feat: make each plugin an npm workspace built by a termix-plugin CLI

plugins/ai/CHANGELOG.md

@@ -0,0 +1,5 @@
+# Changelog
+
+## 1.0.0
+
+Bundled with Termix 2.9.0.

plugins/ai/README.md

@@ -0,0 +1,22 @@
+# AI Assistant
+
+An AI assistant that can read your hosts, snippets, fleets, alerts and automations, and propose changes for you to approve.
+
+Bundled with Termix and enabled by default.
+
+## Layout
+
+    manifest.json        id, capabilities and what this plugin contributes
+    src/backend/         activate(ctx) / deactivate(), plus the routes
+    src/frontend/        the UI it registers
+    locales/en.json      its strings (English only; the rest are translated)
+    tests/backend/       vitest, node
+    tests/frontend/      vitest, jsdom
+
+## Commands
+
+    npm run build     bundle into dist/
+    npm run test      run this plugin's tests
+    npm run typecheck
+
+See packages/plugin-sdk/ARCHITECTURE.md for the contract this follows.

plugins/ai/backend/index.mjs

@@ -1,87 +0,0 @@
-/**
- * AI Assistant - first-party, in-process plugin.
- *
- * This plugin runs on the main thread rather than in a worker for the same
- * reason ssh-terminal, docker and host-metrics do (see
- * src/backend/plugins/first-party.ts): it writes directly across hosts,
- * snippets, fleets and automations through their repositories,
- * runs approved commands over the shared SSH pool via resolveHostById and
- * withConnection, and streams its chat replies as server-sent events. None
- * of that fits the worker ctx -- ctx.hosts is a 13-field read-only view with
- * no write methods and no snippets/fleets/automations surface at all,
- * and ctx.http.route replies with a single postMessage value rather than a
- * stream.
- *
- * Unlike ssh-terminal/docker/host-metrics, this plugin does not own a live
- * transport (no WebSocket server, no long-lived connection that must outlive
- * the main server's route table), so it does not get its own port. Instead
- * it registers its Express router with the always-mounted /ai dispatcher in
- * src/backend/database/routes/ai-dispatch.ts, the same "register a router,
- * a small dispatcher forwards to it" shape /plugin-api uses for worker
- * plugins, just scoped to this one path so /ai/* keeps its existing routes
- * instead of moving under /plugin-api/ai/.
- *
- * routes.ts is what used to be src/backend/ai/index.ts, moved here with its
- * sibling modules (context.ts, egress.ts, engine.ts, gating.ts, redaction.ts,
- * providers/, tools/) and rewritten from `export default router` into
- * startAiService()/stopAiService(), called from activate()/deactivate()
- * below.
- */
-
-import path from "node:path";
-import { fileURLToPath } from "node:url";
-
-/**
- * Loads a module by its path relative to this file's own directory.
- *
- * The plugin ships beside dist/backend, so in a built server the compiled
- * .js for the moved AI files (routes.js, ...) sits right next to this .mjs,
- * and core modules resolve via ../../../backend/backend/... from there
- * (dist/plugins/ai/backend -> dist/backend/backend). Under vitest/dev nothing
- * imports this plugin directly (the loader loads in-process plugins with a
- * plain dynamic import, no tsx loader involved), so the only layout that
- * needs to resolve here is the built one.
- */
-async function loadRelative(relativePath) {
-  const here = path.dirname(fileURLToPath(import.meta.url));
-  const target = path.join(here, `${relativePath}.js`);
-  return import(new URL(`file://${target.replace(/\\/g, "/")}`).href);
-}
-
-let routesModule = null;
-
-export async function activate(ctx) {
-  routesModule = await loadRelative("routes");
-  routesModule.startAiService();
-
-  // Offered by reference, not copied: this hands over a resolver, so the key
-  // stays ours. Rotating or deleting the provider takes effect on the next
-  // read by any borrower, and deactivate() below withdraws the offer entirely.
-  // The resolve is per-user because the keys are stored per-user and encrypted
-  // under that user's data key, so a borrowed read returns the acting user's
-  // own default key, never someone else's.
-  ctx.secrets.offer("api-key", async (userId) => {
-    const { createCurrentAiRepository } = await loadRelative(
-      "../../../backend/backend/database/repositories/factory",
-    );
-    const repository = createCurrentAiRepository();
-    const providers = await repository.listProviders(userId);
-    const preferred = providers.find((provider) => provider.apiKeyPrefix);
-    if (!preferred) return null;
-
-    const withSecret = await repository.findProviderWithSecret(
-      preferred.id,
-      userId,
-    );
-    return withSecret?.apiKey ?? null;
-  });
-
-  ctx.log.info("AI assistant router registered at /ai");
-}
-
-export async function deactivate() {
-  if (routesModule) {
-    routesModule.stopAiService();
-    routesModule = null;
-  }
-}

plugins/ai/locales/en.json

@@ -0,0 +1 @@
+{}

plugins/ai/manifest.json

@@ -69,8 +69,6 @@
       }
     }
   },
-  "backend": "backend/index.mjs",
-  "frontend": "frontend/index.mjs",
   "locales": "locales",
   "platforms": ["linux", "win32", "darwin"]
 }

plugins/ai/package.json

@@ -0,0 +1,15 @@
+{
+  "name": "@termix-plugin/ai",
+  "private": true,
+  "version": "1.0.0",
+  "type": "module",
+  "scripts": {
+    "build": "termix-plugin build",
+    "test": "termix-plugin test",
+    "typecheck": "tsc -p tsconfig.json",
+    "validate": "termix-plugin validate"
+  },
+  "devDependencies": {
+    "@termix/plugin-sdk": "*"
+  }
+}

plugins/ai/src/backend/egress.ts

@@ -1,6 +1,6 @@
 import { isIP } from "net";
-import { isBlockedAddress } from "../../../src/backend/utils/safe-outbound-fetch.js";
-import { createCurrentSettingsRepository } from "../../../src/backend/database/repositories/factory.js";
+import { isBlockedAddress } from "../../../../src/backend/utils/safe-outbound-fetch.js";
+import { createCurrentSettingsRepository } from "../../../../src/backend/database/repositories/factory.js";
 
 /**
  * Where the assistant is allowed to send requests.

plugins/ai/src/backend/engine.ts

@@ -1,4 +1,4 @@
-import { getErrorMessage } from "../../../src/backend/utils/error-message.js";
+import { getErrorMessage } from "../../../../src/backend/utils/error-message.js";
 import { getAdapter } from "./providers/registry.js";
 import type {
   ChatMessage,

plugins/ai/src/backend/gating.ts

@@ -1,9 +1,9 @@
 import type { NextFunction, Response } from "express";
-import type { AuthenticatedRequest } from "../../../src/types/index.js";
+import type { AuthenticatedRequest } from "../../../../src/types/index.js";
 import {
   createCurrentSettingsRepository,
   createCurrentUserPreferenceRepository,
-} from "../../../src/backend/database/repositories/factory.js";
+} from "../../../../src/backend/database/repositories/factory.js";
 
 /**
  * Three gates, all checked on the server.

plugins/ai/src/backend/index.ts

@@ -0,0 +1,32 @@
+import type { PluginContext } from "@termix/plugin-sdk/backend";
+import { startAiService, stopAiService } from "./routes.js";
+import { createCurrentAiRepository } from "../../../../src/backend/database/repositories/factory.js";
+
+export async function activate(ctx: PluginContext) {
+  startAiService();
+
+  // Offered by reference, not copied: this hands over a resolver, so the key
+  // stays ours. Rotating or deleting the provider takes effect on the next
+  // read by any borrower, and deactivate() below withdraws the offer entirely.
+  // The resolve is per-user because the keys are stored per-user and encrypted
+  // under that user's data key, so a borrowed read returns the acting user's
+  // own default key, never someone else's.
+  ctx.secrets.offer("api-key", async (userId: string) => {
+    const repository = createCurrentAiRepository();
+    const providers = await repository.listProviders(userId);
+    const preferred = providers.find((provider) => provider.apiKeyPrefix);
+    if (!preferred) return null;
+
+    const withSecret = await repository.findProviderWithSecret(
+      preferred.id,
+      userId,
+    );
+    return withSecret?.apiKey ?? null;
+  });
+
+  ctx.log.info("AI assistant router registered at /ai");
+}
+
+export async function deactivate() {
+  stopAiService();
+}

