fix: add android platform support for Termux npm install

`npm install -g @gitlawb/zero` failed on Android/Termux with EBADPLATFORM
because `package.json` `os` array lacked `"android"`. Termux provides a
Linux userspace — the existing `linux-arm64` binary works natively.

- Add `"android"` to the `os` array in package.json.
- Map `android` platform to `linux` in postinstall.mjs resolvePlatform().

Fixes #449

README.md

@@ -8,6 +8,8 @@
   <a href="LICENSE"><img alt="license" src="https://img.shields.io/badge/license-MIT-blue"></a>
   <img alt="Go 1.25+" src="https://img.shields.io/badge/Go-1.25+-00ADD8?logo=go&logoColor=white">
   <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
+  <br>
+  <strong>English</strong> | <a href="README_ZH.md">中文</a>
 </p>
 
 Zero is an AI coding agent for your local terminal. It can inspect a repository,
@@ -51,6 +53,27 @@ The npm package installs a small wrapper plus the matching Zero binary for your
 platform from GitHub Releases. It supports Linux, macOS, and Windows on x64 and
 arm64.
 
+### Bun
+
+Bun does not run dependency lifecycle scripts by default, so the `postinstall`
+that fetches the Zero binary is skipped and the first run fails with
+`No native binary found next to the npm wrapper`.
+
+Install with Bun, then fetch the binary in one of two ways:
+
+```bash
+# Option A: run the installer manually
+bun add @gitlawb/zero
+node node_modules/@gitlawb/zero/scripts/postinstall.mjs
+
+# Option B: allow the postinstall to run on install
+# add to your package.json:  "trustedDependencies": ["@gitlawb/zero"]
+bun add @gitlawb/zero
+```
+
+For global installs (`bun add -g @gitlawb/zero`), use Option A against the
+global install path, or prefer the install scripts below.
+
 ### Install scripts
 
 Linux/macOS:
@@ -121,6 +144,13 @@ the key in the wizard:
 export OPENAI_API_KEY=sk-...
 export ANTHROPIC_API_KEY=...
 export GEMINI_API_KEY=...
+export LONGCAT_API_KEY=...
+```
+
+To configure Meituan LongCat (LongCat-2.0) directly, run:
+
+```bash
+zero providers setup longcat --set-active
 ```
 
 For local models, run Ollama or LM Studio and then use `zero setup` or

README_ZH.md

@@ -0,0 +1,297 @@
+<p align="center">
+  <img src="docs/assets/zero-logo.png" alt="Zero" width="385">
+</p>
+
+<p align="center"><strong>一个独属于你的终端编码智能体。</strong></p>
+
+<p align="center">
+  <a href="LICENSE"><img alt="license" src="https://img.shields.io/badge/license-MIT-blue"></a>
+  <img alt="Go 1.25+" src="https://img.shields.io/badge/Go-1.25+-00ADD8?logo=go&logoColor=white">
+  <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
+  <br>
+  <a href="README.md">English</a> | <strong>中文</strong>
+</p>
+
+Zero 是一个用于本地终端的 AI 编码智能体。它可以检查仓库、编辑文件、运行命令、使用浏览器/终端辅助工具，并在你选择模型和权限级别的同时保持持久的本地会话。
+
+```bash
+zero
+zero exec "修复 ./pkg 中失败的测试"
+zero exec --output-format stream-json < turns.jsonl
+```
+
+## 为什么选择 Zero
+
+- **使用你想要的模型。** 支持 OpenAI、Anthropic、Gemini、Groq、OpenRouter、DeepSeek、Mistral、xAI、Qwen、Kimi、GitHub Models、Ollama、LM Studio，或任何 OpenAI/Anthropic 兼容端点。
+- **保持控制权。** 文件写入、Shell 命令、网络访问和工作区外写入都经过 Zero 的权限和沙箱策略。
+- **在终端中工作。** TUI 具有模型/提供商选择器、图片输入、斜杠命令、实时计划/工具渲染、回滚滚动、主题以及恢复/分叉支持。
+- **无 TUI 也能工作。** `zero exec` 可脚本化，支持文本/JSON/stream-JSON I/O、隔离的工作树、规范优先运行，以及用于 CI 的有意义的退出码。
+- **保持上下文本地化。** 会话存储在磁盘上，可搜索、可恢复，且 Zero 不会作为遥测数据上传。
+- **可扩展。** 使用 MCP 服务器、技能、插件、钩子和来自同一 CLI 的专业子智能体。
+
+## 安装
+
+### npm
+
+```bash
+npm install -g @gitlawb/zero
+zero
+```
+
+npm 包安装一个小型包装器以及与你平台匹配的 Zero 二进制文件（从 GitHub Releases 获取）。支持 Linux、macOS 和 Windows 的 x64 和 arm64。
+
+### Bun
+
+Bun 默认不运行依赖的生命周期脚本，因此获取 Zero 二进制文件的 `postinstall` 会被跳过，首次运行会失败并显示 `No native binary found next to the npm wrapper`。
+
+使用 Bun 安装后，通过以下两种方式之一获取二进制文件：
+
+```bash
+# 选项 A：手动运行安装程序
+bun add @gitlawb/zero
+node node_modules/@gitlawb/zero/scripts/postinstall.mjs
+
+# 选项 B：允许 postinstall 在安装时运行
+# 添加到你的 package.json:  "trustedDependencies": ["@gitlawb/zero"]
+bun add @gitlawb/zero
+```
+
+对于全局安装（`bun add -g @gitlawb/zero`），请对全局安装路径使用选项 A，或使用下面的安装脚本。
+
+### 安装脚本
+
+Linux/macOS：
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.sh | bash
+```
+
+Windows PowerShell：
+
+```powershell
+irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | iex
+```
+
+### 从源码构建
+
+源码构建需要 Go 1.25+。
+
+```bash
+git clone https://github.com/Gitlawb/zero.git
+cd zero
+go run ./cmd/zero
+```
+
+发布安装器和 npm 包需要已发布的 GitHub Release 资源。如果你在首次公开发布之前进行测试，请从源码构建：
+
+```bash
+go build -o zero ./cmd/zero
+```
+
+在 Linux 上，如果你需要原生沙箱，还需要构建沙箱辅助程序：
+
+```bash
+go build -o zero-linux-sandbox ./cmd/zero-linux-sandbox
+go build -o zero-seccomp ./cmd/zero-seccomp   # 可选的兼容性包装器
+```
+
+将 `zero` 和 `zero-linux-sandbox` 放在 `PATH` 上的同一目录中（`~/.local/bin` 是一个好的默认选择）。macOS 不需要额外的辅助二进制文件。Windows 源码构建可以使用主 `zero.exe` 作为沙箱辅助程序；发布包仍然附带独立的 Windows 辅助可执行文件。
+
+更多安装细节：[docs/INSTALL.md](docs/INSTALL.md)。
+
+## 首次运行
+
+启动 TUI：
+
+```bash
+zero
+```
+
+设置向导帮助你选择提供商和模型。你也可以从命令行配置提供商：
+
+```bash
+zero setup
+zero providers list
+zero models list
+zero doctor
+```
+
+对于 API 提供商，在设置之前设置匹配的环境变量或在向导中输入密钥：
+
+```bash
+export OPENAI_API_KEY=sk-...
+export ANTHROPIC_API_KEY=...
+export GEMINI_API_KEY=...
+export LONGCAT_API_KEY=...
+```
+
+要直接配置美团 LongCat（LongCat-2.0），运行：
+
+```bash
+zero providers setup longcat --set-active
+```
+
+对于本地模型，运行 Ollama 或 LM Studio，然后使用 `zero setup` 或 `zero providers detect`。
+
+## 日常使用
+
+### 交互式 TUI
+
+```bash
+zero
+```
+
+常用控制：
+
+| 控制 | 操作 |
+|---|---|
+| `Enter` | 发送提示 |
+| `/` | 打开斜杠命令建议 |
+| `Shift+Tab` | 切换权限模式 |
+| `Ctrl+B` | 显示/隐藏侧边栏 |
+| `Ctrl+C` | 取消或退出 |
+
+常用斜杠命令：
+
+| 命令 | 用途 |
+|---|---|
+| `/model`、`/provider` | 切换活动模型/提供商 |
+| `/spec`、`/plan` | 在构建之前起草和审查计划 |
+| `/image` | 为视觉模型附加图片 |
+| `/resume`、`/rewind` | 继续或回滚本地会话 |
+| `/compact`、`/context` | 管理上下文使用 |
+| `/permissions`、`/tools` | 检查可用工具和策略 |
+| `/add-dir` | 为此会话授予额外的写入目录 |
+| `/theme`、`/doctor`、`/config` | 调整外观和检查设置 |
+
+### 无头 `exec` 模式
+
+```bash
+zero exec "解释 internal/agent/loop.go"
+zero exec --model claude-sonnet-4.5 "重构配置加载器"
+zero exec --use-spec "为 API 客户端添加速率限制"
+zero exec --worktree "在隔离的工作树中尝试迁移"
+zero exec --resume
+zero exec --fork <session-id> "尝试另一种方法"
+```
+
+编程使用：
+
+```bash
+zero exec --input-format stream-json --output-format stream-json < turns.jsonl
+```
+
+Stream-JSON 协议文档在 [docs/STREAM_JSON_PROTOCOL.md](docs/STREAM_JSON_PROTOCOL.md)。
+
+## 安全模型
+
+Zero 旨在使副作用可见。
+
+- 工作区读取默认允许。
+- 文件写入限制在工作区内，除非你授予其他目录。
+- Shell 命令、网络访问、破坏性命令和提权操作需要权限授权。
+- `--add-dir <path>` 和 `/add-dir <path>` 授予额外的写入根目录，而不会给智能体整个文件系统。
+- 不安全/自主模式是显式选择加入。
+- 在 Zero 控制的界面上，密钥会从工具输出和日志中被脱敏。
+
+示例：
+
+```bash
+zero --add-dir ../docs-site
+zero exec --add-dir ../shared "更新两个仓库"
+```
+
+沙箱行为可以通过以下方式检查：
+
+```bash
+zero sandbox policy
+zero sandbox grants list
+```
+
+## Web 和本地控制
+
+Zero 包含本地文件/搜索/编辑/Shell 工具、用于公共 URL 的 `web_fetch`，以及用于额外工具的 MCP 支持。
+
+对于本地开发服务器，使用 Shell 命令（如通过 `exec_command` 的 `curl`），这样正常的沙箱和权限策略就会生效。长时间运行的命令保持附加到后台终端会话，可以从 TUI 中列出或停止。
+
+npm 包还包含本地浏览器/终端工具使用的浏览器和终端辅助包。源码构建可以在它们位于 `PATH` 上或在 Zero 的本地控制设置中配置时使用相同的辅助工具。
+
+## 常用命令
+
+```text
+zero                  交互式 TUI
+zero exec             一次性或脚本化智能体运行
+zero setup            首次运行提供商设置
+zero auth             支持提供商的 OAuth/登录辅助
+zero models           模型注册表和能力
+zero providers        提供商配置和检测
+zero doctor           设置、密钥和连接检查
+zero context          上下文预算报告
+zero repo-map         确定性仓库映射
+zero repo-info        本地仓库摘要
+zero search | find    搜索本地会话历史
+zero sessions         检查、恢复、分叉和回滚会话
+zero spec             管理规范模式草稿
+zero specialist       管理专业子智能体
+zero skills           管理 Markdown 指令技能
+zero plugins          管理插件
+zero hooks            管理生命周期钩子
+zero mcp              管理 MCP 服务器和工具
+zero serve --mcp      通过 MCP stdio 暴露 Zero 工具
+zero sandbox          检查沙箱策略和授权
+zero worktrees        准备隔离的 Git 工作树
+zero verify           检测和运行本地验证检查
+zero changes          检查和提交本地 Git 变更
+zero usage            Token 使用量和估算成本
+zero cron             定时智能体任务
+zero update           检查更新版本
+```
+
+## 外观和无障碍
+
+| 控制 | 效果 |
+|---|---|
+| `NO_COLOR=<任意值>` | 禁用颜色输出 |
+| `ZERO_THEME=<名称>` | 选择启动主题（`auto`、`dark`、`light`，或颜色主题如 `dracula`、`nord`、`gruvbox`、`tokyo-night`、`catppuccin`、`one-dark`、`solarized-dark`、`rose-pine`、`everforest`、`solarized-light`） |
+| `--theme <名称>` | 从 CLI 选择 TUI 主题（相同名称） |
+| `/theme` | 在 TUI 中打开主题选择器（实时预览；`/theme <名称>` 直接切换） |
+| `ZERO_NO_FADE=1` | 禁用流式淡入动画 |
+
+含义不仅仅依赖于颜色；差异、权限和状态也使用文本或标记符号。
+
+## 开发
+
+```bash
+go test ./...
+go run ./cmd/zero-release build
+go run ./cmd/zero-release smoke
+go run ./cmd/zero-perf-bench
+```
+
+交叉编译示例：
+
+```bash
+go run ./cmd/zero-release build --goos linux --goarch amd64
+go run ./cmd/zero-release build --goos windows --goarch amd64 --output dist/zero.exe
+```
+
+## 文档
+
+- [安装](docs/INSTALL.md)
+- [更新流程](docs/UPDATE.md)
+- [Stream-JSON 协议](docs/STREAM_JSON_PROTOCOL.md)
+- [专家](docs/SPECIALISTS.md)
+- [GitHub Action](docs/GITHUB_ACTION.md)
+- [基准测试](docs/BENCHMARK.md)
+- [性能](docs/PERFORMANCE.md)
+- [智能体评估](docs/AGENT_EVALS.md)
+
+## 贡献
+
+欢迎贡献。阅读 [CONTRIBUTING.md](CONTRIBUTING.md)，运行相关测试，然后提交一个聚焦的拉取请求。
+
+安全报告应遵循 [SECURITY.md](SECURITY.md)。
+
+## 许可证
+
+Zero 基于 [MIT 许可证](LICENSE) 发布。

bin/zero.js

@@ -54,8 +54,26 @@ const nativePath = join(packageRoot, zeroBinaryName());
 const localControlHelpers = localControlHelperManifest(packageRoot);
 
 if (!existsSync(nativePath)) {
+  const postinstallScript = join(packageRoot, 'scripts', 'postinstall.mjs');
+  const ranByBun = process.execPath.includes('bun') || !!process.versions?.bun;
   console.error(
-    '[zero] No native binary found next to the npm wrapper. Reinstall the zero package or run `go run ./cmd/zero-release build` from the repository.'
+    '[zero] No native binary found next to the npm wrapper.\n' +
+      'The platform binary is fetched at install time by a postinstall script,\n' +
+      'which did not run (or was skipped) for this install.\n' +
+      '\n' +
+      'Fix it now by running the installer manually:\n' +
+      `  node "${postinstallScript}"\n` +
+      '\n' +
+      (ranByBun
+        ? 'You installed with Bun, which does not run dependency lifecycle scripts\n' +
+          'by default. To let it run on future installs, add zero to your\n' +
+          'project\'s trustedDependencies:\n' +
+          '  "trustedDependencies": ["@gitlawb/zero"]\n' +
+          'then reinstall.\n' +
+          '\n'
+        : '') +
+      'If that fails, build from source: https://github.com/Gitlawb/zero\n' +
+      '(go run ./cmd/zero, requires Go 1.25+).',
   );
   process.exit(1);
 }

docs/INSTALL.md

@@ -24,6 +24,38 @@ Requirements:
 - Node.js 18+
 - network access to npm and GitHub Releases
 
+## Bun
+
+Bun is "default-secure" and does not run lifecycle scripts of installed
+dependencies (only the installing project's own scripts), so the `postinstall`
+that fetches the Zero binary is silently skipped. The first run then fails with
+`No native binary found next to the npm wrapper`.
+
+To install with Bun, either run the installer manually after installing:
+
+```bash
+bun add @gitlawb/zero
+node node_modules/@gitlawb/zero/scripts/postinstall.mjs
+```
+
+Or allow the postinstall to run by adding the package to your project's
+`trustedDependencies` before installing:
+
+```json
+{
+  "trustedDependencies": ["@gitlawb/zero"]
+}
+```
+
+```bash
+bun add @gitlawb/zero
+```
+
+For global installs (`bun add -g @gitlawb/zero`), run the installer manually
+against the global install path, or use the install scripts below.
+
+Reference: <https://bun.sh/docs/pm/lifecycle>
+
 ## Linux And macOS Script
 
 Install the latest release:

internal/agent/guardrails.go

@@ -2,7 +2,6 @@ package agent
 
 import (
 	"encoding/json"
-	"fmt"
 	"strconv"
 	"strings"
 
@@ -285,67 +284,12 @@ const (
 	noOutputStopSuffix = "to avoid consuming tokens without making progress."
 )
 
-// providerEmptyStream reports that a cleanly-completed turn carried NOTHING —
-// no visible text, no tool calls (real or dropped), and no reasoning signal.
-// That shape is a transport/backend fault (e.g. the ollama cloud relay
-// answering 200 with an instantly-empty SSE under load), not model behavior:
-// a model that is generating always leaves SOME trace (text, a tool call
-// attempt, or reasoning deltas).
-func providerEmptyStream(collected zeroruntime.CollectedStream) bool {
-	return strings.TrimSpace(collected.Text) == "" &&
-		len(collected.ToolCalls) == 0 &&
-		collected.DroppedToolCalls == 0 &&
-		len(collected.ReasoningBlocks) == 0 &&
-		!collected.ReasoningEmitted
-}
-
-// allEmptyTurnsProviderEmpty reports that every strike toward the no-output
-// guard was a provider-empty stream, so the stop message should name the
-// backend fault rather than generic agent no-progress.
-func (state *guardState) allEmptyTurnsProviderEmpty() bool {
-	return state.emptyTurns > 0 && state.providerEmptyTurns == state.emptyTurns
-}
-
 // noOutputStopAnswer is the final answer returned when the no-output guard
 // stops the run. The turn count is interpolated at the call site.
 func noOutputStopAnswer(turns int) string {
 	return noOutputStopPrefix + strconv.Itoa(turns) + " turns " + noOutputStopMarker + " " + noOutputStopSuffix
 }
 
-// providerEmptyStopAnswer is the final answer when every guard strike was a
-// provider-empty stream: the backend repeatedly answered with contentless
-// completions (already retried with backoff inside each turn). It tells the
-// user the truth — the provider failed, not the agent — and what to do next.
