(
                     id=f"orch_{turns}",
                     name=parsed["tool"],
-                    arguments=parsed["input"] or "{}",
+                    arguments=self._normalize_structured_tool_input(
+                        parsed["tool"],
+                        parsed["input"],
+                    ),
                 )
                 tool_result = self._executor.execute(tool_call)
                 all_tool_results.append(tool_result)
 
-                observation = f"Observation: {tool_result.content}"
+                if tool_result.success:
+                    observation = f"Observation: {tool_result.content}"
+                else:
+                    observation = (
+                        f"Observation: Tool '{tool_result.tool_name}' failed: "
+                        f"{tool_result.content}"
+                    )
                 messages.append(Message(role=Role.USER, content=observation))
                 continue
 
@@ -162,6 +172,75 @@ class OrchestratorAgent(ToolUsingAgent):
         # Max turns exceeded
         return self._max_turns_result(all_tool_results, turns)
 
+    def _normalize_structured_tool_input(
+        self,
+        tool_name: str,
+        raw_input: str,
+    ) -> str:
+        """Map unambiguous structured text input to a string parameter."""
+        if not raw_input:
+            return "{}"
+
+        try:
+            parsed_input = json.loads(raw_input)
+        except json.JSONDecodeError:
+            invalid_json = True
+            string_value = raw_input
+        else:
+            invalid_json = False
+            if isinstance(parsed_input, dict):
+                return raw_input
+            # INPUT is a text protocol. A non-object JSON value such as 42,
+            # true, null, or [1, 2] may still be the intended text for a tool's
+            # string parameter. Quoted JSON strings are decoded to remove only
+            # their surrounding quotes; other values retain their source text.
+            string_value = parsed_input if isinstance(parsed_input, str) else raw_input
+
+        tool_spec = None
+        for candidate in reversed(self._tools):
+            candidate_spec = candidate.spec
+            if candidate_spec.name == tool_name:
+                tool_spec = candidate_spec
+                break
+        if tool_spec is None:
+            return raw_input
+
+        parameters = tool_spec.parameters
+        parameter_container_type = parameters.get("type")
+        if parameter_container_type not in (None, "object"):
+            return raw_input
+
+        properties = parameters.get("properties", {})
+        required = parameters.get("required", [])
+        if not isinstance(properties, dict) or not isinstance(required, list):
+            return raw_input
+
+        if len(required) == 1 and required[0] in properties:
+            parameter_name = required[0]
+        elif not required and len(properties) == 1:
+            parameter_name = next(iter(properties))
+        else:
+            return raw_input
+
+        parameter_schema = properties[parameter_name]
+        if not isinstance(parameter_schema, dict):
+            return raw_input
+        parameter_type = parameter_schema.get("type")
+        accepts_string = parameter_type == "string" or (
+            isinstance(parameter_type, list) and "string" in parameter_type
+        )
+        if not accepts_string:
+            return raw_input
+
+        allow_object_text = (
+            tool_spec.metadata.get("structured_allow_object_text") is True
+        )
+        starts_like_object = raw_input.lstrip("\ufeff \t\r\n").startswith("{")
+        if invalid_json and starts_like_object and not allow_object_text:
+            return raw_input
+
+        return json.dumps({parameter_name: string_value})
+
     @staticmethod
     def _parse_structured_response(text: str) -> dict:
         """Parse THOUGHT/TOOL/INPUT/FINAL_ANSWER from model output."""

src/openjarvis/cli/ask.py

@@ -398,9 +398,8 @@ def _run_agent(
 
     # Wire the SystemPromptBuilder so SOUL.md / MEMORY.md / USER.md persona
     # files actually reach the model. Only passed to agents whose __init__
-    # accepts a `prompt_builder` kwarg (BaseAgent does; agents that override
-    # __init__ without forwarding it, e.g. OrchestratorAgent, opt out
-    # automatically and keep their existing system-prompt machinery).
+    # explicitly accepts a `prompt_builder` kwarg. Agents with specialized
+    # prompt machinery opt in by naming and forwarding the parameter.
     import inspect as _inspect
 
     if "prompt_builder" in _inspect.signature(agent_cls.__init__).parameters:

src/openjarvis/cli/serve.py

@@ -367,6 +367,27 @@ def serve(
                 if getattr(agent_cls, "accepts_tools", False):
                     agent_kwargs["max_turns"] = config.agent.max_turns
 