plugins/ai/src/backend/providers/http.ts

@@ -1,5 +1,5 @@
-import { fetchWithProxy } from "../../../../src/backend/utils/proxy-agent.js";
-import { safeOutboundFetch } from "../../../../src/backend/utils/safe-outbound-fetch.js";
+import { fetchWithProxy } from "../../../../../src/backend/utils/proxy-agent.js";
+import { safeOutboundFetch } from "../../../../../src/backend/utils/safe-outbound-fetch.js";
 import { evaluateEgress, readPrivateAllowlist } from "../egress.js";
 import { AiProviderError } from "./types.js";
 

plugins/ai/src/backend/routes.ts

@@ -1,25 +1,25 @@
-import { getErrorMessage } from "../../../src/backend/utils/error-message.js";
+import { getErrorMessage } from "../../../../src/backend/utils/error-message.js";
 import express from "express";
 import {
   registerAiRouter,
   unregisterAiRouter,
-} from "../../../src/backend/database/routes/ai-dispatch.js";
-import type { AuthenticatedRequest } from "../../../src/types/index.js";
-import { PermissionManager } from "../../../src/backend/utils/permission-manager.js";
-import { AuthManager } from "../../../src/backend/utils/auth-manager.js";
-import { databaseLogger } from "../../../src/backend/utils/logger.js";
+} from "../../../../src/backend/database/routes/ai-dispatch.js";
+import type { AuthenticatedRequest } from "../../../../src/types/index.js";
+import { PermissionManager } from "../../../../src/backend/utils/permission-manager.js";
+import { AuthManager } from "../../../../src/backend/utils/auth-manager.js";
+import { databaseLogger } from "../../../../src/backend/utils/logger.js";
 import {
   getAuditUsername,
   getRequestMeta,
   logAudit,
-} from "../../../src/backend/utils/audit-logger.js";
+} from "../../../../src/backend/utils/audit-logger.js";
 import {
   createCurrentAiRepository,
   createCurrentHostRepository,
   createCurrentSettingsRepository,
   createCurrentUserRepository,
-} from "../../../src/backend/database/repositories/factory.js";
-import type { UserRecord } from "../../../src/backend/database/repositories/user-repository.js";
+} from "../../../../src/backend/database/repositories/factory.js";
+import type { UserRecord } from "../../../../src/backend/database/repositories/user-repository.js";
 import { buildSystemPrompt } from "./context.js";
 import { AI_PRIVATE_ALLOWLIST_KEY, parseAllowlist } from "./egress.js";
 import { runAgent } from "./engine.js";

plugins/ai/src/backend/tools/executor.ts

@@ -3,15 +3,15 @@ import {
   createCurrentFleetRepository,
   createCurrentHostRepository,
   createCurrentSnippetRepository,
-} from "../../../../src/backend/database/repositories/factory.js";
-import { validateDefinition } from "../../../automations/backend/routes.js";
-import { resolveHostById } from "../../../../src/backend/hosts/host-resolver.js";
-import { execCommand } from "../../../../src/backend/hosts/metrics-shared/common-utils.js";
+} from "../../../../../src/backend/database/repositories/factory.js";
+import { validateDefinition } from "../../../../automations/src/backend/routes.js";
+import { resolveHostById } from "../../../../../src/backend/hosts/host-resolver.js";
+import { execCommand } from "../../../../../src/backend/hosts/metrics-shared/common-utils.js";
 import {
   createFleetSshFactory,
   getFleetPoolKey,
-} from "../../../../src/backend/hosts/ssh-client-factory.js";
-import { withConnection } from "../../../../src/backend/hosts/ssh-connection-pool.js";
+} from "../../../../../src/backend/hosts/ssh-client-factory.js";
+import { withConnection } from "../../../../../src/backend/hosts/ssh-connection-pool.js";
 import { getTool } from "./catalog.js";
 
 /** Approved commands get a bounded window rather than hanging the request. */

plugins/ai/src/backend/tools/read-tools.ts

@@ -8,7 +8,7 @@ import {
   createCurrentNetworkTopologyRepository,
   createCurrentSnippetRepository,
   createCurrentWorkspaceRepository,
-} from "../../../../src/backend/database/repositories/factory.js";
+} from "../../../../../src/backend/database/repositories/factory.js";
 import { num, objectSchema, type AiTool } from "./types.js";
 
 /**

plugins/ai/tests/backend/context.test.ts

@@ -0,0 +1,47 @@
+import { describe, expect, it } from "vitest";
+import { buildSystemPrompt } from "../../src/backend/context.js";
+
+const BASE = { hostCount: 3, allowReadOnlyCommands: false };
+
+describe("buildSystemPrompt", () => {
+  it("tells the assistant to stay inside what was asked", () => {
+    // Without these the assistant answered "what is running on this server"
+    // by also proposing a monitoring script and an automation nobody wanted.
+    const prompt = buildSystemPrompt(BASE);
+
+    expect(prompt).toContain("nothing beyond it");
+    expect(prompt).toMatch(/Read, then report\. Do not propose anything\./);
+    expect(prompt).toMatch(/no monitoring automations, no scripts/);
+    expect(prompt).toMatch(/One request means one proposal at most/);
+  });
+
+  it("states that it proposes rather than applies", () => {
+    const prompt = buildSystemPrompt(BASE);
+    expect(prompt).toContain("cannot change anything directly");
+    expect(prompt).toContain("Never claim you have done something");
+  });
+
+  it("never claims credential access", () => {
+    const prompt = buildSystemPrompt(BASE);
+    expect(prompt).toMatch(/no access to passwords, SSH keys, API keys/);
+  });
+
+  it("mentions read-only commands only when the user opted in", () => {
+    expect(buildSystemPrompt(BASE)).not.toMatch(/read-only diagnostic/);
+    expect(buildSystemPrompt({ ...BASE, allowReadOnlyCommands: true })).toMatch(
+      /read-only diagnostic/,
+    );
+  });
+
+  it("pluralises the host count", () => {
+    expect(buildSystemPrompt({ ...BASE, hostCount: 1 })).toContain("1 host ");
+    expect(buildSystemPrompt({ ...BASE, hostCount: 2 })).toContain("2 hosts");
+  });
+
+  it("includes the active tab only when there is one", () => {
+    expect(buildSystemPrompt(BASE)).not.toContain("currently looking at");
+    expect(buildSystemPrompt({ ...BASE, activeTab: "terminal" })).toContain(
+      "currently looking at: terminal",
+    );
+  });
+});

