fix: proxy WebSocket upgrades to the API in dev (#692)

The Vite dev proxy forwards `/v1` as plain HTTP with no `ws: true`, so the
WebSocket upgrade for `/v1/agents/events` is never proxied. The socket does not
open, does not error, and does not close — it just sits silent — so every live
agent view is empty under `npm run dev` while working in a production build,
where the frontend is served from the same origin as the API.

That covers the per-agent live trace on the Agents page, which subscribes via
`useAgentEvents` (`frontend/src/lib/useAgentEvents.ts`).

The silence is what makes it costly: with no error to see, it reads as "no
events are being emitted" rather than "the transport never connected", so the
search starts on the server side.

Verified on Windows 11 with `jarvis start` running: before, a
`new WebSocket('ws://localhost:5173/v1/agents/events')` from the dev page never
fired open, error or close within 6s. After, it opens, and a real agent tick
delivers 8 events (agent_tick_start, inference_start/end, tool_call_start/end,
agent_tick_end).

`changeOrigin` is set alongside so the upgrade request carries the target's
host, which some setups require when the API is not on localhost.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

.github/clone-stats/badge.json

@@ -1,7 +1,7 @@
 {
   "schemaVersion": 1,
   "label": "Git Clones",
-  "message": "159,322",
+  "message": "176,576",
   "color": "green",
   "namedLogo": "git"
 }

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

@@ -1,6 +1,6 @@
 {
-  "total_clones": 159322,
-  "last_updated": "2026-07-16T08:08:14Z",
+  "total_clones": 176576,
+  "last_updated": "2026-07-29T08:35:42Z",
   "daily": {
     "2026-03-27": 2189,
     "2026-03-28": 1874,
@@ -112,6 +112,19 @@
     "2026-07-12": 1917,
     "2026-07-13": 2102,
     "2026-07-14": 2337,
-    "2026-07-15": 2362
+    "2026-07-15": 2362,
+    "2026-07-16": 2497,
+    "2026-07-17": 1773,
+    "2026-07-18": 1542,
+    "2026-07-19": 1445,
+    "2026-07-20": 1481,
+    "2026-07-21": 1528,
+    "2026-07-22": 1529,
+    "2026-07-23": 1209,
+    "2026-07-24": 1118,
+    "2026-07-25": 928,
+    "2026-07-26": 740,
+    "2026-07-27": 799,
+    "2026-07-28": 665
   }
 }

README.md

@@ -4,6 +4,7 @@
   <p><i>Personal AI, On Personal Devices.</i></p>
 
   <p>
+    <a href="https://arxiv.org/abs/2605.17172"><img src="https://img.shields.io/badge/arXiv-2605.17172-b31b1b.svg" alt="arXiv"></a>
     <a href="https://openjarvis.stanford.edu/"><img src="https://img.shields.io/badge/project-OpenJarvis-blue" alt="Project"></a>
     <a href="https://open-jarvis.github.io/OpenJarvis/"><img src="https://img.shields.io/badge/docs-mkdocs-blue" alt="Docs"></a>
     <img src="https://img.shields.io/badge/python-%3E%3D3.10-blue" alt="Python">

configs/openjarvis/examples/full-system-access.toml

@@ -0,0 +1,39 @@
+# Full system access: unrestricted shell and filesystem
+# Copy to ~/.openjarvis/config.toml
+#
+# WARNING: shell_exec runs arbitrary commands as your user. No command
+# allowlist, no denylist, no working-directory restriction. file_read and
+# file_write aren't restricted to any directory either. Only enable what you
+# actually want the agent to have. tools.enabled is the whole permission grant;
+# there's no second allowlist to configure.
+#
+# On macOS, this config alone does not reach TCC-protected data (Messages,
+# Mail, Photos, Safari). That requires Full Disk Access granted to the process
+# hosting the backend. See docs/user-guide/system-access.md.
+#
+# Usage:
+#   jarvis ask "What's using the most disk space in my home directory?"
+#   jarvis chat        # prompts before each shell_exec call
+
+[engine]
+default = "ollama"
+
+[intelligence]
+default_model = "qwen3.5:9b"
+
+[agent]
+default_agent = "orchestrator"
+max_turns = 10
+
+[tools]
+enabled = [
+    "shell_exec",
+    "file_read",
+    "file_write",
+    "apply_patch",
+    "code_interpreter",
+    "git_status",
+    "git_diff",
+    "think",
+    "calculator",
+]

docs/getting-started/configuration.md

@@ -604,7 +604,7 @@ enforce_tool_confirmation = true
 | `scan_output` | bool | `true` | Whether to scan model output. |
 | `secret_scanner` | bool | `true` | Enable secret detection (API keys, tokens, passwords). |
 | `pii_scanner` | bool | `true` | Enable PII detection (emails, SSNs, credit cards). |
-| `enforce_tool_confirmation` | bool | `true` | Require confirmation before executing tools. |
+| `enforce_tool_confirmation` | bool | `true` | Accepted but **not currently enforced**. Whether you get prompts depends on the entry point. See [System Access](../user-guide/system-access.md#confirmation-behaviour). |
 
 !!! tip "Choosing a security mode"
     Use `"warn"` during development to see what would be flagged without disrupting output.

docs/getting-started/quickstart.md

@@ -135,6 +135,19 @@ cd OpenJarvis
 This launches the backend API server and a React frontend at [http://localhost:5173](http://localhost:5173).
 You get a ChatGPT-like interface with streaming responses, tool use, energy monitoring, and a telemetry dashboard — all running locally on your hardware.
 
+Web search is available through the built-in DuckDuckGo fallback. To use
+Tavily, add `TAVILY_API_KEY` under **Settings → Tools → Web Search** after the
+app starts, or export it before starting quickstart:
+
+```bash
+export TAVILY_API_KEY="tvly-..."
+./scripts/quickstart.sh
+```
+
+The script does not automatically source `.env` files. Run `source .env`
+first if that is where you keep the key. Stop any existing OpenJarvis server
+before restarting so it inherits the updated environment.
+
 To stop all services, press ++ctrl+c++ in the terminal.
 
 !!! tip "Environment variable"

docs/user-guide/security.md

@@ -392,7 +392,7 @@ enforce_tool_confirmation = true
 | `secret_scanner` | `bool` | `true` | Run `SecretScanner` on all text |
 | `pii_scanner` | `bool` | `true` | Run `PIIScanner` on all text |
 | `audit_log_path` | `str` | `~/.openjarvis/audit.db` | Path to the SQLite audit log |
-| `enforce_tool_confirmation` | `bool` | `true` | Require explicit confirmation before tool execution |
+| `enforce_tool_confirmation` | `bool` | `true` | Accepted by the loader but **not currently enforced**. See [System Access](system-access.md#confirmation-behaviour) for when prompts actually happen |
 
 !!! tip "Start with warn, tighten later"
     `mode = "warn"` is a good starting point. It lets you observe what patterns are being triggered without disrupting normal usage. Switch to `"redact"` once you are satisfied that the scanner isn't producing too many false positives for your workload.

docs/user-guide/system-access.md

@@ -0,0 +1,198 @@
+# System Access
+
+How to give an agent access to the machine it runs on, and where the real
+limits are.
+
+!!! warning
+    `shell_exec` runs arbitrary commands as your user. There is no command
+    allowlist, no denylist, and no sandbox unless you turn one on. An agent
+    holding this tool can do anything you can do from a terminal.
+
+---
+
+## Start here: you probably have no tools enabled
+
+If the agent tells you it can't run commands or read files, that's usually not
+a permissions problem. It means no tools were enabled in the first place.
+
+Tools come from `tools.enabled`, falling back to `agent.tools`. Both default to
+empty, and an empty value builds the agent with **zero tools**. Nothing is
+enabled by default.
+
+First check whether you have a config file at all:
+
+```bash
+cat ~/.openjarvis/config.toml
+```
+
+If it isn't there, that's your answer. Create it:
+
+```toml
+[engine]
+default = "ollama"
+
+[intelligence]
+default_model = "qwen3.5:9b"
+
+[agent]
+default_agent = "orchestrator"
+
+[tools]
+enabled = ["shell_exec", "file_read", "file_write", "think"]
+```
+
+There's a fuller version at
+`configs/openjarvis/examples/full-system-access.toml`.
+
+Then confirm the list actually resolved:
+
+```bash
+python -c "from openjarvis.core.config import load_config; print(load_config().tools.enabled)"
+```
+
+---
+
+## What the tools reach
+
+| Tool | Scope |
+|------|-------|
+| `shell_exec` | Any command, as your user. 30s default timeout, 300s max, output capped at 100 KB per stream. |
+| `file_read` | Any readable path. 1 MB cap. |
+| `file_write` | Any writable path. 10 MB cap, can create parent directories. |
+| `apply_patch` | Applies unified diffs to any path. |
+| `code_interpreter` | Python in a subprocess, behind a coarse pattern blocklist. |
+
+`file_read` and `file_write` take an `allowed_dirs` argument that limits them to
+a set of directories, but no config key populates it. When it's empty every path
+is allowed. If you want a filesystem jail today, use the container sandbox
+instead of relying on these tools to enforce one.
+
+### Sensitive filenames
+
+`file_read` and `file_write` refuse names matching a short glob list: `.env`,
+`*.pem`, `id_rsa`, `credentials.*` and a dozen or so others. It matches on the
+filename only, not the path or the contents, and only those two tools consult
+it. `shell_exec`, `apply_patch` and `code_interpreter` skip it entirely, so
+`cat ~/.ssh/id_rsa` through `shell_exec` works fine. Treat it as protection
+against fat fingers, not as a security boundary.
+
+---
+
+## Confirmation behaviour
+
+`shell_exec`, `git_commit` and `agent_kill` are marked `requires_confirmation`.
+What that translates to depends entirely on how you launched the agent:
+
+| Entry point | Behaviour |
+|-------------|-----------|
+| `jarvis chat` | Prompts before each call. |
+| `jarvis ask` | Auto-approves. |
+| `jarvis agent ask` | Auto-approves. Pass `--no-yes` if you want prompts. |
+| HTTP server, desktop app | Auto-approves. Tools you added to an agent's toolkit count as pre-approved. |
+| Embedded via `SystemBuilder` | No callback is wired, so these tools fail closed. |
+
+That last row catches people out. If `shell_exec` returns "requires
+confirmation but no confirmation callback is available", you're constructing the
+agent yourself and need to pass a `confirm_callback`.
+
+!!! note "`enforce_tool_confirmation` doesn't do anything"
+    The config loader accepts `security.enforce_tool_confirmation`, but nothing
+    on the tool execution path reads it. Setting it won't change confirmation
+    behaviour anywhere. Use the table above instead.
+
+---
+
+## macOS: Full Disk Access
+
+On macOS the operating system is the real boundary, not the config. Shell
+access and ordinary file access start working as soon as you enable the tools.
+TCC-protected data does not: Messages, Mail, Photos, Safari history, Contacts
+and Calendar all stay locked, and no config key will change that.
+
+Grant Full Disk Access to whichever process hosts the backend. Child processes
+inherit it:
+
+| How you run OpenJarvis | Grant access to |
+|------------------------|-----------------|
+| CLI (`jarvis ask`, `jarvis chat`) | Your terminal (Terminal, iTerm, Warp) |
+| Desktop app | `OpenJarvis.app`, which spawns `jarvis serve` beneath it |
+| launchd (`deploy/launchd/com.openjarvis.plist`) | The `jarvis` binary, as its own entry |
+
+System Settings, then Privacy & Security, then Full Disk Access, then **+**.
+
+A launchd daemon gets its own TCC context, so granting access to Terminal does
+nothing for it. Add `/usr/local/bin/jarvis` separately.
+
+To check whether the grant took:
+
+```bash
+head -c 16 ~/Library/Messages/chat.db >/dev/null 2>&1 \
+  && echo "granted" || echo "denied"
+```
+
+Restart the host process after you change the setting.
+
+### Driving Mac apps
+
+AppleScript works through `shell_exec`:
+
+```
+osascript -e 'tell application "Music" to play'
+```
+
+macOS asks for Automation permission once per target app, the first time you
+touch it.
+
+---
+
+## What you can't do
+
+There's no computer use. OpenJarvis can't see your screen, move the pointer or
+send keystrokes. No tool for it is registered and no input automation library
+appears anywhere in the codebase, so granting Accessibility or Screen Recording
+buys you nothing on its own.
+
+The `click` and `type` actions you'll find are Playwright, scoped to a browser
+page rather than the desktop.
+
+Some of this is reachable through `shell_exec` if you bring the tooling
+yourself. `screencapture` will take screenshots once you've granted Screen
+Recording, and something like `cliclick` will move the pointer. That gets you
+scripted actions. It doesn't get you an agent that looks at the screen and
+works out where to click.
+
+---
+
+## Narrowing access
+
+Access widens and narrows through `tools.enabled`. Drop entries to take
+capabilities away. That list is the whole grant.
+
+Two stronger isolation options exist. Both are off by default:
+
+```toml
+[sandbox]
+enabled = true          # run tools inside a container
+runtime = "docker"
+
+[security.capabilities]
+enabled = true          # RBAC over declared tool capabilities
+policy_path = "~/.openjarvis/policy.yaml"
+```
+
+!!! note "Capabilities are open by default even once enabled"
+    `CapabilityPolicy` is built with `default_deny=False` and no config key
+    exposes that flag, so an agent with no explicit policy entry gets every
+    capability. Write entries for every agent you mean to restrict.
+
+For anything untrusted, reach for `docker_shell_exec` and
+`code_interpreter_docker` rather than the host-side versions.
+
+---
+
+## See also
+
+- [Security](security.md) for scanners, the audit log and guardrails
+- [Tools](tools.md) for the full registry
+- [Code Assistant](code-assistant.md) for a narrower shell-enabled setup
+- [External MCP Servers](mcp-external-servers.md) for capabilities OpenJarvis doesn't ship

