feat(release): prepare v4.0.4 model discovery and runtime compatibility

AGENTS.md

@@ -66,7 +66,7 @@ attachment and optional App-control MCP tools are separate user choices.
 - **cwd must match resume**: a session's jsonl lives at
   `~/.claude/projects/<cwd-with-/-as->/<uuid>.jsonl`. `ClaudeAgentOptions.cwd`
   MUST equal the original session's cwd or `resume` can't find it.
-- **SDK pinned to `claude-agent-sdk==0.2.151`**: message-type shapes and the
+- **SDK pinned to `claude-agent-sdk==0.2.157`**: message-type shapes and the
   interrupt/drain contract can shift between patch versions. Re-run the
   interrupt+drain verification after any upgrade (`SdkHandle.preflight()` guards
   the exact verified patch at startup).
@@ -78,7 +78,7 @@ attachment and optional App-control MCP tools are separate user choices.
 - **`include_partial_messages`** is a `ClaudeAgentOptions` field (set at
   construction, not on `query()`). Streaming events arrive as `StreamEvent`
   (`.event` = raw Anthropic API stream-event dict) — NOT
-  `SDKPartialAssistantMessage` (doesn't exist in 0.2.151). Extract
+  `SDKPartialAssistantMessage` (doesn't exist in 0.2.157). Extract
   `content_block_delta` → `delta.text` from `StreamEvent.event`.
 - **tool_use is batched, not streamed**: emit one `tool_use` event from the
   assembled `AssistantMessage` (full `input`), never as JSON-fragment deltas.

CHANGELOG.md

@@ -2,6 +2,28 @@
 
 [中文](CHANGELOG_zh.md)
 
+## v4.0.4
+
+Discover Claude models from the installed CLI and improve inline code emphasis.
+Wire protocol remains v72. See the bilingual [release notes](docs/releases/v4.0.4.md).
+
+- Populate Web, side-chat and TUI model pickers with the native model catalog and
+  supported effort levels, without sending prompts or resuming sessions.
+- Refresh the catalog when opening model settings, scope its cache by account,
+  directory and CLI build, and preserve explicitly selected historical models.
+- Keep Work discovery within its runtime policy and inherit native Claude model
+  and effort defaults instead of imposing a fixed Opus/max selection.
+- Refresh Codex models on page reload, reconnect and picker open; share concurrent
+  reads per account and retain that account's last good list on discovery failure.
+- Support Codex 0.156's protected socket aliases while preserving account ownership
+  checks and detecting replacement of the underlying app-server listener.
+- Upgrade the verified Python Claude Agent SDK from 0.2.151 to 0.2.157. Wrapper
+  upgrades across this SDK boundary require the documented drain/installer path.
+- Render inline code in softer harness colors without a background badge; actual
+  links retain their link styling and fenced code blocks retain their background.
+- Expand Claude thinking text by default. Pressing `/` outside an editor focuses
+  the composer without inserting a character; pressing it again types normally.
+
 ## v4.0.3
 
 Fix duplicated and misplaced Claude messages when browsing overlapping history pages.

CHANGELOG_zh.md

@@ -2,6 +2,24 @@
 
 [English](CHANGELOG.md)
 
+## v4.0.4
+
+从本机 Claude CLI 自动发现模型，并优化行内代码高亮。通信协议保持 v72。
+详见[双语发布说明](docs/releases/v4.0.4.md)。
+
+- Web、侧聊和 TUI 的模型列表读取原生目录及支持的思考强度，不发送提示词或恢复会话。
+- 打开模型设置时刷新目录，按账号、目录和 CLI 构建缓存；保留用户明确选择的历史模型。
+- Work 模型发现遵循其实际运行策略；Claude 新会话继承原生模型和思考强度默认值，
+  不再强制使用固定的 Opus/max 组合。
+- 刷新页面、重连或打开模型菜单时重新读取 Codex 模型；同账号并发请求合并，读取
+  失败时保留该账号上次成功的列表。
+- 兼容 Codex 0.156 的受保护套接字别名，继续校验账号归属并检测底层监听器替换。
+- 已验证的 Python Claude Agent SDK 从 0.2.151 升级至 0.2.157；跨 SDK 版本升级
+  Wrapper 时，需按文档等待任务结束、迁移服务并使用安装器升级。
+- 行内代码改为更柔和的引擎主题色文字，去掉背景块；真实链接保留链接样式，代码块
+  仍保留背景。
+- Claude 思考文本默认展开；未聚焦编辑器时按 `/` 只聚焦输入框，再按一次才输入字符。
+
 ## v4.0.3
 