+                # Wire the SystemPromptBuilder so SOUL.md / MEMORY.md / USER.md
+                # reach the model on the SERVE path too. ``ask.py`` has done
+                # this since the persona system landed; ``serve.py`` never did,
+                # so an agent served over HTTP silently answered as a generic
+                # assistant while the same agent via the CLI kept its persona.
+                # Guarded so agents with specialized prompt machinery must opt
+                # in by explicitly naming and forwarding the kwarg.
+                import inspect as _inspect
+
+                if (
+                    "prompt_builder"
+                    in _inspect.signature(agent_cls.__init__).parameters
+                ):
+                    from openjarvis.prompt.builder import SystemPromptBuilder
+
+                    agent_kwargs["prompt_builder"] = SystemPromptBuilder(
+                        agent_template=config.agent.default_system_prompt or "",
+                        memory_files_config=config.memory_files,
+                        system_prompt_config=config.system_prompt,
+                    )
+
                 agent = agent_cls(engine, model_name, **agent_kwargs)
                 # Pin MCP transports to the agent's lifetime so HTTP
                 # connections don't close mid-request (#461).

src/openjarvis/sdk.py

@@ -516,6 +516,20 @@ class Jarvis:
             existing = agent_kwargs.get("tools", [])
             agent_kwargs["tools"] = digest_tools + list(existing)
 
+        # Wire the SystemPromptBuilder so SOUL.md / MEMORY.md / USER.md reach
+        # the model — mirrors ``cli/ask.py`` and ``cli/serve.py``. Guarded so
+        # agents whose ``__init__`` doesn't accept the kwarg opt out.
+        import inspect as _inspect
+
+        if "prompt_builder" in _inspect.signature(agent_cls.__init__).parameters:
+            from openjarvis.prompt.builder import SystemPromptBuilder
+
+            agent_kwargs["prompt_builder"] = SystemPromptBuilder(
+                agent_template=self._config.agent.default_system_prompt or "",
+                memory_files_config=self._config.memory_files,
+                system_prompt_config=self._config.system_prompt,
+            )
+
         agent_obj = agent_cls(self._engine, model_name, **agent_kwargs)
         ctx = AgentContext()
 

src/openjarvis/tools/_stubs.py

@@ -136,6 +136,15 @@ class ToolExecutor:
                 content=f"Invalid arguments JSON: {exc}",
                 success=False,
             )
+        if not isinstance(params, dict):
+            return ToolResult(
+                tool_name=tool_call.name,
+                content=(
+                    "Invalid arguments: expected a JSON object, "
+                    f"got {type(params).__name__}."
+                ),
+                success=False,
+            )
 
         # Boundary guard: scan external tool arguments
         if self._boundary_guard is not None and not getattr(tool, "is_local", True):
@@ -143,6 +152,15 @@ class ToolExecutor:
                 tool_call = self._boundary_guard.check_outbound(tool_call)
                 # Re-parse arguments after potential redaction
                 params = json.loads(tool_call.arguments) if tool_call.arguments else {}
+                if not isinstance(params, dict):
+                    return ToolResult(
+                        tool_name=tool_call.name,
+                        content=(
+                            "Invalid arguments: expected a JSON object, "
+                            f"got {type(params).__name__}."
+                        ),
+                        success=False,
+                    )
             except Exception as exc:
                 return ToolResult(
                     tool_name=tool_call.name,

src/openjarvis/tools/code_interpreter.py

@@ -56,6 +56,7 @@ class CodeInterpreterTool(BaseTool):
                 "required": ["code"],
             },
             category="code",
+            metadata={"structured_allow_object_text": True},
         )
 
     def execute(self, **params: Any) -> ToolResult:

