feat(ai-chat): support SDK classifier-driven Auto permission mode

src-node/claude-code-agent.js

@@ -103,11 +103,14 @@ let _planApproved = false;
 let _queuedClarification = null;
 
 // Module-level "runtime" permission mode that hooks read at decision time.
-// Updated on every sendPrompt and via the setPermissionMode peer when the
-// user cycles the panel's permission bar mid-stream — without this, the
-// Bash hook would close over the value at query start and continue
-// prompting for confirmation even after the user has flipped to Full Auto.
-let _runtimePermissionMode = "acceptEdits";
+// One of "plan" | "acceptEdits" | "auto" (SDK classifier-approved) |
+// "bypassPermissions" (Allow Everything). Updated on every sendPrompt and
+// via the setPermissionMode peer when the user cycles the panel's
+// permission bar mid-stream — without this, the Bash hook would close over
+// the value at query start and continue prompting for confirmation even
+// after the user has flipped to Allow Everything. Defaults to "auto" to
+// match the browser's default (see AIChatPanel.js's _permissionMode).
+let _runtimePermissionMode = "auto";
 
 const nodeConnector = global.createNodeConnector(CONNECTOR_ID, exports);
 
@@ -661,7 +664,7 @@ exports.answerPlanModeWriteConfirm = async function (params) {
  * Apply a mid-stream permission-mode change so hooks running for the rest
  * of the turn use the new value. Called from the browser when the user
  * cycles the permission bar (so e.g. Bash stops prompting immediately
- * after switching from Edit Mode to Full Auto). The next sendPrompt also
+ * after switching from Edit Mode to Allow Everything). The next sendPrompt also
  * passes permissionMode in params, so this peer is only strictly required
  * during streaming — but calling it on every cycle keeps the agent's
  * tracker authoritative.
@@ -758,7 +761,7 @@ async function _runQuery(requestId, prompt, projectPath, model, signal, locale,
     // Sync the runtime mutable that hooks read for permission decisions —
     // setPermissionMode (peer) updates this same variable when the user
     // cycles modes mid-stream.
-    _runtimePermissionMode = permissionMode || "acceptEdits";
+    _runtimePermissionMode = permissionMode || "auto";
     let editCount = 0;
     let toolCounter = 0;
     // SDK tool_use id (e.g. "toolu_01...") → our sequential toolCounter so a
@@ -901,7 +904,7 @@ async function _runQuery(requestId, prompt, projectPath, model, signal, locale,
             }
         },
         mcpServers: { "phoenix-editor": editorMcpServer },
-        permissionMode: permissionMode || "acceptEdits",
+        permissionMode: permissionMode || "auto",
         appendSystemPrompt:
             "When modifying an existing file, always prefer the Edit tool " +
             "(find-and-replace) instead of the Write tool. The Write tool should ONLY be used " +
@@ -1288,11 +1291,15 @@ async function _runQuery(requestId, prompt, projectPath, model, signal, locale,
                         async (input) => {
                             // Read from the runtime mutable so mid-stream
                             // permission-mode flips (e.g. user switches Edit
-                            // Mode → Full Auto while bash is in flight) take
-                            // effect on the NEXT bash call without waiting
-                            // for the next prompt.
+                            // Mode → Allow Everything while bash is in flight)
+                            // take effect on the NEXT bash call without
+                            // waiting for the next prompt.
                             if (_runtimePermissionMode !== "acceptEdits") {
-                                // Plan mode: SDK handles. Full Auto: allow freely.
+                                // Plan mode: SDK handles. Auto: SDK's own
+                                // classifier decides. Allow Everything: allow
+                                // freely. Either way, Phoenix's own
+                                // confirm-dialog/safe-bash-allowlist below is
+                                // only for Edit Mode's manual approval flow.
                                 return {};
                             }
                             // Edit Mode: ask user confirmation before running bash