-// detail carries wire-level diagnostics from the last empty attempt
-// (finish reason, token counts) so the condition is debuggable straight from
-// the session log — this exact failure class previously took a live traffic
-// capture to diagnose because nothing about the response shape was recorded.
-func providerEmptyStopAnswer(turns int, detail string) string {
-	answer := providerEmptyStopPrefix +
-		strconv.Itoa(turns) + " consecutive attempts, each already retried with backoff. " +
-		"The backend is likely rate-limiting or degraded right now — try again shortly, or switch providers/models (zero providers use <name>)."
-	if detail != "" {
-		answer += " [last attempt: " + detail + "]"
-	}
-	return answer
-}
-
-// emptyStreamDetail renders the wire-level shape of an empty attempt for
-// providerEmptyStopAnswer.
-func emptyStreamDetail(collected zeroruntime.CollectedStream) string {
-	finish := collected.FinishReason
-	if finish == "" {
-		finish = "stop"
-	}
-	return fmt.Sprintf("finish_reason=%s, output_tokens=%d, reasoning_tokens=%d",
-		finish, collected.Usage.OutputTokens, collected.Usage.ReasoningTokens)
-}
-
-// providerEmptyStopPrefix is the stable head of providerEmptyStopAnswer, used
-// by IsNoProgressStop so session titling / resume filtering treat the
-// provider-empty stop like the other guard answers (not real content).
-const providerEmptyStopPrefix = "The model provider returned an empty response (no text, no tool calls, no reasoning) on "
-
 // IsNoProgressStop reports whether content IS the no-output guardrail stop answer
 // (a run that produced no visible text and no tool calls). It matches the EXACT
 // structure noOutputStopAnswer emits — prefix + "<int> turns " + marker + " " +
@@ -356,15 +300,6 @@ const providerEmptyStopPrefix = "The model provider returned an empty response (
 // and skip its title generation.
 func IsNoProgressStop(content string) bool {
 	trimmed := strings.TrimSpace(content)
-	// The provider-empty variant: stable prefix + a bare integer count.
-	if rest, ok := strings.CutPrefix(trimmed, providerEmptyStopPrefix); ok {
-		if sep := strings.Index(rest, " consecutive attempts"); sep > 0 {
-			if _, err := strconv.Atoi(rest[:sep]); err == nil {
-				return true
-			}
-		}
-		return false
-	}
 	if !strings.HasPrefix(trimmed, noOutputStopPrefix) {
 		return false
 	}
@@ -417,7 +352,6 @@ func toolOnlyProgressReminder(turns int) string {
 // purely from tool-call names and per-turn output, matching what the loop holds.
 type guardState struct {
 	emptyTurns               int
-	providerEmptyTurns       int
 	totalToolCalls           int
 	toolCallsSincePlanUpdate int
 	planEverCalled           bool
@@ -483,18 +417,8 @@ func (state *guardState) observeTurn(collected zeroruntime.CollectedStream) (sto
 
 	if hasToolCalls || hasVisibleText {
 		state.emptyTurns = 0
-		state.providerEmptyTurns = 0
 	} else {
 		state.emptyTurns++
-		// Track separately whether the empty turn was a PROVIDER-empty stream
-		// (nothing at all came back — not even reasoning) vs a behavioral empty
-		// (the model thought/streamed but committed no answer). When every
-		// strike was provider-empty, the stop message names the backend fault
-		// instead of blaming the agent — the difference between the user
-		// retrying/switching providers and filing a "the agent gives up" bug.
-		if providerEmptyStream(collected) {
-			state.providerEmptyTurns++
-		}
 	}
 	if hasToolCalls && !hasVisibleText {
 		state.toolOnlyTurns++

internal/agent/guardrails_test.go

@@ -9,23 +9,11 @@ import (
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
-// emptyTurn is a PROVIDER-empty stream: it completes cleanly carrying nothing
-// at all (the ollama-cloud 200+[DONE] shape). The loop retries these in-turn
-// before counting a no-output strike.
+// emptyTurn is a stream that produces no visible text and no tool calls.
 func emptyTurn() []zeroruntime.StreamEvent {
 	return []zeroruntime.StreamEvent{{Type: zeroruntime.StreamEventDone}}
 }
 
-// reasoningOnlyTurn is a BEHAVIORAL empty turn: the model streamed reasoning
-// (so the provider clearly worked) but committed no text and no tool calls.
-// These are NOT retried in-turn; they count directly toward the no-output guard.
-func reasoningOnlyTurn() []zeroruntime.StreamEvent {
-	return []zeroruntime.StreamEvent{
-		{Type: zeroruntime.StreamEventReasoning, Content: "thinking…"},
-		{Type: zeroruntime.StreamEventDone},
-	}
-}
-
 // textTurn produces a turn with visible assistant text.
 func textTurn(content string) []zeroruntime.StreamEvent {
 	return []zeroruntime.StreamEvent{
@@ -57,9 +45,9 @@ func countUserMessagesContaining(messages []zeroruntime.Message, needle string)
 func TestRunStopsAfterConsecutiveEmptyTurns(t *testing.T) {
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
-			reasoningOnlyTurn(),
-			reasoningOnlyTurn(),
-			reasoningOnlyTurn(),
+			emptyTurn(),
+			emptyTurn(),
+			emptyTurn(),
 			// A 4th turn exists but must never be requested.
 			textTurn("should never reach here"),
 		},
@@ -89,10 +77,10 @@ func TestRunStopsAfterConsecutiveEmptyTurns(t *testing.T) {
 func TestRunResetsEmptyTurnCounterOnVisibleOutput(t *testing.T) {
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
-			reasoningOnlyTurn(),
-			reasoningOnlyTurn(),
+			emptyTurn(),
+			emptyTurn(),
 			textTurn("here is real progress"), // resets the counter and is the final answer
-			reasoningOnlyTurn(),
+			emptyTurn(),
 		},
 	}
 
@@ -121,11 +109,11 @@ func TestRunResetsEmptyTurnCounterOnToolCall(t *testing.T) {
 
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
-			reasoningOnlyTurn(),
-			reasoningOnlyTurn(),
+			emptyTurn(),
+			emptyTurn(),
 			toolTurn("call-1", "read_file", `{"path":"notes.txt"}`), // resets counter
-			reasoningOnlyTurn(),
-			reasoningOnlyTurn(),
+			emptyTurn(),
+			emptyTurn(),
 			textTurn("done"),
 		},
 	}
@@ -467,96 +455,3 @@ func TestRunInjectsToolFailureHintWithSchema(t *testing.T) {
 		t.Fatalf("expected a tool-failure hint on the 3rd turn, messages: %+v", provider.requests[2].Messages)
 	}
 }
-
-// A provider-empty stream (clean completion carrying nothing at all) is
-// retried in-turn with backoff before it may count as a no-output strike; when
-// the backend keeps answering empty, the run stops with a message that names
-// the PROVIDER fault — not the generic "agent made no progress" text. This is
-// the exact live failure observed on the ollama cloud relay (HTTP 200 in ~1s,
-// SSE straight to [DONE]), which silently killed 56% of sessions on the
-// reporting machine.
-func TestRunRetriesProviderEmptyStreamsThenStopsWithProviderMessage(t *testing.T) {
-	turns := make([][]zeroruntime.StreamEvent, 0, 9)
-	for i := 0; i < 9; i++ { // 3 strikes × (1 initial + 2 in-turn retries)
-		turns = append(turns, emptyTurn())
-	}
-	provider := &mockProvider{turns: turns}
-
-	result, err := Run(context.Background(), "go", provider, Options{
-		Registry: tools.NewRegistry(),
-		MaxTurns: 12,
-	})
-	if err != nil {
-		t.Fatal(err)
-	}
-	wantRequests := maxEmptyTurns * (1 + maxEmptyStreamRetries)
-	if len(provider.requests) != wantRequests {
-		t.Fatalf("expected %d requests (%d strikes × %d attempts), got %d",
-			wantRequests, maxEmptyTurns, 1+maxEmptyStreamRetries, len(provider.requests))
-	}
-	if !strings.Contains(result.FinalAnswer, "provider returned an empty response") {
-		t.Fatalf("expected the provider-empty stop message, got %q", result.FinalAnswer)
-	}
-	if strings.Contains(result.FinalAnswer, noOutputStopMarker) {
-		t.Fatalf("provider fault must not be reported as generic agent no-progress: %q", result.FinalAnswer)
-	}
-	if !IsNoProgressStop(result.FinalAnswer) {
-		t.Fatal("the provider-empty stop answer must be recognized by IsNoProgressStop (titling/resume filters)")
-	}
-	// Wire-level diagnostics must land in the message (and therefore the
-	// session log) so this failure class is debuggable without a traffic capture.
-	if !strings.Contains(result.FinalAnswer, "output_tokens=0") || !strings.Contains(result.FinalAnswer, "finish_reason=") {
-		t.Fatalf("stop answer must carry last-attempt diagnostics, got %q", result.FinalAnswer)
-	}
-}
-
-// A transient empty response recovers on the in-turn retry: no strike, no stop,
-// the retried turn's real output is the answer.
-func TestRunRecoversWhenEmptyStreamRetrySucceeds(t *testing.T) {
-	provider := &mockProvider{
-		turns: [][]zeroruntime.StreamEvent{
-			emptyTurn(),           // initial attempt: backend hiccup
-			textTurn("recovered"), // in-turn retry gets the real answer
-		},
-	}
-
-	result, err := Run(context.Background(), "go", provider, Options{
-		Registry: tools.NewRegistry(),
-		MaxTurns: 12,
-	})
-	if err != nil {
-		t.Fatal(err)
-	}
-	if len(provider.requests) != 2 {
-		t.Fatalf("expected 2 requests (initial + one retry), got %d", len(provider.requests))
-	}
-	if result.FinalAnswer != "recovered" {
-		t.Fatalf("expected the retried turn's text as final answer, got %q", result.FinalAnswer)
-	}
-	if result.Turns != 1 {
-		t.Fatalf("an in-turn retry is the SAME turn, want 1 turn, got %d", result.Turns)
-	}
-}
-
-// Mixed strikes (provider-empty + behavioral reasoning-only) must NOT claim a
-// provider fault — the generic no-output message stays.
-func TestRunMixedEmptyTurnsKeepGenericStopMessage(t *testing.T) {
-	provider := &mockProvider{
-		turns: [][]zeroruntime.StreamEvent{
-			reasoningOnlyTurn(),                   // strike 1: behavioral
-			emptyTurn(), emptyTurn(), emptyTurn(), // strike 2: provider-empty ×(1+2 retries)
-			reasoningOnlyTurn(), // strike 3: behavioral
-		},
-	}
-
-	result, err := Run(context.Background(), "go", provider, Options{
-		Registry: tools.NewRegistry(),
-		MaxTurns: 12,
-	})
-	if err != nil {
-		t.Fatal(err)
-	}
-	if !strings.Contains(result.FinalAnswer, noOutputStopMarker) {
-		t.Fatalf("mixed strikes must use the generic no-output message, got %q", result.FinalAnswer)
-	}
-}

internal/agent/loop.go

@@ -25,20 +25,7 @@ const maxTurnsFinalAnswerPrompt = "You have reached the tool-turn limit. Do not
 // WITH NO OUTPUT yet is re-issued on a fresh connection before giving up. Only
 // the no-output case is retried (a partial turn would duplicate), so this is a
 // safe recovery for a stalled/dead pooled connection.
-// maxStreamStallRetries: 3 in-turn re-issues (4 attempts total) approaches
-// Codex's stream_max_retries=5 posture while staying hard-capped (opencode's
-// unbounded session retry has produced infinite loops on this same failure).
-// Affordable now that a mid-stream death is detected at the gap timeout
-// (~2m) instead of the full 5m idle window.
-const maxStreamStallRetries = 3
-
-// maxEmptyStreamRetries caps in-turn re-issues of a request whose stream
-// completed cleanly but carried nothing at all (providerEmptyStream). Observed
-// live on the ollama cloud relay under load: HTTP 200 in ~1s with an SSE that
-// goes straight to [DONE]. The condition is transient backend state, so a
-// short backoff-retry usually recovers; without it three such responses
-// arrived inside two seconds and killed the run via the no-output guard.
-const maxEmptyStreamRetries = 2
+const maxStreamStallRetries = 2
 
 const (
 	toolResultMetaControl       = "control"
@@ -337,35 +324,6 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			}
 			collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, forwardingOpts)
 		}
-		// Provider-empty retry: the stream completed CLEANLY but carried nothing
-		// — no text, no tool calls, no reasoning. That is not model behavior (a
-		// generating model always leaves some trace); it is a backend answering
-		// with a contentless completion (seen on the ollama cloud relay under
-		// load: 200 + instant [DONE]). Transient, so re-issue with backoff before
-		// letting it count as a no-output strike. Nothing from the empty stream
-		// was forwarded or committed to history, so the retry is clean.
-		for attempt := 1; attempt <= maxEmptyStreamRetries &&
-			collected.Error == "" && !forwardedVisibleText &&
-			providerEmptyStream(collected); attempt++ {
-			if notify := emptyRetryNoticeFor(options); notify != nil {
-				notify(attempt, maxEmptyStreamRetries)
-			}
-			if err := sleepWithContext(ctx, backoffFor(attempt)); err != nil {
-				result.Messages = copyMessages(messages)
-				return result, err
-			}
-			retryRequest := zeroruntime.CompletionRequest{
-				Messages:        copyMessages(messages),
-				Tools:           exposed,
-				ReasoningEffort: options.ReasoningEffort,
-			}
-			retryStream, retryErr := streamWithReconnect(ctx, provider, retryRequest, reconnectNoticeFor(options))
-			if retryErr != nil {
-				result.Messages = copyMessages(messages)
-				return result, retryErr
-			}
-			collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, forwardingOpts)
-		}
 		if collected.Error != "" {
 			// Route a reissued stream's non-stall error through the SAME recovery as
 			// the initial stream (image-rejection wrapping / context-limit compaction)
@@ -412,15 +370,6 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			// A truly-empty turn (no text, no tool calls, no dropped calls) is
 			// counted toward the runaway cap so we stop before burning maxTurns.
 			if guards.observeTurn(collected) {
-				// Every strike being a provider-empty stream means the BACKEND
-				// failed repeatedly (each attempt already retried with backoff
-				// above) — say so, instead of the generic no-progress message
-				// that reads as the agent giving up.
-				if guards.allEmptyTurnsProviderEmpty() {
-					result.FinalAnswer = providerEmptyStopAnswer(result.Turns, emptyStreamDetail(collected))
-					result.Messages = copyMessages(messages)
-					return result, nil
-				}
 				result.FinalAnswer = noOutputStopAnswer(result.Turns)
 				result.Messages = copyMessages(messages)
 				return result, nil
@@ -2210,6 +2159,18 @@ func buildPermissionEvent(call ToolCall, tool tools.Tool, args map[string]any, p
 		}
 	}
 
+	// A command that explicitly requests additional sandbox permissions
+	// (sandbox_permissions: with_additional_permissions) is an ELEVATION the user
+	// must consent to, even when the base command's sandbox decision was Allow.
+	// shouldRequestPermission blocks on OnPermissionRequest for exactly this case,
+	// so the event MUST be a prompt — otherwise it carried Action=allow while the
+	// loop waited on a decision, and a UI that renders only prompts (the TUI)
+	// dropped it, deadlocking the run. Keep this consistent with
+	// shouldRequestPermission's own condition.
+	if action == PermissionActionAllow && !permissionGranted && shellCommandAdditionalPermissionsRequested(args) {
+		action = PermissionActionPrompt
+	}
+
 	if safety.Permission == tools.PermissionAllow && action == PermissionActionAllow && !grantMatched && block == nil {
 		return PermissionEvent{}, false
 	}

internal/agent/permission_additional_perms_test.go

@@ -0,0 +1,70 @@
+package agent
+
+import (
+	"context"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+// promptShellTool is a minimal shell-style tool whose Safety requires a prompt,
+// standing in for bash so buildPermissionEvent can be exercised directly.
+type promptShellTool struct{}
+
+func (promptShellTool) Name() string        { return "bash" }
+func (promptShellTool) Description() string { return "run a shell command" }
+func (promptShellTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", AdditionalProperties: false}
+}
+func (promptShellTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tools.SideEffectShell, Permission: tools.PermissionPrompt, Reason: "runs a shell command"}
+}
+func (promptShellTool) Run(context.Context, map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK}
+}
+
+// A command whose sandbox decision is Allow but that explicitly requests
+// additional sandbox permissions is an ELEVATION: it must be surfaced as a
+// prompt, never Action=allow. Previously it carried allow while the loop still
+// blocked on OnPermissionRequest, and the TUI (which renders only prompts)
+// dropped it → the run deadlocked forever. This is the exact gpt-5.5
+// folder-creation shape captured in the reported hang.
+func TestBuildPermissionEventPromptsForAdditionalPermissions(t *testing.T) {
+	call := ToolCall{ID: "call_1", Name: "bash", Arguments: `{}`}
+	args := map[string]any{
+		"command":             "mkdir foo && cd foo && go mod init foo",
+		"sandbox_permissions": string(tools.SandboxPermissionsWithAdditionalPermissions),
+	}
+	// Base command allowed by the sandbox…
+	decision := &sandbox.Decision{Action: sandbox.ActionAllow}
+
+	event, ok := buildPermissionEvent(call, promptShellTool{}, args, false, PermissionModeAsk, Options{}, decision)
+	if !ok {
+		t.Fatal("buildPermissionEvent returned ok=false; the elevation request must produce an event")
+	}
+	if event.Action != PermissionActionPrompt {
+		t.Fatalf("Action = %q, want prompt — an additional-permissions elevation must ask the user, not auto-allow", event.Action)
+	}
+	// Sanity: shouldRequestPermission agrees the loop blocks on this.
+	if !shouldRequestPermission(promptShellTool{}, args, false, decision) {
+		t.Fatal("shouldRequestPermission must be true for an additional-permissions request")
+	}
+}
+
+// Without the additional-permissions flag, a sandbox-allowed command is NOT a
+// prompt (shouldRequestPermission is false; it just runs) — the override must
+// not over-reach and prompt for ordinary allowed commands.