src/openjarvis/tools/code_interpreter_docker.py

@@ -55,6 +55,7 @@ class DockerCodeInterpreterTool(BaseTool):
             },
             category="code",
             timeout_seconds=60.0,
+            metadata={"structured_allow_object_text": True},
         )
 
     def execute(self, **params: Any) -> ToolResult:

src/openjarvis/tools/repl.py

@@ -191,6 +191,7 @@ class ReplTool(BaseTool):
                 "required": ["code"],
             },
             category="code",
+            metadata={"structured_allow_object_text": True},
         )
 
     def execute(self, **params: Any) -> ToolResult:

tests/cli/test_ask_agent.py

@@ -396,11 +396,10 @@ class TestPersonaFilesReachModel:
         assert "MEMORY_SENTINEL" in joined
         assert "USER_SENTINEL" in joined
 
-    def test_orchestrator_keeps_its_own_system_prompt(
+    def test_orchestrator_accepts_persona_prompt_builder(
         self, runner, monkeypatch, tmp_path
     ):
-        """OrchestratorAgent's __init__ doesn't accept ``prompt_builder``;
-        the wiring must skip it silently rather than crash."""
+        """Orchestrator explicitly accepts and applies persona wiring."""
         from openjarvis.core.config import JarvisConfig
 
         soul = tmp_path / "SOUL.md"
@@ -425,5 +424,6 @@ class TestPersonaFilesReachModel:
         ):
             result = runner.invoke(cli, ["ask", "--agent", "orchestrator", "Hello"])
 
-        # Pass condition: doesn't crash with TypeError on prompt_builder kwarg.
         assert result.exit_code == 0, result.output
+        messages = engine.generate.call_args.args[0]
+        assert "ORCH_PERSONA_SENTINEL" in messages[0].content