plugins/ai/tests/backend/egress.test.ts

@@ -0,0 +1,122 @@
+import { describe, expect, it } from "vitest";
+import {
+  DEFAULT_PRIVATE_ALLOWLIST,
+  evaluateEgress,
+  isPrivateDestination,
+  parseAllowlist,
+} from "../../src/backend/egress.js";
+
+/**
+ * The rule that lets a self-hosted Ollama work without turning the backend
+ * into an authenticated probe of its own network: private destinations are
+ * refused unless an admin named the host.
+ */
+describe("isPrivateDestination", () => {
+  it("recognises loopback and private ranges", () => {
+    for (const url of [
+      "http://localhost:11434",
+      "http://127.0.0.1:11434",
+      "http://10.0.0.5:11434",
+      "http://192.168.1.10:11434",
+      "http://172.16.4.4:11434",
+      "http://169.254.169.254/latest/meta-data",
+      "http://[::1]:11434",
+    ]) {
+      expect(isPrivateDestination(url), url).toBe(true);
+    }
+  });
+
+  it("treats public hosts as public", () => {
+    for (const url of [
+      "https://api.openai.com/v1",
+      "https://api.anthropic.com",
+      "https://generativelanguage.googleapis.com",
+      "http://8.8.8.8",
+    ]) {
+      expect(isPrivateDestination(url), url).toBe(false);
+    }
+  });
+});
+
+describe("evaluateEgress", () => {
+  it("allows public destinations without an allowlist entry", () => {
+    const decision = evaluateEgress("https://api.openai.com/v1", []);
+    expect(decision.allowed).toBe(true);
+    expect(decision.isPrivate).toBe(false);
+  });
+
+  it("refuses a private destination that is not allowlisted", () => {
+    const decision = evaluateEgress("http://192.168.1.50:11434", ["localhost"]);
+    expect(decision.allowed).toBe(false);
+    expect(decision.isPrivate).toBe(true);
+    expect(decision.reason).toContain("allowlist");
+  });
+
+  it("allows a private destination once its host is allowlisted", () => {
+    const decision = evaluateEgress("http://localhost:11434", [
+      "localhost",
+      "127.0.0.1",
+    ]);
+    expect(decision.allowed).toBe(true);
+    expect(decision.isPrivate).toBe(true);
+  });
+
+  it.each(["llm.internal", "host.docker.internal"])(
+    "routes allowlisted private DNS name %s through the private path",
+    (host) => {
+      expect(evaluateEgress(`http://${host}:8000/v1`, [host])).toEqual({
+        allowed: true,
+        isPrivate: true,
+      });
+    },
+  );
+
+  it("matches the allowlist case-insensitively", () => {
+    expect(
+      evaluateEgress("http://LOCALHOST:11434", ["localhost"]).allowed,
+    ).toBe(true);
+  });
+
+  it("does not let one allowlisted host authorise another", () => {
+    // A cloud metadata endpoint is the classic target, so an allowlist for
+    // localhost must not open 169.254.169.254.
+    const decision = evaluateEgress("http://169.254.169.254/latest/meta-data", [
+      "localhost",
+      "127.0.0.1",
+    ]);
+    expect(decision.allowed).toBe(false);
+  });
+
+  it("refuses non-http protocols and embedded credentials", () => {
+    expect(evaluateEgress("file:///etc/passwd", []).allowed).toBe(false);
+    expect(evaluateEgress("ftp://example.com", []).allowed).toBe(false);
+    expect(evaluateEgress("https://user:pass@api.openai.com", []).allowed).toBe(
+      false,
+    );
+  });
+
+  it("refuses a malformed url", () => {
+    expect(evaluateEgress("not a url", []).allowed).toBe(false);
+  });
+});
+
+describe("parseAllowlist", () => {
+  it("falls back to the defaults when unset or malformed", () => {
+    expect(parseAllowlist(null)).toEqual(DEFAULT_PRIVATE_ALLOWLIST);
+    expect(parseAllowlist("not json")).toEqual(DEFAULT_PRIVATE_ALLOWLIST);
+    expect(parseAllowlist('{"a":1}')).toEqual(DEFAULT_PRIVATE_ALLOWLIST);
+  });
+
+  it("normalises stored entries", () => {
+    expect(parseAllowlist('[" LocalHost ", "", "10.0.0.5"]')).toEqual([
+      "localhost",
+      "10.0.0.5",
+    ]);
+  });
+
+  it("honours a deliberately empty allowlist", () => {
+    // An admin clearing the list must actually block everything private,
+    // not silently get the defaults back.
+    expect(parseAllowlist("[]")).toEqual([]);
+  });
+});

plugins/ai/tests/backend/engine.test.ts

@@ -0,0 +1,208 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import type { ChatChunk } from "../../src/backend/providers/types.js";
+
+const streamChat = vi.fn();
+const handler = vi.fn();
+
+vi.mock("../../src/backend/providers/registry.js", () => ({
+  getAdapter: () => ({ streamChat, listModels: async () => [] }),
+}));
+
+vi.mock("../../src/backend/tools/catalog.js", () => ({
+  getTool: (name: string) =>
+    name === "list_hosts"
+      ? {
+          name: "list_hosts",
+          description: "List hosts",
+          category: "read",
+          parameters: { type: "object", properties: {} },
+          handler,
+        }
+      : undefined,
+  toolDefinitions: () => [
+    { name: "list_hosts", description: "List hosts", parameters: {} },
+  ],
+}));
+
+const { runAgent } = await import("../../src/backend/engine.js");
+
+function chunks(...values: ChatChunk[]) {
+  return (async function* () {
+    for (const value of values) yield value;
+  })();
+}
+
+const BASE = {
+  config: { providerType: "ollama" as const },
+  model: "test",
+  system: "system",
+  context: {
+    userId: "user-1",
+    conversationId: 1,
+    allowReadOnlyCommands: false,
+  },
+};
+
+async function collect(history: any[] = []) {
+  const events: any[] = [];
+  for await (const event of runAgent({ ...BASE, history })) {
+    events.push(event);
+  }
+  return events;
+}
+
+describe("runAgent", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+  });
+
+  it("streams text and finishes when no tools are called", async () => {
+    streamChat.mockReturnValueOnce(
+      chunks({ type: "text", text: "hello" }, { type: "done" }),
+    );
+
+    const events = await collect();
+
+    expect(events.filter((e) => e.type === "token")).toHaveLength(1);
+    expect(events.at(-1).type).toBe("done");
+  });
+
+  it("runs a known tool and feeds the result back", async () => {
+    handler.mockResolvedValue({ hosts: [{ id: 1, name: "web-1" }] });
+    streamChat
+      .mockReturnValueOnce(
+        chunks(
+          {
+            type: "tool_call",
+            call: { id: "c1", name: "list_hosts", arguments: {} },
+          },
+          { type: "done" },
+        ),
+      )
+      .mockReturnValueOnce(
+        chunks({ type: "text", text: "ok" }, { type: "done" }),
+      );
+
+    const events = await collect();
+
+    expect(handler).toHaveBeenCalledOnce();
+    expect(events.some((e) => e.type === "tool_result")).toBe(true);
+    expect(streamChat).toHaveBeenCalledTimes(2);
+  });
+
+  it("refuses a tool that is not in the catalog", async () => {
+    streamChat
+      .mockReturnValueOnce(
+        chunks(
+          {
+            type: "tool_call",
+            call: { id: "c1", name: "read_credentials", arguments: {} },
+          },
+          { type: "done" },
+        ),
+      )
+      .mockReturnValueOnce(
+        chunks({ type: "text", text: "ok" }, { type: "done" }),
+      );
+
+    const events = await collect();
+
+    // A model can emit any name it likes; only the catalog decides what runs.
+    expect(handler).not.toHaveBeenCalled();
+    const result = events.find((e) => e.type === "tool_result");
+    expect(JSON.stringify(result.result)).toContain("Unknown tool");
+  });
+
+  it("surfaces a handler failure without ending the run", async () => {
+    handler.mockRejectedValue(new Error("database is down"));
+    streamChat
+      .mockReturnValueOnce(
+        chunks(
+          {
+            type: "tool_call",
+            call: { id: "c1", name: "list_hosts", arguments: {} },
+          },
+          { type: "done" },
+        ),
+      )
+      .mockReturnValueOnce(
+        chunks({ type: "text", text: "ok" }, { type: "done" }),
+      );
+
+    const events = await collect();
+
+    const result = events.find((e) => e.type === "tool_result");
+    expect(JSON.stringify(result.result)).toContain("database is down");
+    expect(events.at(-1).type).toBe("done");
+  });
+
+  it("closes the tool call when a tool returns a proposal", async () => {
+    // Without a matching tool_result the call rendered as permanently
+    // running, even though the work was done and awaiting the user.
+    handler.mockResolvedValue({
+      __proposal: true,
+      kind: "propose_create_host",
+      summary: "Add host web-1",
+      payload: {},
+    });
+    streamChat
+      .mockReturnValueOnce(
+        chunks(
+          {
+            type: "tool_call",
+            call: { id: "c1", name: "list_hosts", arguments: {} },
+          },
+          { type: "done" },
+        ),
+      )
+      .mockReturnValueOnce(
+        chunks({ type: "text", text: "ok" }, { type: "done" }),
+      );
+
+    const events = await collect();
+
+    const callIndex = events.findIndex((e) => e.type === "tool_call");
+    const resultIndex = events.findIndex((e) => e.type === "tool_result");
+    const proposalIndex = events.findIndex((e) => e.type === "proposal");
+
+    expect(resultIndex).toBeGreaterThan(callIndex);
+    expect(proposalIndex).toBeGreaterThan(resultIndex);
+    expect(events[resultIndex]).toMatchObject({
+      name: "list_hosts",
+      result: { status: "awaiting_user_approval" },
+    });
+  });
+
+  it("reports a provider failure as an error and stops", async () => {
+    streamChat.mockImplementationOnce(() => {
+      throw new Error("provider unreachable");
+    });
+
+    const events = await collect();
+
+    expect(events).toHaveLength(1);
+    expect(events[0]).toMatchObject({
+      type: "error",
+      message: "provider unreachable",
+    });
+  });
+
+  it("stops after too many tool turns", async () => {
+    handler.mockResolvedValue({ ok: true });
+    streamChat.mockImplementation(() =>
+      chunks(
+        {
+          type: "tool_call",
+          call: { id: "c", name: "list_hosts", arguments: {} },
+        },
+        { type: "done" },
+      ),
+    );
+
+    const events = await collect();
+
+    // A model that never stops calling tools must not spin forever.
+    expect(events.at(-1)).toMatchObject({ type: "error" });
+    expect(streamChat.mock.calls.length).toBeLessThanOrEqual(8);
+  });
+});