+func TestBuildPermissionEventKeepsAllowForOrdinaryAllowedCommand(t *testing.T) {
+	call := ToolCall{ID: "call_2", Name: "bash", Arguments: `{}`}
+	args := map[string]any{"command": "ls"}
+	decision := &sandbox.Decision{Action: sandbox.ActionAllow}
+
+	event, _ := buildPermissionEvent(call, promptShellTool{}, args, false, PermissionModeAsk, Options{}, decision)
+	if event.Action == PermissionActionPrompt {
+		t.Fatalf("Action = prompt for an ordinary allowed command; want allow (no spurious prompt)")
+	}
+	if shouldRequestPermission(promptShellTool{}, args, false, decision) {
+		t.Fatal("shouldRequestPermission must be false for an ordinary sandbox-allowed command")
+	}
+}

internal/agent/reconnect.go

@@ -55,22 +55,6 @@ func stallRetryNoticeFor(options Options) reconnectNotifier {
 	}
 }
 
-// emptyRetryNoticeFor builds a notifier for the loop's provider-empty retry
-// (a stream that completed cleanly but carried NOTHING — no text, no tool
-// calls, no reasoning). Distinct wording from the stall notice because nothing
-// timed out: the backend answered instantly with an empty completion (observed
-// on the ollama cloud relay under load, and possible on any gateway having a
-// bad moment). Surfaced through OnReasoning, the non-content channel. Nil when
-// there is no reasoning sink (the retry still happens silently).
-func emptyRetryNoticeFor(options Options) reconnectNotifier {
-	if options.OnReasoning == nil {
-		return nil
-	}
-	return func(attempt, max int) {
-		options.OnReasoning(fmt.Sprintf("\n[provider returned an empty response — retrying %d/%d…]\n", attempt, max))
-	}
-}
-
 // streamWithReconnect issues request via provider.StreamCompletion and, on a
 // transient disconnect error, retries the connect up to maxStreamReconnects
 // times with exponential backoff. It returns the live stream on success, or the

internal/cli/provider_catalog_onboarding_test.go

@@ -18,6 +18,7 @@ func TestProviderCatalogRuntimeProvidersShowOnboardingSetupHints(t *testing.T) {
 	}{
 		{id: "openai", name: "OpenAI", transport: "openai", defaultModel: "gpt-4.1"},
 		{id: "groq", name: "Groq", transport: "openai-compatible", defaultModel: "llama-3.3-70b-versatile"},
+		{id: "longcat", name: "LongCat", transport: "openai-compatible", defaultModel: "LongCat-2.0"},
 		{id: "ollama-cloud", name: "Ollama Cloud", transport: "openai-compatible", defaultModel: "qwen3-coder:480b"},
 		{id: "ollama", name: "Ollama Local", transport: "openai-compatible", defaultModel: "llama3.1"},
 	}
@@ -46,6 +47,21 @@ func TestProviderCatalogRuntimeProvidersShowOnboardingSetupHints(t *testing.T) {
 	}
 }
 
+func TestProviderCatalogLongCatRequiresAPIKey(t *testing.T) {
+	output := runProviderCatalogOnboarding(t)
+	block := providerCatalogOnboardingBlock(t, output, "longcat")
+
+	for _, want := range []string{
+		"requiresAuth=true",
+		"authEnvVars=LONGCAT_API_KEY",
+		"setup: zero providers setup longcat --set-active",
+	} {
+		if !strings.Contains(block, want) {
+			t.Fatalf("expected LongCat catalog block to contain %q, got:\n%s", want, block)
+		}
+	}
+}
+
 func TestProviderCatalogOllamaCloudRequiresAPIKey(t *testing.T) {
 	output := runProviderCatalogOnboarding(t)
 	block := providerCatalogOnboardingBlock(t, output, "ollama-cloud")

internal/config/resolver_test.go

@@ -1059,6 +1059,130 @@ func TestResolveAppliesProviderCatalogDefaults(t *testing.T) {
 	}
 }
 
+func TestResolveAppliesMiniMaxCNCatalogDefaults(t *testing.T) {
+	path := writeConfig(t, `{
+		"activeProvider": "mini-cn",
+		"providers": [{
+			"name": "mini-cn",
+			"catalog_id": "minimaxi-cn"
+		}]
+	}`)
+
+	resolved, err := Resolve(ResolveOptions{
+		ProjectConfigPath: path,
+		Env: map[string]string{
+			"MINIMAXI_API_KEY": "sk-cn-mini",
+		},
+	})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+
+	if resolved.Provider.CatalogID != "minimaxi-cn" {
+		t.Fatalf("CatalogID = %q, want minimaxi-cn", resolved.Provider.CatalogID)
+	}
+	if resolved.Provider.ProviderKind != ProviderKindAnthropicCompat {
+		t.Fatalf("ProviderKind = %q, want %q", resolved.Provider.ProviderKind, ProviderKindAnthropicCompat)
+	}
+	if resolved.Provider.BaseURL != "https://api.minimaxi.com/anthropic" {
+		t.Fatalf("BaseURL = %q, want MiniMax CN default", resolved.Provider.BaseURL)
+	}
+	if resolved.Provider.Model != "MiniMax-M3" {
+		t.Fatalf("Model = %q, want MiniMax-M3", resolved.Provider.Model)
+	}
+	if resolved.Provider.APIKeyEnv != "MINIMAXI_API_KEY" {
+		t.Fatalf("APIKeyEnv = %q, want MINIMAXI_API_KEY", resolved.Provider.APIKeyEnv)
+	}
+	if resolved.Provider.APIKey != "sk-cn-mini" {
+		t.Fatalf("APIKey = %q, want env-resolved secret", resolved.Provider.APIKey)
+	}
+}
+
+func TestResolveAppliesZaiCNCatalogDefaults(t *testing.T) {
+	// The "zai-cn" catalog entry preserves the legacy open.bigmodel.cn endpoint
+	// for users who still target the China hosting; "zai" itself now points at
+	// api.z.ai (international). The two env-vars are intentionally distinct
+	// (ZHIPU_API_KEY for China, ZAI_API_KEY for international) so an account on
+	// one side cannot be silently used against the other.
+	path := writeConfig(t, `{
+		"activeProvider": "zai-cn",
+		"providers": [{
+			"name": "zai-cn",
+			"catalog_id": "zai-cn"
+		}]
+	}`)
+
+	resolved, err := Resolve(ResolveOptions{
+		ProjectConfigPath: path,
+		Env: map[string]string{
+			"ZHIPU_API_KEY": "sk-zhipu-cn",
+		},
+	})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+
+	if resolved.Provider.CatalogID != "zai-cn" {
+		t.Fatalf("CatalogID = %q, want zai-cn", resolved.Provider.CatalogID)
+	}
+	if resolved.Provider.ProviderKind != ProviderKindOpenAICompatible {
+		t.Fatalf("ProviderKind = %q, want %q", resolved.Provider.ProviderKind, ProviderKindOpenAICompatible)
+	}
+	if resolved.Provider.BaseURL != "https://open.bigmodel.cn/api/paas/v4" {
+		t.Fatalf("BaseURL = %q, want Z.ai China default", resolved.Provider.BaseURL)
+	}
+	if resolved.Provider.Model != "glm-4.5" {
+		t.Fatalf("Model = %q, want glm-4.5 (China default)", resolved.Provider.Model)
+	}
+	if resolved.Provider.APIKeyEnv != "ZHIPU_API_KEY" {
+		t.Fatalf("APIKeyEnv = %q, want ZHIPU_API_KEY", resolved.Provider.APIKeyEnv)
+	}
+	if resolved.Provider.APIKey != "sk-zhipu-cn" {
+		t.Fatalf("APIKey = %q, want env-resolved secret", resolved.Provider.APIKey)
+	}
+}
+
+func TestResolveAppliesZaiInternationalCatalogDefaults(t *testing.T) {
+	// "zai" now resolves to the international endpoint; the catalog default
+	// model stays glm-4.5 (live discovery upgrades to glm-5.2 at runtime).
+	path := writeConfig(t, `{
+		"activeProvider": "zai-intl",
+		"providers": [{
+			"name": "zai-intl",
+			"catalog_id": "zai"
+		}]
+	}`)
+
+	resolved, err := Resolve(ResolveOptions{
+		ProjectConfigPath: path,
+		Env: map[string]string{
+			"ZAI_API_KEY": "sk-zai-intl",
+		},
+	})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+
+	if resolved.Provider.CatalogID != "zai" {
+		t.Fatalf("CatalogID = %q, want zai", resolved.Provider.CatalogID)
+	}
+	if resolved.Provider.ProviderKind != ProviderKindOpenAICompatible {
+		t.Fatalf("ProviderKind = %q, want %q", resolved.Provider.ProviderKind, ProviderKindOpenAICompatible)
+	}
+	if resolved.Provider.BaseURL != "https://api.z.ai/api/paas/v4" {
+		t.Fatalf("BaseURL = %q, want Z.ai international default", resolved.Provider.BaseURL)
+	}
+	if resolved.Provider.Model != "glm-4.5" {
+		t.Fatalf("Model = %q, want glm-4.5 (preserved default; live discovery upgrades to glm-5.2 at runtime)", resolved.Provider.Model)
+	}
+	if resolved.Provider.APIKeyEnv != "ZAI_API_KEY" {
+		t.Fatalf("APIKeyEnv = %q, want ZAI_API_KEY", resolved.Provider.APIKeyEnv)
+	}
+	if resolved.Provider.APIKey != "sk-zai-intl" {
+		t.Fatalf("APIKey = %q, want env-resolved secret", resolved.Provider.APIKey)
+	}
+}
+
 func TestResolveProviderCatalogAliasAndLocalNoAuth(t *testing.T) {
 	path := writeConfig(t, `{
 		"activeProvider": "local",

internal/mcp/client_test.go

@@ -11,6 +11,7 @@ import (
 	"os"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"testing"
 	"time"
 
@@ -211,6 +212,85 @@ func TestHTTPClientListsAndCallsTools(t *testing.T) {
 	}
 }
 
+func TestHTTPClientFollowsSameOriginRedirect(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	testServer := httptest.NewServer(http.HandlerFunc(func(response http.ResponseWriter, request *http.Request) {
+		if request.URL.Path == "/mcp" {
+			http.Redirect(response, request, "/redirected", http.StatusTemporaryRedirect)
+			return
+		}
+		if request.Method != http.MethodPost || request.URL.Path != "/redirected" {
+			t.Errorf("request = %s %s, want POST /redirected", request.Method, request.URL.Path)
+			http.Error(response, "bad request", http.StatusNotFound)
+			return
+		}
+		if got := request.Header.Get("X-Api-Key"); got != "secret" {
+			t.Errorf("X-Api-Key = %q, want secret", got)
+			http.Error(response, "missing auth", http.StatusUnauthorized)
+			return
+		}
+
+		message := readHTTPRPCMessage(t, request)
+		switch message.Method {
+		case "initialize":
+			writeHTTPRPCResponse(t, response, message.ID, map[string]any{"protocolVersion": "2024-11-05"})
+		case "notifications/initialized":
+			response.WriteHeader(http.StatusAccepted)
+		default:
+			writeHTTPRPCResponse(t, response, message.ID, map[string]any{})
+		}
+	}))
+	defer testServer.Close()
+
+	client, err := Connect(ctx, Server{
+		Name:    "docs",
+		Type:    ServerTypeHTTP,
+		URL:     testServer.URL + "/mcp",
+		Headers: map[string]string{"X-Api-Key": "secret"},
+	})
+	if err != nil {
+		t.Fatalf("Connect() error = %v", err)
+	}
+	if err := client.Close(); err != nil {
+		t.Fatalf("Close() error = %v", err)
+	}
+}
+
+func TestHTTPClientRejectsCrossOriginRedirectBeforeSendingHeaders(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	var targetHits int32
+	target := httptest.NewServer(http.HandlerFunc(func(response http.ResponseWriter, request *http.Request) {
+		atomic.AddInt32(&targetHits, 1)
+		if got := request.Header.Get("X-Api-Key"); got != "" {
+			t.Errorf("redirect target received X-Api-Key = %q", got)
+		}
+		response.WriteHeader(http.StatusOK)
+	}))
+	defer target.Close()
+
+	redirector := httptest.NewServer(http.HandlerFunc(func(response http.ResponseWriter, request *http.Request) {
+		http.Redirect(response, request, target.URL+"/mcp", http.StatusTemporaryRedirect)
+	}))
+	defer redirector.Close()
+
+	_, err := Connect(ctx, Server{
+		Name:    "docs",
+		Type:    ServerTypeHTTP,
+		URL:     redirector.URL + "/mcp",
+		Headers: map[string]string{"X-Api-Key": "secret"},
+	})
+	if err == nil || !strings.Contains(err.Error(), "cross-origin redirect") {
+		t.Fatalf("Connect() error = %v, want cross-origin redirect error", err)
+	}
+	if got := atomic.LoadInt32(&targetHits); got != 0 {
+		t.Fatalf("redirect target hits = %d, want 0", got)
+	}
+}
+
 func TestSSEClientListsAndCallsToolsFromRemoteStream(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()
@@ -340,6 +420,53 @@ func TestSSEClientListsAndCallsToolsFromRemoteStream(t *testing.T) {
 	}
 }
 
+func TestSSEClientRejectsCrossOriginEndpointBeforeSendingHeaders(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	var targetHits int32
+	target := httptest.NewServer(http.HandlerFunc(func(response http.ResponseWriter, request *http.Request) {
+		atomic.AddInt32(&targetHits, 1)
+		if got := request.Header.Get("X-Api-Key"); got != "" {
+			t.Errorf("SSE endpoint target received X-Api-Key = %q", got)
+		}
+		response.WriteHeader(http.StatusAccepted)
+	}))
+	defer target.Close()
+
+	stream := httptest.NewServer(http.HandlerFunc(func(response http.ResponseWriter, request *http.Request) {
+		if got := request.Header.Get("X-Api-Key"); got != "secret" {
+			t.Errorf("X-Api-Key = %q, want secret on configured SSE server", got)
+			http.Error(response, "missing auth", http.StatusUnauthorized)
+			return
+		}
+		if request.Method != http.MethodGet || request.URL.Path != "/sse" {
+			t.Errorf("request = %s %s, want GET /sse", request.Method, request.URL.Path)
+			http.Error(response, "bad request", http.StatusNotFound)
+			return
+		}
+		response.Header().Set("Content-Type", "text/event-stream")
+		_, _ = fmt.Fprintf(response, "event: endpoint\ndata: %s/messages\n\n", target.URL)
+		if flusher, ok := response.(http.Flusher); ok {
+			flusher.Flush()
+		}
+	}))
+	defer stream.Close()
+
+	_, err := Connect(ctx, Server{
+		Name:    "docs",
+		Type:    ServerTypeSSE,
+		URL:     stream.URL + "/sse",
+		Headers: map[string]string{"X-Api-Key": "secret"},
+	})
+	if err == nil || !strings.Contains(err.Error(), "endpoint origin") {
+		t.Fatalf("Connect() error = %v, want cross-origin endpoint error", err)
+	}
+	if got := atomic.LoadInt32(&targetHits); got != 0 {
+		t.Fatalf("SSE endpoint target hits = %d, want 0", got)
+	}
+}
+
 func TestHTTPClientReportsNonOKStatus(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()

internal/mcp/network_client.go

@@ -733,7 +733,11 @@ func resolveSSEEndpointURL(baseValue string, endpointValue string) (string, erro
 	if err != nil {
 		return "", fmt.Errorf("parse MCP SSE endpoint URL: %w", err)
 	}
-	return baseURL.ResolveReference(endpointURL).String(), nil
+	resolvedURL := baseURL.ResolveReference(endpointURL)
+	if !sameMCPOrigin(baseURL, resolvedURL) {
+		return "", fmt.Errorf("MCP SSE endpoint origin %s differs from configured server origin %s", mcpOrigin(resolvedURL), mcpOrigin(baseURL))
+	}
+	return resolvedURL.String(), nil
 }
 
 func rpcResponseKey(id any) string {
@@ -873,10 +877,10 @@ func (source *storeTokenSource) Refresh(ctx context.Context) (string, error) {
 
 // oauthHTTPClient returns an HTTP client whose transport attaches OAuth bearer
 // tokens and refreshes them on 401 for OAuth-configured servers. Servers that do
-// not declare OAuth get the default client unchanged.