tests/cli/test_serve_persona.py

@@ -0,0 +1,129 @@
+"""Regression: ``jarvis serve`` must wire the ``SystemPromptBuilder`` into the
+agent it constructs, so SOUL.md / MEMORY.md / USER.md reach the model over HTTP.
+
+``cli/ask.py`` (and ``cli/chat_cmd.py`` and the managed-agent executor) have
+wired the builder since the persona system landed. The serve path never did, so
+an agent served over HTTP silently answered as a generic assistant — explicitly
+denying the persona — while the same agent via the CLI kept it. Found deploying
+a personal assistant: SOUL.md was correct on disk the whole time; no error, no
+warning.
+
+This test boots ``serve`` just far enough to capture the agent handed to
+``create_app`` and asserts the builder (and thus the persona content) is
+present. It fails on the unpatched serve path: ``agent._prompt_builder`` is
+``None``, so the persona files never reach the model.
+"""
+
+from __future__ import annotations
+
+import importlib
+from unittest.mock import MagicMock, patch
+
+import pytest
+from click.testing import CliRunner
+
+from openjarvis.cli import cli
+
+pytest.importorskip("fastapi")
+pytest.importorskip("uvicorn")
+
+# ``openjarvis.cli.serve`` as a package attribute resolves to the click
+# *command* (re-exported); grab the real module to monkeypatch its globals.
+serve_mod = importlib.import_module("openjarvis.cli.serve")
+
+
+def _fake_engine() -> MagicMock:
+    engine = MagicMock()
+    engine.list_models.return_value = ["test-model"]
+    engine.health.return_value = True
+    engine.name = "mock"
+    engine.engine_id = "mock"
+    return engine
+
+
+@pytest.mark.parametrize(
+    "agent_name",
+    ["simple", "orchestrator", "monitor_operative", "operative"],
+)
+def test_serve_wires_persona_builder_into_served_agent(
+    tmp_path, monkeypatch, agent_name
+):
+    """The agent built on the serve path must carry a SystemPromptBuilder whose
+    assembled prompt includes SOUL.md content (regression for the HTTP persona
+    loss)."""
+    from openjarvis.agents.monitor_operative import MonitorOperativeAgent
+    from openjarvis.agents.operative import OperativeAgent
+    from openjarvis.agents.orchestrator import OrchestratorAgent
+    from openjarvis.agents.simple import SimpleAgent
+    from openjarvis.core.config import JarvisConfig
+    from openjarvis.core.registry import AgentRegistry
+
+    # Persona file with a unique sentinel we can grep for in the built prompt.
+    soul = tmp_path / "SOUL.md"
+    soul.write_text("SERVE_PERSONA_SENTINEL", encoding="utf-8")
+
+    # conftest clears registries per-test; re-register the agent we exercise.
+    agent_classes = {
+        "simple": SimpleAgent,
+        "orchestrator": OrchestratorAgent,
+        "monitor_operative": MonitorOperativeAgent,
+        "operative": OperativeAgent,
+    }
+    if not AgentRegistry.contains(agent_name):
+        AgentRegistry.register_value(agent_name, agent_classes[agent_name])
+
+    config = JarvisConfig()
+    config.server.host = "127.0.0.1"
+    config.server.port = 8123
+    config.intelligence.default_model = "test-model"
+    config.memory_files.soul_path = str(soul)
+    # Keep the heavy optional subsystems off so we reach create_app cleanly.
+    config.telemetry.enabled = False
+    config.agent_manager.enabled = False
+    config.sessions.enabled = False
+    config.channel.enabled = False
+    config.skills.enabled = False
+    config.agent.context_from_memory = False
+
+    engine = _fake_engine()
+    monkeypatch.setattr(serve_mod, "load_config", lambda *a, **k: config)
+    monkeypatch.setattr(serve_mod, "get_engine", lambda *a, **k: ("mock", engine))
+    monkeypatch.setattr(serve_mod, "discover_engines", lambda *a, **k: {})
+    monkeypatch.setattr(serve_mod, "discover_models", lambda *a, **k: {})
+
+    # setup_security returns its own context; pass the engine straight through.
+    sec = MagicMock()
+    sec.engine = engine
+    sec.capability_policy = None
+    sec.audit_logger = None
+    monkeypatch.setattr("openjarvis.security.setup_security", lambda *a, **k: sec)
+
+    captured: dict = {}
+
+    def _capture_create_app(*args, **kwargs):
+        captured["agent"] = kwargs.get("agent")
+        return MagicMock(name="app")
+
+    with (
+        patch("openjarvis.server.app.create_app", side_effect=_capture_create_app),
+        patch("uvicorn.run", lambda *a, **k: None),
+    ):
+        result = CliRunner().invoke(
+            cli, ["serve", "--agent", agent_name], catch_exceptions=False
+        )
+
+    assert result.exit_code == 0, result.output
+    agent = captured.get("agent")
+    assert agent is not None, (
+        "serve did not construct an agent or never reached create_app; "
+        f"output:\n{result.output}"
+    )
+    # The regression: without the fix ``agent._prompt_builder`` is None and the
+    # persona files never reach the model over HTTP.
+    assert agent._prompt_builder is not None, (
+        f"serve constructed {agent_name} without a prompt_builder — SOUL.md / "
+        "MEMORY.md / USER.md would be silently dropped on the HTTP path."
+    )
+    assert "SERVE_PERSONA_SENTINEL" in agent._prompt_builder.build(), (
+        "prompt_builder is wired on serve, but its built prompt omits SOUL.md"
+    )

tests/deployment/test_desktop_network_policy.py

@@ -0,0 +1,34 @@
+"""Regression guards for the desktop app's outbound network policy."""
+
+from __future__ import annotations
+
+import json
+import plistlib
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent.parent.parent
+TAURI_CONFIG = ROOT / "frontend" / "src-tauri" / "tauri.conf.json"
+MACOS_INFO_PLIST = ROOT / "frontend" / "src-tauri" / "Info.plist"
+
+
+def _csp_sources(directive: str) -> set[str]:
+    config = json.loads(TAURI_CONFIG.read_text(encoding="utf-8"))
+    csp = config["app"]["security"]["csp"]
+    directives = {
+        parts[0]: set(parts[1:]) for item in csp.split(";") if (parts := item.split())
+    }
+    return directives[directive]
+
+
+def test_desktop_csp_allows_remote_api_servers() -> None:
+    """The user-configured API URL may point beyond localhost (#649)."""
+    connect_sources = _csp_sources("connect-src")
+
+    assert {"http:", "https:", "ws:", "wss:"} <= connect_sources
+
+
+def test_macos_webview_allows_user_configured_http_servers() -> None:
+    """CSP alone cannot override App Transport Security for public hosts."""
+    info = plistlib.loads(MACOS_INFO_PLIST.read_bytes())
+
+    assert info["NSAppTransportSecurity"]["NSAllowsArbitraryLoadsInWebContent"] is True