plugins/ai/tests/backend/gating.test.ts

@@ -0,0 +1,94 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+const settingsRepository = { getBoolean: vi.fn() };
+const userPreferenceRepository = { findByUserId: vi.fn() };
+
+vi.mock("../../../../src/backend/database/repositories/factory.js", () => ({
+  createCurrentSettingsRepository: () => settingsRepository,
+  createCurrentUserPreferenceRepository: () => userPreferenceRepository,
+}));
+
+const { isAiGloballyEnabled, resolveAiAccess } =
+  await import("../../src/backend/gating.js");
+
+describe("AI gating", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+  });
+
+  it("defaults to off so upgrading an install enables nothing", async () => {
+    settingsRepository.getBoolean.mockResolvedValue(false);
+    await isAiGloballyEnabled();
+    expect(settingsRepository.getBoolean).toHaveBeenCalledWith(
+      "ai_globally_enabled",
+      false,
+    );
+  });
+
+  it("blocks everyone when the admin global is off", async () => {
+    settingsRepository.getBoolean.mockResolvedValue(false);
+    userPreferenceRepository.findByUserId.mockResolvedValue({
+      aiAssistantEnabled: true,
+      aiReadOnlyCommands: true,
+    });
+
+    const access = await resolveAiAccess("user-1");
+
+    expect(access.enabled).toBe(false);
+    expect(access.allowReadOnlyCommands).toBe(false);
+    // The kill switch short-circuits, so the preference is never consulted.
+    expect(userPreferenceRepository.findByUserId).not.toHaveBeenCalled();
+  });
+
+  it("blocks a user who has not enabled it", async () => {
+    settingsRepository.getBoolean.mockResolvedValue(true);
+    userPreferenceRepository.findByUserId.mockResolvedValue({
+      aiAssistantEnabled: false,
+    });
+
+    expect((await resolveAiAccess("user-1")).enabled).toBe(false);
+  });
+
+  it("treats never-asked as not enabled", async () => {
+    // Null means the user was never shown the choice, which is not consent.
+    settingsRepository.getBoolean.mockResolvedValue(true);
+    userPreferenceRepository.findByUserId.mockResolvedValue({
+      aiAssistantEnabled: null,
+    });
+
+    expect((await resolveAiAccess("user-1")).enabled).toBe(false);
+  });
+
+  it("treats a missing preference row as not enabled", async () => {
+    settingsRepository.getBoolean.mockResolvedValue(true);
+    userPreferenceRepository.findByUserId.mockResolvedValue(null);
+
+    expect((await resolveAiAccess("user-1")).enabled).toBe(false);
+  });
+
+  it("allows only when both gates are open", async () => {
+    settingsRepository.getBoolean.mockResolvedValue(true);
+    userPreferenceRepository.findByUserId.mockResolvedValue({
+      aiAssistantEnabled: true,
+      aiReadOnlyCommands: true,
+    });
+
+    const access = await resolveAiAccess("user-1");
+
+    expect(access.enabled).toBe(true);
+    expect(access.allowReadOnlyCommands).toBe(true);
+  });
+
+  it("keeps read-only commands off unless separately opted in", async () => {
+    settingsRepository.getBoolean.mockResolvedValue(true);
+    userPreferenceRepository.findByUserId.mockResolvedValue({
+      aiAssistantEnabled: true,
+      aiReadOnlyCommands: null,
+    });
+
+    const access = await resolveAiAccess("user-1");
+
+    expect(access.enabled).toBe(true);
+    expect(access.allowReadOnlyCommands).toBe(false);
+  });
+});

plugins/ai/tests/backend/providers/provider-http.test.ts

@@ -0,0 +1,49 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+const fetchWithProxy = vi.fn();
+const getFetchDispatcher = vi.fn();
+const safeOutboundFetch = vi.fn();
+const readPrivateAllowlist = vi.fn();
+const evaluateEgress = vi.fn();
+const globalFetch = vi.fn();
+
+vi.mock("../../../../../src/backend/utils/proxy-agent.js", () => ({
+  fetchWithProxy,
+  getFetchDispatcher,
+}));
+vi.mock("../../../../../src/backend/utils/safe-outbound-fetch.js", () => ({
+  safeOutboundFetch,
+}));
+vi.mock("../../../src/backend/egress.js", () => ({
+  readPrivateAllowlist,
+  evaluateEgress,
+}));
+
+const { providerFetch } =
+  await import("../../../src/backend/providers/http.js");
+
+describe("providerFetch", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    vi.stubGlobal("fetch", globalFetch);
+    readPrivateAllowlist.mockResolvedValue(["192.168.1.50"]);
+  });
+
+  it("uses the matching Undici fetch implementation for private providers", async () => {
+    const response = { ok: true, status: 200 } as Response;
+    const init = { method: "GET" };
+    evaluateEgress.mockReturnValue({ allowed: true, isPrivate: true });
+    fetchWithProxy.mockResolvedValue(response);
+
+    await expect(
+      providerFetch("http://192.168.1.50:11434/api/tags", init),
+    ).resolves.toBe(response);
+
+    expect(fetchWithProxy).toHaveBeenCalledWith(
+      "http://192.168.1.50:11434/api/tags",
+      init,
+    );
+    expect(globalFetch).not.toHaveBeenCalled();
+    expect(safeOutboundFetch).not.toHaveBeenCalled();
+  });
+});