+// not declare OAuth use the default transport with the same MCP redirect guard.
 func oauthHTTPClient(server Server) (*http.Client, error) {
 	if !strings.EqualFold(strings.TrimSpace(server.Auth), ServerAuthOAuth) {
-		return http.DefaultClient, nil
+		return mcpHTTPClient(server, nil), nil
 	}
 	store, err := NewTokenStore(TokenStoreOptions{})
 	if err != nil {
@@ -888,5 +892,5 @@ func oauthHTTPClient(server Server) (*http.Client, error) {
 		httpClient: http.DefaultClient,
 		now:        time.Now,
 	}
-	return &http.Client{Transport: newOAuthRoundTripper(http.DefaultTransport, source, server.Name)}, nil
+	return mcpHTTPClient(server, newOAuthRoundTripper(http.DefaultTransport, source, server.Name)), nil
 }

internal/mcp/network_client_test.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"net/http"
 	"net/http/httptest"
+	"net/url"
 	"strings"
 	"sync/atomic"
 	"testing"
@@ -149,6 +150,141 @@ func TestOAuthRoundTripperDoesNotRetryAfterSuccessfulNon401(t *testing.T) {
 	}
 }
 
+func TestOAuthRoundTripperRejectsCrossOriginRedirectBeforeBearerLeak(t *testing.T) {
+	source := &fakeTokenSource{}
+	source.access.Store("access-secret")
+	source.refreshFunc = func() (string, error) { return "access-fresh", nil }
+
+	var targetHits int32
+	target := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		atomic.AddInt32(&targetHits, 1)
+		if got := r.Header.Get("Authorization"); got != "" {
+			t.Errorf("redirect target received Authorization = %q", got)
+		}
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer target.Close()
+
+	redirector := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		http.Redirect(w, r, target.URL+"/mcp", http.StatusTemporaryRedirect)
+	}))
+	defer redirector.Close()
+
+	client := mcpHTTPClient(
+		Server{Name: "oauth-docs", Type: ServerTypeHTTP},
+		newOAuthRoundTripper(http.DefaultTransport, source, "oauth-docs"),
+	)
+	resp, err := client.Get(redirector.URL + "/mcp")
+	if resp != nil {
+		_ = resp.Body.Close()
+	}
+	if err == nil || !strings.Contains(err.Error(), "cross-origin redirect") {
+		t.Fatalf("Get() error = %v, want cross-origin redirect error", err)
+	}
+	if got := atomic.LoadInt32(&targetHits); got != 0 {
+		t.Fatalf("redirect target hits = %d, want 0", got)
+	}
+}
+
+func TestSameMCPOriginNormalizesDefaultPorts(t *testing.T) {
+	for _, tc := range []struct {
+		name  string
+		left  string
+		right string
+		want  bool
+	}{
+		{
+			name:  "https default port",
+			left:  "https://example.test/mcp",
+			right: "https://EXAMPLE.test:443/other",
+			want:  true,
+		},
+		{
+			name:  "http default port",
+			left:  "http://example.test/mcp",
+			right: "http://example.test:80/other",
+			want:  true,
+		},
+		{
+			name:  "different scheme",
+			left:  "https://example.test/mcp",
+			right: "http://example.test:443/other",
+			want:  false,
+		},
+		{
+			name:  "different port",
+			left:  "https://example.test/mcp",
+			right: "https://example.test:8443/other",
+			want:  false,
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			left, err := url.Parse(tc.left)
+			if err != nil {
+				t.Fatal(err)
+			}
+			right, err := url.Parse(tc.right)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if got := sameMCPOrigin(left, right); got != tc.want {
+				t.Fatalf("sameMCPOrigin(%q, %q) = %v, want %v", tc.left, tc.right, got, tc.want)
+			}
+		})
+	}
+}
+
+func TestResolveSSEEndpointURLRestrictsOrigin(t *testing.T) {
+	for _, tc := range []struct {
+		name     string
+		base     string
+		endpoint string
+		want     string
+		wantErr  string
+	}{
+		{
+			name:     "relative endpoint",
+			base:     "https://example.test/sse",
+			endpoint: "/messages",
+			want:     "https://example.test/messages",
+		},
+		{
+			name:     "absolute same origin",
+			base:     "https://example.test/sse",
+			endpoint: "https://example.test/messages",
+			want:     "https://example.test/messages",
+		},
+		{
+			name:     "default port equivalent",
+			base:     "https://example.test/sse",
+			endpoint: "https://example.test:443/messages",
+			want:     "https://example.test:443/messages",
+		},
+		{
+			name:     "cross origin",
+			base:     "https://example.test/sse",
+			endpoint: "https://evil.test/messages",
+			wantErr:  "differs from configured server origin",
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := resolveSSEEndpointURL(tc.base, tc.endpoint)
+			if tc.wantErr != "" {
+				if err == nil || !strings.Contains(err.Error(), tc.wantErr) {
+					t.Fatalf("resolveSSEEndpointURL() error = %v, want %q", err, tc.wantErr)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("resolveSSEEndpointURL() error = %v", err)
+			}
+			if got != tc.want {
+				t.Fatalf("resolveSSEEndpointURL() = %q, want %q", got, tc.want)
+			}
+		})
+	}
+}
+
 func TestNonOAuthServerIsUnaffected(t *testing.T) {
 	// Regression: a server without auth must not get a bearer header or any
 	// OAuth round tripper, and must connect normally.

internal/mcp/network_redirect.go

@@ -0,0 +1,74 @@
+package mcp
+
+import (
+	"fmt"
+	"net"
+	"net/http"
+	"net/url"
+	"strings"
+)
+
+const maxMCPRedirects = 10
+
+func mcpHTTPClient(server Server, transport http.RoundTripper) *http.Client {
+	return &http.Client{
+		Transport:     transport,
+		CheckRedirect: checkMCPRedirect(server),
+	}
+}
+
+func checkMCPRedirect(server Server) func(*http.Request, []*http.Request) error {
+	return func(request *http.Request, via []*http.Request) error {
+		if len(via) >= maxMCPRedirects {
+			return fmt.Errorf("MCP %s server %s stopped after %d redirects", server.Type, server.Name, maxMCPRedirects)
+		}
+		if len(via) == 0 {
+			return nil
+		}
+		if !sameMCPOrigin(via[0].URL, request.URL) {
+			return fmt.Errorf("MCP %s server %s refused cross-origin redirect to %s", server.Type, server.Name, mcpOrigin(request.URL))
+		}
+		return nil
+	}
+}
+
+func sameMCPOrigin(left *url.URL, right *url.URL) bool {
+	if left == nil || right == nil {
+		return false
+	}
+	return strings.EqualFold(left.Scheme, right.Scheme) &&
+		strings.EqualFold(left.Hostname(), right.Hostname()) &&
+		effectiveMCPPort(left) == effectiveMCPPort(right)
+}
+
+func effectiveMCPPort(value *url.URL) string {
+	if value == nil {
+		return ""
+	}
+	if port := value.Port(); port != "" {
+		return port
+	}
+	switch strings.ToLower(value.Scheme) {
+	case "http":
+		return "80"
+	case "https":
+		return "443"
+	default:
+		return ""
+	}
+}
+
+func mcpOrigin(value *url.URL) string {
+	if value == nil {
+		return "<nil>"
+	}
+	host := value.Hostname()
+	port := effectiveMCPPort(value)
+	if port != "" {
+		host = net.JoinHostPort(host, port)
+	}
+	if value.Scheme == "" {
+		return host
+	}
+	return strings.ToLower(value.Scheme) + "://" + host
+}

internal/providercatalog/catalog.go

@@ -123,8 +123,10 @@ var descriptors = []Descriptor{
 	openAICompat("together", "Together AI", "https://api.together.xyz/v1", "meta-llama/Llama-3.3-70B-Instruct-Turbo", []string{"TOGETHER_API_KEY"}),
 	openAICompat("dashscope", "DashScope", "https://dashscope-intl.aliyuncs.com/compatible-mode/v1", "qwen-plus", []string{"DASHSCOPE_API_KEY", "QWEN_API_KEY"}, "qwen"),
 	openAICompat("moonshot", "Moonshot AI", "https://api.moonshot.ai/v1", "kimi-k2-0905-preview", []string{"MOONSHOT_API_KEY"}, "kimi"),
+	openAICompat("longcat", "LongCat", "https://api.longcat.chat/openai", "LongCat-2.0", []string{"LONGCAT_API_KEY"}, "meituan longcat", "meituan", "longcat-2.0"),
 	openAICompat("nvidia-nim", "NVIDIA NIM", "https://integrate.api.nvidia.com/v1", "nvidia/llama-3.1-nemotron-70b-instruct", []string{"NVIDIA_API_KEY"}, "nvidia nim"),
 	anthropicCompat("minimax", "MiniMax", "https://api.minimax.io/anthropic", "MiniMax-M3", []string{"MINIMAX_API_KEY"}, "mini-max", "mini_max"),
+	anthropicCompat("minimaxi-cn", "MiniMax CN", "https://api.minimaxi.com/anthropic", "MiniMax-M3", []string{"MINIMAXI_API_KEY"}, "minimax cn", "minimax-cn"),
 	openAICompat("mistral", "Mistral", "https://api.mistral.ai/v1", "mistral-large-latest", []string{"MISTRAL_API_KEY"}),
 	openAICompat("github", "GitHub Models", "https://models.inference.ai.azure.com", "openai/gpt-4.1", []string{"GITHUB_TOKEN"}, "github-models"),
 	transportDescriptor("bedrock", "Amazon Bedrock", TransportBedrock, "https://bedrock-runtime.${AWS_REGION}.amazonaws.com", "anthropic.claude-3-5-sonnet-20241022-v2:0", []string{"AWS_ACCESS_KEY_ID", "AWS_PROFILE"}, []APIFormat{APIFormatBedrockConverse}, true),
@@ -133,7 +135,8 @@ var descriptors = []Descriptor{
 	openAICompat("venice", "Venice AI", "https://api.venice.ai/api/v1", "qwen-2.5-qwq-32b", []string{"VENICE_API_KEY"}),
 	openAICompat("xiaomi-mimo", "Xiaomi MiMo", "https://api.mimo.xiaomi.com/openai/v1", "mimo-vl", []string{"MIMO_API_KEY", "XIAOMI_API_KEY"}, "xiaomi mimo"),
 	openAICompat("bankr", "Bankr", "https://api.bankr.bot/v1", "bankr-large", []string{"BANKR_API_KEY"}),
-	openAICompat("zai", "Z.ai", "https://open.bigmodel.cn/api/paas/v4", "glm-4.5", []string{"ZAI_API_KEY", "ZHIPU_API_KEY"}),
+	openAICompat("zai", "Z.ai", "https://api.z.ai/api/paas/v4", "glm-4.5", []string{"ZAI_API_KEY"}, "z.ai"),
+	openAICompat("zai-cn", "Z.ai CN", "https://open.bigmodel.cn/api/paas/v4", "glm-4.5", []string{"ZHIPU_API_KEY"}, "z.ai cn"),
 	openAICompat("kilocode", "KiloCode", "https://api.kilo.ai/api/gateway", "anthropic/claude-sonnet-4.6", []string{"KILO_API_KEY"}, "kilo", "kilo gateway"),
 	openAICompat("opencode", "OpenCode Zen", "https://opencode.ai/zen/v1", "deepseek-v4-flash", []string{"OPENCODE_API_KEY"}, "opencode zen"),
 	openAICompat("opencode-go", "OpenCode Go", "https://opencode.ai/zen/go/v1", "deepseek-v4-pro", []string{"OPENCODE_API_KEY"}, "opencode go"),

internal/providercatalog/catalog_test.go

@@ -23,8 +23,10 @@ var expectedCatalogIDs = []string{
 	"together",
 	"dashscope",
 	"moonshot",
+	"longcat",
 	"nvidia-nim",
 	"minimax",
+	"minimaxi-cn",
 	"mistral",
 	"github",
 	"bedrock",
@@ -34,6 +36,7 @@ var expectedCatalogIDs = []string{
 	"xiaomi-mimo",
 	"bankr",
 	"zai",
+	"zai-cn",
 	"kilocode",
 	"opencode",
 	"opencode-go",
@@ -99,6 +102,28 @@ func TestRecommendedProviderEndpoint(t *testing.T) {
 	}
 }
 
+func TestLongCatDescriptor(t *testing.T) {
+	descriptor, err := Require("longcat")
+	if err != nil {
+		t.Fatalf("Require(longcat) error = %v", err)
+	}
+	if descriptor.Name != "LongCat" {
+		t.Fatalf("Name = %q, want LongCat", descriptor.Name)
+	}
+	if descriptor.DefaultBaseURL != "https://api.longcat.chat/openai" {
+		t.Fatalf("DefaultBaseURL = %q, want LongCat OpenAI-compatible endpoint", descriptor.DefaultBaseURL)
+	}
+	if descriptor.DefaultModel != "LongCat-2.0" {
+		t.Fatalf("DefaultModel = %q, want LongCat-2.0", descriptor.DefaultModel)
+	}
+	if descriptor.Transport != TransportOpenAICompatible {
+		t.Fatalf("Transport = %q, want %q", descriptor.Transport, TransportOpenAICompatible)
+	}
+	if !reflect.DeepEqual(descriptor.AuthEnvVars, []string{"LONGCAT_API_KEY"}) {
+		t.Fatalf("AuthEnvVars = %#v, want LONGCAT_API_KEY", descriptor.AuthEnvVars)
+	}
+}
+
 func TestCatalogDescriptorsExposeRequiredDefaults(t *testing.T) {
 	for _, descriptor := range All() {
 		if descriptor.ID == "" {
@@ -266,8 +291,8 @@ func TestListByTransportPreservesCatalogOrder(t *testing.T) {
 		TransportGoogle:          {"google"},
 		TransportBedrock:         {"bedrock"},
 		TransportVertex:          {"vertex"},
-		TransportAnthropicCompat: {"minimax", "custom-anthropic-compatible"},
-		TransportOpenAICompat:    {"gitlawb-opengateway", "ollama-cloud", "ollama", "lmstudio", "openrouter", "huggingface", "chatgpt", "groq", "deepseek", "together", "dashscope", "moonshot", "nvidia-nim", "mistral", "github", "xai", "venice", "xiaomi-mimo", "bankr", "zai", "kilocode", "opencode", "opencode-go", "atomic-chat", "chatgpt-proxy", "custom-openai-compatible"},
+		TransportAnthropicCompat: {"minimax", "minimaxi-cn", "custom-anthropic-compatible"},
+		TransportOpenAICompat:    {"gitlawb-opengateway", "ollama-cloud", "ollama", "lmstudio", "openrouter", "huggingface", "chatgpt", "groq", "deepseek", "together", "dashscope", "moonshot", "longcat", "nvidia-nim", "mistral", "github", "xai", "venice", "xiaomi-mimo", "bankr", "zai", "zai-cn", "kilocode", "opencode", "opencode-go", "atomic-chat", "chatgpt-proxy", "custom-openai-compatible"},
 	}
 
 	for transport, wantIDs := range cases {

internal/providermodelcatalog/catalog.go

@@ -21,6 +21,22 @@ type Model struct {
 	Source           string
 }
 
+// Shared by both "minimax" and "minimaxi-cn"; Models() rebuilds a fresh slice
+// per call, so the shared backing slice cannot be mutated by callers.
+var minimaxCuratedModels = []Model{
+	{ID: "MiniMax-M3", Description: "catalog default"},
+	{ID: "MiniMax-M2.1", Description: "agentic coding model"},
+}
+
+// Shared by both "zai" (international, api.z.ai) and "zai-cn" (China,
+// open.bigmodel.cn); same model lineup on both endpoints.
+var zaiCuratedModels = []Model{
+	{ID: "glm-4.5", Description: "catalog default"},
+	{ID: "glm-4.5-air", Description: "fast model"},
+	{ID: "glm-4.6", Description: "latest general model"},
+	{ID: "glm-z1-air", Description: "reasoning model"},
+}
+
 var curatedModels = map[string][]Model{
 	"ollama-cloud": {
 		{ID: "qwen3-coder:480b", Description: "catalog default"},
@@ -88,10 +104,8 @@ var curatedModels = map[string][]Model{
 		{ID: "meta/llama-3.1-70b-instruct", Description: "general model"},
 		{ID: "mistralai/mixtral-8x7b-instruct-v0.1", Description: "mixture model"},
 	},
-	"minimax": {
-		{ID: "MiniMax-M3", Description: "catalog default"},
-		{ID: "MiniMax-M2.1", Description: "agentic coding model"},
-	},
+	"minimax":     minimaxCuratedModels,
+	"minimaxi-cn": minimaxCuratedModels,
 	"mistral": {
 		{ID: "mistral-large-latest", Description: "catalog default"},
 		{ID: "codestral-latest", Description: "coding model"},
@@ -123,12 +137,8 @@ var curatedModels = map[string][]Model{
 	"bankr": {
 		{ID: "bankr-large", Description: "catalog default"},
 	},
-	"zai": {
-		{ID: "glm-4.5", Description: "catalog default"},
-		{ID: "glm-4.5-air", Description: "fast model"},
-		{ID: "glm-4.6", Description: "latest general model"},
-		{ID: "glm-z1-air", Description: "reasoning model"},
-	},
+	"zai":    zaiCuratedModels,
+	"zai-cn": zaiCuratedModels,
 	// OpenGateway smart-routes by model id across its upstream providers
 	// (see /health: xiaomi-mimo, minimax, qwen, google, nvidia, z-ai). These are
 	// the curated coding defaults; the gateway accepts any model its upstreams

internal/providermodelcatalog/catalog_test.go

@@ -37,6 +37,16 @@ func TestModelsAreProviderScoped(t *testing.T) {
 			want:     []string{"mistral-large-latest", "codestral-latest"},
 			notWant:  []string{"gpt-4.1", "claude-sonnet-4.5"},
 		},
+		{
+			provider: "minimaxi-cn",
+			want:     []string{"MiniMax-M3", "MiniMax-M2.1"},
+			notWant:  []string{"gpt-4.1", "claude-sonnet-4.5"},
+		},
+		{
+			provider: "zai-cn",
+			want:     []string{"glm-4.5", "glm-4.6", "glm-4.5-air", "glm-z1-air"},
+			notWant:  []string{"gpt-4.1", "claude-sonnet-4.5"},
+		},
 	}
 
 	for _, tt := range tests {
@@ -88,3 +98,66 @@ func modelIDs(models []Model) []string {
 	}
 	return ids
 }
+
+// Pins the contract that MiniMax CN exposes exactly the same model IDs as the
+// international provider; a divergence here is almost always a mistake.
+func TestMiniMaxCNModelsMirrorInternational(t *testing.T) {
+	intl, ok := providercatalog.Get("minimax")
+	if !ok {
+		t.Fatal("provider minimax missing from catalog")
+	}
+	cn, ok := providercatalog.Get("minimaxi-cn")
+	if !ok {
+		t.Fatal("provider minimaxi-cn missing from catalog")
+	}
+	intlModels := Models(intl)
+	cnModels := Models(cn)
+	if len(intlModels) != len(cnModels) {
+		t.Fatalf("model count mismatch: international=%d china=%d", len(intlModels), len(cnModels))
+	}
+	intlIDs := map[string]string{}
+	for _, m := range intlModels {
+		intlIDs[m.ID] = m.Description
+	}
+	for _, m := range cnModels {
+		wantDesc, ok := intlIDs[m.ID]
+		if !ok {
+			t.Fatalf("china provider exposes %q which is not in the international catalog", m.ID)
+		}
+		if m.Description != wantDesc {
+			t.Fatalf("model %q description diverged: international=%q china=%q", m.ID, wantDesc, m.Description)
+		}
+	}
+}
+
+// Same contract as TestMiniMaxCNModelsMirrorInternational, for Z.ai: the
+// international (api.z.ai) and China (open.bigmodel.cn) endpoints expose the
+// same model lineup.