tests/sdk/test_sdk.py

@@ -95,6 +95,30 @@ class TestJarvisAsk:
             assert result == "Agent response"
             j.close()
 
+    def test_ask_with_agent_wires_persona(self, tmp_path):
+        from openjarvis.agents.simple import SimpleAgent
+        from openjarvis.core.registry import AgentRegistry
+
+        soul = tmp_path / "SOUL.md"
+        soul.write_text("SDK_PERSONA_SENTINEL", encoding="utf-8")
+
+        cfg = JarvisConfig()
+        cfg.memory_files.soul_path = str(soul)
+        cfg.memory_files.memory_path = ""
+        cfg.memory_files.user_path = ""
+        cfg.agent.context_from_memory = False
+
+        if not AgentRegistry.contains("simple"):
+            AgentRegistry.register_value("simple", SimpleAgent)
+
+        engine = _make_engine()
+        with patch("openjarvis.sdk.get_engine", return_value=("mock", engine)):
+            j = Jarvis(config=cfg, model="test-model")
+            j.ask("Hello", agent="simple")
+            messages = engine.generate.call_args.args[0]
+            assert "SDK_PERSONA_SENTINEL" in messages[0].content
+            j.close()
+
     def test_ask_no_engine_raises(self):
         with patch("openjarvis.sdk.get_engine", return_value=None):
             j = Jarvis(config=JarvisConfig())

tests/test_orchestrator_learning/test_orchestrator_structured.py

@@ -2,6 +2,8 @@
 
 from __future__ import annotations
 
+import pytest
+
 from openjarvis.agents.orchestrator import OrchestratorAgent
 from openjarvis.core.types import ToolResult
 from openjarvis.engine._stubs import InferenceEngine
@@ -18,8 +20,10 @@ class _MockEngine(InferenceEngine):
     def __init__(self, responses: list[str]) -> None:
         self._responses = list(responses)
         self._call_idx = 0
+        self.calls = []
 
     def generate(self, messages, **kwargs) -> dict:
+        self.calls.append(list(messages))
         if self._call_idx < len(self._responses):
             content = self._responses[self._call_idx]
             self._call_idx += 1
@@ -58,6 +62,109 @@ class _MockTool(BaseTool):
         return ToolResult(tool_name="calculator", content=str(expr), success=True)
 