src/nls/root/strings.js

@@ -2660,9 +2660,9 @@ define({
     "AI_CHAT_FILE_NOT_FOUND_MSG": "Could not open <span class=\"dialog-filename\">{0}</span>. The file may have been moved or deleted.",
     "AI_CHAT_UNDO_RESTORE_WARNING_TITLE": "AI Undo & Restore",
     "AI_CHAT_UNDO_RESTORE_WARNING_BODY": "This will only undo changes made by the AI. Changes made outside the AI won’t be restored and may be lost. For full version history, use version control like Git.",
-    "AI_CHAT_FULL_AUTO_WARNING_TITLE": "Switch to Full Auto Mode?",
-    "AI_CHAT_FULL_AUTO_WARNING_BODY": "Full Auto mode lets the AI run any tool — Bash commands, file edits, file deletions, web fetches — without asking you first.<br><br>This is convenient for trusted scratch projects, but can be risky: a misjudged step could overwrite or delete files, run a destructive shell command, or push unintended changes. Use version control (Git) so you can recover if something goes wrong.<br><br>Only enable Full Auto in projects you trust. You can switch back to Edit Mode at any time using <kbd>Shift+Tab</kbd> or by clicking the mode bar.",
-    "AI_CHAT_FULL_AUTO_WARNING_PROCEED": "Enable Full Auto",
+    "AI_CHAT_FULL_AUTO_WARNING_TITLE": "Switch to Allow Everything Mode?",
+    "AI_CHAT_FULL_AUTO_WARNING_BODY": "Allow Everything mode lets the AI run any tool — Bash commands, file edits, file deletions, web fetches — without asking you first.<br><br>This is convenient for trusted scratch projects, but can be risky: a misjudged step could overwrite or delete files, run a destructive shell command, or push unintended changes. Use version control (Git) so you can recover if something goes wrong.<br><br>Only enable Allow Everything in projects you trust. You can switch back to Edit Mode at any time using <kbd>Shift+Tab</kbd> or by clicking the mode bar.",
+    "AI_CHAT_FULL_AUTO_WARNING_PROCEED": "Enable Allow Everything",
     "AI_CHAT_ONBOARDING_REVIEW_PROMPT": "Ready to build",
     "AI_CHAT_ONBOARDING_SEND": "Send",
     "AI_CHAT_SHOW_DIFF": "Show diff",
@@ -2699,9 +2699,11 @@ define({
     "AI_CHAT_PLAN_REVISE_DEFAULT": "Please revise the plan.",
     "AI_CHAT_MODE_PLAN": "Plan Mode",
     "AI_CHAT_MODE_EDIT": "AI Edit Mode",
-    "AI_CHAT_MODE_FULL_AUTO": "Full Auto",
+    "AI_CHAT_MODE_AUTO": "Auto",
+    "AI_CHAT_MODE_FULL_AUTO": "Allow Everything",
     "AI_CHAT_MODE_INFO_PLAN": "AI will propose a plan before making changes (Click to switch)",
     "AI_CHAT_MODE_INFO_EDIT": "AI can edit files. Shell commands need approval (Click to switch)",
+    "AI_CHAT_MODE_INFO_AUTO": "AI uses judgment to auto-approve safe actions and asks about risky ones (Click to switch)",
     "AI_CHAT_MODE_INFO_FULL_AUTO": "AI can edit files and run commands without approval (Click to switch)",
     "AI_CHAT_PERMISSION_SELECT_TITLE": "Choose how much the AI can do without approval [or Shift+Tab]",
     "AI_CHAT_AUTH_ERROR_NOTICE": "Claude Code is signed out or your login has expired.",

src/styles/Extn-AIChatPanel.less

@@ -454,6 +454,12 @@
 .dropdownbutton-popup.ai-permission-dropdown-popup {
     min-width: 220px;
     max-width: 320px;
+    // Override the shared .dropdownbutton-popup.dropdown-menu 160px cap
+    // (brackets_patterns_override.less) — 4 two-line items run ~200px,
+    // past that cap, which forced a scrollbar for a list that otherwise
+    // fits easily. Cap against the viewport instead so it only scrolls
+    // if it would genuinely run past the screen.
+    max-height: 90vh;
 
     li a {
         white-space: normal;
@@ -2726,6 +2732,10 @@
                 &.mode-plan {
                     background-color: #3498db;
                 }
+
+                &.mode-classifier {
+                    background-color: #9b59b6;
+                }
             }
 
             .ai-permission-label {

tracking-repos.json

@@ -1,5 +1,5 @@
 {
   "phoenixPro": {
-    "commitID": "4c5eea43d26daf438d288eb12ff968d1a9780b38"
+    "commitID": "ab180272ae33a5abbd53dcdc60eddb0c4d035a6c"
   }
 }