frontend/src/App.tsx

@@ -89,7 +89,7 @@ export default function App() {
           setSavings(data);
           if (optInEnabled && optInDisplayName && data) {
             const claudeEntry = data.per_provider.find(
-              (p) => p.provider === 'claude-opus-4.6',
+              (p) => p.provider === 'claude-fable-5',
             );
             const dollarSavings = claudeEntry ? claudeEntry.total_cost : 0;
             const energySaved = data.per_provider.reduce(

frontend/src/components/Chat/ChatArea.tsx

@@ -22,6 +22,8 @@ export function ChatArea() {
   const navigate = useNavigate();
   const listRef = useRef<HTMLDivElement>(null);
   const shouldAutoScroll = useRef(true);
+  const wasStreaming = useRef(false);
+  const lastScrollTop = useRef(0);
 
   // Check if any data sources are connected
   const [hasConnectedSources, setHasConnectedSources] = useState<boolean | null>(null);
@@ -34,15 +36,34 @@ export function ChatArea() {
   }, []);
 
   useEffect(() => {
+    // Sending a message always pins the view to the bottom, even if the
+    // user had scrolled up to read earlier messages.
+    if (streamState.isStreaming && !wasStreaming.current) {
+      shouldAutoScroll.current = true;
+    }
+    wasStreaming.current = streamState.isStreaming;
     if (shouldAutoScroll.current && listRef.current) {
       listRef.current.scrollTop = listRef.current.scrollHeight;
     }
-  }, [messages, streamState.content]);
+  }, [messages, streamState.content, streamState.isStreaming]);
 
   const handleScroll = () => {
     if (!listRef.current) return;
     const { scrollTop, scrollHeight, clientHeight } = listRef.current;
-    shouldAutoScroll.current = scrollHeight - scrollTop - clientHeight < 100;
+    const distance = scrollHeight - scrollTop - clientHeight;
+    const scrolledUp = scrollTop < lastScrollTop.current;
+    lastScrollTop.current = scrollTop;
+    if (scrolledUp && distance >= 1) {
+      // Any upward scroll away from the bottom stops autoscroll immediately,
+      // so streaming content never fights the user (no jitter). Sub-1px
+      // upward movement (elastic bounce settling at the bottom) is ignored.
+      shouldAutoScroll.current = false;
+    } else if (!scrolledUp) {
+      // Re-engage when scrolled back to the bottom. < 2 rather than < 1:
+      // at fractional zoom levels the at-bottom residual can reach 1px,
+      // which would otherwise leave autoscroll permanently disengaged.
+      shouldAutoScroll.current = distance < 2;
+    }
   };
 
   const isEmpty = messages.length === 0 && !streamState.isStreaming;

frontend/src/components/CommandPalette.tsx

@@ -149,8 +149,7 @@ export function CommandPalette() {
     setCommandPaletteOpen(false);
 
     if (modelId !== previousModel) {
-      const { createConversation, setModelLoading, addLogEntry } = useAppStore.getState();
-      createConversation(modelId);
+      const { setModelLoading, addLogEntry } = useAppStore.getState();
       setModelLoading(true);
       addLogEntry({ timestamp: Date.now(), level: 'info', category: 'model', message: `Switching to ${modelId}...` });
       try {

frontend/src/lib/api.auth.test.ts

@@ -6,6 +6,7 @@ import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
 // authHeaders) that source the key and build the header.
 
 const SETTINGS_KEY = 'openjarvis-settings';
+const fetchMock = vi.fn<typeof fetch>();
 
 // Minimal in-memory localStorage stub so the helpers can run under node
 // (no jsdom dependency).
@@ -28,6 +29,8 @@ class MemoryStorage {
 beforeEach(() => {
   vi.resetModules();
   vi.stubEnv('VITE_SUPABASE_ANON_KEY', 'test-anon-key');
+  fetchMock.mockReset();
+  globalThis.fetch = fetchMock;
   (globalThis as unknown as { localStorage: MemoryStorage }).localStorage =
     new MemoryStorage();
 });
@@ -86,3 +89,50 @@ describe('authHeaders', () => {
     });
   });
 });
+
+describe('tool credentials', () => {
+  it('reads credential status from the local server', async () => {
+    fetchMock.mockResolvedValue(
+      new Response(JSON.stringify({ TAVILY_API_KEY: true }), {
+        status: 200,
+        headers: { 'Content-Type': 'application/json' },
+      }),
+    );
+    const { fetchToolCredentialStatus } = await freshApi();
+
+    await expect(fetchToolCredentialStatus('web_search')).resolves.toEqual({
+      TAVILY_API_KEY: true,
+    });
+    expect(fetchMock).toHaveBeenCalledWith(
+      '/v1/tools/web_search/credentials/status',
+      { headers: {} },
+    );
+  });
+
+  it('saves a tool credential through the local server', async () => {
+    fetchMock.mockResolvedValue(new Response('{}', { status: 200 }));
+    const { saveToolCredentials } = await freshApi();
+
+    await saveToolCredentials('web_search', {
+      TAVILY_API_KEY: 'tvly-test',
+    });
+
+    expect(fetchMock).toHaveBeenCalledWith('/v1/tools/web_search/credentials', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({ TAVILY_API_KEY: 'tvly-test' }),
+    });
+  });
+
+  it('deletes a tool credential through the local server', async () => {
+    fetchMock.mockResolvedValue(new Response('{}', { status: 200 }));
+    const { deleteToolCredential } = await freshApi();
+
+    await deleteToolCredential('web_search', 'TAVILY_API_KEY');
+
+    expect(fetchMock).toHaveBeenCalledWith(
+      '/v1/tools/web_search/credentials/TAVILY_API_KEY',
+      { method: 'DELETE', headers: {} },
+    );
+  });
+});

frontend/src/lib/api.ts

@@ -885,6 +885,25 @@ export async function saveToolCredentials(
   if (!res.ok) throw new Error(`Failed: ${res.status}`);
 }
 
+export async function fetchToolCredentialStatus(
+  toolName: string,
+): Promise<Record<string, boolean>> {
+  const res = await apiFetch(`/v1/tools/${toolName}/credentials/status`);
+  if (!res.ok) throw new Error(`Failed: ${res.status}`);
+  return await res.json();
+}
+
+export async function deleteToolCredential(
+  toolName: string,
+  keyName: string,
+): Promise<void> {
+  const res = await apiFetch(
+    `/v1/tools/${encodeURIComponent(toolName)}/credentials/${encodeURIComponent(keyName)}`,
+    { method: 'DELETE' },
+  );
+  if (!res.ok) throw new Error(`Failed: ${res.status}`);
+}
+
 export interface AgentTraceDetail {
   id: string;
   agent: string;

frontend/src/pages/AgentsPage.tsx

@@ -23,7 +23,6 @@ import {
   fetchAgentTrace,
   fetchManagedAgent,
   fetchAvailableTools,
-  saveToolCredentials,
   fetchModels,
   updateManagedAgent,
   fetchRecommendedModel,

frontend/src/pages/SettingsPage.tsx

@@ -27,6 +27,9 @@ import {
   setInferenceSource,
   getCloudKeyStatus,
   saveCloudKey,
+  fetchToolCredentialStatus,
+  saveToolCredentials,
+  deleteToolCredential,
   isTauri,
   type InferenceSource,
 } from '../lib/api';
@@ -56,25 +59,37 @@ function OllamaModelList() {
   );
 }
 