 
+class _FileReadLikeTool(BaseTool):
+    """Small test double with one required and one optional parameter."""
+
+    tool_id = "file_read"
+
+    @property
+    def spec(self) -> ToolSpec:
+        return ToolSpec(
+            name="file_read",
+            description="Read a file",
+            parameters={
+                "type": "object",
+                "properties": {
+                    "path": {"type": "string"},
+                    "max_lines": {"type": "integer"},
+                },
+                "required": ["path"],
+            },
+        )
+
+    def execute(self, **params) -> ToolResult:
+        return ToolResult(
+            tool_name="file_read",
+            content=params.get("path", ""),
+            success=True,
+        )
+
+
+class _AmbiguousTool(BaseTool):
+    """Test double whose bare input cannot map to one parameter safely."""
+
+    tool_id = "copy"
+
+    @property
+    def spec(self) -> ToolSpec:
+        return ToolSpec(
+            name="copy",
+            description="Copy a value",
+            parameters={
+                "type": "object",
+                "properties": {
+                    "source": {"type": "string"},
+                    "destination": {"type": "string"},
+                },
+                "required": ["source", "destination"],
+            },
+        )
+
+    def execute(self, **params) -> ToolResult:
+        return ToolResult(tool_name="copy", content="copied", success=True)
+
+
+class _CodeLikeTool(BaseTool):
+    """Test double that explicitly accepts object-prefixed source text."""
+
+    tool_id = "code"
+
+    @property
+    def spec(self) -> ToolSpec:
+        return ToolSpec(
+            name="code",
+            description="Execute source code",
+            parameters={
+                "type": "object",
+                "properties": {"code": {"type": "string"}},
+                "required": ["code"],
+            },
+            metadata={"structured_allow_object_text": True},
+        )
+
+    def execute(self, **params) -> ToolResult:
+        return ToolResult(
+            tool_name="code",
+            content=params.get("code", ""),
+            success=True,
+        )
+
+
+class _UnionStringTool(BaseTool):
+    """Test double with a JSON Schema union that accepts strings."""
+
+    tool_id = "union_file_read"
+
+    @property
+    def spec(self) -> ToolSpec:
+        return ToolSpec(
+            name="union_file_read",
+            description="Read a nullable path",
+            parameters={
+                "type": "object",
+                "properties": {"path": {"type": ["string", "null"]}},
+                "required": ["path"],
+            },
+        )
+
+    def execute(self, **params) -> ToolResult:
+        return ToolResult(
+            tool_name="union_file_read",
+            content=params.get("path", ""),
+            success=True,
+        )
+
+
 # -- Tests -------------------------------------------------------------------
 