 修复 Claude 历史翻页时对话重复、后续问题被放到开头的问题。通信协议保持 v72。

CLAUDE.md

@@ -49,7 +49,7 @@ a separate choice; sharing alone does not authorize them.
 - **cwd must match resume**: a session's jsonl lives at
   `~/.claude/projects/<cwd-with-/-as->/<uuid>.jsonl`. `ClaudeAgentOptions.cwd`
   MUST equal the original session's cwd or `resume` can't find it.
-- **SDK pinned to `claude-agent-sdk==0.2.151`**: message-type shapes and the
+- **SDK pinned to `claude-agent-sdk==0.2.157`**: message-type shapes and the
   interrupt/drain contract can shift between patch versions. Re-run the
   interrupt+drain verification after any upgrade (`SdkHandle.preflight()` guards
   the exact verified patch at startup).
@@ -61,7 +61,7 @@ a separate choice; sharing alone does not authorize them.
 - **`include_partial_messages`** is a `ClaudeAgentOptions` field (set at
   construction, not on `query()`). Streaming events arrive as `StreamEvent`
   (`.event` = raw Anthropic API stream-event dict) — NOT
-  `SDKPartialAssistantMessage` (doesn't exist in 0.2.151). Extract
+  `SDKPartialAssistantMessage` (doesn't exist in 0.2.157). Extract
   `content_block_delta` → `delta.text` from `StreamEvent.event`.
 - **tool_use is batched, not streamed**: emit one `tool_use` event from the
   assembled `AssistantMessage` (full `input`), never as JSON-fragment deltas.

README.md

@@ -4,7 +4,7 @@
 
 自托管 · 多会话 · 多设备 · 实时工具过程 · Code / Work · Web / PWA / TUI
 
-**产品版本：v4.0.3** · Wire protocol v72
+**产品版本：v4.0.4** · Wire protocol v72
 
 [English](README_en.md) · [功能对照](#引擎与功能) · [快速开始](#快速开始) ·
 [终端工作台](#terminal-workspace) · [安装与升级](#安装与升级) · [文档](#文档) · [更新记录](CHANGELOG_zh.md)
@@ -152,7 +152,7 @@ Wrapper 主动出站连接 Relay，设备不需要开放公网入站端口。Rel
 至少准备一个可用引擎：
 
 - **Claude**：日常 Claude Code `>= 2.1.263`，默认路径 `~/.local/bin/claude`。
-  Wrapper 使用该 CLI；Python Agent SDK 固定为 `0.2.151`。
+  Wrapper 使用该 CLI；Python Agent SDK 固定为 `0.2.157`。
 - **Codex**：已登录的官方 CLI。共享控制需要同时支持
   `codex app-server daemon --help` 和 `codex app-server proxy --help`。
 

README_en.md

@@ -4,7 +4,7 @@
 
 Self-hosted · Multiple sessions and devices · Live tool activity · Code / Work · Web / PWA / TUI
 
-**Product version: v4.0.3** · Wire protocol v72
+**Product version: v4.0.4** · Wire protocol v72
 
 [中文](README.md) · [Engine comparison](#engines-and-features) · [Quick start](#quick-start) ·
 [Terminal workspace](#terminal-workspace) · [Install and upgrade](#install-and-upgrade) · [Documentation](#documentation) · [Changelog](CHANGELOG.md)
@@ -173,7 +173,7 @@ development/builds use **Python 3.13 and Node 24**, matching CI and [`.nvmrc`](.
 Prepare at least one working engine:
 