-function ApiKeyInput({ keyName, placeholder }: { keyName: string; placeholder: string }) {
+function ApiKeyInput({
+  keyName,
+  placeholder,
+  toolName,
+}: {
+  keyName: string;
+  placeholder: string;
+  toolName?: string;
+}) {
   const [value, setValue] = useState('');
   const [saved, setSaved] = useState(false);
   const [hasKey, setHasKey] = useState(false);
   const [error, setError] = useState('');
   const desktopKeyStorage = isTauri();
+  const serverToolStorage = !desktopKeyStorage && !!toolName;
+  const canManage = desktopKeyStorage || serverToolStorage;
 
   const refresh = useCallback(async () => {
-    if (!desktopKeyStorage) {
+    if (!canManage) {
       setHasKey(false);
       return;
     }
     try {
-      const status = await getCloudKeyStatus();
+      const status = desktopKeyStorage
+        ? await getCloudKeyStatus()
+        : await fetchToolCredentialStatus(toolName!);
       setHasKey(!!status[keyName]);
     } catch {
       setHasKey(false);
     }
-  }, [desktopKeyStorage, keyName]);
+  }, [canManage, desktopKeyStorage, keyName, toolName]);
 
   useEffect(() => {
     void refresh();
@@ -87,7 +102,13 @@ function ApiKeyInput({ keyName, placeholder }: { keyName: string; placeholder: s
     if (!next) return;
     setError('');
     try {
-      await saveCloudKey(keyName, next);
+      if (desktopKeyStorage) {
+        await saveCloudKey(keyName, next);
+      } else if (toolName) {
+        await saveToolCredentials(toolName, { [keyName]: next });
+      } else {
+        return;
+      }
       setValue('');
       setHasKey(true);
       setSaved(true);
@@ -101,7 +122,13 @@ function ApiKeyInput({ keyName, placeholder }: { keyName: string; placeholder: s
   const remove = async () => {
     setError('');
     try {
-      await saveCloudKey(keyName, '');
+      if (desktopKeyStorage) {
+        await saveCloudKey(keyName, '');
+      } else if (toolName) {
+        await deleteToolCredential(toolName, keyName);
+      } else {
+        return;
+      }
       setValue('');
       setHasKey(false);
       setSaved(true);
@@ -119,8 +146,8 @@ function ApiKeyInput({ keyName, placeholder }: { keyName: string; placeholder: s
         value={value}
         onChange={e => setValue(e.target.value)}
         onBlur={() => { if (value.trim()) void save(value); }}
-        placeholder={hasKey ? 'Saved in secure storage' : placeholder}
-        disabled={!desktopKeyStorage}
+        placeholder={hasKey ? (desktopKeyStorage ? 'Saved in secure storage' : 'Saved by local server') : placeholder}
+        disabled={!canManage}
         className="w-48 px-2 py-1 rounded text-xs"
         style={{ background: 'var(--color-bg)', border: '1px solid var(--color-border)', color: 'var(--color-text)' }} />
       {hasKey && (
@@ -542,7 +569,7 @@ export function SettingsPage() {
           {/* Tools */}
           <Section title="Tools">
             <SettingRow label="Web Search" description="Tavily key for web search tool">
-              <ApiKeyInput keyName="TAVILY_API_KEY" placeholder="tvly-..." />
+              <ApiKeyInput keyName="TAVILY_API_KEY" placeholder="tvly-..." toolName="web_search" />
             </SettingRow>
           </Section>
 

frontend/vite.config.ts

@@ -54,7 +54,15 @@ export default defineConfig({
   server: {
     port: 5173,
     proxy: {
-      '/v1': process.env.VITE_API_URL || 'http://localhost:8000',
+      // ws: true is required for the /v1/agents/events WebSocket. Without it
+      // Vite proxies the HTTP request but not the upgrade, so the socket never
+      // opens — no error, no close event, just silence — and every live agent
+      // view sits empty in dev while working in a production build.
+      '/v1': {
+        target: process.env.VITE_API_URL || 'http://localhost:8000',
+        changeOrigin: true,
+        ws: true,
+      },
       '/health': process.env.VITE_API_URL || 'http://localhost:8000',
       '/api': process.env.VITE_API_URL || 'http://localhost:8000',
     },

mkdocs.yml

@@ -196,6 +196,7 @@ nav:
       - Telemetry: user-guide/telemetry.md
       - Evaluations: user-guide/evaluations.md
       - Benchmarks: user-guide/benchmarks.md
+      - System Access: user-guide/system-access.md
       - Security: user-guide/security.md
       - LLM-guided spec search: user-guide/llm-guided-spec-search.md
   - Leaderboard: leaderboard.md

scripts/quickstart.sh

@@ -148,7 +148,8 @@ fi
 
 # ── 7. Install Python dependencies ──────────────────────────────────
 info "Installing Python dependencies..."
-uv sync --extra desktop --quiet 2>/dev/null || uv sync --extra desktop
+uv sync --extra desktop --extra tools-search --quiet 2>/dev/null \
+  || uv sync --extra desktop --extra tools-search
 ok "Python dependencies installed"
 
 # ── 7b. Build Rust extension ──────────────────────────────────────
@@ -164,11 +165,17 @@ ok "Frontend dependencies installed"
 
 # ── 9. Start backend ────────────────────────────────────────────────
 info "Starting backend API server on port 8000..."
+if curl -sf http://localhost:8000/health &>/dev/null; then
+  fail "An OpenJarvis server is already running on port 8000. Stop it before re-running quickstart so updated environment variables are applied."
+fi
 uv run jarvis serve --port 8000 &>/dev/null &
-CLEANUP_PIDS+=($!)
+BACKEND_PID=$!
+CLEANUP_PIDS+=("$BACKEND_PID")
 sleep 3
 
-if curl -sf http://localhost:8000/health &>/dev/null; then
+if ! kill -0 "$BACKEND_PID" 2>/dev/null; then
+  fail "Backend exited during startup. Run 'uv run jarvis serve --port 8000' to see the error."
+elif curl -sf http://localhost:8000/health &>/dev/null; then
   ok "Backend running at http://localhost:8000"
 else
   warn "Backend may still be starting..."

src/openjarvis/cli/daemon_cmd.py

@@ -81,14 +81,28 @@ def start(
     if agent_name:
         cmd.extend(["--agent", agent_name])
 
-    # Start as background process
+    # Start as background process, fully detached from the launching terminal.
+    #
+    # ``start_new_session`` is POSIX-only: CPython's Windows ``_execute_child``
+    # names the parameter ``unused_start_new_session`` and ignores it. Relying
+    # on it there leaves the server sharing its parent's console, so closing
+    # that console — or logging off — delivers CTRL_CLOSE_EVENT and kills the
+    # daemon. DETACHED_PROCESS gives it no console at all; the new process
+    # group additionally stops a Ctrl-C in the parent reaching it.
     DEFAULT_CONFIG_DIR.mkdir(parents=True, exist_ok=True)
     log_fh = open(_LOG_FILE, "a")  # noqa: SIM115
+    spawn_kwargs: dict = {}
+    if sys.platform == "win32":
+        spawn_kwargs["creationflags"] = (
+            subprocess.DETACHED_PROCESS | subprocess.CREATE_NEW_PROCESS_GROUP
+        )
+    else:
+        spawn_kwargs["start_new_session"] = True
     proc = subprocess.Popen(
         cmd,
         stdout=log_fh,
         stderr=log_fh,
-        start_new_session=True,
+        **spawn_kwargs,
     )
     _write_pid(proc.pid)
 

src/openjarvis/cli/serve.py

@@ -10,6 +10,7 @@ from rich.console import Console
 
 from openjarvis.cli._banner import print_banner
 from openjarvis.core.config import load_config
+from openjarvis.core.credentials import inject_credentials
 from openjarvis.core.events import EventBus
 from openjarvis.core.paths import get_config_dir
 from openjarvis.engine import (
@@ -122,6 +123,11 @@ def serve(
         )
         sys.exit(1)
 
+    # Tool credentials saved through the browser UI live in the OpenJarvis
+    # credential store. Restore them before engines and tools are constructed
+    # so availability checks and tool instances see the same environment.
+    inject_credentials()
+
     config = load_config()
 
     # Resolve host/port from CLI args or config

src/openjarvis/cli/skill_cmd.py

@@ -199,7 +199,16 @@ def _get_resolver(source: str, url: str = ""):
     default="",
     help="Repo URL (required when source is 'github').",
 )
-def install(query: str, with_scripts: bool, force: bool, url: str):
+@click.option(
+    "--yes-dangerous",
+    is_flag=True,
+    default=False,
+    help=(
+        "Confirm installing an unreviewed skill that requests dangerous "
+        "capabilities (shell/network-listen/filesystem-write)."
+    ),
+)
+def install(query: str, with_scripts: bool, force: bool, url: str, yes_dangerous: bool):
     """Install a skill from a source.
 
     Example: ``jarvis skill install hermes:apple-notes``
@@ -233,7 +242,12 @@ def install(query: str, with_scripts: bool, force: bool, url: str):
     from openjarvis.skills.tool_translator import ToolTranslator
 
     importer = SkillImporter(parser=SkillParser(), tool_translator=ToolTranslator())
-    result = importer.import_skill(matches[0], with_scripts=with_scripts, force=force)
+    result = importer.import_skill(
+        matches[0],
+        with_scripts=with_scripts,
+        force=force,
+        confirm_dangerous=yes_dangerous,
+    )
 
     if result.success:
         if result.skipped:
@@ -270,6 +284,15 @@ def install(query: str, with_scripts: bool, force: bool, url: str):
     help="Import scripts/ directories.",
 )
 @click.option("--force", is_flag=True, default=False, help="Re-import existing skills.")
+@click.option(
+    "--yes-dangerous",
+    is_flag=True,
+    default=False,
+    help=(
+        "Confirm installing unreviewed skills that request dangerous "
+        "capabilities (shell/network-listen/filesystem-write)."
+    ),
+)
 def sync(
     source: str,
     category: str,
@@ -277,6 +300,7 @@ def sync(
     search: str,
     with_scripts: bool,
     force: bool,
+    yes_dangerous: bool,
 ):
     """Bulk install + update from a source (or all configured sources)."""
     console = Console()
@@ -343,9 +367,20 @@ def sync(
 
         installed_count = 0
         for resolved in skills_to_import:
-            r = importer.import_skill(resolved, with_scripts=with_scripts, force=force)
+            r = importer.import_skill(
+                resolved,
+                with_scripts=with_scripts,
+                force=force,
+                confirm_dangerous=yes_dangerous,
+            )
             if r.success and not r.skipped:
                 installed_count += 1
+            elif not r.success and r.requires_confirmation:
+                console.print(
+                    f"  [yellow]Skipped {resolved.name}: requests dangerous "
+                    f"capabilities {r.dangerous_capabilities} "
+                    "(re-run with --yes-dangerous to install)[/yellow]"
+                )
         console.print(f"  Imported {installed_count}/{len(skills_to_import)} skills")
         total_installed += installed_count
 

src/openjarvis/core/credentials.py

@@ -67,6 +67,24 @@ def load_credentials(path: Path | None = None) -> dict[str, dict[str, str]]:
         return tomllib.load(f)
 