+func TestZaiCNModelsMirrorInternational(t *testing.T) {
+	intl, ok := providercatalog.Get("zai")
+	if !ok {
+		t.Fatal("provider zai missing from catalog")
+	}
+	cn, ok := providercatalog.Get("zai-cn")
+	if !ok {
+		t.Fatal("provider zai-cn missing from catalog")
+	}
+	intlModels := Models(intl)
+	cnModels := Models(cn)
+	if len(intlModels) != len(cnModels) {
+		t.Fatalf("model count mismatch: international=%d china=%d", len(intlModels), len(cnModels))
+	}
+	intlIDs := map[string]string{}
+	for _, m := range intlModels {
+		intlIDs[m.ID] = m.Description
+	}
+	for _, m := range cnModels {
+		wantDesc, ok := intlIDs[m.ID]
+		if !ok {
+			t.Fatalf("china provider exposes %q which is not in the international catalog", m.ID)
+		}
+		if m.Description != wantDesc {
+			t.Fatalf("model %q description diverged: international=%q china=%q", m.ID, wantDesc, m.Description)
+		}
+	}
+}

internal/providermodelcatalog/remote.go

@@ -128,6 +128,10 @@ func ModelsDevProviderID(provider providercatalog.Descriptor) string {
 		return "nvidia"
 	case "xiaomi-mimo":
 		return "xiaomi"
+	case "zai-cn":
+		return "zai"
+	case "minimaxi-cn":
+		return "minimax"
 	default:
 		return strings.TrimSpace(provider.ID)
 	}

internal/providermodelcatalog/remote_test.go

@@ -118,6 +118,8 @@ func TestModelsDevProviderIDMapsZeroAliases(t *testing.T) {
 		"xiaomi-mimo":  "xiaomi",
 		"dashscope":    "alibaba",
 		"ollama-cloud": "ollama-cloud",
+		"zai-cn":       "zai",
+		"minimaxi-cn":  "minimax",
 	}
 	for zeroID, want := range tests {
 		provider, ok := providercatalog.Get(zeroID)

internal/providers/openai/provider.go

@@ -276,16 +276,6 @@ func (provider *Provider) emitPayload(ctx context.Context, data string, state *t
 	return true
 }
 
-// firstNonEmptyString returns the first argument that is not empty.
-func firstNonEmptyString(values ...string) string {
-	for _, value := range values {
-		if value != "" {
-			return value
-		}
-	}
-	return ""
-}
-
 func (provider *Provider) emitChunk(
 	ctx context.Context,
 	chunk streamChunk,
@@ -293,12 +283,10 @@ func (provider *Provider) emitChunk(
 	events chan<- zeroruntime.StreamEvent,
 ) {
 	for _, choice := range chunk.Choices {
-		// Two dialects for the same thing (see streamDelta): prefer
-		// `reasoning_content`, fall back to `reasoning`. No backend sends both.
-		if reasoning := firstNonEmptyString(choice.Delta.ReasoningContent, choice.Delta.Reasoning); reasoning != "" {
+		if choice.Delta.ReasoningContent != "" {
 			sendEvent(ctx, events, zeroruntime.StreamEvent{
 				Type:    zeroruntime.StreamEventReasoning,
-				Content: reasoning,
+				Content: choice.Delta.ReasoningContent,
 			})
 		}
 		if choice.Delta.Content != "" {

internal/providers/openai/provider_test.go

@@ -227,33 +227,6 @@ func TestStreamCompletionEmitsReasoningContentDeltas(t *testing.T) {
 	}
 }
 
-// Ollama's OpenAI-compat endpoint, OpenRouter, and most gateways stream
-// thinking under `reasoning` (not DeepSeek's `reasoning_content`). Dropping it
-// made every thinking token on those backends vanish: the model looked frozen
-// for minutes and reasoning-only turns were indistinguishable from a dead
-// provider. The exact shape below is a live capture from ollama glm-5.2:cloud.
-func TestStreamCompletionEmitsOllamaStyleReasoningDeltas(t *testing.T) {
-	provider := newTestProvider(t, func(w http.ResponseWriter, r *http.Request) {
-		writeSSE(w, `{"choices":[{"index":0,"delta":{"role":"assistant","content":"","reasoning":"The user"},"finish_reason":null}]}`)
-		writeSSE(w, `{"choices":[{"index":0,"delta":{"role":"assistant","content":"","reasoning":" wants hello"},"finish_reason":null}]}`)
-		writeSSE(w, `{"choices":[{"index":0,"delta":{"role":"assistant","content":"hello"},"finish_reason":"stop"}]}`)
-		writeSSE(w, `[DONE]`)
-	})
-
-	events := collectProviderEvents(t, provider)
-	reasoning := eventsOfType(events, zeroruntime.StreamEventReasoning)
-	if len(reasoning) != 2 {
-		t.Fatalf("reasoning events = %#v, want two reasoning deltas (ollama `reasoning` field regression)", reasoning)
-	}
-	if reasoning[0].Content != "The user" || reasoning[1].Content != " wants hello" {
-		t.Fatalf("unexpected reasoning events: %#v", reasoning)
-	}
-	text := eventsOfType(events, zeroruntime.StreamEventText)
-	if len(text) != 1 || text[0].Content != "hello" {
-		t.Fatalf("content after thinking must still stream as text, got %#v", text)
-	}
-}
-
 func TestStreamCompletionEmitsReasoningBeforeRegularContent(t *testing.T) {
 	provider := newTestProvider(t, func(w http.ResponseWriter, r *http.Request) {
 		writeSSE(w, `{"choices":[{"delta":{"reasoning_content":"Thinking. ","content":"Answer."}}]}`)

internal/providers/openai/types.go

@@ -76,16 +76,8 @@ type streamChoice struct {
 }
 
 type streamDelta struct {
-	Content string `json:"content"`
-	// Reasoning/thinking deltas arrive under DIFFERENT keys depending on the
-	// backend dialect: DeepSeek-style servers emit `reasoning_content`, while
-	// ollama's OpenAI-compat endpoint, OpenRouter, and most gateways emit
-	// `reasoning`. Parse both — dropping `reasoning` made every thinking token
-	// on those backends silently vanish (the model looked frozen for minutes,
-	// its thinking was never replayed into the next turn's context, and a
-	// reasoning-only turn was indistinguishable from a dead provider).
+	Content          string                `json:"content"`
 	ReasoningContent string                `json:"reasoning_content"`
-	Reasoning        string                `json:"reasoning"`
 	ToolCalls        []streamToolCallDelta `json:"tool_calls"`
 }
 

internal/providers/providerio/providerio.go

@@ -82,53 +82,6 @@ func ContentStallTimeout(idleTimeout time.Duration) time.Duration {
 	return idleTimeout + extra
 }
 
-// DefaultStreamGapTimeout bounds the byte-silent gap AFTER a stream has
-// started delivering payloads. Once ANY payload (data or keep-alive) has
-// arrived, the backend has proven it heartbeats/streams — so a subsequent gap
-// with zero bytes is almost always a dead connection (live capture of the
-// reported "stuck for 33 minutes": ~119KB streamed, then the socket went
-// byte-frozen forever). Detecting that at 2 minutes instead of the 5-minute
-// idle default turns a ~15-minute blind wait (3 × 5m attempts) into a ~6-minute
-// bounded recovery, without touching the pre-first-byte window that slow cloud
-// proxies legitimately need (they may withhold output until the upstream model
-// produces its first token). Override with ZERO_STREAM_GAP_TIMEOUT. For
-// comparison: Codex ships the same 5m idle with 5 stream retries; opencode
-// suffers this exact freeze (idle HTTPS socket in "working" state) with no
-// mid-stream gap detection at all.
-const DefaultStreamGapTimeout = 2 * time.Minute
-
-// streamGapTimeoutEnv overrides the post-first-payload gap timeout. Accepts the
-// same forms as ZERO_STREAM_IDLE_TIMEOUT; "0"/"off" disables the tightened gap
-// (the plain idle timeout then applies for the whole stream).
-const streamGapTimeoutEnv = "ZERO_STREAM_GAP_TIMEOUT"
-
-// ResolveStreamGapTimeout selects the effective post-first-payload gap timeout
-// for a stream whose idle timeout resolved to idleTimeout. It is clamped to
-// never EXCEED the idle timeout (a gap larger than idle would be dead config),
-// and disabled entirely when the idle watchdog is disabled.
-func ResolveStreamGapTimeout(idleTimeout time.Duration) time.Duration {
-	if idleTimeout <= 0 {
-		return 0
-	}
-	gap := DefaultStreamGapTimeout
-	if raw := strings.TrimSpace(os.Getenv(streamGapTimeoutEnv)); raw != "" {
-		switch strings.ToLower(raw) {
-		case "0", "off", "none", "disabled":
-			return idleTimeout // no tightened gap: idle applies throughout
-		default:
-			if d, err := time.ParseDuration(raw); err == nil && d > 0 {
-				gap = d
-			} else if secs, err := strconv.Atoi(raw); err == nil && secs > 0 {
-				gap = time.Duration(secs) * time.Second
-			}
-		}
-	}
-	if gap > idleTimeout {
-		return idleTimeout
-	}
-	return gap
-}
-
 // streamIdleTimeoutEnv is the global override for the stream idle timeout. It
 // accepts a Go duration ("5m", "300s", "90s") or a bare number of seconds
 // ("300"). A value of "0", "off", "none", or "disabled" turns the watchdog off
@@ -364,14 +317,7 @@ func ScanSSEDataWithContext(
 		idle := time.NewTimer(idleTimeout)
 		defer idle.Stop()
 		idleC = idle.C
-		// The timer was armed above with the FULL idle window, which governs
-		// until the first payload (slow cloud proxies may withhold output until
-		// the upstream model produces its first token). resetIdle only runs when
-		// a payload HAS arrived — the backend has proven it streams/heartbeats —
-		// so every re-arm uses the tighter gap window: a byte-frozen socket
-		// mid-stream is a dead connection, and waiting the full idle for it just
-		// multiplies the user's blind wait per retry.
-		resetIdle = reset(idle, ResolveStreamGapTimeout(idleTimeout))
+		resetIdle = reset(idle, idleTimeout)
 
 		// Content watchdog: only real data lines reset it (keep-alives do not), so a
 		// stream that heartbeats without producing output is bounded instead of

internal/providers/providerio/providerio_test.go

@@ -371,86 +371,3 @@ func TestHTTPClientReturnsStallHardenedSharedClient(t *testing.T) {
 		t.Fatal("an explicit client must be returned unchanged")
 	}
 }
-
-// Once a stream has delivered its first payload, a byte-silent gap is bounded
-// by the TIGHTER gap timeout, not the full idle window — a mid-stream frozen
-// socket (live-captured "stuck" shape: bytes flowed, then nothing forever)
-// must be detected fast so retries aren't multiplied by five-minute waits.
-func TestScanSSETightensGapAfterFirstPayload(t *testing.T) {
-	t.Setenv("ZERO_STREAM_GAP_TIMEOUT", "80ms")
-	pr, pw := io.Pipe()
-	defer func() { _ = pw.Close() }()
-
-	go func() {
-		_, _ = io.WriteString(pw, "data: first\n\n")
-		// then: byte-frozen forever
-	}()
-
-	start := time.Now()
-	done := make(chan error, 1)
-	go func() {
-		done <- ScanSSEDataWithContext(context.Background(), func() {}, pr, 2*time.Second, func(string) bool { return true })
-	}()
-	select {
-	case err := <-done:
-		if !errors.Is(err, ErrStreamIdle) {
-			t.Fatalf("err = %v, want ErrStreamIdle", err)
-		}
-		if elapsed := time.Since(start); elapsed > time.Second {
-			t.Fatalf("gap abort took %s — the tightened window (80ms) did not apply after the first payload", elapsed)
-		}
-	case <-time.After(5 * time.Second):
-		t.Fatal("stream not aborted")
-	}
-}
-
-// BEFORE the first payload the FULL idle window applies: slow cloud proxies may
-// withhold output until the upstream model emits its first token, so the gap
-// timeout must NOT govern the wait for the first byte.
-func TestScanSSEKeepsFullIdleBeforeFirstPayload(t *testing.T) {
-	t.Setenv("ZERO_STREAM_GAP_TIMEOUT", "30ms")
-	pr, pw := io.Pipe()
-	defer func() { _ = pw.Close() }()
-	// never write anything
-
-	start := time.Now()
-	done := make(chan error, 1)
-	go func() {
-		done <- ScanSSEDataWithContext(context.Background(), func() {}, pr, 300*time.Millisecond, func(string) bool { return true })
-	}()
-	select {
-	case err := <-done:
-		if !errors.Is(err, ErrStreamIdle) {
-			t.Fatalf("err = %v, want ErrStreamIdle", err)
-		}
-		if elapsed := time.Since(start); elapsed < 250*time.Millisecond {
-			t.Fatalf("aborted after %s — the 30ms gap timeout must not apply before the first payload", elapsed)
-		}
-	case <-time.After(5 * time.Second):
-		t.Fatal("stream not aborted")
-	}
-}
-
-func TestResolveStreamGapTimeout(t *testing.T) {
-	t.Setenv("ZERO_STREAM_GAP_TIMEOUT", "")
-	if got := ResolveStreamGapTimeout(5 * time.Minute); got != DefaultStreamGapTimeout {
-		t.Fatalf("default gap = %s, want %s", got, DefaultStreamGapTimeout)
-	}
-	// Clamped: the gap never exceeds the idle window.
-	if got := ResolveStreamGapTimeout(time.Minute); got != time.Minute {
-		t.Fatalf("gap = %s, want clamp to the 1m idle", got)
-	}
-	// Disabled idle disables the gap too.
-	if got := ResolveStreamGapTimeout(0); got != 0 {
-		t.Fatalf("gap with idle off = %s, want 0", got)
-	}
-	// Env override wins; "off" falls back to the idle window for the whole stream.
-	t.Setenv("ZERO_STREAM_GAP_TIMEOUT", "45s")
-	if got := ResolveStreamGapTimeout(5 * time.Minute); got != 45*time.Second {
-		t.Fatalf("env gap = %s, want 45s", got)
-	}
-	t.Setenv("ZERO_STREAM_GAP_TIMEOUT", "off")
-	if got := ResolveStreamGapTimeout(5 * time.Minute); got != 5*time.Minute {
-		t.Fatalf("off gap = %s, want the idle window", got)
-	}
-}

internal/sandbox/manager.go

@@ -19,6 +19,13 @@ var errWindowsSandboxNotInitialized = errors.New(
 const windowsSetupDowngradeReason = "Windows OS sandbox inactive — commands run with workspace path-confinement and " +
 	"per-command approval only. Run `zero sandbox setup` from an elevated (Administrator) terminal for full filesystem and network isolation."
 
+// windowsUnelevatedDowngradeReason is surfaced when Windows runs the sandbox in
+// the unelevated tier: the restricted-token write-jail is enforced (the command
+// runner applies the workspace ACLs itself, which needs no Administrator
+// rights), but the WFP network filters are Administrator-only and absent.
+const windowsUnelevatedDowngradeReason = "Windows sandbox running unelevated — the filesystem write-jail is enforced, but network " +
+	"isolation still relies on per-command approval. Run `zero sandbox setup` from an elevated (Administrator) terminal for OS-level network enforcement."
+
 // windowsSandboxInitialized reports whether the per-host Windows sandbox setup
 // marker exists. A missing marker is the common fresh-install state, so the
 // caller degrades rather than bricking the command. Indirected through a var so
@@ -174,19 +181,28 @@ func (manager SandboxManager) BuildExecutionRequest(request SandboxManagerReques
 	if request.ValidateExecution && preference == SandboxPreferenceRequire && backend.SupportLevel() != BackendSupportNative {
 		return SandboxExecutionRequest{}, nativeSandboxUnavailableError(backend)
 	}
-	// Windows: the OS sandbox needs a one-time elevated `zero sandbox setup` (it
-	// applies WFP network filters + workspace ACLs and writes a marker). Without
-	// it the command runner hard-fails — so on the default preference, DEGRADE to
-	// the in-process policy gate + per-command approval (matching Linux/macOS when
-	// their sandbox is unavailable) instead of bricking every command. A strict
-	// caller (--sandbox require) still gets a clear error pointing at setup.
+	// Windows: the FULL OS sandbox needs a one-time elevated `zero sandbox setup`
+	// (it applies WFP network filters + workspace ACLs and writes a marker).
+	// Without it, a restricted-filesystem profile can still run in the UNELEVATED
+	// tier: the command runner applies the workspace ACL plan itself (DACL edits
+	// on user-owned roots need no Administrator rights) and wraps the command in
+	// the write-restricted token, so the filesystem write-jail holds. Only the
+	// WFP network filters are Administrator-only, so network stays with the
+	// in-process approval gate at that tier. A profile that needs the sandbox
+	// solely for network isolation gains nothing from the unelevated tier and
+	// DEGRADES to the policy gate as before. A strict caller (--sandbox require)
+	// still gets a clear error pointing at setup.
 	windowsNeedsSetup := false
 	if manager.goos == "windows" && requiresPlatformSandbox && enforcementLevel == EnforcementNative && !windowsSandboxInitialized() {
 		if preference == SandboxPreferenceRequire {
 			return SandboxExecutionRequest{}, errWindowsSandboxNotInitialized
 		}
-		enforcementLevel = EnforcementDegraded
 		windowsNeedsSetup = true
+		if profile.FileSystem.Kind == FileSystemRestricted {
+			enforcementLevel = EnforcementUnelevated
+		} else {
+			enforcementLevel = EnforcementDegraded
+		}
 	}
 	targetBackend := manager.targetBackend(preference, policy, requiresPlatformSandbox)
 	downgradeReason := ""
@@ -196,7 +212,11 @@ func (manager SandboxManager) BuildExecutionRequest(request SandboxManagerReques
 			downgradeReason = windowsSetupDowngradeReason
 		}
 	}
-	wrapped := backend.CommandWrapping && backend.Available && enforcementLevel == EnforcementNative
+	if enforcementLevel == EnforcementUnelevated {
+		downgradeReason = windowsUnelevatedDowngradeReason
+	}
+	wrapped := backend.CommandWrapping && backend.Available &&
+		(enforcementLevel == EnforcementNative || enforcementLevel == EnforcementUnelevated)
 	markers := backend.SandboxEnvMarkers(policy)
 	if !wrapped && backend.Name != BackendWSL {
 		markers = nil

internal/sandbox/runner_windows_integration_test.go

@@ -132,6 +132,64 @@ exit 0
 	}
 }
 
+// TestWindowsUnelevatedRealSandboxSmoke exercises the unelevated tier end to
+// end WITHOUT running the elevated setup helper: the command runner applies
+// the workspace ACLs itself, a write inside the workspace succeeds under the
+// restricted token, and a write outside every granted root is denied. Unlike
+// the elevated smoke above it needs no Administrator terminal.
