✨ feat(vscode): codex 命令统一组装，模型/思考级别可在设置面板配置

新增 cc/codexCommand.ts 作为 codex 调用的唯一出口（buildCodexArgs /
buildCodexCommandString），把 exec-review / resume / interactive 三种形态
的 flag 拼装收敛到一处；模型与思考级别作为自由文本项写入全局设置并贯穿
webview 消息通道，空值则不传 -c，沿用 codex config.toml 默认。

vscode/src/panel/handlers/sessions.ts

@@ -9,6 +9,7 @@ import * as path from 'node:path'
 import { window, workspace } from 'vscode'
 import { getToken } from '../../auth/secrets'
 import { buildCcCommand } from '../../cc/ccCommand'
+import { buildCodexCommandString } from '../../cc/codexCommand'
 import { watchForNewCodexSession } from '../../cc/codexSessionWatcher'
 import { getBrainstormContinuePrompt, getImplementPlanPrompt } from '../../cc/prompts'
 import { projectsDirFor, watchForNewSession } from '../../cc/sessionWatcher'
@@ -287,7 +288,13 @@ export async function handleResumeReviewSession(panel: KanbanWebviewPanel, sessi
     panel.reviewTerminals.set(sessionId, terminal)
     panel.trackSessionTerminal(terminal, issueNumber, 'review')
     terminal.show(false)
-    terminal.sendText(`codex resume -c model_reasoning_effort=xhigh --dangerously-bypass-approvals-and-sandbox ${sessionId}`)
+    const settings = getSettings(panel.context)
+    terminal.sendText(buildCodexCommandString({
+      mode: 'resume',
+      sessionId,
+      model: settings.codexModel,
+      reasoningEffort: settings.codexReasoningEffort,
+    }))
     logger.add({
       level: 'info',
       source: 'terminal',
@@ -418,7 +425,13 @@ export async function triggerAutoReviewTab(panel: KanbanWebviewPanel, opts: {
     panel.reviewTerminals.set(existingSessionId, terminal)
     panel.trackSessionTerminal(terminal, opts.issueNumber, 'review')
     terminal.show(false)
-    terminal.sendText(`codex resume -c model_reasoning_effort=xhigh --dangerously-bypass-approvals-and-sandbox ${existingSessionId}`)
+    const settings = getSettings(panel.context)
+    terminal.sendText(buildCodexCommandString({
+      mode: 'resume',
+      sessionId: existingSessionId,
+      model: settings.codexModel,
+      reasoningEffort: settings.codexReasoningEffort,
+    }))
     // codex resume rebuilds the conversation from the jsonl rollout
     // before the TUI accepts input — empirically much slower than the
     // 250ms gap the reuse path uses. 8s is a conservative guess that
@@ -494,7 +507,13 @@ export async function triggerAutoReviewTab(panel: KanbanWebviewPanel, opts: {
   // the terminal alive after the run so users can follow up with codex,
   // mirroring how claude implementation/brainstorm sessions stay open.
   // 提示词模板自带 /review 前缀，不在这里再拼。
-  const cmd = `codex -c model_reasoning_effort=xhigh --dangerously-bypass-approvals-and-sandbox '${opts.prompt}'`
+  const settings = getSettings(panel.context)
+  const cmd = buildCodexCommandString({
+    mode: 'interactive',
+    prompt: opts.prompt,
+    model: settings.codexModel,
+    reasoningEffort: settings.codexReasoningEffort,
+  })
   terminal.sendText(cmd)
   logger.add({
     level: 'info',