 
+def _validate_credential_key(tool_name: str, key: str) -> None:
+    allowed = TOOL_CREDENTIALS.get(tool_name, [])
+    if key not in allowed:
+        raise ValueError(f"Unknown credential key '{key}' for tool '{tool_name}'")
+
+
+def _write_credentials(creds: dict[str, dict[str, str]], path: Path) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    lines: list[str] = []
+    for section, kvs in creds.items():
+        lines.append(f"[{section}]")
+        for k, v in kvs.items():
+            lines.append(f'{k} = "{v}"')
+        lines.append("")
+    path.write_text("\n".join(lines))
+    os.chmod(path, 0o600)
+
+
 def save_credential(
     tool_name: str,
     key: str,
@@ -75,9 +93,7 @@ def save_credential(
     path: Path | None = None,
 ) -> None:
     """Save a single credential key, validate, write file, and set os.environ."""
-    allowed = TOOL_CREDENTIALS.get(tool_name, [])
-    if key not in allowed:
-        raise ValueError(f"Unknown credential key '{key}' for tool '{tool_name}'")
+    _validate_credential_key(tool_name, key)
     stripped = value.strip()
     if not stripped:
         raise ValueError("Credential value must not be empty")
@@ -88,20 +104,32 @@ def save_credential(
         if tool_name not in creds:
             creds[tool_name] = {}
         creds[tool_name][key] = stripped
-
-        p.parent.mkdir(parents=True, exist_ok=True)
-        lines: list[str] = []
-        for section, kvs in creds.items():
-            lines.append(f"[{section}]")
-            for k, v in kvs.items():
-                lines.append(f'{k} = "{v}"')
-            lines.append("")
-        p.write_text("\n".join(lines))
-        os.chmod(p, 0o600)
+        _write_credentials(creds, p)
 
     os.environ[key] = stripped
 
 
+def delete_credential(
+    tool_name: str,
+    key: str,
+    *,
+    path: Path | None = None,
+) -> None:
+    """Delete a persisted credential and remove it from the running process."""
+    _validate_credential_key(tool_name, key)
+    p = Path(path) if path else _default_path()
+    with _LOCK:
+        creds = load_credentials(path=p)
+        tool_creds = creds.get(tool_name)
+        if tool_creds is not None:
+            tool_creds.pop(key, None)
+            if not tool_creds:
+                creds.pop(tool_name, None)
+            _write_credentials(creds, p)
+
+    os.environ.pop(key, None)
+
+
 def get_credential_status(tool_name: str) -> dict[str, bool]:
     """Return {KEY: bool} for each required key indicating if set in env."""
     keys = TOOL_CREDENTIALS.get(tool_name, [])

src/openjarvis/server/agent_manager_routes.py

@@ -285,16 +285,17 @@ def build_tools_list() -> List[Dict[str, Any]]:
             logger.debug("Could not instantiate tool %s: %s", name, exc)
             spec = None
         cred_keys = TOOL_CREDENTIALS.get(name, [])
+        has_fallback = bool(spec and spec.metadata.get("fallback"))
         items.append(
             {
                 "name": name,
                 "description": spec.description if spec else "",
                 "category": spec.category if spec else "",
                 "source": "tool",
-                "requires_credentials": len(cred_keys) > 0,
+                "requires_credentials": len(cred_keys) > 0 and not has_fallback,
                 "credential_keys": cred_keys,
                 "configured": (
-                    all(bool(os.environ.get(k)) for k in cred_keys)
+                    has_fallback or all(bool(os.environ.get(k)) for k in cred_keys)
                     if cred_keys
                     else True
                 ),
@@ -2238,6 +2239,13 @@ def create_agent_manager_router(
             saved.append(key)
         return {"saved": saved}
 
+    @tools_router.delete("/{tool_name}/credentials/{key}")
+    def remove_tool_credential(tool_name: str, key: str):
+        from openjarvis.core.credentials import delete_credential
+
+        delete_credential(tool_name, key)
+        return {"deleted": key}
+
     @tools_router.get("/{tool_name}/credentials/status")
     def credential_status(tool_name: str):
         from openjarvis.core.credentials import get_credential_status

src/openjarvis/server/session_store.py

@@ -25,9 +25,11 @@ class SessionStore:
     def __init__(self, db_path: str = "") -> None:
         if not db_path:
             db_path = str(get_config_dir() / "sessions.db")
-        from openjarvis.security.file_utils import secure_create
+        # Ensure the parent directory exists (skip for :memory:)
+        if db_path != ":memory:":
+            from openjarvis.security.file_utils import secure_create
 
-        secure_create(Path(db_path))
+            secure_create(Path(db_path))
         self._db = sqlite3.connect(db_path, check_same_thread=False)
         self._db.row_factory = sqlite3.Row
         self._create_tables()

src/openjarvis/skills/executor.py

@@ -5,10 +5,11 @@ from __future__ import annotations
 import json
 import re
 from dataclasses import dataclass, field
-from typing import Any, Callable, Dict, List, Optional
+from typing import Any, Callable, Dict, List, Optional, Set
 
 from openjarvis.core.events import EventBus, EventType
 from openjarvis.core.types import ToolCall, ToolResult
+from openjarvis.skills.security import validate_capabilities
 from openjarvis.skills.types import SkillManifest
 from openjarvis.tools._stubs import ToolExecutor
 
@@ -37,10 +38,16 @@ class SkillExecutor:
         tool_executor: ToolExecutor,
         *,
         bus: Optional[EventBus] = None,
+        allowed_capabilities: Optional[Set[str]] = None,
     ) -> None:
         self._tool_executor = tool_executor
         self._bus = bus
         self._skill_resolver: Optional[SkillResolver] = None
+        # None means "no capability policy" — every skill runs, matching the
+        # behavior before capability enforcement existed. Pass a set (even an
+        # empty one) to enforce: skills whose required_capabilities are not a
+        # subset of it are blocked before any step runs.
+        self._allowed_capabilities: Optional[Set[str]] = allowed_capabilities
 
     def set_skill_resolver(self, resolver: SkillResolver) -> None:
         """Register a callback used to delegate ``skill_name`` steps."""
@@ -53,6 +60,38 @@ class SkillExecutor:
         initial_context: Optional[Dict[str, Any]] = None,
     ) -> SkillResult:
         """Execute all steps in a skill manifest."""
+        missing = (
+            validate_capabilities(manifest, self._allowed_capabilities)
+            if self._allowed_capabilities is not None
+            else []
+        )
+        if missing:
+            if self._bus:
+                self._bus.publish(
+                    EventType.SKILL_EXECUTE_START,
+                    {"skill": manifest.name, "steps": len(manifest.steps)},
+                )
+                self._bus.publish(
+                    EventType.SKILL_EXECUTE_END,
+                    {"skill": manifest.name, "success": False},
+                )
+            return SkillResult(
+                skill_name=manifest.name,
+                success=False,
+                step_results=[
+                    ToolResult(
+                        tool_name=manifest.name,
+                        content=(
+                            f"Blocked: skill '{manifest.name}' requires "
+                            f"capabilities {missing} that were not granted "
+                            "for this session."
+                        ),
+                        success=False,
+                    )
+                ],
+                context=dict(initial_context or {}),
+            )
+
         ctx: Dict[str, Any] = dict(initial_context or {})
         all_results: List[ToolResult] = []
 

src/openjarvis/skills/importer.py

@@ -25,6 +25,11 @@ import yaml
 
 from openjarvis.core.paths import get_config_dir
 from openjarvis.skills.parser import SkillParser
+from openjarvis.skills.security import (
+    TrustTier,
+    classify_trust_tier,
+    has_dangerous_capabilities,
+)
 from openjarvis.skills.sources.base import ResolvedSkill
 from openjarvis.skills.tool_translator import ToolTranslator
 
@@ -43,6 +48,9 @@ class ImportResult:
     untranslated_tools: List[str] = field(default_factory=list)
     scripts_imported: bool = False
     warnings: List[str] = field(default_factory=list)
+    trust_tier: TrustTier = TrustTier.UNREVIEWED
+    dangerous_capabilities: List[str] = field(default_factory=list)
+    requires_confirmation: bool = False
 
 
 class SkillImporter:
@@ -66,6 +74,7 @@ class SkillImporter:
         *,
         with_scripts: bool = False,
         force: bool = False,
+        confirm_dangerous: bool = False,
     ) -> ImportResult:
         """Install *resolved* into ``<target_root>/<source>/<name>/``.
 
@@ -95,12 +104,45 @@ class SkillImporter:
 
         try:
             frontmatter, body = self._read_skill_md(source_md)
-            self._parser.parse_frontmatter(frontmatter, markdown_content=body)
+            manifest = self._parser.parse_frontmatter(
+                frontmatter, markdown_content=body
+            )
         except Exception as exc:
             result.success = False
             result.warnings.append(f"Parse error: {exc}")
             return result
 
+        # 1a. Classify trust and check for dangerous capabilities *before*
+        # writing anything to disk. Everything the importer handles comes from
+        # an external source (github/hermes/openclaw), so the BUNDLED and
+        # WORKSPACE tiers never apply here, and no resolver verifies index
+        # membership yet — a signature alone still classifies as UNREVIEWED.
+        # Community skills get no special treatment just because they came
+        # from a named source.
+        result.trust_tier = classify_trust_tier(
+            has_signature=bool(manifest.signature),
+        )
+        result.dangerous_capabilities = has_dangerous_capabilities(manifest)
+
+        if result.dangerous_capabilities and result.trust_tier == TrustTier.UNREVIEWED:
+            result.requires_confirmation = True
+            if not confirm_dangerous:
+                result.success = False
+                result.warnings.append(
+                    "Refusing to install: this unreviewed skill requests "
+                    f"dangerous capabilities {result.dangerous_capabilities}. "
+                    "Re-run with confirm_dangerous=True (or `--yes-dangerous` "
+                    "on the CLI) only if you trust the source and have "
+                    "reviewed what it does."
+                )
+                return result
+            result.warnings.append(
+                "Installed with dangerous capabilities "
+                f"{result.dangerous_capabilities} — confirmed by caller. "
+                "This skill can run shell commands, open network listeners, "
+                "and/or write to the filesystem."
+            )
+
         # 2. Translate tool references
         translated_body, untranslated = self._translator.translate_markdown(body)
         result.untranslated_tools = untranslated
@@ -180,6 +222,7 @@ class SkillImporter:
         translated_str = ", ".join(f'"{t}"' for t in result.translated_tools)
         missing_str = ", ".join(f'"{t}"' for t in result.untranslated_tools)
         scripts_lower = "true" if result.scripts_imported else "false"
+        dangerous_str = ", ".join(f'"{c}"' for c in result.dangerous_capabilities)
 
         content = (
             f'source = "{resolved.source}:{resolved.name}"\n'
@@ -189,6 +232,8 @@ class SkillImporter:
             f"translated_tools = [{translated_str}]\n"
             f"missing_tools = [{missing_str}]\n"
             f"scripts_imported = {scripts_lower}\n"
+            f'trust_tier = "{result.trust_tier.value}"\n'
+            f"dangerous_capabilities = [{dangerous_str}]\n"
         )
         (target_dir / ".source").write_text(content, encoding="utf-8")
 

src/openjarvis/telemetry/store.py

@@ -108,6 +108,13 @@ INSERT INTO telemetry (
 )
 """
 