+func TestWindowsUnelevatedRealSandboxSmoke(t *testing.T) {
+	if os.Getenv("ZERO_SANDBOX_REAL_SMOKE") != "1" {
+		t.Skip("set ZERO_SANDBOX_REAL_SMOKE=1 to run real Windows sandbox smoke tests")
+	}
+	runnerExe := realSmokeExecutable(t, "ZERO_WINDOWS_COMMAND_RUNNER_EXE", WindowsSandboxCommandRunnerName)
+
+	root := t.TempDir()
+	outside := t.TempDir()
+	sandboxHome := filepath.Join(root, ".zero-sandbox")
+	profile := PermissionProfile{
+		FileSystem: FileSystemPolicy{
+			Kind:                 FileSystemRestricted,
+			ReadRoots:            []string{root},
+			WriteRoots:           []WritableRoot{{Root: root, ProtectedMetadataNames: []string{".git", ".zero", ".agents"}}},
+			IncludePlatformRoots: true,
+			AllowTemp:            true,
+		},
+		Network: NetworkPolicy{Mode: NetworkDeny},
+	}
+	config := WindowsSandboxCommandArgsOptions{
+		SandboxHome:       sandboxHome,
+		CommandCWD:        root,
+		WorkspaceRoots:    []string{root},
+		PermissionProfile: profile,
+		SandboxLevel:      WindowsSandboxLevelUnelevated,
+	}
+
+	// cmd.exe rather than powershell.exe: managed PowerShell cannot initialize
+	// its crypto provider under a write-restricted token on some hosts
+	// (0x8009001d, the same restricted-token limitation the runner documents for
+	// Schannel), and the write-jail assertion only needs a native shell.
+	insideMarker := filepath.Join(root, "unelevated-write-ok.txt")
+	runWindowsRealSmokeCommand(t, runnerExe, config, []string{
+		"cmd.exe", "/d", "/s", "/c", "echo ok>" + insideMarker,
+	}, 0)
+	if bytes, err := os.ReadFile(insideMarker); err != nil || strings.TrimSpace(string(bytes)) != "ok" {
+		t.Fatalf("unelevated sandboxed write marker = %q, %v; want ok", bytes, err)
+	}
+	if _, err := os.Stat(WindowsUnelevatedSetupMarkerPath(sandboxHome)); err != nil {
+		t.Fatalf("expected the unelevated setup marker to be recorded: %v", err)
+	}
+
+	outsideMarker := filepath.Join(outside, "unelevated-write-denied.txt")
+	runWindowsRealSmokeCommand(t, runnerExe, config, []string{
+		"cmd.exe", "/d", "/s", "/c", "echo leaked>" + outsideMarker,
+	}, 1)
+	if _, err := os.Stat(outsideMarker); err == nil {
+		t.Fatalf("unelevated sandbox allowed a write outside every granted root")
+	} else if !os.IsNotExist(err) {
+		t.Fatalf("stat outside marker: %v", err)
+	}
+}
+
 func realSmokeExecutable(t *testing.T, envKey string, fallbackName string) string {
 	t.Helper()
 	if path := os.Getenv(envKey); path != "" {

internal/sandbox/types.go

@@ -143,9 +143,16 @@ const (
 )
 
 const (
-	EnforcementNative   EnforcementLevel = "native"
-	EnforcementDegraded EnforcementLevel = "degraded"
-	EnforcementDisabled EnforcementLevel = "disabled"
+	EnforcementNative EnforcementLevel = "native"
+	// EnforcementUnelevated is the Windows-only middle tier used when the
+	// elevated `zero sandbox setup` has not run: commands still wrap through the
+	// command runner with a write-restricted token and workspace ACLs (which the
+	// runner can apply without Administrator rights), but the WFP network
+	// filters are absent, so network isolation stays with the in-process
+	// approval gate.
+	EnforcementUnelevated EnforcementLevel = "unelevated"
+	EnforcementDegraded   EnforcementLevel = "degraded"
+	EnforcementDisabled   EnforcementLevel = "disabled"
 )
 
 type Policy struct {

internal/sandbox/windows_command_runner_windows.go

@@ -8,14 +8,21 @@ import (
 )
 
 func runWindowsSandboxCommand(config WindowsSandboxCommandConfig, stderr io.Writer) int {
-	if config.SandboxLevel != WindowsSandboxLevelRestrictedToken {
+	switch config.SandboxLevel {
+	case WindowsSandboxLevelRestrictedToken:
+		if err := ValidateWindowsSandboxSetupMarker(WindowsSandboxSetupConfigFromCommand(config)); err != nil {
+			fmt.Fprintln(stderr, WindowsSandboxCommandRunnerName+": "+err.Error())
+			return 1
+		}
+	case WindowsSandboxLevelUnelevated:
+		if err := ensureWindowsUnelevatedSetup(config); err != nil {
+			fmt.Fprintln(stderr, WindowsSandboxCommandRunnerName+": "+err.Error())
+			return 1
+		}
+	default:
 		fmt.Fprintf(stderr, "%s: unsupported Windows sandbox level %q\n", WindowsSandboxCommandRunnerName, config.SandboxLevel)
 		return 1
 	}
-	if err := ValidateWindowsSandboxSetupMarker(WindowsSandboxSetupConfigFromCommand(config)); err != nil {
-		fmt.Fprintln(stderr, WindowsSandboxCommandRunnerName+": "+err.Error())
-		return 1
-	}
 	if err := ValidateWindowsNetworkPolicy(config.PermissionProfile.Network); err != nil {
 		fmt.Fprintln(stderr, WindowsSandboxCommandRunnerName+": "+err.Error())
 		return 1
@@ -57,3 +64,32 @@ func runWindowsSandboxCommand(config WindowsSandboxCommandConfig, stderr io.Writ
 	}
 	return exitCode
 }
+
+// ensureWindowsUnelevatedSetup applies the workspace ACL plan from the current
+// (non-elevated) process so the write-restricted token has somewhere its
+// capability SIDs are granted. DACL edits on user-owned workspace and temp
+// roots need no Administrator rights; the WFP network filters DO, so this tier
+// provisions no network enforcement — the offline-marker SID composed into the
+// token stays inert until an elevated `zero sandbox setup` installs the block
+// filters. Applied plans are recorded by hash so repeat commands skip the
+// re-apply; like the elevated setup, grants are left in place (the rollback is
+// deliberately discarded) because they only name synthetic capability SIDs
+// that no other token carries.
+func ensureWindowsUnelevatedSetup(config WindowsSandboxCommandConfig) error {
+	applied, plan, err := buildWindowsUnelevatedAppliedPlan(config)
+	if err != nil {
+		return err
+	}
+	marker, err := loadWindowsUnelevatedSetupMarker(config.SandboxHome)
+	if err != nil {
+		return err
+	}
+	if marker.contains(applied) {
+		return nil
+	}
+	if _, err := applyWindowsACLPlan(plan); err != nil {
+		return fmt.Errorf("apply unelevated workspace ACLs: %w — the workspace may be on a filesystem the current user does not own; "+
+			"run `zero sandbox setup` from an elevated (Administrator) terminal, or re-run with `--sandbox forbid` to skip OS sandboxing", err)
+	}
+	return recordWindowsUnelevatedAppliedPlan(config.SandboxHome, applied)
+}

internal/sandbox/windows_degrade_test.go

@@ -2,14 +2,16 @@ package sandbox
 
 import (
 	"errors"
+	"strings"
 	"testing"
 )
 
-// On Windows, a not-yet-set-up sandbox must DEGRADE (run unwrapped with a
-// downgrade reason) on the default preference — never brick the command — while
-// a strict --sandbox require still errors with a setup hint. Once the marker
-// exists, enforcement is native and the command wraps.
-func TestWindowsDegradesWhenSandboxNotInitialized(t *testing.T) {
+// On Windows, a not-yet-set-up sandbox must fall back to the UNELEVATED tier
+// (wrapped through the runner with the write-restricted token, network left to
+// the approval gate) on the default preference — never brick the command —
+// while a strict --sandbox require still errors with a setup hint. Once the
+// marker exists, enforcement is native and the command wraps.
+func TestWindowsFallsBackUnelevatedWhenSandboxNotInitialized(t *testing.T) {
 	restore := windowsSandboxInitialized
 	t.Cleanup(func() { windowsSandboxInitialized = restore })
 
@@ -24,13 +26,70 @@ func TestWindowsDegradesWhenSandboxNotInitialized(t *testing.T) {
 		ValidateExecution: true,
 	}
 
-	// Marker missing + auto -> degrade (unwrapped, with a reason), no error.
+	// Marker missing + auto -> unelevated tier: still wrapped, with a reason
+	// documenting the missing network enforcement, no error.
 	windowsSandboxInitialized = func() bool { return false }
 	auto := base
 	auto.Preference = SandboxPreferenceAuto
 	req, err := mgr.BuildExecutionRequest(auto)
 	if err != nil {
-		t.Fatalf("auto must degrade, not error: %v", err)
+		t.Fatalf("auto must fall back, not error: %v", err)
+	}
+	if req.EnforcementLevel != EnforcementUnelevated {
+		t.Fatalf("enforcement = %v, want unelevated", req.EnforcementLevel)
+	}
+	if !req.CommandWrapped {
+		t.Fatal("unelevated command must still be wrapped through the sandbox runner")
+	}
+	if req.DowngradeReason == "" || !strings.Contains(req.DowngradeReason, "network") {
+		t.Fatalf("downgrade reason = %q, want an explanation of the missing network enforcement", req.DowngradeReason)
+	}
+
+	// Marker missing + require -> hard error pointing at setup.
+	require := base
+	require.Preference = SandboxPreferenceRequire
+	if _, err := mgr.BuildExecutionRequest(require); !errors.Is(err, errWindowsSandboxNotInitialized) {
+		t.Fatalf("require must error with the setup hint, got %v", err)
+	}
+
+	// Marker present -> native, wrapped, no downgrade reason.
+	windowsSandboxInitialized = func() bool { return true }
+	req, err = mgr.BuildExecutionRequest(auto)
+	if err != nil {
+		t.Fatalf("initialized auto: %v", err)
+	}
+	if req.EnforcementLevel != EnforcementNative || !req.CommandWrapped {
+		t.Fatalf("initialized -> native wrapped, got enforcement=%v wrapped=%v", req.EnforcementLevel, req.CommandWrapped)
+	}
+	if req.DowngradeReason != "" {
+		t.Fatalf("native downgrade reason = %q, want empty", req.DowngradeReason)
+	}
+}
+
+// A profile that needs the platform sandbox ONLY for network isolation gains
+// nothing from the unelevated tier (which cannot enforce network), so it must
+// keep the old degrade-to-policy-gate behavior.
+func TestWindowsNetworkOnlyProfileStillDegradesWhenNotInitialized(t *testing.T) {
+	restore := windowsSandboxInitialized
+	t.Cleanup(func() { windowsSandboxInitialized = restore })
+	windowsSandboxInitialized = func() bool { return false }
+
+	mgr := NewSandboxManager(SandboxManagerOptions{
+		GOOS:    "windows",
+		Backend: Backend{Name: BackendWindowsRestrictedToken, Available: true, Executable: "zero.exe", Platform: "windows"},
+	})
+	req, err := mgr.BuildExecutionRequest(SandboxManagerRequest{
+		WorkspaceRoot: `C:\ws`,
+		Command:       CommandSpec{Name: "bash", Args: []string{"-c", "echo hi"}},
+		Policy:        DefaultPolicy(),
+		Profile: PermissionProfile{
+			FileSystem: FileSystemPolicy{Kind: FileSystemUnrestricted, IncludePlatformRoots: true, AllowTemp: true},
+			Network:    NetworkPolicy{Mode: NetworkDeny},
+		},
+		Preference: SandboxPreferenceAuto,
+	})
+	if err != nil {
+		t.Fatalf("network-only profile must degrade, not error: %v", err)
 	}
 	if req.EnforcementLevel != EnforcementDegraded {
 		t.Fatalf("enforcement = %v, want degraded", req.EnforcementLevel)
@@ -41,21 +100,4 @@ func TestWindowsDegradesWhenSandboxNotInitialized(t *testing.T) {
 	if req.DowngradeReason == "" {
 		t.Fatal("expected a downgrade reason explaining the missing setup")
 	}
-
-	// Marker missing + require -> hard error pointing at setup.
-	require := base
-	require.Preference = SandboxPreferenceRequire
-	if _, err := mgr.BuildExecutionRequest(require); !errors.Is(err, errWindowsSandboxNotInitialized) {
-		t.Fatalf("require must error with the setup hint, got %v", err)
-	}
-
-	// Marker present -> native, wrapped.
-	windowsSandboxInitialized = func() bool { return true }
-	req, err = mgr.BuildExecutionRequest(auto)
-	if err != nil {
-		t.Fatalf("initialized auto: %v", err)
-	}
-	if req.EnforcementLevel != EnforcementNative || !req.CommandWrapped {
-		t.Fatalf("initialized -> native wrapped, got enforcement=%v wrapped=%v", req.EnforcementLevel, req.CommandWrapped)
-	}
 }

internal/sandbox/windows_runner.go

@@ -95,8 +95,14 @@ type WindowsSandboxLevel string
 
 const (
 	WindowsSandboxLevelRestrictedToken WindowsSandboxLevel = "restricted-token"
-	WindowsSandboxLevelElevated        WindowsSandboxLevel = "elevated"
-	WindowsSandboxLevelDisabled        WindowsSandboxLevel = "disabled"
+	// WindowsSandboxLevelUnelevated is the restricted-token tier without the
+	// elevated setup: the runner applies the workspace ACL plan itself before
+	// launching (no Administrator rights needed) and skips the elevated setup
+	// marker check. Network is NOT enforced at this level — the WFP filters need
+	// the elevated setup — so callers must keep network behind the approval gate.
+	WindowsSandboxLevelUnelevated WindowsSandboxLevel = "unelevated"
+	WindowsSandboxLevelElevated   WindowsSandboxLevel = "elevated"
+	WindowsSandboxLevelDisabled   WindowsSandboxLevel = "disabled"
 )
 
 type WindowsSandboxCommandArgsOptions struct {
@@ -271,13 +277,20 @@ func windowsRestrictedTokenCommandPlan(execRequest SandboxExecutionRequest, poli
 		return CommandPlan{}, err
 	}
 	childEnv := windowsSandboxChildEnv(spec.Env, policy, execRequest.WorkspaceRoot)
+	// The unelevated enforcement tier maps to the runner's unelevated level: same
+	// restricted token, but the runner applies the workspace ACLs itself instead
+	// of requiring the elevated setup marker.
+	level := WindowsSandboxLevelRestrictedToken
+	if execRequest.EnforcementLevel == EnforcementUnelevated {
+		level = WindowsSandboxLevelUnelevated
+	}
 	args, err := BuildWindowsSandboxCommandArgs(WindowsSandboxCommandArgsOptions{
 		SandboxHome:       sandboxHome,
 		CommandCWD:        spec.Dir,
 		WorkspaceRoots:    []string{execRequest.WorkspaceRoot},
 		PermissionProfile: execRequest.PermissionProfile,
 		Env:               childEnv,
-		SandboxLevel:      WindowsSandboxLevelRestrictedToken,
+		SandboxLevel:      level,
 		Command:           append([]string{spec.Name}, spec.Args...),
 	})
 	if err != nil {
@@ -381,7 +394,7 @@ func nextWindowsSandboxFlagValue(args []string, index int) (string, int, error)
 
 func validWindowsSandboxLevel(level WindowsSandboxLevel) bool {
 	switch level {
-	case WindowsSandboxLevelRestrictedToken, WindowsSandboxLevelElevated, WindowsSandboxLevelDisabled:
+	case WindowsSandboxLevelRestrictedToken, WindowsSandboxLevelUnelevated, WindowsSandboxLevelElevated, WindowsSandboxLevelDisabled:
 		return true
 	default:
 		return false

internal/sandbox/windows_unelevated.go

@@ -0,0 +1,144 @@
+package sandbox
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+const windowsUnelevatedSetupMarkerSchemaVersion = 1
+
+// windowsUnelevatedSetupMarkerMaxPlans bounds the applied-plan history so the
+// marker cannot grow without limit when a user hops between many workspaces.
+// Oldest entries are dropped first; re-applying an already-granted plan is
+// harmless, so eviction only costs a redundant ACL apply.
+const windowsUnelevatedSetupMarkerMaxPlans = 64
+
+// WindowsUnelevatedAppliedPlan fingerprints one workspace ACL plan the
+// unelevated tier has applied. Unlike the elevated setup marker it carries no
+// network infrastructure hash: the unelevated tier never provisions WFP
+// filters, so there is nothing network-side to fingerprint.
+type WindowsUnelevatedAppliedPlan struct {
+	ACLPlanHash    string `json:"aclPlanHash"`
+	ACLPlanEntries int    `json:"aclPlanEntries"`
+}
+
+// WindowsUnelevatedSetupMarker records which workspace ACL plans the command
+// runner has applied without elevation, keyed by plan hash, so repeat commands
+// in a known workspace skip the re-apply. One file serves every workspace under
+// a sandbox home (the elevated marker is instead a single-plan snapshot,
+// because its WFP half genuinely is machine-global).
+type WindowsUnelevatedSetupMarker struct {
+	SchemaVersion int                            `json:"schemaVersion"`
+	AppliedPlans  []WindowsUnelevatedAppliedPlan `json:"appliedPlans,omitempty"`
+}
+
+func WindowsUnelevatedSetupMarkerPath(sandboxHome string) string {
+	return filepath.Join(filepath.Clean(sandboxHome), "windows-unelevated-setup.json")
+}
+
+// buildWindowsUnelevatedAppliedPlan derives the ACL plan for the command's
+// permission profile plus its marker fingerprint. The same BuildWindowsACLPlan
+// output later feeds the apply step, so the fingerprint and the applied grants
+// can never drift apart.