plugins/ai/tests/backend/providers/providers.test.ts

@@ -0,0 +1,308 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import type { ChatChunk } from "../../../src/backend/providers/types.js";
+
+const providerFetch = vi.fn();
+
+vi.mock("../../../src/backend/providers/http.js", async () => {
+  const actual = await vi.importActual<
+    typeof import("../../../src/backend/providers/http.js")
+  >("../../../src/backend/providers/http.js");
+  return { ...actual, providerFetch };
+});
+
+const { openAiAdapter } =
+  await import("../../../src/backend/providers/openai.js");
+const { ollamaAdapter } =
+  await import("../../../src/backend/providers/ollama.js");
+const { geminiAdapter } =
+  await import("../../../src/backend/providers/gemini.js");
+
+/** Builds a Response whose body streams the given text chunks. */
+function streamingResponse(lines: string[]): Response {
+  const encoder = new TextEncoder();
+  return {
+    ok: true,
+    status: 200,
+    body: {
+      getReader() {
+        let index = 0;
+        return {
+          async read() {
+            if (index >= lines.length) return { done: true, value: undefined };
+            return { done: false, value: encoder.encode(lines[index++]) };
+          },
+          releaseLock() {},
+        };
+      },
+    },
+  } as unknown as Response;
+}
+
+/** One SSE frame, built from an object so the JSON stays readable. */
+function sseFrame(payload: unknown): string {
+  return `data: ${JSON.stringify(payload)}\n`;
+}
+
+async function collect(iterable: AsyncIterable<ChatChunk>) {
+  const chunks: ChatChunk[] = [];
+  for await (const chunk of iterable) chunks.push(chunk);
+  return chunks;
+}
+
+const REQUEST = {
+  model: "test-model",
+  system: "system",
+  messages: [{ role: "user" as const, content: "hi" }],
+  tools: [],
+};
+
+describe("openAiAdapter", () => {
+  beforeEach(() => vi.clearAllMocks());
+
+  it("normalises streamed text", async () => {
+    providerFetch.mockResolvedValue(
+      streamingResponse([
+        'data: {"choices":[{"delta":{"content":"Hel"}}]}\n',
+        'data: {"choices":[{"delta":{"content":"lo"}}]}\n',
+        "data: [DONE]\n",
+      ]),
+    );
+
+    const chunks = await collect(
+      openAiAdapter.streamChat({ providerType: "openai" }, REQUEST),
+    );
+
+    expect(chunks.filter((c) => c.type === "text")).toEqual([
+      { type: "text", text: "Hel" },
+      { type: "text", text: "lo" },
+    ]);
+    expect(chunks.at(-1)?.type).toBe("done");
+  });
+
+  it("reassembles tool arguments split across deltas", async () => {
+    providerFetch.mockResolvedValue(
+      streamingResponse([
+        'data: {"choices":[{"delta":{"tool_calls":[{"index":0,"id":"c1","function":{"name":"list_hosts","arguments":"{\\"a"}}]}}]}\n',
+        'data: {"choices":[{"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\\":1}"}}]}}]}\n',
+        "data: [DONE]\n",
+      ]),
+    );
+
+    const chunks = await collect(
+      openAiAdapter.streamChat({ providerType: "openai" }, REQUEST),
+    );
+
+    const call = chunks.find((c) => c.type === "tool_call");
+    expect(call).toMatchObject({
+      type: "tool_call",
+      call: { id: "c1", name: "list_hosts", arguments: { a: 1 } },
+    });
+  });
+
+  it("survives malformed tool arguments", async () => {
+    // A model that emits broken JSON gets an empty object; the tool's own
+    // validation then reports the problem back to it.
+    providerFetch.mockResolvedValue(
+      streamingResponse([
+        'data: {"choices":[{"delta":{"tool_calls":[{"index":0,"id":"c1","function":{"name":"list_hosts","arguments":"{not json"}}]}}]}\n',
+        "data: [DONE]\n",
+      ]),
+    );
+
+    const chunks = await collect(
+      openAiAdapter.streamChat({ providerType: "openai" }, REQUEST),
+    );
+
+    expect(chunks.find((c) => c.type === "tool_call")).toMatchObject({
+      call: { arguments: {} },
+    });
+  });
+
+  it("needs a base url for an openai-compatible provider", async () => {
+    await expect(
+      collect(
+        openAiAdapter.streamChat(
+          { providerType: "openai_compatible" },
+          REQUEST,
+        ),
+      ),
+    ).rejects.toThrow("base URL");
+  });
+});
+
+describe("ollamaAdapter", () => {
+  beforeEach(() => vi.clearAllMocks());
+
+  it("reads newline-delimited json rather than sse", async () => {
+    providerFetch.mockResolvedValue(
+      streamingResponse([
+        '{"message":{"content":"Hel"}}\n',
+        '{"message":{"content":"lo"}}\n',
+        '{"done":true,"done_reason":"stop"}\n',
+      ]),
+    );
+
+    const chunks = await collect(
+      ollamaAdapter.streamChat({ providerType: "ollama" }, REQUEST),
+    );
+
+    expect(chunks.filter((c) => c.type === "text")).toHaveLength(2);
+    expect(chunks.at(-1)).toMatchObject({ type: "done", stopReason: "stop" });
+  });
+
+  it("accepts tool arguments as an object or a json string", async () => {
+    providerFetch.mockResolvedValue(
+      streamingResponse([
+        '{"message":{"tool_calls":[{"function":{"name":"list_hosts","arguments":{"a":1}}}]}}\n',
+        '{"message":{"tool_calls":[{"function":{"name":"get_host","arguments":"{\\"hostId\\":2}"}}]}}\n',
+        '{"done":true}\n',
+      ]),
+    );
+
+    const chunks = await collect(
+      ollamaAdapter.streamChat({ providerType: "ollama" }, REQUEST),
+    );
+
+    const calls = chunks.filter((c) => c.type === "tool_call") as any[];
+    expect(calls[0].call.arguments).toEqual({ a: 1 });
+    expect(calls[1].call.arguments).toEqual({ hostId: 2 });
+  });
+});
+
+describe("geminiAdapter", () => {
+  beforeEach(() => vi.clearAllMocks());
+
+  it("carries thoughtSignature off a function call", async () => {
+    // Gemini 2.5+ 400s a follow-up whose functionCall parts lost their
+    // signature, which broke every conversation on the second turn.
+    providerFetch.mockResolvedValue(
+      streamingResponse([
+        sseFrame({
+          candidates: [
+            {
+              content: {
+                parts: [
+                  {
+                    functionCall: { name: "list_hosts", args: {} },
+                    thoughtSignature: "sig-abc",
+                  },
+                ],
+              },
+            },
+          ],
+        }),
+      ]),
+    );
+
+    const chunks = await collect(
+      geminiAdapter.streamChat(
+        { providerType: "gemini", apiKey: "k" },
+        REQUEST,
+      ),
+    );
+
+    expect(chunks.find((c) => c.type === "tool_call")).toMatchObject({
+      call: { name: "list_hosts", providerSignature: "sig-abc" },
+    });
+  });
+
+  it("echoes the signature back on the next turn", async () => {
+    providerFetch.mockResolvedValue(streamingResponse([]));
+
+    await collect(
+      geminiAdapter.streamChat(
+        { providerType: "gemini", apiKey: "k" },
+        {
+          ...REQUEST,
+          messages: [
+            { role: "user", content: "hi" },
+            {
+              role: "assistant",
+              content: "",
+              toolCalls: [
+                {
+                  id: "c1",
+                  name: "list_hosts",
+                  arguments: {},
+                  providerSignature: "sig-abc",
+                },
+              ],
+            },
+            {
+              role: "tool",
+              content: "{}",
+              toolCallId: "c1",
+              toolName: "list_hosts",
+            },
+          ],
+        },
+      ),
+    );
+
+    const body = JSON.parse(providerFetch.mock.calls[0][1].body as string);
+    const modelTurn = body.contents.find((c: any) => c.role === "model");
+    expect(modelTurn.parts[0].thoughtSignature).toBe("sig-abc");
+  });
+});
+
+describe("assertOk error messages", () => {
+  beforeEach(() => vi.clearAllMocks());
+
+  function errorResponse(status: number, body: string): Response {
+    return {
+      ok: false,
+      status,
+      text: async () => body,
+    } as unknown as Response;
+  }
+
+  it("pulls the message out of a nested error body", async () => {
+    // Slicing the raw JSON used to cut the text off mid-sentence.
+    providerFetch.mockResolvedValue(
+      errorResponse(
+        400,
+        JSON.stringify({
+          error: {
+            code: 400,
+            message: "Function call is missing a signature.",
+          },
+        }),
+      ),
+    );
+
+    await expect(
+      collect(openAiAdapter.streamChat({ providerType: "openai" }, REQUEST)),
+    ).rejects.toThrow("Function call is missing a signature.");
+  });
+
+  it("explains a rate limit instead of dumping the body", async () => {
+    providerFetch.mockResolvedValue(
+      errorResponse(
+        429,
+        JSON.stringify({ error: { message: "You exceeded your quota." } }),
+      ),
+    );
+
+    await expect(
+      collect(openAiAdapter.streamChat({ providerType: "openai" }, REQUEST)),
+    ).rejects.toThrow(/rate limit reached/i);
+  });
+
+  it("explains a rejected key", async () => {
+    providerFetch.mockResolvedValue(
+      errorResponse(401, JSON.stringify({ error: { message: "Bad key" } })),
+    );
+
+    await expect(
+      collect(openAiAdapter.streamChat({ providerType: "openai" }, REQUEST)),
+    ).rejects.toThrow(/rejected the API key/i);
+  });
+
+  it("falls back to a trimmed snippet for a non-JSON body", async () => {
+    providerFetch.mockResolvedValue(errorResponse(500, "upstream exploded"));
+
+    await expect(
+      collect(openAiAdapter.streamChat({ providerType: "openai" }, REQUEST)),
+    ).rejects.toThrow("upstream exploded");
+  });
+});