+_INSERT_MINING = """\
+INSERT INTO mining_stats (
+    recorded_at, provider_id, shares_submitted, shares_accepted, blocks_found,
+    hashrate, uptime_seconds, last_share_at, last_error, payout_target, fees_owed
+) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+"""
+
 _MIGRATE_COLUMNS = [
     ("gpu_utilization_pct", "REAL NOT NULL DEFAULT 0.0"),
     ("gpu_memory_used_gb", "REAL NOT NULL DEFAULT 0.0"),
@@ -144,10 +151,35 @@ _MIGRATE_COLUMNS = [
 
 
 class TelemetryStore:
-    """Append-only SQLite store for inference telemetry records."""
+    """Append-only SQLite store for inference telemetry records.
+
+    Writes are batched in memory and flushed to SQLite when a batch reaches
+    ``batch_size``, when ``flush_interval_seconds`` elapses (a background
+    flusher thread guarantees this even with no further writes), on any read
+    through this store, and on ``close()``. Readers that open their OWN
+    connection to the database file (e.g. ``TelemetryAggregator``) therefore
+    see new rows within ``flush_interval_seconds`` at the latest; call
+    ``flush()`` first for immediate visibility. Pass
+    ``flush_interval_seconds=0`` to disable time-based flushing (batch-size
+    and read/close flushes still apply).
+    """
+
+    def __init__(
+        self,
+        db_path: str | Path,
+        batch_size: int = 50,
+        flush_interval_seconds: float = 5.0,
+    ) -> None:
+        if batch_size < 1:
+            raise ValueError("batch_size must be >= 1")
+        if flush_interval_seconds < 0:
+            raise ValueError("flush_interval_seconds must be >= 0")
 
-    def __init__(self, db_path: str | Path) -> None:
         self._db_path = str(db_path)
+        if self._db_path != ":memory:":
+            from openjarvis.security.file_utils import secure_create
+
+            secure_create(Path(self._db_path))
         self._conn = sqlite3.connect(self._db_path, check_same_thread=False)
         self._lock = threading.Lock()
         self._conn.execute("PRAGMA journal_mode=WAL")
@@ -158,6 +190,38 @@ class TelemetryStore:
         self._conn.commit()
         self._migrate_schema()
 
+        self._batch_size = batch_size
+        self._flush_interval_seconds = flush_interval_seconds
+        self._last_flush_time = time.monotonic()
+        self._telemetry_batch: list[tuple[Any, ...]] = []
+        self._mining_batch: list[tuple[Any, ...]] = []
+        self._closed = False
+
+        # Background flusher: without it, a partial batch written just before
+        # traffic stops would stay invisible to other connections until the
+        # NEXT write arrived (the stale check in ``_maybe_flush_unlocked``
+        # only runs inside record calls). Daemon so it never blocks exit.
+        self._stop_flusher = threading.Event()
+        self._flusher: threading.Thread | None = None
+        if flush_interval_seconds > 0:
+            self._flusher = threading.Thread(
+                target=self._flush_loop,
+                name="telemetry-store-flusher",
+                daemon=True,
+            )
+            self._flusher.start()
+
+    def _flush_loop(self) -> None:
+        """Periodically flush pending batches until ``close()`` stops us."""
+        while not self._stop_flusher.wait(self._flush_interval_seconds):
+            with self._lock:
+                # ``close()`` sets the event BEFORE taking the lock, so seeing
+                # it unset here means the connection is still open.
+                if self._stop_flusher.is_set():
+                    break
+                if self._telemetry_batch or self._mining_batch:
+                    self._flush_unlocked()
+
     def _migrate_schema(self) -> None:
         """Add new columns to existing databases (idempotent)."""
         for col_name, col_def in _MIGRATE_COLUMNS:
@@ -171,54 +235,52 @@ class TelemetryStore:
 
     def record(self, rec: TelemetryRecord) -> None:
         """Persist a single telemetry record."""
+        row = (
+            rec.timestamp,
+            rec.model_id,
+            rec.engine,
+            rec.agent,
+            rec.prompt_tokens,
+            rec.prompt_tokens_evaluated,
+            rec.completion_tokens,
+            rec.total_tokens,
+            rec.latency_seconds,
+            rec.ttft,
+            rec.cost_usd,
+            rec.energy_joules,
+            rec.power_watts,
+            rec.gpu_utilization_pct,
+            rec.gpu_memory_used_gb,
+            rec.gpu_temperature_c,
+            rec.throughput_tok_per_sec,
+            rec.prefill_latency_seconds,
+            rec.decode_latency_seconds,
+            rec.energy_method,
+            rec.energy_vendor,
+            rec.batch_id,
+            1 if rec.is_warmup else 0,
+            rec.cpu_energy_joules,
+            rec.gpu_energy_joules,
+            rec.dram_energy_joules,
+            rec.tokens_per_joule,
+            rec.energy_per_output_token_joules,
+            rec.throughput_per_watt,
+            rec.prefill_energy_joules,
+            rec.decode_energy_joules,
+            rec.mean_itl_ms,
+            rec.median_itl_ms,
+            rec.p90_itl_ms,
+            rec.p95_itl_ms,
+            rec.p99_itl_ms,
+            rec.std_itl_ms,
+            1 if rec.is_streaming else 0,
+            rec.token_counting_version,
+            rec.mining_session_id,
+            json.dumps(rec.metadata),
+        )
         with self._lock:
-            self._conn.execute(
-                _INSERT,
-                (
-                    rec.timestamp,
-                    rec.model_id,
-                    rec.engine,
-                    rec.agent,
-                    rec.prompt_tokens,
-                    rec.prompt_tokens_evaluated,
-                    rec.completion_tokens,
-                    rec.total_tokens,
-                    rec.latency_seconds,
-                    rec.ttft,
-                    rec.cost_usd,
-                    rec.energy_joules,
-                    rec.power_watts,
-                    rec.gpu_utilization_pct,
-                    rec.gpu_memory_used_gb,
-                    rec.gpu_temperature_c,
-                    rec.throughput_tok_per_sec,
-                    rec.prefill_latency_seconds,
-                    rec.decode_latency_seconds,
-                    rec.energy_method,
-                    rec.energy_vendor,
-                    rec.batch_id,
-                    1 if rec.is_warmup else 0,
-                    rec.cpu_energy_joules,
-                    rec.gpu_energy_joules,
-                    rec.dram_energy_joules,
-                    rec.tokens_per_joule,
-                    rec.energy_per_output_token_joules,
-                    rec.throughput_per_watt,
-                    rec.prefill_energy_joules,
-                    rec.decode_energy_joules,
-                    rec.mean_itl_ms,
-                    rec.median_itl_ms,
-                    rec.p90_itl_ms,
-                    rec.p95_itl_ms,
-                    rec.p99_itl_ms,
-                    rec.std_itl_ms,
-                    1 if rec.is_streaming else 0,
-                    rec.token_counting_version,
-                    rec.mining_session_id,
-                    json.dumps(rec.metadata),
-                ),
-            )
-            self._conn.commit()
+            self._telemetry_batch.append(row)
+            self._maybe_flush_unlocked()
 
     def record_mining_stats(self, stats: Any) -> None:
         """Persist one mining stats snapshot.
@@ -226,43 +288,70 @@ class TelemetryStore:
         ``stats`` is duck-typed to keep telemetry usable without importing the
         optional mining package at module import time.
         """
+        row = (
+            time.time(),
+            stats.provider_id,
+            stats.shares_submitted,
+            stats.shares_accepted,
+            stats.blocks_found,
+            stats.hashrate,
+            stats.uptime_seconds,
+            stats.last_share_at,
+            stats.last_error,
+            stats.payout_target,
+            stats.fees_owed,
+        )
         with self._lock:
-            self._conn.execute(
-                """\
-INSERT INTO mining_stats (
-    recorded_at, provider_id, shares_submitted, shares_accepted, blocks_found,
-    hashrate, uptime_seconds, last_share_at, last_error, payout_target, fees_owed
-) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
-""",
-                (
-                    time.time(),
-                    stats.provider_id,
-                    stats.shares_submitted,
-                    stats.shares_accepted,
-                    stats.blocks_found,
-                    stats.hashrate,
-                    stats.uptime_seconds,
-                    stats.last_share_at,
-                    stats.last_error,
-                    stats.payout_target,
-                    stats.fees_owed,
-                ),
-            )
-            self._conn.commit()
+            self._mining_batch.append(row)
+            self._maybe_flush_unlocked()
+
+    def flush(self) -> None:
+        """Write all pending records to the database."""
+        with self._lock:
+            self._flush_unlocked()
+
+    def _flush_unlocked(self) -> None:
+        if self._telemetry_batch:
+            self._conn.executemany(_INSERT, self._telemetry_batch)
+            self._telemetry_batch.clear()
+        if self._mining_batch:
+            self._conn.executemany(_INSERT_MINING, self._mining_batch)
+            self._mining_batch.clear()
+        self._conn.commit()
+        self._last_flush_time = time.monotonic()
+
+    def _maybe_flush_unlocked(self) -> None:
+        """Flush when the batch is full or has been pending too long."""
+        if not self._telemetry_batch and not self._mining_batch:
+            return
+        batch_full = (
+            len(self._telemetry_batch) >= self._batch_size
+            or len(self._mining_batch) >= self._batch_size
+        )
+        stale = (
+            self._flush_interval_seconds > 0
+            and time.monotonic() - self._last_flush_time >= self._flush_interval_seconds
+        )
+        if batch_full or stale:
+            self._flush_unlocked()
 
     def list_recent(self, limit: int = 50) -> list[dict[str, Any]]:
         """Return recent telemetry rows as dictionaries."""
-        return self._select_dicts(
-            "SELECT * FROM telemetry ORDER BY timestamp DESC LIMIT ?",
-            (limit,),
-        )
+        with self._lock:
+            self._flush_unlocked()
+            return self._select_dicts_unlocked(
+                "SELECT * FROM telemetry ORDER BY timestamp DESC LIMIT ?",
+                (limit,),
+            )
 
     def list_recent_mining_stats(self, limit: int = 50) -> list[dict[str, Any]]:
         """Return recent mining stats snapshots as dictionaries."""
-        return self._select_dicts(
-            "SELECT * FROM mining_stats ORDER BY recorded_at DESC LIMIT ?",
-            (limit,),
-        )
+        with self._lock:
+            self._flush_unlocked()
+            return self._select_dicts_unlocked(
+                "SELECT * FROM mining_stats ORDER BY recorded_at DESC LIMIT ?",
+                (limit,),
+            )
 
     def subscribe_to_bus(self, bus: EventBus) -> None:
         """Subscribe to ``TELEMETRY_RECORD`` events on *bus*."""
@@ -277,15 +366,36 @@ INSERT INTO mining_stats (
                 logger.debug("Failed to record telemetry event: %s", exc)
 
     def close(self) -> None:
-        """Close the underlying SQLite connection."""
-        self._conn.close()
+        """Flush pending records and close the underlying SQLite connection."""
+        # Set the stop event BEFORE taking the lock: a flusher iteration
+        # already waiting on the lock re-checks the event after acquiring it
+        # and exits instead of touching the closed connection.
+        self._stop_flusher.set()
+        with self._lock:
+            if self._closed:
+                return
+            self._flush_unlocked()
+            self._conn.close()
+            self._closed = True
+        if self._flusher is not None:
+            self._flusher.join(timeout=1.0)
+            self._flusher = None
 