+func buildWindowsUnelevatedAppliedPlan(config WindowsSandboxCommandConfig) (WindowsUnelevatedAppliedPlan, WindowsACLPlan, error) {
+	plan, err := BuildWindowsACLPlan(config)
+	if err != nil {
+		return WindowsUnelevatedAppliedPlan{}, WindowsACLPlan{}, err
+	}
+	hash, err := WindowsACLPlanHash(plan)
+	if err != nil {
+		return WindowsUnelevatedAppliedPlan{}, WindowsACLPlan{}, err
+	}
+	return WindowsUnelevatedAppliedPlan{
+		ACLPlanHash:    hash,
+		ACLPlanEntries: len(plan.Entries),
+	}, plan, nil
+}
+
+// loadWindowsUnelevatedSetupMarker reads the marker; a missing file is the
+// common first-run state and returns an empty marker, not an error. A marker
+// with an unexpected schema version is treated as empty so the runner re-applies
+// (and rewrites) rather than failing.
+func loadWindowsUnelevatedSetupMarker(sandboxHome string) (WindowsUnelevatedSetupMarker, error) {
+	sandboxHome = strings.TrimSpace(sandboxHome)
+	if sandboxHome == "" {
+		return WindowsUnelevatedSetupMarker{}, errors.New("windows sandbox home is required")
+	}
+	path := WindowsUnelevatedSetupMarkerPath(sandboxHome)
+	bytes, err := os.ReadFile(path)
+	if err != nil {
+		if errors.Is(err, os.ErrNotExist) {
+			return WindowsUnelevatedSetupMarker{SchemaVersion: windowsUnelevatedSetupMarkerSchemaVersion}, nil
+		}
+		return WindowsUnelevatedSetupMarker{}, fmt.Errorf("read windows unelevated setup marker: %w", err)
+	}
+	var marker WindowsUnelevatedSetupMarker
+	if err := json.Unmarshal(bytes, &marker); err != nil {
+		return WindowsUnelevatedSetupMarker{}, fmt.Errorf("parse windows unelevated setup marker %s: %w", path, err)
+	}
+	if marker.SchemaVersion != windowsUnelevatedSetupMarkerSchemaVersion {
+		return WindowsUnelevatedSetupMarker{SchemaVersion: windowsUnelevatedSetupMarkerSchemaVersion}, nil
+	}
+	return marker, nil
+}
+
+func (marker WindowsUnelevatedSetupMarker) contains(applied WindowsUnelevatedAppliedPlan) bool {
+	if strings.TrimSpace(applied.ACLPlanHash) == "" {
+		return false
+	}
+	for _, plan := range marker.AppliedPlans {
+		if plan.ACLPlanHash == applied.ACLPlanHash && plan.ACLPlanEntries == applied.ACLPlanEntries {
+			return true
+		}
+	}
+	return false
+}
+
+// recordWindowsUnelevatedAppliedPlan appends the plan to the marker (dropping
+// the oldest entries past the cap) and writes it atomically, following the
+// temp-file-then-rename pattern of the sibling marker writers.
+func recordWindowsUnelevatedAppliedPlan(sandboxHome string, applied WindowsUnelevatedAppliedPlan) error {
+	marker, err := loadWindowsUnelevatedSetupMarker(sandboxHome)
+	if err != nil {
+		return err
+	}
+	if marker.contains(applied) {
+		return nil
+	}
+	marker.SchemaVersion = windowsUnelevatedSetupMarkerSchemaVersion
+	marker.AppliedPlans = append(marker.AppliedPlans, applied)
+	if overflow := len(marker.AppliedPlans) - windowsUnelevatedSetupMarkerMaxPlans; overflow > 0 {
+		marker.AppliedPlans = append([]WindowsUnelevatedAppliedPlan(nil), marker.AppliedPlans[overflow:]...)
+	}
+	path := WindowsUnelevatedSetupMarkerPath(sandboxHome)
+	if err := os.MkdirAll(filepath.Dir(path), 0o700); err != nil {
+		return fmt.Errorf("create windows unelevated setup marker dir: %w", err)
+	}
+	bytes, err := json.MarshalIndent(marker, "", "  ")
+	if err != nil {
+		return fmt.Errorf("marshal windows unelevated setup marker: %w", err)
+	}
+	tmp, err := os.CreateTemp(filepath.Dir(path), ".windows-unelevated-setup-*.tmp")
+	if err != nil {
+		return fmt.Errorf("create windows unelevated setup marker temp file: %w", err)
+	}
+	tmpPath := tmp.Name()
+	if _, err := tmp.Write(bytes); err != nil {
+		_ = tmp.Close()
+		_ = os.Remove(tmpPath)
+		return fmt.Errorf("write windows unelevated setup marker temp file: %w", err)
+	}
+	if err := tmp.Close(); err != nil {
+		_ = os.Remove(tmpPath)
+		return fmt.Errorf("close windows unelevated setup marker temp file: %w", err)
+	}
+	if err := os.Rename(tmpPath, path); err != nil {
+		_ = os.Remove(tmpPath)
+		return fmt.Errorf("replace windows unelevated setup marker: %w", err)
+	}
+	return nil
+}

internal/sandbox/windows_unelevated_test.go

@@ -0,0 +1,157 @@
+package sandbox
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func testWindowsUnelevatedCommandConfig(t *testing.T) WindowsSandboxCommandConfig {
+	t.Helper()
+	root := t.TempDir()
+	return WindowsSandboxCommandConfig{
+		SandboxHome:    filepath.Join(root, ".zero-sandbox"),
+		CommandCWD:     root,
+		WorkspaceRoots: []string{root},
+		PermissionProfile: PermissionProfile{
+			FileSystem: FileSystemPolicy{
+				Kind:       FileSystemRestricted,
+				ReadRoots:  []string{root},
+				WriteRoots: []WritableRoot{{Root: root, ProtectedMetadataNames: []string{".git", ".zero", ".agents"}}},
+				AllowTemp:  true,
+			},
+			Network: NetworkPolicy{Mode: NetworkDeny},
+		},
+		SandboxLevel: WindowsSandboxLevelUnelevated,
+		Command:      []string{"cmd.exe", "/c", "echo hi"},
+	}
+}
+
+func TestWindowsSandboxCommandArgsRoundTripUnelevatedLevel(t *testing.T) {
+	config := testWindowsUnelevatedCommandConfig(t)
+	args, err := BuildWindowsSandboxCommandArgs(WindowsSandboxCommandArgsOptions{
+		SandboxHome:       config.SandboxHome,
+		CommandCWD:        config.CommandCWD,
+		WorkspaceRoots:    config.WorkspaceRoots,
+		PermissionProfile: config.PermissionProfile,
+		SandboxLevel:      WindowsSandboxLevelUnelevated,
+		Command:           config.Command,
+	})
+	if err != nil {
+		t.Fatalf("BuildWindowsSandboxCommandArgs: %v", err)
+	}
+	parsed, err := ParseWindowsSandboxCommandArgs(args)
+	if err != nil {
+		t.Fatalf("ParseWindowsSandboxCommandArgs: %v", err)
+	}
+	if parsed.SandboxLevel != WindowsSandboxLevelUnelevated {
+		t.Fatalf("round-tripped sandbox level = %q, want %q", parsed.SandboxLevel, WindowsSandboxLevelUnelevated)
+	}
+}
+
+func TestBuildWindowsUnelevatedAppliedPlanFingerprint(t *testing.T) {
+	config := testWindowsUnelevatedCommandConfig(t)
+	applied, plan, err := buildWindowsUnelevatedAppliedPlan(config)
+	if err != nil {
+		t.Fatalf("buildWindowsUnelevatedAppliedPlan: %v", err)
+	}
+	if applied.ACLPlanHash == "" || applied.ACLPlanEntries == 0 || applied.ACLPlanEntries != len(plan.Entries) {
+		t.Fatalf("applied plan = %+v with %d plan entries, want a non-empty consistent fingerprint", applied, len(plan.Entries))
+	}
+	// Same config -> same fingerprint (the marker's dedupe key).
+	again, _, err := buildWindowsUnelevatedAppliedPlan(config)
+	if err != nil {
+		t.Fatalf("rebuild: %v", err)
+	}
+	if again != applied {
+		t.Fatalf("fingerprint not stable: %+v then %+v", applied, again)
+	}
+	// An unrestricted profile has no ACL plan to apply and must error.
+	config.PermissionProfile.FileSystem.Kind = FileSystemUnrestricted
+	if _, _, err := buildWindowsUnelevatedAppliedPlan(config); err == nil {
+		t.Fatal("expected an error for an unrestricted filesystem profile")
+	}
+}
+
+func TestWindowsUnelevatedSetupMarkerRecordAndLoad(t *testing.T) {
+	config := testWindowsUnelevatedCommandConfig(t)
+	applied, _, err := buildWindowsUnelevatedAppliedPlan(config)
+	if err != nil {
+		t.Fatalf("buildWindowsUnelevatedAppliedPlan: %v", err)
+	}
+
+	// Missing marker file is the fresh-install state: empty, no error, no match.
+	marker, err := loadWindowsUnelevatedSetupMarker(config.SandboxHome)
+	if err != nil {
+		t.Fatalf("load missing marker: %v", err)
+	}
+	if marker.contains(applied) {
+		t.Fatal("empty marker must not contain the plan")
+	}
+
+	if err := recordWindowsUnelevatedAppliedPlan(config.SandboxHome, applied); err != nil {
+		t.Fatalf("record: %v", err)
+	}
+	marker, err = loadWindowsUnelevatedSetupMarker(config.SandboxHome)
+	if err != nil {
+		t.Fatalf("load recorded marker: %v", err)
+	}
+	if !marker.contains(applied) || len(marker.AppliedPlans) != 1 {
+		t.Fatalf("marker = %+v, want exactly the recorded plan", marker)
+	}
+
+	// Recording the same plan again must not duplicate it.
+	if err := recordWindowsUnelevatedAppliedPlan(config.SandboxHome, applied); err != nil {
+		t.Fatalf("re-record: %v", err)
+	}
+	marker, err = loadWindowsUnelevatedSetupMarker(config.SandboxHome)
+	if err != nil {
+		t.Fatalf("reload marker: %v", err)
+	}
+	if len(marker.AppliedPlans) != 1 {
+		t.Fatalf("marker has %d plans after duplicate record, want 1", len(marker.AppliedPlans))
+	}
+}
+
+func TestWindowsUnelevatedSetupMarkerEvictsOldestPastCap(t *testing.T) {
+	home := filepath.Join(t.TempDir(), ".zero-sandbox")
+	first := WindowsUnelevatedAppliedPlan{ACLPlanHash: "hash-0", ACLPlanEntries: 1}
+	if err := recordWindowsUnelevatedAppliedPlan(home, first); err != nil {
+		t.Fatalf("record first: %v", err)
+	}
+	for index := 1; index <= windowsUnelevatedSetupMarkerMaxPlans; index++ {
+		plan := WindowsUnelevatedAppliedPlan{ACLPlanHash: fmt.Sprintf("hash-%d", index), ACLPlanEntries: index}
+		if err := recordWindowsUnelevatedAppliedPlan(home, plan); err != nil {
+			t.Fatalf("record %d: %v", index, err)
+		}
+	}
+	marker, err := loadWindowsUnelevatedSetupMarker(home)
+	if err != nil {
+		t.Fatalf("load: %v", err)
+	}
+	if len(marker.AppliedPlans) != windowsUnelevatedSetupMarkerMaxPlans {
+		t.Fatalf("marker has %d plans, want the cap of %d", len(marker.AppliedPlans), windowsUnelevatedSetupMarkerMaxPlans)
+	}
+	if marker.contains(first) {
+		t.Fatal("oldest plan should have been evicted past the cap")
+	}
+}
+
+func TestLoadWindowsUnelevatedSetupMarkerResetsUnknownSchema(t *testing.T) {
+	home := filepath.Join(t.TempDir(), ".zero-sandbox")
+	if err := os.MkdirAll(home, 0o700); err != nil {
+		t.Fatalf("mkdir sandbox home: %v", err)
+	}
+	path := WindowsUnelevatedSetupMarkerPath(home)
+	if err := os.WriteFile(path, []byte(`{"schemaVersion":99,"appliedPlans":[{"aclPlanHash":"h","aclPlanEntries":1}]}`), 0o600); err != nil {
+		t.Fatalf("write marker: %v", err)
+	}
+	marker, err := loadWindowsUnelevatedSetupMarker(home)
+	if err != nil {
+		t.Fatalf("load: %v", err)
+	}
+	if len(marker.AppliedPlans) != 0 || marker.SchemaVersion != windowsUnelevatedSetupMarkerSchemaVersion {
+		t.Fatalf("unknown schema must reset to empty current-schema marker, got %+v", marker)
+	}
+}

internal/tui/keybinding_help.go

@@ -124,9 +124,9 @@ func formatKeybindingLine(binding keybinding, keyColumn int, innerWidth int) str
 	return "  " + keyCell + "  " + desc
 }
 
-// renderKeybindingHelpOverlay renders the framed, centered `?` help overlay for
-// the given terminal dimensions.
-func (m model) renderKeybindingHelpOverlay(width int, height int) string {
+// the given terminal width. Vertical centering is handled by the caller's
+// overlay compositing pipeline (overlayViewportLines in transcriptView).
+func (m model) renderKeybindingHelpOverlay(width int) string {
 	overlayWidth := keybindingHelpOverlayWidth(width)
 	lines := m.renderKeybindingHelpLines(overlayWidth - 4)
 	block := styledBlockFillTitle(overlayWidth, "Keyboard Shortcuts", lines, zeroTheme.line, zeroTheme.panel)

internal/tui/keybinding_help_test.go

@@ -105,3 +105,55 @@ func TestKeybindingGroupsAreWellFormed(t *testing.T) {
 		}
 	}
 }
+
+// #419: the `?` help overlay must render ON TOP of the chat (like the model
+// picker), not REPLACE the whole screen. The old full-screen replace produced
+// only the centered shortcut block on a blank canvas — no title bar, no
+// composer. With the overlay open, surrounding chat chrome that is NOT covered
+// by the centered box (the model title bar at top, the composer at bottom) must
+// still be present alongside "Keyboard Shortcuts".
+func TestHelpOverlayCompositesOverChatNotReplacingIt(t *testing.T) {
+	m := newModel(context.Background(), Options{ModelName: "gpt-4o"})
+	m.width = 100
+	m.height = 40
+	m.altScreen = true
+
+	base := plainRender(t, m.View()) // no overlay: baseline chrome
+	m.helpOverlay = true
+	over := plainRender(t, m.View())
+
+	if !strings.Contains(over, "Keyboard Shortcuts") {
+		t.Fatalf("help overlay not rendered:\n%s", over)
+	}
+	// Chrome that renders in the baseline (and sits outside the centered overlay
+	// box) must survive behind the overlay. The full-screen replace showed none.
+	for _, marker := range []string{"gpt-4o", "describe a task"} {
+		if !strings.Contains(base, marker) {
+			t.Fatalf("precondition: baseline chat should contain %q:\n%s", marker, base)
+		}
+		if !strings.Contains(over, marker) {
+			t.Fatalf("#419: help replaced the chat instead of overlaying it; %q is gone:\n%s", marker, over)
+		}
+	}
+}
+
+// A populated transcript row also survives behind the overlay (peeking out to
+// the left of the centered box), proving the chat body — not just the chrome —
+// is composited under the overlay rather than discarded.
+func TestHelpOverlayKeepsTranscriptBodyBehindIt(t *testing.T) {
+	m := newModel(context.Background(), Options{ModelName: "gpt-4o"})
+	m.width = 120
+	m.height = 40
+	m.altScreen = true
+	m.transcript = appendTranscriptRow(m.transcript, transcriptRow{kind: rowUser, text: "hello there this is a chat line"})
+	m.helpOverlay = true
+
+	view := plainRender(t, m.View())
+	if !strings.Contains(view, "Keyboard Shortcuts") {
+		t.Fatalf("help overlay not rendered:\n%s", view)
+	}
+	// The start of the transcript line peeks to the left of the centered box.
+	if !strings.Contains(view, "hello") {
+		t.Fatalf("#419: transcript body was replaced by the help overlay:\n%s", view)
+	}
+}

internal/tui/model.go

@@ -1669,17 +1669,38 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 		// when the loop is already running or there is nothing to animate.
 		return m, m.ensureSpinnerTick()
 	case permissionRequestMsg:
+		// The agent goroutine that raised this request is BLOCKED waiting on the
+		// decision callback, so every branch below must resolve it exactly once —
+		// or store it in pendingPermission, which resolves it on the user's reply.
+		// Dropping a request without resolving parks the run forever (the reported
+		// "stuck for 33 minutes" deadlock): the agent waits on a decision channel
+		// nothing will ever signal, with no visible prompt and no network activity.
 		if msg.runID != m.activeRunID {
+			// A superseded/stale run: unblock its parked goroutine now rather than
+			// relying on that run's context cancel to fire first.
+			if msg.decide != nil {
+				msg.decide(agent.PermissionDecision{Action: agent.PermissionDecisionCancel, Reason: "run superseded"})
+			}
+			return m, nil
+		}
+		if msg.request.Action != agent.PermissionActionPrompt {
+			// Not a user-facing prompt (e.g. a sandbox-allowed command that still
+			// blocked because it requested additional permissions). The UI has
+			// nothing to ask, so resolve immediately and FAIL CLOSED — never
+			// silently grant access that was never surfaced to the user. (The agent
+			// now marks such elevation requests as prompts, so in practice this is a
+			// defensive backstop; matches the ACP handler's fail-closed contract.)