plugins/ai/tests/backend/redaction.test.ts

@@ -0,0 +1,76 @@
+import { describe, expect, it } from "vitest";
+import { REDACTED, redact, redactString } from "../../src/backend/redaction.js";
+
+describe("redact", () => {
+  it("drops secret-named fields at any depth", () => {
+    const input = {
+      name: "web-1",
+      password: "hunter2",
+      nested: { privateKey: "abc", apiKey: "def", port: 22 },
+      list: [{ keyPassword: "xyz", label: "ok" }],
+    };
+
+    const output = redact(input) as any;
+
+    expect(output.name).toBe("web-1");
+    expect(output.password).toBe(REDACTED);
+    expect(output.nested.privateKey).toBe(REDACTED);
+    expect(output.nested.apiKey).toBe(REDACTED);
+    expect(output.nested.port).toBe(22);
+    expect(output.list[0].keyPassword).toBe(REDACTED);
+    expect(output.list[0].label).toBe("ok");
+  });
+
+  it("keeps a null secret null so absence stays distinguishable", () => {
+    const output = redact({ password: null }) as any;
+    expect(output.password).toBeNull();
+  });
+
+  it("leaves ordinary values untouched", () => {
+    const input = { id: 4, enabled: true, tags: ["a", "b"], note: null };
+    expect(redact(input)).toEqual(input);
+  });
+
+  it("does not recurse forever on a cyclic object", () => {
+    const cyclic: Record<string, unknown> = { name: "loop" };
+    cyclic.self = cyclic;
+    expect(() => redact(cyclic)).not.toThrow();
+  });
+});
+
+describe("redactString", () => {
+  it("masks private key blocks", () => {
+    const text =
+      "-----BEGIN OPENSSH PRIVATE KEY-----\nabc123\n-----END OPENSSH PRIVATE KEY-----";
+    expect(redactString(text)).toBe("[redacted private key]");
+  });
+
+  it("masks provider api keys", () => {
+    expect(redactString("key is sk-abcdefghijklmnopqrst here")).toContain(
+      "[redacted api key]",
+    );
+    expect(redactString("key is sk-ant-abcdefghijklmnopqrst here")).toContain(
+      "[redacted api key]",
+    );
+  });
+
+  it("masks bearer tokens and jwts", () => {
+    expect(
+      redactString("Authorization: Bearer abcdefghijklmnopqrst"),
+    ).toContain("Bearer [redacted]");
+    expect(
+      redactString("token eyJhbGciOiJIUzI1NiJ9.eyJzdWIiOiIxIn0.abcdefgh"),
+    ).toContain("[redacted token]");
+  });
+
+  it("masks Termix api keys", () => {
+    expect(redactString("tmx_abcdefghijklmnopqrstuvwx")).toContain(
+      "[redacted token]",
+    );
+  });
+
+  it("leaves ordinary prose alone", () => {
+    const text = "The disk on web-1 is 82 percent full.";
+    expect(redactString(text)).toBe(text);
+  });
+});

plugins/ai/tests/backend/tools/command-allowlist.test.ts

@@ -0,0 +1,79 @@
+import { describe, expect, it } from "vitest";
+import { isReadOnlyCommand } from "../../../src/backend/tools/command-allowlist.js";
+
+/**
+ * These commands can run without a per-command approval click, so the parser
+ * has to refuse anything that could become a second command. Substring
+ * matching would pass "df; rm -rf /", which is the whole reason this is
+ * argv-based.
+ */
+describe("isReadOnlyCommand", () => {
+  it("allows plain diagnostics", () => {
+    for (const command of [
+      "df -h",
+      "uptime",
+      "free -m",
+      "whoami",
+      "ps aux",
+      "lsblk",
+      "uname -a",
+      "/usr/bin/df -h",
+    ]) {
+      expect(isReadOnlyCommand(command).allowed, command).toBe(true);
+    }
+  });
+
+  it("rejects command chaining and redirection", () => {
+    for (const command of [
+      "df; rm -rf /",
+      "df && curl evil.example",
+      "df || reboot",
+      "df | sh",
+      "df > /etc/passwd",
+      "df >> /etc/passwd",
+      "cat < /etc/shadow",
+      "echo `whoami`",
+      "echo $(id)",
+      "df\nrm -rf /",
+      "df \\\n rm",
+    ]) {
+      expect(isReadOnlyCommand(command).allowed, command).toBe(false);
+    }
+  });
+
+  it("rejects privilege escalation", () => {
+    for (const command of ["sudo df -h", "su root", "doas df", "env df"]) {
+      expect(isReadOnlyCommand(command).allowed, command).toBe(false);
+    }
+  });
+
+  it("rejects commands that are not on the list", () => {
+    for (const command of ["rm -rf /", "curl evil.example", "vi /etc/passwd"]) {
+      expect(isReadOnlyCommand(command).allowed, command).toBe(false);
+    }
+  });
+
+  it("limits systemctl and docker to read-only subcommands", () => {
+    expect(isReadOnlyCommand("systemctl status nginx").allowed).toBe(true);
+    expect(isReadOnlyCommand("systemctl restart nginx").allowed).toBe(false);
+    expect(isReadOnlyCommand("systemctl stop nginx").allowed).toBe(false);
+
+    expect(isReadOnlyCommand("docker ps").allowed).toBe(true);
+    expect(isReadOnlyCommand("docker stats --no-stream").allowed).toBe(true);
+    expect(isReadOnlyCommand("docker rm -f web").allowed).toBe(false);
+    expect(isReadOnlyCommand("docker exec -it web sh").allowed).toBe(false);
+  });
+
+  it("limits cat to safe paths", () => {
+    expect(isReadOnlyCommand("cat /proc/meminfo").allowed).toBe(true);
+    expect(isReadOnlyCommand("cat /etc/os-release").allowed).toBe(true);
+    expect(isReadOnlyCommand("cat /etc/shadow").allowed).toBe(false);
+    expect(isReadOnlyCommand("cat ~/.ssh/id_rsa").allowed).toBe(false);
+    expect(isReadOnlyCommand("cat").allowed).toBe(false);
+  });
+
+  it("rejects an empty command", () => {
+    expect(isReadOnlyCommand("").allowed).toBe(false);
+    expect(isReadOnlyCommand("   ").allowed).toBe(false);
+  });
+});