 
@@ -80,6 +187,207 @@ class TestStructuredMode:
         assert result.content == "4"
         assert result.turns == 2
         assert len(result.tool_results) == 1
+        assert result.tool_results[0].success is True
+        assert result.tool_results[0].content == "2+2"
+
+    @pytest.mark.parametrize(
+        ("raw_input", "expected"),
+        [
+            ("notes/today.md", "notes/today.md"),
+            ('"notes/today.md"', "notes/today.md"),
+            ('{"path": "notes/today.md"}', "notes/today.md"),
+            ("42", "42"),
+            ("true", "true"),
+            ("false", "false"),
+            ("null", "null"),
+            ("1e3", "1e3"),
+            ('["notes/today.md"]', '["notes/today.md"]'),
+            ('["notes/today.md",]', '["notes/today.md",]'),
+            ("[draft] notes.md", "[draft] notes.md"),
+            (
+                "[x * 2 for x in range(3)]",
+                "[x * 2 for x in range(3)]",
+            ),
+            ('"notes/today.md', '"notes/today.md'),
+            ('""', ""),
+            ('"{draft} notes.md"', "{draft} notes.md"),
+            (r'"C:\\Users\\me\\notes.txt"', r"C:\Users\me\notes.txt"),
+        ],
+    )
+    def test_single_string_parameter_accepts_text_input(self, raw_input, expected):
+        """Structured text maps to the unambiguous string parameter."""
+        engine = _MockEngine(
+            [
+                f"TOOL: file_read\nINPUT: {raw_input}",
+                "FINAL_ANSWER: done",
+            ]
+        )
+        agent = OrchestratorAgent(
+            engine=engine,
+            model="test",
+            tools=[_FileReadLikeTool()],
+            mode="structured",
+        )
+
+        result = agent.run("Read my notes")
+
+        assert result.tool_results[0].success is True
+        assert result.tool_results[0].content == expected
+
+    @pytest.mark.parametrize(
+        "raw_input",
+        [
+            '{"path": "notes/today.md",}',
+            '{path: "notes/today.md"}',
+            "{'path': 'notes/today.md'}",
+            '{"unknown": 1, "path": "notes/today.md",}',
+            r'{"pa\u0074h": "notes/today.md",}',
+            '\ufeff{"path": "notes/today.md",}',
+        ],
+    )
+    def test_malformed_json_like_input_remains_an_argument_error(self, raw_input):
+        """Malformed JSON-looking text is not reclassified as a tool value."""
+        engine = _MockEngine(
+            [
+                f"TOOL: file_read\nINPUT: {raw_input}",
+                "FINAL_ANSWER: done",
+            ]
+        )
+        agent = OrchestratorAgent(
+            engine=engine,
+            model="test",
+            tools=[_FileReadLikeTool()],
+            mode="structured",
+        )
+
+        result = agent.run("Read my notes")
+
+        assert result.tool_results[0].success is False
+        assert "Invalid arguments JSON" in result.tool_results[0].content
+        assert "Tool 'file_read' failed: Invalid arguments JSON" in (
+            engine.calls[1][-1].content
+        )
+
+    @pytest.mark.parametrize(
+        "raw_input",
+        [
+            "{'code': value}",
+            '{"code": object()}',
+            "{'nested': {'value': 1}}",
+        ],
+    )
+    def test_opted_in_tool_accepts_object_prefixed_text(self, raw_input):
+        """Explicit raw-text metadata disambiguates dict-shaped source code."""
+        engine = _MockEngine(
+            [
+                f"TOOL: code\nINPUT: {raw_input}",
+                "FINAL_ANSWER: done",
+            ]
+        )
+        agent = OrchestratorAgent(
+            engine=engine,
+            model="test",
+            tools=[_CodeLikeTool()],
+            mode="structured",
+        )
+
+        result = agent.run("Execute code")
+
+        assert result.tool_results[0].success is True
+        assert result.tool_results[0].content == raw_input
+
+    def test_valid_object_remains_arguments_for_opted_in_tool(self):
+        """Raw-text metadata does not override valid JSON argument objects."""
+        engine = _MockEngine(
+            [
+                'TOOL: code\nINPUT: {"code": "print(1)"}',
+                "FINAL_ANSWER: done",
+            ]
+        )
+        agent = OrchestratorAgent(
+            engine=engine,
+            model="test",
+            tools=[_CodeLikeTool()],
+            mode="structured",
+        )
+
+        result = agent.run("Execute code")
+
+        assert result.tool_results[0].success is True
+        assert result.tool_results[0].content == "print(1)"
+
+    def test_union_string_schema_accepts_json_scalar_text(self):
+        """String unions normalize text that also parses as a JSON scalar."""
+        engine = _MockEngine(
+            [
+                "TOOL: union_file_read\nINPUT: null",
+                "FINAL_ANSWER: done",
+            ]
+        )
+        agent = OrchestratorAgent(
+            engine=engine,
+            model="test",
+            tools=[_UnionStringTool()],
+            mode="structured",
+        )
+
+        result = agent.run("Read the path named null")
+
+        assert result.tool_results[0].success is True
+        assert result.tool_results[0].content == "null"
+
+    def test_ambiguous_json_scalar_gets_stable_object_error(self):
+        """Ambiguous valid JSON is rejected before tool dispatch."""
+        engine = _MockEngine(
+            [
+                "TOOL: copy\nINPUT: 42",
+                "FINAL_ANSWER: done",
+            ]
+        )
+        agent = OrchestratorAgent(
+            engine=engine,
+            model="test",
+            tools=[_AmbiguousTool()],
+            mode="structured",
+        )
+
+        result = agent.run("Copy my notes")
+
+        assert result.tool_results[0].success is False
+        assert result.tool_results[0].content == (
+            "Invalid arguments: expected a JSON object, got int."
+        )
+        assert "Tool 'copy' failed: Invalid arguments" in (engine.calls[1][-1].content)
+
+    @pytest.mark.parametrize(
+        "raw_input",
+        [
+            "notes/today.md",
+            '{"source": "notes/today.md",}',
+        ],
+    )
+    def test_multiple_required_parameters_do_not_guess_string_mapping(
+        self,
+        raw_input,
+    ):
+        """Ambiguous bare input remains invalid instead of choosing a field."""
+        engine = _MockEngine(
+            [
+                f"TOOL: copy\nINPUT: {raw_input}",
+                "FINAL_ANSWER: done",
+            ]
+        )
+        agent = OrchestratorAgent(
+            engine=engine,
+            model="test",
+            tools=[_AmbiguousTool()],
+            mode="structured",
+        )
+
+        result = agent.run("Copy my notes")
+
+        assert result.tool_results[0].success is False
+        assert "Invalid arguments JSON" in result.tool_results[0].content
 
     def test_direct_final_answer(self):
         """Test that FINAL_ANSWER on first turn works."""

