Add temp Claude config isolation to daemon E2E tests

- Create shared useTempClaudeConfigDir() helper in test-utils/claude-config.ts
- Export from test-utils/index.ts for reuse across test files
- Update daemon.e2e.test.ts to use beforeAll/afterAll hooks with temp config
- Remove .skip from "permission flow: Claude" describe block
- Refactor claude-agent.test.ts to use shared helper instead of local copy

This fixes the daemon E2E Claude permission tests that were skipped due to
the user's real ~/.claude/settings.json having "Bash(rm:*)" in the allow
list, which caused rm commands to auto-execute without permission prompts.

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

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

packages/server/src/server/agent/providers/claude-agent.test.ts

@@ -2,7 +2,6 @@ import { afterAll, beforeAll, describe, expect, test, vi } from "vitest";
 import { createServer } from "http";
 import { randomUUID } from "node:crypto";
 import {
-  copyFileSync,
   existsSync,
   mkdtempSync,
   readFileSync,
@@ -17,6 +16,7 @@ import { StreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/
 import { isInitializeRequest } from "@modelcontextprotocol/sdk/types.js";
 
 import { ClaudeAgentClient, convertClaudeHistoryEntry } from "./claude-agent.js";
+import { useTempClaudeConfigDir } from "../../test-utils/claude-config.js";
 import {
   hydrateStreamState,
   type StreamItem,
@@ -46,46 +46,6 @@ function tmpCwd(): string {
   }
 }
 
-function copyClaudeCredentials(sourceDir: string, targetDir: string): void {
-  const sourceCredentials = path.join(sourceDir, ".credentials.json");
-  if (!existsSync(sourceCredentials)) {
-    return;
-  }
-  copyFileSync(sourceCredentials, path.join(targetDir, ".credentials.json"));
-}
-
-function useTempClaudeConfigDir(): () => void {
-  const previousConfigDir = process.env.CLAUDE_CONFIG_DIR;
-  const sourceConfigDir =
-    previousConfigDir ?? path.join(os.homedir(), ".claude");
-  const configDir = mkdtempSync(path.join(os.tmpdir(), "claude-config-"));
-  const settings = {
-    permissions: {
-      allow: [],
-      deny: [],
-      ask: ["Bash(rm:*)"],
-      additionalDirectories: [],
-    },
-    sandbox: {
-      enabled: true,
-      autoAllowBashIfSandboxed: false,
-    },
-  };
-  const settingsText = `${JSON.stringify(settings, null, 2)}\n`;
-  writeFileSync(path.join(configDir, "settings.json"), settingsText, "utf8");
-  writeFileSync(path.join(configDir, "settings.local.json"), settingsText, "utf8");
-  copyClaudeCredentials(sourceConfigDir, configDir);
-  process.env.CLAUDE_CONFIG_DIR = configDir;
-  return () => {
-    if (previousConfigDir === undefined) {
-      delete process.env.CLAUDE_CONFIG_DIR;
-    } else {
-      process.env.CLAUDE_CONFIG_DIR = previousConfigDir;
-    }
-    rmSync(configDir, { recursive: true, force: true });
-  };
-}
-
 async function autoApprove(session: Awaited<ReturnType<ClaudeAgentClient["createSession"]>>, event: AgentStreamEvent) {
   if (event.type === "permission_requested") {
     await session.respondToPermission(event.request.id, { behavior: "allow" });

packages/server/src/server/daemon.e2e.test.ts

@@ -1,10 +1,11 @@
-import { describe, test, expect, beforeEach, afterEach } from "vitest";
+import { describe, test, expect, beforeEach, afterEach, beforeAll, afterAll } from "vitest";
 import { mkdtempSync, writeFileSync, existsSync, rmSync } from "fs";
 import { tmpdir } from "os";
 import path from "path";
 import {
   createDaemonTestContext,
   type DaemonTestContext,
+  useTempClaudeConfigDir,
 } from "./test-utils/index.js";
 import type { AgentTimelineItem } from "./agent/agent-sdk-types.js";
 import type { AgentSnapshotPayload } from "./messages.js";
@@ -900,14 +901,19 @@ describe("daemon E2E", () => {
     );
   });
 