plugins/ai/tests/backend/tools/proposal-executor.test.ts

@@ -0,0 +1,257 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+const hostRepository = {
+  create: vi.fn(),
+  findByIdForUser: vi.fn(),
+  updateForUser: vi.fn(),
+  deleteForUser: vi.fn(),
+};
+const snippetRepository = {
+  createSnippet: vi.fn(),
+  findOwnedById: vi.fn(),
+  updateSnippet: vi.fn(),
+  deleteSnippet: vi.fn(),
+};
+const fleetRepository = { create: vi.fn(), addMember: vi.fn() };
+const automationRepository = { create: vi.fn() };
+
+vi.mock("../../../../../src/backend/database/repositories/factory.js", () => ({
+  createCurrentHostRepository: () => hostRepository,
+  createCurrentSnippetRepository: () => snippetRepository,
+  createCurrentFleetRepository: () => fleetRepository,
+  createCurrentAutomationRepository: () => automationRepository,
+}));
+
+const resolveHostById = vi.fn();
+vi.mock("../../../../../src/backend/hosts/host-resolver.js", () => ({
+  resolveHostById: (...args: unknown[]) => resolveHostById(...args),
+}));
+
+const execCommand = vi.fn();
+vi.mock(
+  "../../../../../src/backend/hosts/metrics-shared/common-utils.js",
+  () => ({
+    execCommand: (...args: unknown[]) => execCommand(...args),
+  }),
+);
+vi.mock("../../../../../src/backend/hosts/ssh-client-factory.js", () => ({
+  createFleetSshFactory: () => () => ({}),
+  getFleetPoolKey: () => "pool",
+}));
+vi.mock("../../../../../src/backend/hosts/ssh-connection-pool.js", () => ({
+  withConnection: async (
+    _key: string,
+    _factory: unknown,
+    run: (client: unknown) => Promise<unknown>,
+  ) => run({}),
+}));
+
+const { applyProposal } =
+  await import("../../../src/backend/tools/executor.js");
+
+describe("applyProposal", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+  });
+
+  it("refuses a kind that is not a real tool", async () => {
+    // The stored payload is treated as untrusted even though the server wrote
+    // it, because a proposal can outlive the release that created it.
+    await expect(
+      applyProposal("propose_delete_everything", {}, "user-1"),
+    ).rejects.toThrow("Unknown proposal kind");
+  });
+
+  it("validates an automation definition before creating it", async () => {
+    // Reuses the automations route's own validator, so a definition the model
+    // invented is held to the same standard as a hand-written one.
+    await expect(
+      applyProposal(
+        "propose_create_automation",
+        {
+          name: "bad",
+          definition: { trigger: { kind: "nonsense" }, steps: [] },
+        },
+        "user-1",
+      ),
+    ).rejects.toThrow();
+    expect(automationRepository.create).not.toHaveBeenCalled();
+  });
+
+  it("creates a valid automation disabled so it cannot fire unwatched", async () => {
+    automationRepository.create.mockResolvedValue({ id: 5, name: "nightly" });
+
+    const result = await applyProposal(
+      "propose_create_automation",
+      {
+        name: "nightly",
+        definition: {
+          trigger: { kind: "schedule", intervalSeconds: 3600 },
+          steps: [{ id: "s1", type: "wait", seconds: 1 }],
+        },
+      },
+      "user-1",
+    );
+
+    expect(result.ok).toBe(true);
+    expect(automationRepository.create).toHaveBeenCalledWith(
+      expect.objectContaining({ userId: "user-1", enabled: false }),
+    );
+  });
+
+  it("runs an approved command only on a host the user can reach", async () => {
+    // resolveHostById returns null when the connect-level permission check
+    // fails, so an unreachable host never gets as far as an SSH attempt.
+    resolveHostById.mockResolvedValue(null);
+
+    await expect(
+      applyProposal(
+        "propose_run_command",
+        { hostId: 9, command: "uptime" },
+        "user-1",
+      ),
+    ).rejects.toThrow("Host not found");
+    expect(execCommand).not.toHaveBeenCalled();
+  });
+
+  it("returns command output on success", async () => {
+    resolveHostById.mockResolvedValue({ id: 9, ip: "10.0.0.9" });
+    execCommand.mockResolvedValue({ stdout: "up 3 days", stderr: "", code: 0 });
+
+    const result = await applyProposal(
+      "propose_run_command",
+      { hostId: 9, command: "uptime" },
+      "user-1",
+    );
+
+    expect(result.ok).toBe(true);
+    expect(result.summary).toContain("up 3 days");
+  });
+
+  it("resolves the host rather than using a raw repository row", async () => {
+    // A raw row has no decrypted auth and an unresolved jumpHosts field, which
+    // made the SSH factory fail with a jump host error on hosts that have none.
+    resolveHostById.mockResolvedValue({ id: 9, ip: "10.0.0.9" });
+    execCommand.mockResolvedValue({ stdout: "ok", stderr: "", code: 0 });
+
+    await applyProposal(
+      "propose_run_command",
+      { hostId: 9, command: "uptime" },
+      "user-1",
+    );
+
+    expect(resolveHostById).toHaveBeenCalledWith(9, "user-1");
+    expect(hostRepository.findByIdForUser).not.toHaveBeenCalled();
+  });
+
+  it("surfaces a non-zero exit rather than reporting success", async () => {
+    resolveHostById.mockResolvedValue({ id: 9, ip: "10.0.0.9" });
+    execCommand.mockResolvedValue({ stdout: "", stderr: "denied", code: 1 });
+
+    await expect(
+      applyProposal(
+        "propose_run_command",
+        { hostId: 9, command: "cat /etc/shadow" },
+        "user-1",
+      ),
+    ).rejects.toThrow("code 1");
+  });
+
+  it("creates a host through the normal repository", async () => {
+    hostRepository.create.mockResolvedValue({ id: 7, name: "web-1" });
+
+    const result = await applyProposal(
+      "propose_create_host",
+      { name: "web-1", ip: "10.0.0.5", port: 22, tags: ["prod"] },
+      "user-1",
+    );
+
+    expect(result.ok).toBe(true);
+    expect(hostRepository.create).toHaveBeenCalledWith(
+      expect.objectContaining({
+        userId: "user-1",
+        name: "web-1",
+        ip: "10.0.0.5",
+      }),
+    );
+  });
+
+  it("rejects a host payload missing required fields", async () => {
+    await expect(
+      applyProposal("propose_create_host", { name: "web-1" }, "user-1"),
+    ).rejects.toThrow("ip is required");
+    expect(hostRepository.create).not.toHaveBeenCalled();
+  });
+
+  it("scopes an update to the approving user", async () => {
+    hostRepository.findByIdForUser.mockResolvedValue({ id: 7 });
+    hostRepository.updateForUser.mockResolvedValue({ id: 7 });
+
+    await applyProposal(
+      "propose_update_host",
+      { hostId: 7, changes: { name: "renamed" } },
+      "user-1",
+    );
+
+    expect(hostRepository.findByIdForUser).toHaveBeenCalledWith("user-1", 7);
+    expect(hostRepository.updateForUser).toHaveBeenCalledWith(
+      "user-1",
+      7,
+      expect.objectContaining({ name: "renamed" }),
+    );
+  });
+
+  it("refuses to update a host the user does not own", async () => {
+    hostRepository.findByIdForUser.mockResolvedValue(null);
+
+    await expect(
+      applyProposal(
+        "propose_update_host",
+        { hostId: 999, changes: { name: "x" } },
+        "user-1",
+      ),
+    ).rejects.toThrow("Host not found");
+    expect(hostRepository.updateForUser).not.toHaveBeenCalled();
+  });
+
+  it("rejects a non-numeric id rather than coercing it", async () => {
+    await expect(
+      applyProposal(
+        "propose_update_host",
+        { hostId: "7; DROP TABLE hosts", changes: { name: "x" } },
+        "user-1",
+      ),
+    ).rejects.toThrow("hostId must be a positive integer");
+  });
+
+  it("only adds fleet members the approving user owns", async () => {
+    fleetRepository.create.mockResolvedValue({ id: 3, name: "prod" });
+    hostRepository.findByIdForUser.mockImplementation(
+      async (_userId: string, hostId: number) =>
+        hostId === 1 ? { id: 1 } : null,
+    );
+
+    const result = await applyProposal(
+      "propose_create_fleet",
+      { name: "prod", hostIds: [1, 2] },
+      "user-1",
+    );
+
+    expect(fleetRepository.addMember).toHaveBeenCalledTimes(1);
+    expect(fleetRepository.addMember).toHaveBeenCalledWith(3, 1);
+    expect(result.summary).toContain("1 host");
+  });
+
+  it("reports nothing to change on an empty update", async () => {
+    hostRepository.findByIdForUser.mockResolvedValue({ id: 7 });
+
+    const result = await applyProposal(
+      "propose_update_host",
+      { hostId: 7, changes: {} },
+      "user-1",
+    );
+
+    expect(result.ok).toBe(false);
+    expect(hostRepository.updateForUser).not.toHaveBeenCalled();
+  });
+});

plugins/ai/tests/backend/tools/tool-catalog.test.ts

@@ -0,0 +1,112 @@
+import { describe, expect, it } from "vitest";
+import {
+  AI_TOOLS,
+  FORBIDDEN_DOMAINS,
+  getTool,
+  listToolNames,
+  toolDefinitions,
+} from "../../../src/backend/tools/catalog.js";
+
+/**
+ * The security regression test for the whole feature.
+ *
+ * Tools run in-process and never pass through the RBAC-gated routers, so "the
+ * assistant cannot reach credentials or user administration" is a property of
+ * this catalog, and nothing else. If a future change adds a tool that touches a
+ * forbidden domain, this test is what catches it.
+ */
+describe("AI tool catalog", () => {
+  it("exposes no tool naming a forbidden domain", () => {
+    for (const tool of AI_TOOLS) {
+      for (const domain of FORBIDDEN_DOMAINS) {
+        expect(
+          tool.name.includes(domain),
+          `${tool.name} references the forbidden domain "${domain}"`,
+        ).toBe(false);
+      }
+    }
+  });
+
+  it("has no tool that could read a credential", () => {
+    const banned = [
+      "get_credential",
+      "list_credentials",
+      "get_password",
+      "get_private_key",
+      "get_api_key",
+      "create_user",
+      "delete_user",
+      "grant_permission",
+      "update_settings",
+    ];
+    for (const name of banned) {
+      expect(getTool(name), `${name} must not exist`).toBeUndefined();
+    }
+  });
+
+  it("only allows read or propose categories", () => {
+    for (const tool of AI_TOOLS) {
+      expect(["read", "propose"]).toContain(tool.category);
+    }
+  });
+
+  it("names every tool by its category", () => {
+    // A propose tool that does not say "propose" would read as a direct action
+    // in the transcript, which is exactly the confusion this feature avoids.
+    for (const tool of AI_TOOLS) {
+      if (tool.category === "propose") {
+        expect(
+          tool.name.startsWith("propose_"),
+          `${tool.name} is a propose tool but is not named propose_*`,
+        ).toBe(true);
+      } else {
+        expect(
+          tool.name.startsWith("propose_"),
+          `${tool.name} is a read tool but is named propose_*`,
+        ).toBe(false);
+      }
+    }
+  });
+
+  it("has unique tool names", () => {
+    const names = listToolNames();
+    expect(new Set(names).size).toBe(names.length);
+  });
+
+  it("gives every tool a described object schema", () => {
+    for (const definition of toolDefinitions()) {
+      expect(definition.description.length, definition.name).toBeGreaterThan(
+        20,
+      );
+      expect(definition.parameters.type, definition.name).toBe("object");
+      // additionalProperties:false keeps a model from smuggling extra fields
+      // past the handler's explicit reads.
+      expect(definition.parameters.additionalProperties, definition.name).toBe(
+        false,
+      );
+    }
+  });
+
+  it("never takes a userId from the model", () => {
+    // Ownership is always derived from the verified JWT. A userId parameter
+    // would let the model ask for another account's data.
+    for (const tool of AI_TOOLS) {
+      const properties = (tool.parameters.properties ?? {}) as Record<
+        string,
+        unknown
+      >;
+      for (const key of Object.keys(properties)) {
+        expect(
+          /^user_?id$/i.test(key),
+          `${tool.name} accepts a model-supplied ${key}`,
+        ).toBe(false);
+      }
+    }
+  });
+
+  it("still offers the read tools the assistant needs to be useful", () => {
+    for (const name of ["list_hosts", "list_snippets", "list_automations"]) {
+      expect(getTool(name), name).toBeDefined();
+    }
+  });
+});