tests/tools/test_code_interpreter.py

@@ -127,6 +127,10 @@ class TestCodeInterpreterTool:
         tool = CodeInterpreterTool()
         assert tool.tool_id == "code_interpreter"
 
+    def test_structured_object_text_opt_in(self):
+        tool = CodeInterpreterTool()
+        assert tool.spec.metadata["structured_allow_object_text"] is True
+
     def test_registry_registration(self):
         ToolRegistry.register_value("code_interpreter", CodeInterpreterTool)
         assert ToolRegistry.contains("code_interpreter")

tests/tools/test_code_interpreter_docker.py

@@ -25,6 +25,7 @@ class TestDockerCodeInterpreterTool:
         assert spec.name == "code_interpreter_docker"
         assert "code" in spec.parameters["properties"]
         assert spec.category == "code"
+        assert spec.metadata["structured_allow_object_text"] is True
 
     def test_empty_code(self):
         from openjarvis.tools.code_interpreter_docker import (

tests/tools/test_repl.py

@@ -17,6 +17,10 @@ class TestReplSpec:
         tool = ReplTool()
         assert tool.spec.category == "code"
 
+    def test_structured_object_text_opt_in(self):
+        tool = ReplTool()
+        assert tool.spec.metadata["structured_allow_object_text"] is True
+
     def test_spec_parameters(self):
         tool = ReplTool()
         params = tool.spec.parameters

tests/tools/test_stubs.py

@@ -2,6 +2,10 @@
 
 from __future__ import annotations
 
+import json
+
+import pytest
+
 from openjarvis.core.events import EventBus, EventType
 from openjarvis.core.types import ToolCall, ToolResult
 from openjarvis.tools._stubs import BaseTool, ToolExecutor, ToolSpec
@@ -50,6 +54,17 @@ class _ErrorTool(BaseTool):
         raise RuntimeError("boom")
 
 
+class _ScalarBoundaryGuard:
+    """Test guard that rewrites outbound arguments to a JSON scalar."""
+
+    def check_outbound(self, tool_call: ToolCall) -> ToolCall:
+        return ToolCall(
+            id=tool_call.id,
+            name=tool_call.name,
+            arguments=json.dumps("redacted"),
+        )
+
+
 # ---------------------------------------------------------------------------
 # ToolSpec tests
 # ---------------------------------------------------------------------------
@@ -134,6 +149,38 @@ class TestToolExecutor:
         assert result.success is False
         assert "Invalid arguments JSON" in result.content
 
+    @pytest.mark.parametrize(
+        ("arguments", "decoded_type"),
+        [
+            ("42", "int"),
+            ("true", "bool"),
+            ("null", "NoneType"),
+            ("[]", "list"),
+            ('"text"', "str"),
+        ],
+    )
+    def test_execute_rejects_non_object_json(self, arguments, decoded_type):
+        executor = ToolExecutor([_EchoTool()])
+        call = ToolCall(id="1", name="echo", arguments=arguments)
+
+        result = executor.execute(call)
+
+        assert result.success is False
+        assert result.content == (
+            f"Invalid arguments: expected a JSON object, got {decoded_type}."
+        )
+
+    def test_execute_revalidates_boundary_guard_arguments(self):
+        tool = _EchoTool()
+        tool.is_local = False
+        executor = ToolExecutor([tool], boundary_guard=_ScalarBoundaryGuard())
+        call = ToolCall(id="1", name="echo", arguments='{"text":"safe"}')
+
+        result = executor.execute(call)
+
+        assert result.success is False
+        assert result.content == ("Invalid arguments: expected a JSON object, got str.")
+
     def test_execute_empty_arguments(self):
         executor = ToolExecutor([_EchoTool()])
         call = ToolCall(id="1", name="echo", arguments="")