-  // Claude permission tests are skipped due to SDK behavior:
-  // - The sandbox config IS passed correctly to Claude SDK
-  // - Claude executes tool calls without requesting permission
-  // - This appears to be related to user/project settings or SDK behavior
-  // - The direct claude-agent.test.ts permission tests pass
-  // - Codex permission tests through the daemon work correctly
-  // TODO: Investigate Claude SDK permission behavior in daemon context
-  describe.skip("permission flow: Claude", () => {
+  describe("permission flow: Claude", () => {
+    // Use isolated Claude config to ensure permission prompts are triggered
+    // (user's real config may have allow rules that auto-approve commands)
+    let restoreClaudeConfig: () => void;
+
+    beforeAll(() => {
+      restoreClaudeConfig = useTempClaudeConfigDir();
+    });
+
+    afterAll(() => {
+      restoreClaudeConfig();
+    });
+
     test(
       "approves permission and executes command",
       async () => {

packages/server/src/server/test-utils/claude-config.ts

@@ -0,0 +1,53 @@
+import { mkdtempSync, writeFileSync, copyFileSync, existsSync, rmSync } from "fs";
+import { tmpdir, homedir } from "os";
+import path from "path";
+
+function copyClaudeCredentials(sourceDir: string, targetDir: string): void {
+  const sourceCredentials = path.join(sourceDir, ".credentials.json");
+  if (!existsSync(sourceCredentials)) {
+    return;
+  }
+  copyFileSync(sourceCredentials, path.join(targetDir, ".credentials.json"));
+}
+
+/**
+ * Sets up an isolated Claude config directory for testing.
+ * Creates a temp directory with:
+ * - settings.json with ask: ["Bash(rm:*)"] to trigger permission prompts
+ * - settings.local.json with the same settings
+ * - .credentials.json copied from user's real config
+ *
+ * Sets CLAUDE_CONFIG_DIR env var to point to the temp directory.
+ * Returns a cleanup function that restores the original env and removes the temp dir.
+ */
+export function useTempClaudeConfigDir(): () => void {
+  const previousConfigDir = process.env.CLAUDE_CONFIG_DIR;
+  const sourceConfigDir =
+    previousConfigDir ?? path.join(homedir(), ".claude");
+  const configDir = mkdtempSync(path.join(tmpdir(), "claude-config-"));
+  const settings = {
+    permissions: {
+      allow: [],
+      deny: [],
+      ask: ["Bash(rm:*)"],
+      additionalDirectories: [],
+    },
+    sandbox: {
+      enabled: true,
+      autoAllowBashIfSandboxed: false,
+    },
+  };
+  const settingsText = `${JSON.stringify(settings, null, 2)}\n`;
+  writeFileSync(path.join(configDir, "settings.json"), settingsText, "utf8");
+  writeFileSync(path.join(configDir, "settings.local.json"), settingsText, "utf8");
+  copyClaudeCredentials(sourceConfigDir, configDir);
+  process.env.CLAUDE_CONFIG_DIR = configDir;
+  return () => {
+    if (previousConfigDir === undefined) {
+      delete process.env.CLAUDE_CONFIG_DIR;
+    } else {
+      process.env.CLAUDE_CONFIG_DIR = previousConfigDir;
+    }
+    rmSync(configDir, { recursive: true, force: true });
+  };
+}

packages/server/src/server/test-utils/index.ts

@@ -11,3 +11,4 @@ export {
   createDaemonTestContext,
   type DaemonTestContext,
 } from "./daemon-test-context.js";
+export { useTempClaudeConfigDir } from "./claude-config.js";

plan.md

@@ -1153,3 +1153,87 @@ Build a new Codex MCP provider side‑by‑side with the existing Codex SDK prov
   - Claude permission tests pass (or clear explanation of why they can't)
 