     # -- helpers for querying (used by tests) --------------------------------
 
     def _fetchall(self, sql: str = "SELECT * FROM telemetry") -> list:
-        return self._conn.execute(sql).fetchall()
+        with self._lock:
+            self._flush_unlocked()
+            return self._conn.execute(sql).fetchall()
 
     def _select_dicts(self, sql: str, params: tuple[Any, ...]) -> list[dict[str, Any]]:
+        with self._lock:
+            self._flush_unlocked()
+            return self._select_dicts_unlocked(sql, params)
+
+    def _select_dicts_unlocked(
+        self, sql: str, params: tuple[Any, ...]
+    ) -> list[dict[str, Any]]:
         cur = self._conn.execute(sql, params)
         columns = [desc[0] for desc in cur.description]
         return [dict(zip(columns, row)) for row in cur.fetchall()]

src/openjarvis/tools/git_tool.py

@@ -139,8 +139,8 @@ class GitStatusTool(BaseTool):
 
     def execute(self, **params: Any) -> ToolResult:
         repo_path = params.get("repo_path", ".")
-        _rust = get_rust_module()
         try:
+            _rust = get_rust_module()
             output = _rust.GitStatusTool().execute(repo_path)
             return ToolResult(
                 tool_name="git_status",
@@ -148,6 +148,8 @@ class GitStatusTool(BaseTool):
                 success=True,
                 metadata={"returncode": 0},
             )
+        except ImportError as exc:
+            logger.debug("Rust git_status fallback to CLI: %s", exc)
         except Exception as exc:
             return ToolResult(
                 tool_name="git_status",
@@ -155,6 +157,8 @@ class GitStatusTool(BaseTool):
                 success=False,
             )
 
+        return _run_git(["git", "status", "--porcelain"], cwd=repo_path)
+
 
 # ---------------------------------------------------------------------------
 # GitDiffTool
@@ -208,9 +212,9 @@ class GitDiffTool(BaseTool):
         staged = params.get("staged", False)
         file_path = params.get("path")
 
-        _rust = get_rust_module()
         if not staged and not file_path:
             try:
+                _rust = get_rust_module()
                 output = _rust.GitDiffTool().execute(repo_path)
                 return ToolResult(
                     tool_name="git_diff",
@@ -218,6 +222,8 @@ class GitDiffTool(BaseTool):
                     success=True,
                     metadata={"returncode": 0},
                 )
+            except ImportError as exc:
+                logger.debug("Rust git_diff fallback to CLI: %s", exc)
             except Exception as exc:
                 return ToolResult(
                     tool_name="git_diff",
@@ -371,8 +377,8 @@ class GitLogTool(BaseTool):
         count = params.get("count", 10)
         oneline = params.get("oneline", True)
 
-        _rust = get_rust_module()
         try:
+            _rust = get_rust_module()
             output = _rust.GitLogTool().execute(repo_path, count)
             return ToolResult(
                 tool_name="git_log",

tests/cli/test_daemon_cmd.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import subprocess
 from pathlib import Path
 from unittest.mock import MagicMock, patch
 
@@ -79,3 +80,80 @@ class TestDaemonCommands:
             result = CliRunner().invoke(cli, ["start"])
         assert result.exit_code != 0
         assert "already running" in result.output
+
+
+class TestDaemonDetachment:
+    """The spawned server must outlive the console that started it.
+
+    ``start_new_session`` is POSIX-only — CPython's Windows ``_execute_child``
+    names the parameter ``unused_start_new_session``. Relying on it there leaves
+    the server sharing its parent's console, so closing that console (or logging
+    off) delivers CTRL_CLOSE_EVENT and kills the daemon.
+    """
+
+    @staticmethod
+    def _spawn_kwargs(platform: str) -> dict:
+        """Return the kwargs ``start`` passes to Popen when spawning the server.
+
+        ``load_config`` is stubbed because it shells out for GPU detection —
+        patching Popen wholesale would otherwise break config loading before
+        the spawn is reached.
+        """
+        with (
+            patch("openjarvis.cli.daemon_cmd._read_pid", return_value=None),
+            patch("openjarvis.cli.daemon_cmd._write_pid"),
+            patch("openjarvis.cli.daemon_cmd.load_config"),
+            patch("openjarvis.cli.daemon_cmd.sys.platform", platform),
+            patch("openjarvis.cli.daemon_cmd.subprocess.Popen") as popen,
+            patch("builtins.open", MagicMock()),
+        ):
+            popen.return_value = MagicMock(pid=4321)
+            result = CliRunner().invoke(cli, ["start"])
+            assert result.exit_code == 0, result.output
+            spawns = [
+                c for c in popen.call_args_list if c.args and "serve" in c.args[0]
+            ]
+            assert spawns, f"start did not spawn the server: {popen.call_args_list}"
+            return spawns[-1].kwargs
+
+    def test_windows_spawn_is_detached_from_the_console(self) -> None:
+        # These constants are only exported by ``subprocess`` on Windows.
+        # Supply their documented values so the simulated Windows branch is
+        # still exercised by the POSIX test job.
+        detached_process = getattr(subprocess, "DETACHED_PROCESS", 0x00000008)
+        create_new_process_group = getattr(
+            subprocess, "CREATE_NEW_PROCESS_GROUP", 0x00000200
+        )
+        with (
+            patch.object(
+                subprocess,
+                "DETACHED_PROCESS",
+                detached_process,
+                create=True,
+            ),
+            patch.object(
+                subprocess,
+                "CREATE_NEW_PROCESS_GROUP",
+                create_new_process_group,
+                create=True,
+            ),
+        ):
+            kwargs = self._spawn_kwargs("win32")
+
+        flags = kwargs.get("creationflags", 0)
+        assert flags & detached_process, (
+            "server must be spawned with DETACHED_PROCESS on Windows, otherwise "
+            "closing the launching console kills it"
+        )
+        assert flags & create_new_process_group, (
+            "server must be in its own process group so Ctrl-C in the parent "
+            "console does not propagate to it"
+        )
+        assert not kwargs.get("start_new_session"), (
+            "start_new_session is ignored on Windows; it must not be relied on"
+        )
+
+    def test_posix_spawn_still_uses_start_new_session(self) -> None:
+        kwargs = self._spawn_kwargs("linux")
+        assert kwargs.get("start_new_session") is True
+        assert "creationflags" not in kwargs or kwargs["creationflags"] == 0

tests/cli/test_serve_single_build.py

@@ -159,6 +159,8 @@ def test_serve_does_not_call_systembuilder_build(tmp_path, monkeypatch):
         )
     )
     set_system_spy = MagicMock()
+    inject_spy = MagicMock()
+    monkeypatch.setattr(serve_mod, "inject_credentials", inject_spy)
 
     result = _run_serve(
         tmp_path,
@@ -169,6 +171,7 @@ def test_serve_does_not_call_systembuilder_build(tmp_path, monkeypatch):
 
     assert result.exit_code == 0, result.output
     build_spy.assert_not_called()
+    inject_spy.assert_called_once_with()
 
 
 def test_executor_receives_required_system_attrs(tmp_path, monkeypatch):

tests/conftest.py

@@ -30,6 +30,19 @@ from openjarvis.core.registry import (
 )
 
 
+@pytest.fixture(autouse=True)
+def _no_update_check(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Never let the CLI's PyPI update-check nag run during tests.
+
+    ``check_for_updates`` writes its banner to stderr, which ``CliRunner``
+    merges into ``result.output`` — polluting JSON/CSV output of any test
+    that invokes a CLI command. It already self-disables when ``CI`` is
+    set, but that only helps in CI; locally (e.g. a dev with a stale
+    version-check cache and network access) it fires for real.
+    """
+    monkeypatch.setenv("OPENJARVIS_NO_UPDATE_CHECK", "1")
+
+
 @pytest.fixture(autouse=True)
 def _clean_registries() -> None:
     """Ensure each test starts with empty registries and a fresh event bus."""

tests/core/test_credentials.py

@@ -5,7 +5,9 @@ import os
 import pytest
 
 from openjarvis.core.credentials import (
+    delete_credential,
     get_credential_status,
+    inject_credentials,
     load_credentials,
     save_credential,
 )
@@ -54,3 +56,26 @@ def test_file_permissions(cred_path):
     save_credential("web_search", "TAVILY_API_KEY", "tvly-x", path=cred_path)
     mode = oct(cred_path.stat().st_mode & 0o777)
     assert mode == "0o600"
+
+
+def test_inject_credentials_restores_saved_value(cred_path, monkeypatch):
+    save_credential("web_search", "TAVILY_API_KEY", "tvly-persisted", path=cred_path)
+    monkeypatch.delenv("TAVILY_API_KEY")
+
+    inject_credentials(path=cred_path)
+
+    assert os.environ["TAVILY_API_KEY"] == "tvly-persisted"
+
+
+def test_delete_credential_removes_file_value_and_env(cred_path, monkeypatch):
+    save_credential("web_search", "TAVILY_API_KEY", "tvly-delete", path=cred_path)
+
+    delete_credential("web_search", "TAVILY_API_KEY", path=cred_path)
+
+    assert load_credentials(path=cred_path) == {}
+    assert "TAVILY_API_KEY" not in os.environ
+
+
+def test_delete_rejects_unknown_key(cred_path):
+    with pytest.raises(ValueError, match="Unknown credential key"):
+        delete_credential("web_search", "BOGUS_KEY", path=cred_path)

tests/deployment/test_packaging.py

@@ -18,6 +18,7 @@ ROOT = Path(__file__).resolve().parent.parent.parent
 PYPROJECT = ROOT / "pyproject.toml"
 DESKTOP_LIB_RS = ROOT / "frontend" / "src-tauri" / "src" / "lib.rs"
 WINDOWS_INSTALL_PS1 = ROOT / "deploy" / "windows" / "install.ps1"
+QUICKSTART_SH = ROOT / "scripts" / "quickstart.sh"
 
 
 def _pyproject() -> dict:
@@ -59,3 +60,9 @@ def test_windows_installer_syncs_the_native_group() -> None:
         "the Windows installer must include `--group desktop-native` so "
         "openjarvis_rust is built during source install."
     )
+
+
+def test_quickstart_installs_web_search_dependencies() -> None:
+    quickstart = QUICKSTART_SH.read_text()
+    assert "--extra tools-search" in quickstart
+    assert "already running on port 8000" in quickstart

tests/server/test_api_routes.py

@@ -37,16 +37,22 @@ class TestAgentRoutes:
 
 
 class TestMemoryRoutes:
+    # 503 is the documented response when the native ``openjarvis_rust``
+    # extension is absent from the venv (see TestMemoryRustMissing below).
+    # These tests are only asserting "the route is wired up", so a backend
+    # that cannot be built is tolerated the same way a 500 is.
+    _BACKEND_OPTIONAL = (200, 500, 503)
+
     def test_search(self):
         client = TestClient(_make_app())
         resp = client.post("/v1/memory/search", json={"query": "test"})
         # May fail if SQLite not set up, that's ok
-        assert resp.status_code in (200, 500)
+        assert resp.status_code in self._BACKEND_OPTIONAL
 
     def test_stats(self):
         client = TestClient(_make_app())
         resp = client.get("/v1/memory/stats")
-        assert resp.status_code in (200, 500)
+        assert resp.status_code in self._BACKEND_OPTIONAL
 
 
 class TestMemoryRustMissing:

tests/server/test_session_store.py

@@ -115,3 +115,29 @@ class TestLastActiveChannel:
 
     def test_returns_none_for_unknown_user(self, store):
         assert store.get_last_active_channel("nobody") is None
+
+
+class TestInMemoryDatabase:
+    """``:memory:`` is a SQLite sentinel, not a path — it must not be created.
+
+    ``secure_create`` treats it as a filename, which fails outright on Windows
+    (``:`` is illegal there) and litters the working directory elsewhere.
+    """
+
+    def test_in_memory_store_is_usable(self):
+        s = SessionStore(db_path=":memory:")
+        try:
+            session = s.get_or_create("user1", "twilio")
+            assert session["sender_id"] == "user1"
+        finally:
+            s.close()
+
+    def test_in_memory_store_creates_no_file(self, tmp_path, monkeypatch):
+        monkeypatch.chdir(tmp_path)
+        before = set(tmp_path.iterdir())
+        s = SessionStore(db_path=":memory:")
+        try:
+            assert set(tmp_path.iterdir()) == before
+            assert not (tmp_path / ":memory:").exists()
+        finally:
+            s.close()

tests/server/test_tools_endpoint.py

@@ -1,5 +1,7 @@
 """Tests for GET /v1/tools endpoint."""
 