 - **Claude:** daily Claude Code `>= 2.1.263`, normally at `~/.local/bin/claude`.
-  Wrapper launches that CLI; the Python Agent SDK is pinned to `0.2.151`.
+  Wrapper launches that CLI; the Python Agent SDK is pinned to `0.2.157`.
 - **Codex:** an authenticated official CLI. Shared control requires both
   `codex app-server daemon --help` and `codex app-server proxy --help`.
 

cc_remote/__init__.py

@@ -5,4 +5,4 @@
   - control link: client <-> relay(WS) <-> wrapper <-> ClaudeSDKClient <-> cc
 """
 
-__version__ = "4.0.3"
+__version__ = "4.0.4"

cc_remote/claude_paths.py

@@ -16,7 +16,7 @@ def claude_config_dir() -> Path:
     """Return the active native Claude Code config root."""
     configured = os.environ.get("CLAUDE_CONFIG_DIR")
     if configured:
-        # Match claude-agent-sdk 0.2.151 exactly: the environment value is a
+        # Match claude-agent-sdk 0.2.157 exactly: the environment value is a
         # filesystem path, not a shell expression, so do not expand "~" or
         # resolve symlinks behind the SDK's back.
         return Path(unicodedata.normalize("NFC", configured))

cc_remote/protocol.py

@@ -1797,9 +1797,8 @@ class DirList(_Base):
 class GetModels(_Command):
     """client -> wrapper: what models does this engine actually offer?
 
-    `codex` answers with app-server's real catalog. Claude has no equivalent
-    catalog RPC, but can resolve explicit no-override settings for a cwd;
-    its model list therefore remains empty and the client keeps the static table.
+    Codex answers with app-server's model/list. Claude reads the native picker
+    from a prompt-free initialization and resolves no-override settings for cwd.
     """
     type: Literal["get_models"] = "get_models"
     engine: Optional[Literal["cc", "claude", "codex"]] = None

cc_remote/tui_settings.py

@@ -131,6 +131,8 @@ def args(self, kind):
         values = {"cwd": self.values["cwd"], **self.profiles}
         if kind == "models":
             values["engine"] = self.engine
+            if self.space == "work":
+                values.pop("cwd", None)
         else:
             values.pop("claude_profile_id", None)
         return values
@@ -237,8 +239,8 @@ def choices(self, field):
         choices = [("Engine default (no override)", None)] if self.new else []
         if field == "model":
             models = self.catalog("models").get("models", [])
-            if self.engine == "claude":
-                # Claude has no model/list RPC; same curated choices as Web.
+            if self.engine == "claude" and not models:
+                # Offline fallback only; the native CLI catalog wins.
                 models = [{"id": value} for value in CLAUDE_MODELS]
             for model in models:
                 identity = model.get("id")
@@ -258,7 +260,7 @@ def choices(self, field):
                 {},
             )
             efforts = model.get("efforts", [])
-            if self.engine == "claude":
+            if self.engine == "claude" and not model:
                 from cc_remote.wrapper.claude_controls import CLAUDE_EFFORTS
 
                 efforts = sorted(CLAUDE_EFFORTS)
@@ -467,9 +469,5 @@ def created(self, success):
 }
 
 CLAUDE_MODELS = (
-    "claude-opus-5[1m]",
-    "claude-mythos-5-1",
-    "claude-sonnet-5",
-    "claude-haiku-4-5",
-    "claude-fable-5-1",
+    "opus[1m]", "sonnet", "haiku",
 )