plugins/ai/tests/frontend/AiProviderSettings.test.tsx

@@ -0,0 +1,98 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+import {
+  cleanup,
+  fireEvent,
+  render,
+  screen,
+  waitFor,
+} from "@testing-library/react";
+import type { AiProvider } from "@/api/ai-api";
+
+const api = vi.hoisted(() => ({
+  createAiProvider: vi.fn(),
+  deleteAiProvider: vi.fn(),
+  getAiProviderModels: vi.fn(),
+  probeAiModels: vi.fn(),
+  updateAiProvider: vi.fn(),
+}));
+
+vi.mock("@/api/ai-api", () => api);
+vi.mock("react-i18next", () => ({
+  useTranslation: () => ({ t: (key: string) => key }),
+}));
+vi.mock("sonner", () => ({
+  toast: { success: vi.fn(), error: vi.fn() },
+}));
+
+import { AiProviderSettings } from "@/features/ai/AiProviderSettings";
+
+const provider: AiProvider = {
+  id: 7,
+  providerType: "ollama",
+  label: "Local Ollama",
+  baseUrl: "http://localhost:11434",
+  apiKeyPrefix: null,
+  defaultModel: "llama3.1",
+  enabled: true,
+  createdAt: "2026-08-21T00:00:00Z",
+};
+
+beforeEach(() => {
+  api.getAiProviderModels.mockReset();
+  api.getAiProviderModels.mockResolvedValue([]);
+  api.updateAiProvider.mockReset();
+  api.updateAiProvider.mockResolvedValue(provider);
+});
+
+afterEach(cleanup);
+
+describe("AiProviderSettings", () => {
+  it("edits an existing provider name and configured model", async () => {
+    const onChanged = vi.fn();
+    render(<AiProviderSettings providers={[provider]} onChanged={onChanged} />);
+
+    expect(screen.getByText(/llama3\.1/)).toBeTruthy();
+    fireEvent.click(screen.getByRole("button", { name: "ai.editProvider" }));
+
+    const label = screen.getByLabelText("ai.providerLabel");
+    const model = screen.getByLabelText("ai.defaultModel");
+    fireEvent.change(label, { target: { value: "Production Ollama" } });
+    fireEvent.change(model, { target: { value: "qwen3:32b" } });
+    fireEvent.click(screen.getByRole("button", { name: "ai.save" }));
+
+    await waitFor(() => {
+      expect(api.updateAiProvider).toHaveBeenCalledWith(7, {
+        label: "Production Ollama",
+        defaultModel: "qwen3:32b",
+      });
+    });
+    expect(onChanged).toHaveBeenCalledWith(7);
+  });
+
+  it("loads the saved provider model list when editing starts", async () => {
+    api.getAiProviderModels.mockResolvedValue(["llama3.1", "qwen3:32b"]);
+    render(<AiProviderSettings providers={[provider]} onChanged={() => {}} />);
+
+    fireEvent.click(screen.getByRole("button", { name: "ai.editProvider" }));
+
+    await waitFor(() => {
+      expect(api.getAiProviderModels).toHaveBeenCalledWith(7);
+    });
+    expect(
+      screen.getByRole("combobox", { name: "ai.defaultModel" }),
+    ).toBeTruthy();
+  });
+
+  it("shows the provider error when refreshing models fails", async () => {
+    api.getAiProviderModels.mockRejectedValue(
+      new Error("Add llm.internal to the AI endpoint allowlist"),
+    );
+    render(<AiProviderSettings providers={[provider]} onChanged={() => {}} />);
+
+    fireEvent.click(screen.getByRole("button", { name: "ai.editProvider" }));
+
+    expect(
+      await screen.findByText("Add llm.internal to the AI endpoint allowlist"),
+    ).toBeTruthy();
+  });
+});

plugins/ai/tests/frontend/labels.test.ts

@@ -0,0 +1,73 @@
+import { describe, expect, it } from "vitest";
+import {
+  fieldLabel,
+  humanize,
+  mentionLabel,
+  toolLabel,
+} from "@/features/ai/labels";
+import { AI_TOOLS } from "../../src/backend/tools/catalog.js";
+
+describe("humanize", () => {
+  it("splits camelCase into words", () => {
+    // The case that shipped as "thresholddurationseconds".
+    expect(humanize("thresholdDurationSeconds")).toBe(
+      "Threshold duration seconds",
+    );
+  });
+
+  it("splits snake_case into words", () => {
+    expect(humanize("propose_create_automation")).toBe(
+      "Propose create automation",
+    );
+  });
+
+  it("leaves a single word alone apart from casing", () => {
+    expect(humanize("port")).toBe("Port");
+  });
+
+  it("returns the input when there is nothing to split", () => {
+    expect(humanize("")).toBe("");
+  });
+});
+
+describe("fieldLabel", () => {
+  it("uses the mapped name where there is one", () => {
+    expect(fieldLabel("thresholdDurationSeconds")).toBe("Duration (seconds)");
+    expect(fieldLabel("cooldownMinutes")).toBe("Cooldown (minutes)");
+    expect(fieldLabel("ip")).toBe("Address");
+  });
+
+  it("falls back to humanizing an unmapped key", () => {
+    expect(fieldLabel("someFutureField")).toBe("Some future field");
+  });
+
+  it("never returns a run-on identifier", () => {
+    for (const key of ["thresholdDurationSeconds", "hostIds", "triggerType"]) {
+      expect(fieldLabel(key)).not.toBe(key.toLowerCase());
+    }
+  });
+});
+
+describe("toolLabel", () => {
+  it("phrases tools as actions rather than function names", () => {
+    expect(toolLabel("list_hosts")).toBe("Reading hosts");
+    expect(toolLabel("propose_create_fleet")).toBe("Create fleet");
+  });
+
+  it("covers every tool in the catalog", () => {
+    // A new tool without a label would surface its raw snake_case name.
+    for (const tool of AI_TOOLS) {
+      const label = toolLabel(tool.name);
+      expect(label, tool.name).not.toContain("_");
+      expect(label, tool.name).not.toBe(tool.name);
+    }
+  });
+});
+
+describe("mentionLabel", () => {
+  it("capitalises the mention kinds", () => {
+    expect(mentionLabel("host")).toBe("Host");
+    expect(mentionLabel("snippet")).toBe("Snippet");
+    expect(mentionLabel("automation")).toBe("Automation");
+  });
+});

plugins/ai/tsconfig.json

@@ -0,0 +1,10 @@
+{
+  "extends": "@termix/plugin-sdk/tsconfig.plugin.json",
+  "compilerOptions": {
+    "paths": {
+      "@/types/*": ["../../src/types/*"],
+      "@/*": ["../../src/ui/*"]
+    }
+  },
+  "include": ["src/**/*.ts", "src/**/*.tsx", "tests/**/*.ts", "tests/**/*.tsx"]
+}

plugins/ai/vitest.config.ts

@@ -0,0 +1,3 @@
+import { pluginVitestConfig } from "@termix/plugin-sdk/vitest-preset";
+
+export default pluginVitestConfig(import.meta.url);