+from unittest.mock import MagicMock
+
 import pytest
 
 try:
@@ -45,3 +47,49 @@ def test_browser_meta_group():
     names = {t["name"] for t in tools}
     assert "browser" in names
     assert "browser_navigate" not in names
+
+
+def test_web_search_available_without_tavily_key(monkeypatch):
+    """DuckDuckGo fallback keeps web search usable without Tavily."""
+    from openjarvis.server.agent_manager_routes import build_tools_list
+
+    monkeypatch.delenv("TAVILY_API_KEY", raising=False)
+    tools = build_tools_list()
+    web_search = next(t for t in tools if t["name"] == "web_search")
+
+    assert web_search["configured"] is True
+    assert web_search["requires_credentials"] is False
+
+
+def test_tool_credentials_browser_lifecycle(tmp_path, monkeypatch):
+    """The browser API can save, report, and remove a Tavily key."""
+    from fastapi import FastAPI
+    from fastapi.testclient import TestClient
+
+    from openjarvis.server.agent_manager_routes import create_agent_manager_router
+
+    monkeypatch.setenv("OPENJARVIS_HOME", str(tmp_path / "openjarvis-home"))
+    monkeypatch.delenv("TAVILY_API_KEY", raising=False)
+    app = FastAPI()
+    tools_router = create_agent_manager_router(MagicMock())[3]
+    app.include_router(tools_router)
+    client = TestClient(app)
+
+    saved = client.post(
+        "/v1/tools/web_search/credentials",
+        json={"TAVILY_API_KEY": "tvly-browser-test"},
+    )
+    assert saved.status_code == 200
+    assert saved.json() == {"saved": ["TAVILY_API_KEY"]}
+    assert client.get("/v1/tools/web_search/credentials/status").json() == {
+        "TAVILY_API_KEY": True
+    }
+
+    deleted = client.delete(
+        "/v1/tools/web_search/credentials/TAVILY_API_KEY",
+    )
+    assert deleted.status_code == 200
+    assert deleted.json() == {"deleted": "TAVILY_API_KEY"}
+    assert client.get("/v1/tools/web_search/credentials/status").json() == {
+        "TAVILY_API_KEY": False
+    }

tests/skills/test_importer.py

@@ -202,3 +202,71 @@ class TestImportSkill:
 
         installed = target_root / "hermes" / "my-skill" / "SKILL.md"
         assert "Original" in installed.read_text()
+
+
+class TestDangerousCapabilityGate:
+    def _make_importer(self, tmp_path: Path) -> SkillImporter:
+        return SkillImporter(
+            parser=SkillParser(),
+            tool_translator=ToolTranslator(),
+            target_root=tmp_path / "skills",
+        )
+
+    def _make_resolved_with_caps(
+        self, tmp_path: Path, caps: list[str]
+    ) -> ResolvedSkill:
+        src_dir = tmp_path / "source" / "my-skill"
+        src_dir.mkdir(parents=True)
+        caps_yaml = "".join(f"  - {c}\n" for c in caps)
+        (src_dir / "SKILL.md").write_text(
+            "---\n"
+            "name: my-skill\n"
+            "description: A test skill\n"
+            f"required_capabilities:\n{caps_yaml}"
+            "---\n"
+            "Body"
+        )
+        return ResolvedSkill(
+            name="my-skill",
+            source="hermes",
+            path=src_dir,
+            category="testing",
+            description="A test skill",
+            commit="abc123",
+        )
+
+    def test_refuses_unreviewed_dangerous_skill(self, tmp_path: Path):
+        importer = self._make_importer(tmp_path)
+        resolved = self._make_resolved_with_caps(tmp_path, ["shell:execute"])
+        result = importer.import_skill(resolved)
+
+        assert not result.success
+        assert result.requires_confirmation
+        assert result.dangerous_capabilities == ["shell:execute"]
+        assert any("dangerous" in w.lower() for w in result.warnings)
+        # Nothing may be written to disk on refusal
+        assert not (tmp_path / "skills" / "hermes" / "my-skill").exists()
+
+    def test_confirm_dangerous_installs_and_records_tier(self, tmp_path: Path):
+        importer = self._make_importer(tmp_path)
+        resolved = self._make_resolved_with_caps(tmp_path, ["shell:execute"])
+        result = importer.import_skill(resolved, confirm_dangerous=True)
+
+        assert result.success
+        assert result.requires_confirmation
+        assert any("confirmed by caller" in w for w in result.warnings)
+        content = (tmp_path / "skills" / "hermes" / "my-skill" / ".source").read_text()
+        assert 'trust_tier = "unreviewed"' in content
+        assert 'dangerous_capabilities = ["shell:execute"]' in content
+
+    def test_benign_capabilities_need_no_confirmation(self, tmp_path: Path):
+        importer = self._make_importer(tmp_path)
+        resolved = self._make_resolved_with_caps(tmp_path, ["network:fetch"])
+        result = importer.import_skill(resolved)
+
+        assert result.success
+        assert not result.requires_confirmation
+        assert result.dangerous_capabilities == []
+        content = (tmp_path / "skills" / "hermes" / "my-skill" / ".source").read_text()
+        assert 'trust_tier = "unreviewed"' in content
+        assert "dangerous_capabilities = []" in content

tests/skills/test_skills.py

@@ -148,6 +148,56 @@ class TestSkillExecutor:
         assert EventType.SKILL_EXECUTE_END in event_types
 
 
+class TestSkillExecutorCapabilities:
+    def _manifest(self):
+        return SkillManifest(
+            name="capskill",
+            required_capabilities=["network:fetch"],
+            steps=[
+                SkillStep(
+                    tool_name="echo",
+                    arguments_template='{"text": "hello"}',
+                    output_key="result",
+                )
+            ],
+        )
+
+    def test_no_policy_runs_capability_skills(self):
+        """Default construction (no allowed_capabilities) must not enforce —
+        this is the pre-enforcement behavior every manager.py call site relies on."""
+        executor = SkillExecutor(ToolExecutor([EchoTool()]))
+        result = executor.run(self._manifest())
+        assert result.success
+        assert result.context.get("result") == "hello"
+
+    def test_policy_blocks_missing_capability(self):
+        executor = SkillExecutor(ToolExecutor([EchoTool()]), allowed_capabilities=set())
+        result = executor.run(self._manifest())
+        assert not result.success
+        assert len(result.step_results) == 1
+        assert "Blocked" in result.step_results[0].content
+        assert "network:fetch" in result.step_results[0].content
+
+    def test_policy_allows_granted_capability(self):
+        executor = SkillExecutor(
+            ToolExecutor([EchoTool()]),
+            allowed_capabilities={"network:fetch"},
+        )
+        result = executor.run(self._manifest())
+        assert result.success
+        assert result.context.get("result") == "hello"
+
+    def test_blocked_run_publishes_events(self):
+        bus = EventBus(record_history=True)
+        executor = SkillExecutor(
+            ToolExecutor([EchoTool()]), bus=bus, allowed_capabilities=set()
+        )
+        executor.run(self._manifest())
+        event_types = {e.event_type for e in bus.history}
+        assert EventType.SKILL_EXECUTE_START in event_types
+        assert EventType.SKILL_EXECUTE_END in event_types
+
+
 class TestSkillStepExtended:
     def test_step_with_skill_name(self):
         step = SkillStep(skill_name="summarize", output_key="result")

tests/telemetry/test_derived_metrics.py

@@ -164,6 +164,7 @@ class TestDerivedMetricsInStore:
             throughput_per_watt=0.5,
         )
         store.record(rec)
+        store.flush()
 
         agg = TelemetryAggregator(tmp_path / "test.db")
         stats = agg.per_model_stats()
@@ -185,6 +186,8 @@ class TestDerivedMetricsInStore:
                     throughput_per_watt=1.0 * (i + 1),
                 )
             )
+        store.flush()
+
         agg = TelemetryAggregator(tmp_path / "test.db")
         summary = agg.summary()
         assert summary.avg_energy_per_output_token_joules > 0

tests/telemetry/test_itl_metrics.py

@@ -345,6 +345,7 @@ class TestItlStorage:
             )
         )
 
+        store.flush()
         agg = TelemetryAggregator(tmp_path / "test.db")
         stats = agg.per_model_stats()
         assert len(stats) == 1

tests/telemetry/test_phase_energy.py

@@ -222,6 +222,7 @@ class TestPhaseEnergyStorage:
             )
         )
 