cc_remote/wrapper/claude_models.py

@@ -0,0 +1,181 @@
+"""Discover models from the daily Claude CLI's initialization response.
+
+This is the same native catalog exposed by the SDK's get_server_info(). No
+prompt, resume, or model request is sent. A short-lived, customization-disabled
+child preserves native account/model settings without running hooks or MCPs.
+"""
+from __future__ import annotations
+
+import asyncio
+from collections import OrderedDict
+from contextlib import ExitStack
+import json
+import os
+from pathlib import Path
+import tempfile
+import time
+
+from cc_remote.log import logger
+from cc_remote.wrapper.child_env import claude_profile_process_env
+from cc_remote.wrapper.claude_controls import valid_claude_model
+from cc_remote.wrapper.claude_runtime import resolve_claude_cli
+from cc_remote.workspaces import _claude_runtime_settings
+
+log = logger("cc_remote.wrapper.claude_models")
+_TTL = 60.0
+_TIMEOUT = 20.0
+_MAX_SCOPES = 64
+_cache: OrderedDict[tuple, tuple[float, list[dict]]] = OrderedDict()
+_lock = asyncio.Lock()
+
+
+def _normalize(raw: object) -> list[dict]:
+    if not isinstance(raw, list):
+        return []
+    models: dict[str, dict] = {}
+    for item in raw[:256]:
+        if not isinstance(item, dict):
+            continue
+        # Keep Default as a native selection: availableModels can permit the
+        # tier default while forbidding an explicit selection of its resolved
+        # ID. Named rows use resolved IDs to expose newly shipped versions.
+        value = valid_claude_model(item.get("value"))
+        identity = ("default" if value == "default" else
+                    valid_claude_model(item.get("resolvedModel")) or value)
+        if not identity:
+            continue
+        levels = item.get("supportedEffortLevels")
+        levels = levels if isinstance(levels, list) else []
+        efforts = list(dict.fromkeys(
+            level for level in levels[:16]
+            if isinstance(level, str) and 0 < len(level) <= 64
+        )) if item.get("supportsEffort") is not False else []
+        is_default = item.get("value") == "default"
+        previous = models.get(identity)
+        # Repeated rows must not create duplicate choices.
+        if previous and is_default:
+            previous["is_default"] = True
+            continue
+        models[identity] = {
+            "id": identity,
+            "display_name": str(item.get("displayName") or identity)[:4096],
+            "description": str(item.get("description") or "")[:4096],
+            "efforts": efforts,
+            "default_effort": None,
+            "is_default": is_default or bool(previous and previous["is_default"]),
+        }
+    return list(models.values())
+
+
+async def _read_catalog(
+    binary: str, cwd: str, config_dir: str | None, isolate_account_env: bool,
+    work_only: bool,
+) -> list[dict]:
+    env = claude_profile_process_env(
+        config_dir, isolate_account_env=isolate_account_env)
+    env.pop("CLAUDECODE", None)
+    args = [
+        binary, "--print", "--input-format", "stream-json",
+        "--output-format", "stream-json", "--verbose",
+        "--no-session-persistence", "--safe-mode", "--strict-mcp-config",
+        "--tools", "",
+    ]
+    with ExitStack() as scope:
+        if work_only:
+            # Match Work's sole policy source and provider allowlist. Never load
+            # user/project customizations or put provider credentials in argv.
+            settings = scope.enter_context(tempfile.NamedTemporaryFile(
+                mode="w+", prefix="cc-remote-models-", suffix=".json"))
+            json.dump(_claude_runtime_settings(config_dir), settings)
+            settings.flush()
+            args.extend(["--setting-sources", "", "--settings", settings.name])
+        elif isolate_account_env:
+            args.extend(["--setting-sources", "user"])
+        return await _initialize_catalog(args, cwd, env)
+
+
+async def _initialize_catalog(args: list[str], cwd: str, env: dict[str, str]) -> list[dict]:
+    proc = await asyncio.create_subprocess_exec(
+        *args, stdin=asyncio.subprocess.PIPE, stdout=asyncio.subprocess.PIPE,
+        stderr=asyncio.subprocess.DEVNULL, cwd=cwd, env=env,