+			if msg.decide != nil {
+				msg.decide(autoResolvedPermissionDecision(msg.request.Action))
+			}
 			return m, nil
 		}
 		promptRow := permissionTranscriptRow(permissionEventFromRequest(msg.request))
 		promptRow.runID = msg.runID
 		m.transcript = appendTranscriptRow(m.transcript, promptRow)
-		if msg.request.Action == agent.PermissionActionPrompt {
-			m.pendingPermission = &pendingPermissionPrompt{
-				request: msg.request,
-				decide:  msg.decide,
-			}
+		m.pendingPermission = &pendingPermissionPrompt{
+			request: msg.request,
+			decide:  msg.decide,
 		}
 		return m, nil
 	case askUserRequestMsg:
@@ -2097,12 +2118,13 @@ func (m model) View() tea.View {
 	var content string
 	if m.setup.visible {
 		content = m.setupView(chatWidth(m.width))
-	} else if m.helpOverlay {
-		content = m.renderKeybindingHelpOverlay(chatWidth(m.width), m.height)
-	} else if m.transcriptDetailed {
-		content = m.detailedTranscriptView()
-	} else {
+	} else if m.helpOverlay || !m.transcriptDetailed {
+		// When helpOverlay is active the help panel is composited into the normal
+		// transcript view as a true overlay (scrim + vertical centering), matching
+		// how the suggestion picker / provider wizard / pickers are drawn.
 		content = m.transcriptView()
+	} else {
+		content = m.detailedTranscriptView()
 	}
 
 	view := tea.NewView(content)
@@ -2173,6 +2195,11 @@ func (m model) transcriptView() string {
 		return body + footer
 	}
 
+	helpOverlayContent := ""
+	if m.helpOverlay {
+		helpOverlayContent = m.renderKeybindingHelpOverlay(width)
+	}
+
 	suggestionOverlay := m.suggestionOverlay(width)
 	providerOverlay := m.providerWizardOverlay(width)
 	mcpAddOverlay := m.mcpAddWizardOverlay(width)
@@ -2180,6 +2207,8 @@ func (m model) transcriptView() string {
 	pickerOverlay := m.pickerOverlay(width)
 	viewportOverlay := ""
 	switch {
+	case helpOverlayContent != "":
+		viewportOverlay = helpOverlayContent
 	case providerOverlay != "":
 		viewportOverlay = providerOverlay
 	case mcpAddOverlay != "":
@@ -4525,6 +4554,21 @@ func (m model) sendPermissionRequest(runID int, request agent.PermissionRequest,
 	m.runtimeMessageSink(permissionRequestMsg{runID: runID, request: request, decide: decide})
 }
 
+// autoResolvedPermissionDecision resolves a permission request the TUI cannot
+// turn into a user prompt (Action != prompt). The agent is blocked awaiting a
+// decision, so one must ALWAYS be produced. Only an explicit Cancel is honored
+// as such; every other non-prompt action — including allow — is DENIED, so the
+// UI never silently grants access it did not surface for approval.
+func autoResolvedPermissionDecision(action agent.PermissionAction) agent.PermissionDecision {
+	if action == agent.PermissionActionCancel {
+		return agent.PermissionDecision{Action: agent.PermissionDecisionCancel, Reason: "run cancelled"}
+	}
+	return agent.PermissionDecision{
+		Action: agent.PermissionDecisionDeny,
+		Reason: "permission request could not be surfaced for approval",
+	}
+}
+
 func (m model) sendAskUserRequest(runID int, request agent.AskUserRequest, answer func([]string)) {
 	if m.runtimeMessageSink == nil {
 		return

internal/tui/model_picker_display_test.go

@@ -0,0 +1,57 @@
+package tui
+
+import "testing"
+
+// The model picker must title each row with the model NAME, not the models.dev
+// marketing description. Regression: Ollama Cloud models resolved full-sentence
+// descriptions from the remote catalog, so rows showed sentences and several
+// models sharing one description looked like exact duplicates (issue: /model
+// list rendered descriptions as titles).
+func TestModelPickerDisplayNamePrefersModelName(t *testing.T) {
+	sentence := "DeepSeek chat model for instruction following, coding, and analysis"
+	cases := []struct {
+		id          string
+		description string
+		want        string
+	}{
+		{"deepseek-v3.2", sentence, "Deepseek V3.2"},
+		{"deepseek-v3.2-thinking", sentence, "Deepseek V3.2 Thinking"},
+		{"glm-5.2", "GLM flagship model", "GLM 5.2"},
+		{"qwen3-coder:480b", "Qwen coding agent model for repository tasks", "Qwen3 Coder 480b"},
+		{"anthropic/claude-sonnet-4.6", "some blurb", "Claude Sonnet 4.6"},
+	}
+	for _, tc := range cases {
+		if got := modelPickerDisplayName(tc.id, tc.description); got != tc.want {
+			t.Errorf("modelPickerDisplayName(%q, %q) = %q, want %q", tc.id, tc.description, got, tc.want)
+		}
+	}
+}
+
+// Two distinct model ids that share the same catalog description must render as
+// distinct rows (the "duplicate-looking rows" symptom), never as the shared
+// sentence.
+func TestModelPickerDisplayNameDistinctForSharedDescription(t *testing.T) {
+	shared := "Mistral coding agent model for repository tasks and software engineering"
+	a := modelPickerDisplayName("codestral-2", shared)
+	b := modelPickerDisplayName("devstral-2", shared)
+	if a == b {
+		t.Fatalf("shared description collapsed rows to the same title: %q", a)
+	}
+	if a == shared || b == shared {
+		t.Fatalf("row title used the description sentence: a=%q b=%q", a, b)
+	}
+}
+
+// With no id to name the row, a non-generic description is still an acceptable
+// fallback; a generic placeholder (or nothing) falls back to "model".
+func TestModelPickerDisplayNameFallsBackToDescriptionOnlyWithoutID(t *testing.T) {
+	if got := modelPickerDisplayName("", "Friendly Name"); got != "Friendly Name" {
+		t.Fatalf("empty id should fall back to the description, got %q", got)
+	}
+	if got := modelPickerDisplayName("", "catalog default"); got != "model" {
+		t.Fatalf("generic description with no id should yield %q, got %q", "model", got)
+	}
+	if got := modelPickerDisplayName("", ""); got != "model" {
+		t.Fatalf("empty id and description should yield %q, got %q", "model", got)
+	}
+}

internal/tui/permission_deadlock_test.go

@@ -0,0 +1,109 @@
+package tui
+
+import (
+	"context"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/agent"
+)
+
+func permissionModel(t *testing.T) model {
+	t.Helper()
+	m := newModel(context.Background(), Options{})
+	m.pending = true
+	m.activeRunID = 7
+	m.width = 96
+	return m
+}
+
+// The deadlock: a permission request whose Action is NOT prompt (e.g. a
+// sandbox-allowed command that still blocked to request additional permissions)
+// must be resolved immediately — the agent goroutine is parked awaiting a
+// decision. Before the fix the handler dropped it, and the run hung forever
+// (the reported "stuck for 33 minutes"). Fail closed: a non-prompt allow denies.
+func TestPermissionRequestNonPromptResolvesFailClosed(t *testing.T) {
+	m := permissionModel(t)
+	var got *agent.PermissionDecision
+	decide := func(d agent.PermissionDecision) { got = &d }
+
+	updated, _ := m.Update(permissionRequestMsg{
+		runID:   7,
+		request: agent.PermissionRequest{ToolCallID: "c1", ToolName: "bash", Action: agent.PermissionActionAllow},
+		decide:  decide,
+	})
+	next := updated.(model)
+
+	if got == nil {
+		t.Fatal("DEADLOCK: decide was never called for a non-prompt request — the agent goroutine would park forever")
+	}
+	if got.Action != agent.PermissionDecisionDeny {
+		t.Fatalf("decision = %q, want deny (fail closed — never silently grant an unsurfaced permission)", got.Action)
+	}
+	if next.pendingPermission != nil {
+		t.Fatal("a non-prompt request must not open a pending prompt")
+	}
+}
+
+// A superseded/stale run's request (runID mismatch) must also be resolved
+// (cancelled) so its parked goroutine unblocks instead of depending on
+// context-cancel ordering.
+func TestPermissionRequestStaleRunResolvesCancel(t *testing.T) {
+	m := permissionModel(t) // activeRunID = 7
+	var got *agent.PermissionDecision
+	decide := func(d agent.PermissionDecision) { got = &d }
+
+	m.Update(permissionRequestMsg{
+		runID:   3, // != activeRunID
+		request: agent.PermissionRequest{ToolCallID: "c1", ToolName: "bash", Action: agent.PermissionActionPrompt},
+		decide:  decide,
+	})
+
+	if got == nil {
+		t.Fatal("stale-run request was dropped without resolving — its agent goroutine would park")
+	}
+	if got.Action != agent.PermissionDecisionCancel {
+		t.Fatalf("decision = %q, want cancel for a superseded run", got.Action)
+	}
+}
+
+// An explicit-cancel action is honored as a cancel (not denied).
+func TestPermissionRequestCancelActionResolvesCancel(t *testing.T) {
+	m := permissionModel(t)
+	var got *agent.PermissionDecision
+	decide := func(d agent.PermissionDecision) { got = &d }
+
+	m.Update(permissionRequestMsg{
+		runID:   7,
+		request: agent.PermissionRequest{ToolCallID: "c1", ToolName: "bash", Action: agent.PermissionActionCancel},
+		decide:  decide,
+	})
+	if got == nil || got.Action != agent.PermissionDecisionCancel {
+		t.Fatalf("cancel action must resolve as cancel, got %#v", got)
+	}
+}
+
+// A genuine prompt request stores the pending prompt and does NOT resolve yet —
+// the decision comes from the user's reply. (Regression guard: the fix must not
+// accidentally auto-resolve real prompts.)
+func TestPermissionRequestPromptDefersToUser(t *testing.T) {
+	m := permissionModel(t)
+	resolved := false
+	decide := func(agent.PermissionDecision) { resolved = true }
+
+	updated, _ := m.Update(permissionRequestMsg{
+		runID: 7,
+		request: agent.PermissionRequest{
+			ToolCallID: "c1", ToolName: "bash", Action: agent.PermissionActionPrompt,
+			Reason: "runs a shell command", AvailableDecisions: []agent.PermissionDecisionAction{agent.PermissionDecisionAllow},
+		},
+		decide: decide,
+	})
+	next := updated.(model)
+
+	if resolved {
+		t.Fatal("a real prompt must not be auto-resolved; it awaits the user")
+	}
+	if next.pendingPermission == nil {
+		t.Fatal("a prompt request must open a pending permission prompt")
+	}
+}

internal/tui/picker.go

@@ -452,11 +452,17 @@ func registryModelPickerMeta(entry modelregistry.ModelEntry) string {
 }
 
 func modelPickerDisplayName(id string, description string) string {
-	if description = strings.TrimSpace(description); description != "" && !providerWizardGenericModelDescription(description) {
-		return description
-	}
+	// The row title is the model NAME, not its marketing blurb. Catalog/discovery
+	// descriptions (from models.dev) are full sentences, and several models share
+	// the same one — using them as titles showed sentence-long rows that looked like
+	// exact duplicates and hid the actual id (which only survived in the meta line).
+	// Prefer the (prettified) id; fall back to a non-generic description only when
+	// there is no id to name the row.
 	id = strings.TrimSpace(id)
 	if id == "" {
+		if description = strings.TrimSpace(description); description != "" && !providerWizardGenericModelDescription(description) {
+			return description
+		}
 		return "model"
 	}
 	name := id

internal/tui/provider_wizard_discovery.go

@@ -44,11 +44,19 @@ func (m model) advanceProviderWizard() (model, tea.Cmd) {
 	// keep/replace/remove before re-entering credentials.
 	if m.providerWizard.step == providerWizardStepProvider && !m.providerWizard.oauthMode {
 		if name, ok := m.wizardProviderStoredKey(m.providerWizard.currentProvider()); ok {
-			m.providerWizard.manageProviderName = name
-			m.providerWizard.manageKeyCursor = 0
-			m.providerWizard.err = ""
-			m.providerWizard.step = providerWizardStepManageKey
-			return m, nil
+			// Generic/custom providers (custom-openai-compatible etc.) all share
+			// the same CatalogID — matching on CatalogID would block creating a
+			// second instance. Skip ManageKey and fall through to the shared
+			// advance() path below; the user can overwrite by re-entering the
+			// same name or create a new one with a different name.
+			if !providerWizardNeedsEndpoint(m.providerWizard.currentProvider()) {
+				m.providerWizard.manageProviderName = name
+				m.providerWizard.manageKeyCursor = 0
+				m.providerWizard.err = ""
+				m.providerWizard.step = providerWizardStepManageKey
+				return m, nil
+			}
+			m.providerWizard.manageProviderName = ""
 		}
 	}
 	previous := m.providerWizard.step

internal/tui/provider_wizard_test.go

@@ -1167,6 +1167,71 @@ func TestProviderWizardManageKeyReplaceAndKeep(t *testing.T) {
 	}
 }
 
+func TestAdvanceProviderWizardCustomSkipsManageKey(t *testing.T) {
+	tests := []struct {
+		name      string
+		catalogID string
+		transport providercatalog.Transport
+	}{
+		{
+			name:      "custom-openai-compatible",
+			catalogID: "custom-openai-compatible",
+			transport: providercatalog.TransportOpenAICompatible,
+		},
+		{
+			name:      "custom-anthropic-compatible",
+			catalogID: "custom-anthropic-compatible",
+			transport: providercatalog.TransportAnthropicCompatible,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			m := model{savedProviders: []config.ProviderProfile{
+				{Name: "my-gateway", CatalogID: tt.catalogID, APIKeyStored: true},
+			}}
+			m.providerWizard = &providerWizardState{
+				step: providerWizardStepProvider,
+				providers: []providercatalog.Descriptor{
+					{ID: tt.catalogID, Name: tt.name, Transport: tt.transport},
+				},
+				selectedProvider: 0,
+			}
+
+			next, _ := m.advanceProviderWizard()
+			if next.providerWizard == nil {
+				t.Fatal("wizard should not be nil after advancing from provider step")
+			}
+			if next.providerWizard.step == providerWizardStepManageKey {
+				t.Fatal("custom provider should skip ManageKey and route to endpoint, got ManageKey")
+			}
+			if next.providerWizard.step != providerWizardStepEndpoint {
+				t.Fatalf("custom provider should route to endpoint step, got step: %v", next.providerWizard.step)
+			}
+		})
+	}
+}
+
+func TestAdvanceProviderWizardNamedShowsManageKey(t *testing.T) {
+	m := model{savedProviders: []config.ProviderProfile{
+		{Name: "openai", CatalogID: "openai", APIKeyStored: true},
+	}}
+	m.providerWizard = &providerWizardState{
+		step: providerWizardStepProvider,
+		providers: []providercatalog.Descriptor{
+			{ID: "openai", Name: "OpenAI", Transport: providercatalog.TransportOpenAI},
+		},
+		selectedProvider: 0,
+	}
+
+	next, _ := m.advanceProviderWizard()
+	if next.providerWizard == nil {
+		t.Fatal("wizard should not be nil after advancing from provider step")
+	}
+	if next.providerWizard.step != providerWizardStepManageKey {
+		t.Fatalf("named provider should route to ManageKey step, got step: %v", next.providerWizard.step)
+	}
+}
+
 func readProviderWizardConfigFixture(t *testing.T, path string) config.FileConfig {
 	t.Helper()
 

internal/tui/sidebar.go

@@ -70,7 +70,7 @@ func (m model) sidebarAvailable() bool {
 	// list) take over the chat column and render at full width; suppress the
 	// second column while any is active so their geometry and mouse hit-testing
 	// stay full-width as before.
-	if m.setup.visible || m.providerWizard != nil || m.mcpAddWizard != nil ||
+	if m.setup.visible || m.helpOverlay || m.providerWizard != nil || m.mcpAddWizard != nil ||
 		m.mcpManager != nil || m.picker != nil || m.suggestionsActive() {
 		return false
 	}

internal/zeroruntime/helpers.go

@@ -21,12 +21,6 @@ type CollectedStream struct {
 	// thinking blocks) that must be replayed on the next turn. Empty for providers
 	// or runs without extended thinking.
 	ReasoningBlocks []ReasoningBlock
-	// ReasoningEmitted reports that the stream carried ANY reasoning signal
-	// (live reasoning deltas or preserved blocks). It distinguishes "the model
-	// thought but produced no answer" (a behavioral empty turn) from "the
-	// provider returned a contentless stream" (a transport/backend fault worth
-	// retrying) — the two need different handling in the agent loop.
-	ReasoningEmitted bool
 }
 
 // Truncated reports whether the response ended for a non-normal reason (the
@@ -126,7 +120,6 @@ func CollectStreamWithOptions(ctx context.Context, events <-chan StreamEvent, op
 			// accumulate them regardless of type so they survive for replay.
 			if len(event.ReasoningBlocks) > 0 {
 				collected.ReasoningBlocks = append(collected.ReasoningBlocks, event.ReasoningBlocks...)
-				collected.ReasoningEmitted = true
 			}
 
 			switch event.Type {
@@ -139,9 +132,6 @@ func CollectStreamWithOptions(ctx context.Context, events <-chan StreamEvent, op
 					options.OnText(event.Content)
 				}
 			case StreamEventReasoning:
-				if event.Content != "" {
-					collected.ReasoningEmitted = true
-				}
 				if options.OnReasoning != nil {
 					options.OnReasoning(event.Content)
 				}

package.json

@@ -24,7 +24,8 @@
   "os": [
     "linux",
     "darwin",
-    "win32"
+    "win32",
+    "android"
   ],
   "cpu": [
     "x64",

scripts/postinstall.mjs

@@ -14,7 +14,7 @@
 // Env overrides (testing / mirrors / locked-down installs):
 //   ZERO_SKIP_DOWNLOAD=1      skip entirely, exit 0 (wrapper will guide if run)
 //   ZERO_INSTALL_DRY_RUN=1    print the resolved plan as JSON, no network, exit 0
-//   ZERO_INSTALL_PLATFORM=…   override process.platform (linux|darwin|win32)
+//   ZERO_INSTALL_PLATFORM=…   override process.platform (linux|darwin|win32|android)
 //   ZERO_INSTALL_ARCH=…       override process.arch (x64|arm64)
 //   ZERO_REPO=owner/name      override the GitHub repo (default Gitlawb/zero)
 //   ZERO_GITHUB_BASE_URL=…    override the download host (default https://github.com)
@@ -65,6 +65,7 @@ function warnSkip(message) {
 function resolvePlatform(p) {
   switch (p) {
     case 'linux':
+    case 'android':
       return 'linux';
     case 'darwin':
       return 'macos';