+        store.flush()
         agg = TelemetryAggregator(tmp_path / "test.db")
         stats = agg.per_model_stats()
         assert len(stats) == 1

tests/telemetry/test_store.py

@@ -178,3 +178,60 @@ class TestTelemetryRecordFields:
     def test_tokens_per_joule_set(self):
         rec = TelemetryRecord(timestamp=1.0, model_id="test", tokens_per_joule=80.0)
         assert rec.tokens_per_joule == 80.0
+
+    def test_batching_delays_commit(self, tmp_path: Path) -> None:
+        db_path = tmp_path / "test.db"
+        # flush_interval_seconds=0 disables the background flusher so the
+        # buffered/flushed states below are deterministic, not a race.
+        store = TelemetryStore(db_path, batch_size=2, flush_interval_seconds=0)
+        rec = TelemetryRecord(timestamp=time.time(), model_id="m1", engine="e1")
+
+        store.record(rec)
+        # Should not be in DB yet because batch_size is 2 and we haven't flushed
+        # Need a separate connection to check because store._fetchall() calls flush()!
+        assert _count_rows_via_own_connection(db_path) == 0
+
+        # Hit batch size
+        store.record(rec)
+        assert _count_rows_via_own_connection(db_path) == 2
+        store.close()
+
+    def test_background_flush_makes_records_visible(self, tmp_path: Path) -> None:
+        # A partial batch must become visible to OTHER connections within the
+        # flush interval even if no further write ever arrives — the background
+        # flusher covers the "traffic stopped mid-batch" case that per-record
+        # stale checks cannot.
+        db_path = tmp_path / "test.db"
+        store = TelemetryStore(db_path, batch_size=50, flush_interval_seconds=0.05)
+        rec = TelemetryRecord(timestamp=time.time(), model_id="m1", engine="e1")
+        store.record(rec)
+
+        deadline = time.time() + 5.0
+        rows = 0
+        while time.time() < deadline:
+            rows = _count_rows_via_own_connection(db_path)
+            if rows:
+                break
+            time.sleep(0.02)
+        assert rows == 1
+        store.close()
+
+    def test_close_flushes_and_is_idempotent(self, tmp_path: Path) -> None:
+        db_path = tmp_path / "test.db"
+        store = TelemetryStore(db_path, batch_size=50, flush_interval_seconds=0)
+        rec = TelemetryRecord(timestamp=time.time(), model_id="m1", engine="e1")
+        store.record(rec)
+        store.close()
+        assert _count_rows_via_own_connection(db_path) == 1
+        store.close()  # second close must be a no-op, not a ProgrammingError
+
+
+def _count_rows_via_own_connection(db_path: Path) -> int:
+    """Count telemetry rows through a separate connection (like the aggregator)."""
+    import contextlib
+    import sqlite3
+
+    # NB: sqlite3's ``with conn`` is a TRANSACTION context (it does not close);
+    # ``contextlib.closing`` actually closes the connection.
+    with contextlib.closing(sqlite3.connect(db_path)) as conn:
+        return len(conn.execute("SELECT * FROM telemetry").fetchall())

tests/tools/storage/test_dense.py

@@ -14,9 +14,9 @@ skipped if the server is unreachable.
 from __future__ import annotations
 
 import os
-import socket
 from pathlib import Path
 
+import httpx
 import pytest
 
 from openjarvis.tools.storage.dense import (
@@ -25,20 +25,58 @@ from openjarvis.tools.storage.dense import (
     chunk_markdown,
     dedupe_chunks,
 )
+from openjarvis.tools.storage.embeddings import OllamaEmbedder
 
 _FIXTURE_DIR = Path(__file__).resolve().parents[2] / "fixtures" / "docs"
-_OLLAMA_HOST = os.environ.get("OLLAMA_HOST", "localhost")
-_OLLAMA_PORT = int(os.environ.get("OLLAMA_PORT", "11434"))
+_EMBED_MODEL = "nomic-embed-text"
+
+
+def _ollama_base_url() -> str:
+    """Resolve the Ollama base URL from ``OLLAMA_HOST``.
+
+    The project documents ``OLLAMA_HOST`` as a full URL
+    (``http://<remote-ip>:11434``), but Ollama's own convention also
+    allows bare ``host`` / ``host:port`` forms — accept all three so
+    the probe and the embedder under test agree on one endpoint.
+    """
+    host = os.environ.get("OLLAMA_HOST", "")
+    if not host:
+        return "http://localhost:11434"
+    if host.startswith(("http://", "https://")):
+        return host.rstrip("/")
+    if ":" in host:
+        return f"http://{host}"
+    return f"http://{host}:11434"
 
 
 def _ollama_up() -> bool:
+    """True only if Ollama is reachable *and* the embed model is pulled.
+
+    A bare TCP connect isn't enough — a machine can run Ollama for chat
+    models without ever having pulled ``nomic-embed-text``, which makes
+    ``/api/embed`` 404 instead of the tests skipping as intended.
+    """
     try:
-        with socket.create_connection((_OLLAMA_HOST, _OLLAMA_PORT), timeout=1.0):
-            return True
-    except OSError:
+        resp = httpx.get(f"{_ollama_base_url()}/api/tags", timeout=1.0)
+        resp.raise_for_status()
+        models = {m.get("model", "") for m in resp.json().get("models", [])}
+        return any(m.startswith(_EMBED_MODEL) for m in models)
+    except (httpx.HTTPError, ValueError):
         return False
 
 
+def _make_backend() -> DenseMemory:
+    """DenseMemory wired to the same Ollama endpoint the probe checked.
+
+    ``DenseMemory()`` alone would build an ``OllamaEmbedder`` with its
+    hard-coded localhost default, so a remote ``OLLAMA_HOST`` could pass
+    the probe and then have every test call the wrong server.
+    """
+    return DenseMemory(
+        embedder=OllamaEmbedder(model=_EMBED_MODEL, base_url=_ollama_base_url())
+    )
+
+
 ollama_required = pytest.mark.skipif(
     not _ollama_up(),
     reason=(
@@ -48,6 +86,65 @@ ollama_required = pytest.mark.skipif(
 )
 
 
+# ---------------------------------------------------------------------------
+# Skip-guard unit tests (no Ollama required)
+# ---------------------------------------------------------------------------
+
+
+class _FakeTagsResponse:
+    def __init__(self, payload: dict) -> None:
+        self._payload = payload
+
+    def raise_for_status(self) -> None:
+        pass
+
+    def json(self) -> dict:
+        return self._payload
+
+
+class TestOllamaProbe:
+    def test_base_url_accepts_all_documented_forms(self, monkeypatch):
+        cases = {
+            "http://remote:11434": "http://remote:11434",
+            "http://remote:11434/": "http://remote:11434",
+            "https://ollama.internal": "https://ollama.internal",
+            "remote:8080": "http://remote:8080",
+            "remote": "http://remote:11434",
+        }
+        for raw, expected in cases.items():
+            monkeypatch.setenv("OLLAMA_HOST", raw)
+            assert _ollama_base_url() == expected, raw
+        monkeypatch.delenv("OLLAMA_HOST", raising=False)
+        assert _ollama_base_url() == "http://localhost:11434"
+
+    def test_probe_false_when_server_down(self, monkeypatch):
+        def _refuse(url, timeout):
+            raise httpx.ConnectError("connection refused")
+
+        monkeypatch.setattr(httpx, "get", _refuse)
+        assert _ollama_up() is False
+
+    def test_probe_false_when_model_missing(self, monkeypatch):
+        monkeypatch.setattr(
+            httpx,
+            "get",
+            lambda url, timeout: _FakeTagsResponse({"models": [{"model": "llama3"}]}),
+        )
+        assert _ollama_up() is False
+
+    def test_probe_true_with_tagged_model_on_configured_url(self, monkeypatch):
+        monkeypatch.setenv("OLLAMA_HOST", "http://remote:9999")
+        seen = {}
+
+        def _get(url, timeout):
+            seen["url"] = url
+            return _FakeTagsResponse({"models": [{"model": "nomic-embed-text:latest"}]})
+
+        monkeypatch.setattr(httpx, "get", _get)
+        assert _ollama_up() is True
+        assert seen["url"] == "http://remote:9999/api/tags"
+
+
 # ---------------------------------------------------------------------------
 # Chunking unit tests (no Ollama required)
 # ---------------------------------------------------------------------------
@@ -283,7 +380,7 @@ def indexed_backend():
     if not _ollama_up():
         pytest.skip("Ollama not reachable")
 
-    backend = DenseMemory()
+    backend = _make_backend()
     md_files = sorted(_FIXTURE_DIR.glob("*.md"))
     assert md_files, f"no fixtures at {_FIXTURE_DIR}"
 
@@ -452,7 +549,7 @@ def test_score_distribution_vs_threshold(indexed_backend, capsys):
 class TestDenseMemoryAPI:
     @ollama_required
     def test_store_and_delete(self):
-        backend = DenseMemory()
+        backend = _make_backend()
         doc_id = backend.store("the cat sat on the mat", source="a.txt")
         assert backend.count() == 1
         hits = backend.retrieve("where is the cat", top_k=1)
@@ -463,12 +560,12 @@ class TestDenseMemoryAPI:
 
     @ollama_required
     def test_empty_retrieve(self):
-        backend = DenseMemory()
+        backend = _make_backend()
         assert backend.retrieve("anything", top_k=3) == []
 
     @ollama_required
     def test_clear(self):
-        backend = DenseMemory()
+        backend = _make_backend()
         backend.store("foo")
         backend.store("bar")
         backend.clear()

tests/tools/test_git_tool.py

@@ -537,3 +537,51 @@ class TestGitLogTool:
         fn = tool.to_openai_function()
         assert fn["type"] == "function"
         assert fn["function"]["name"] == "git_log"
+
+
+# ---------------------------------------------------------------------------
+# CLI fallback when the Rust extension is missing
+# ---------------------------------------------------------------------------
+
+
+class TestCliFallbackWhenRustMissing:
+    """When ``get_rust_module`` raises ImportError (extension not built,
+    e.g. a plain pip install), the read-only git tools must fall back to
+    the git CLI instead of letting the ImportError escape ``execute()``."""
+
+    def _patch_no_rust(self):
+        return patch(
+            "openjarvis.tools.git_tool.get_rust_module",
+            side_effect=ImportError("No module named 'openjarvis_rust'"),
+        )
+
+    def test_git_status_falls_back_to_cli(self, tmp_path):
+        _init_repo(tmp_path)
+        (tmp_path / "new_file.txt").write_text("hello")
+        with self._patch_no_rust():
+            result = GitStatusTool().execute(repo_path=str(tmp_path))
+        assert result.success is True
+        assert "new_file.txt" in result.content
+
+    def test_git_diff_falls_back_to_cli(self, tmp_path):
+        _init_repo(tmp_path)
+        (tmp_path / "README.md").write_text("# Modified\n")
+        with self._patch_no_rust():
+            result = GitDiffTool().execute(repo_path=str(tmp_path))
+        assert result.success is True
+        assert "README.md" in result.content
+
+    def test_git_log_falls_back_to_cli(self, tmp_path):
+        _init_repo(tmp_path)
+        with self._patch_no_rust():
+            result = GitLogTool().execute(repo_path=str(tmp_path))
+        assert result.success is True
+        assert "Initial commit" in result.content
+
+    def test_fallback_failure_is_a_tool_result_not_an_exception(self, tmp_path):
+        # Even when the fallback itself fails (not a git repo), the tool
+        # must return a failed ToolResult rather than raising.
+        with self._patch_no_rust():
+            result = GitStatusTool().execute(repo_path=str(tmp_path))
+        assert result.success is False
+        assert "not a git repository" in result.content

tests/tools/test_knowledge_sql.py

@@ -70,9 +70,7 @@ def test_allows_select_with_keyword_substring(store: KnowledgeStore) -> None:
     from openjarvis.tools.knowledge_sql import KnowledgeSQLTool
 
     tool = KnowledgeSQLTool(store=store)
-    result = tool.execute(
-        query="SELECT author AS created_author FROM knowledge_chunks"
-    )
+    result = tool.execute(query="SELECT author AS created_author FROM knowledge_chunks")
     assert result.success, result.content
     assert "Alice" in result.content