   - **Done (2025-12-25 19:12)**: WHAT: Analyzed permission handling differences between `claude-agent.test.ts` and `daemon.e2e.test.ts`. RESULT: Root cause identified - daemon E2E tests read the user's real `~/.claude/settings.json` which has `allow: ["Bash(rm:*)"]`, causing `rm` commands to auto-execute without permission prompts. In contrast, direct tests use `useTempClaudeConfigDir()` (`claude-agent.test.ts:57-87`) which creates a temp config dir with `ask: ["Bash(rm:*)"]` and sets `CLAUDE_CONFIG_DIR` env var. EVIDENCE: User's `~/.claude/settings.json` contains `"Bash(rm:*)"` in allow list. SDK uses `settingSources: ["user", "project"]` (`claude-agent.ts:665`) to read settings from disk. FIX: Add temp config setup to daemon tests (same pattern as direct tests) or use `settingSources: []` for SDK isolation mode. Full report: `REPORT-claude-permission-tests.md`.
+
+## DaemonClient Surface Expansion (App Parity)
+
+Goal: Expand DaemonClient to cover all daemon WebSocket capabilities so the app can eventually use it instead of raw WebSocket code.
+
+- [x] **Fix**: Add temp Claude config isolation to daemon E2E tests and unskip Claude permission tests.
+
+  **Context**: Investigation found Claude permission tests fail because they read user's real `~/.claude/settings.json` which auto-allows `rm`. Direct tests use `useTempClaudeConfigDir()` for isolation.
+
+  **Implementation**:
+  1. Export `useTempClaudeConfigDir()` helper from `claude-agent.test.ts` to `test-utils/`
+  2. Update daemon test context to optionally use temp Claude config
+  3. Update Claude permission tests to use temp config
+  4. Unskip tests and verify they pass
+
+  **Acceptance criteria**:
+  - Claude permission approve/deny tests pass (not skipped)
+  - Same isolation pattern as direct claude-agent tests
+
+  - **Done (2025-12-25 18:06)**: WHAT: Created shared `useTempClaudeConfigDir()` helper in `test-utils/claude-config.ts:22-49`, exported from `test-utils/index.ts:14`. Updated `daemon.e2e.test.ts:904-915` to add beforeAll/afterAll hooks using this helper and removed `.skip` from `describe("permission flow: Claude")`. Refactored `claude-agent.test.ts:19` to import shared helper and removed duplicate local implementation. RESULT: Both daemon Claude permission tests now pass. EVIDENCE: `npx vitest run --testNamePattern="permission flow: Claude"` shows: ✓ approves permission and executes command (8424ms), ✓ denies permission and prevents execution (7945ms). Direct `claude-agent.test.ts` permission tests also pass (4 tests). Typecheck passes.
+
+- [ ] **Implement**: Add `getGitDiff()` to DaemonClient + E2E test.
+
+  **Context**: App calls `git_diff_request` to get file diffs. DaemonClient doesn't support this yet.
+
+  **Implementation**:
+  1. Add `getGitDiff(agentId: string, filepath?: string)` method to DaemonClient
+  2. Send `git_diff_request` message, wait for `git_diff_response`
+  3. Add E2E test that creates agent in a git repo, modifies a file, calls getGitDiff
+
+  **Acceptance criteria**:
+  - Method returns diff content
+  - E2E test passes
+
+- [ ] **Implement**: Add `getGitRepoInfo()` to DaemonClient + E2E test.
+
+  **Context**: App calls `git_repo_info_request` to get repo info (branch, status, etc).
+
+  **Implementation**:
+  1. Add `getGitRepoInfo(agentId: string)` method to DaemonClient
+  2. Send `git_repo_info_request` message, wait for `git_repo_info_response`
+  3. Add E2E test in a git repo
+
+  **Acceptance criteria**:
+  - Method returns repo info (branch, status, remotes)
+  - E2E test passes
+
+- [ ] **Implement**: Add `exploreFileSystem()` to DaemonClient + E2E test.
+
+  **Context**: App calls `file_explorer_request` to browse filesystem.
+
+  **Implementation**:
+  1. Add `exploreFileSystem(agentId: string, path: string)` method to DaemonClient
+  2. Send `file_explorer_request` message, wait for `file_explorer_response`
+  3. Add E2E test that lists a directory
+
+  **Acceptance criteria**:
+  - Method returns file/directory listing
+  - E2E test passes
+
+- [ ] **Implement**: Add `listProviderModels()` to DaemonClient + E2E test.
+
+  **Context**: App calls `list_provider_models_request` to get available models.
+
+  **Implementation**:
+  1. Add `listProviderModels(provider: AgentProvider)` method to DaemonClient
+  2. Send `list_provider_models_request` message, wait for `list_provider_models_response`
+  3. Add E2E test for Codex and Claude providers
+
+  **Acceptance criteria**:
+  - Method returns model list
+  - E2E test passes for at least one provider
+
+- [ ] **Implement**: Add `sendImages()` support to DaemonClient + E2E test.
+
+  **Context**: App can send messages with image attachments. DaemonClient `sendMessage()` has `images` option but it's not tested.
+
+  **Implementation**:
+  1. Add E2E test that sends a message with an image attachment
+  2. Verify agent receives and processes the image
+
+  **Acceptance criteria**:
+  - E2E test passes with image attachment
+  - Claude provider correctly receives image (multimodal support)