+        limit=4 * 1024 * 1024,
+    )
+    try:
+        async with asyncio.timeout(_TIMEOUT):
+            proc.stdin.write((json.dumps({
+                "type": "control_request", "request_id": "model-catalog",
+                "request": {"subtype": "initialize"},
+            }) + "\n").encode())
+            await proc.stdin.drain()
+            while line := await proc.stdout.readline():
+                message = json.loads(line)
+                if not isinstance(message, dict):
+                    continue
+                response = message.get("response")
+                if (message.get("type") != "control_response"
+                        or not isinstance(response, dict)
+                        or response.get("request_id") != "model-catalog"):
+                    continue
+                if response.get("subtype") != "success":
+                    raise RuntimeError("Claude initialization rejected")
+                result = response.get("response")
+                return _normalize(result.get("models") if isinstance(result, dict) else None)
+            raise RuntimeError("Claude closed before model discovery")
+    finally:
+        proc.stdin.close()
+        if proc.returncode is None:
+            try:
+                proc.terminate()
+            except ProcessLookupError:
+                pass
+        try:
+            await asyncio.wait_for(proc.wait(), 3.0)
+        except asyncio.TimeoutError:
+            try:
+                proc.kill()
+            except ProcessLookupError:
+                pass
+            await proc.wait()
+
+
+async def claude_model_catalog(
+    *, claude_bin: str, cwd: str | None = None,
+    config_dir: str | None = None, isolate_account_env: bool = False,
+) -> list[dict]:
+    """Bounded account/cwd cache; CLI replacement bypasses the TTL immediately.
+
+    Missing cwd means Work discovery using its filtered runtime settings, never
+    borrowing customizations from the Wrapper's Code project or user settings.
+    """
+    async with _lock:
+        try:
+            binary, _ = resolve_claude_cli(
+                claude_bin or str(Path.home() / ".local/bin/claude"))
+            binary = os.path.realpath(binary)
+            stat = os.stat(binary)
+            target = os.path.realpath(os.path.expanduser(cwd or str(Path.home())))
+            key = (
+                binary, stat.st_mtime_ns, stat.st_size, target,
+                os.path.realpath(config_dir) if config_dir else None,
+                isolate_account_env, cwd is None,
+            )
+        except (OSError, RuntimeError):
+            return []
+        cached_at, cached = _cache.get(key, (0.0, []))
+        if key in _cache and time.monotonic() - cached_at < _TTL:
+            _cache.move_to_end(key)
+            return cached
+        try:
+            models = await _read_catalog(
+                binary, target, config_dir, isolate_account_env, cwd is None)
+        except Exception as exc:
+            # Native errors can contain provider/account information; report
+            # only their class, never dump raw initialization or stderr.
+            log.warning("Claude model discovery failed", reason=type(exc).__name__)
+            models = []
+        result = models or cached
+        _cache[key] = (time.monotonic(), result)
+        _cache.move_to_end(key)
+        while len(_cache) > _MAX_SCOPES:
+            _cache.popitem(last=False)
+        return result

cc_remote/wrapper/claude_runtime.py

@@ -16,7 +16,7 @@
 import claude_agent_sdk
 
 
-VERIFIED_SDK_VERSION = "0.2.151"
+VERIFIED_SDK_VERSION = "0.2.157"
 MINIMUM_CLAUDE_CLI_VERSION = "2.1.263"
 _CLI_VERSION_TIMEOUT = 3.0
 _VERSION_RE = re.compile(

cc_remote/wrapper/claude_transport.py

@@ -14,7 +14,7 @@ class AccountIsolatedSubprocessCLITransport(
 ):
     """Spawn one Claude child without inherited account selectors.
 
-    Agent SDK 0.2.151 treats ``ClaudeAgentOptions.env`` as an overlay on the
+    Agent SDK 0.2.157 treats ``ClaudeAgentOptions.env`` as an overlay on the
     Wrapper process environment. Empty strings are not equivalent to unsetting
     modern Claude authentication selectors, while mutating ``os.environ`` would
     race unrelated Codex and tool spawns. This pinned adapter mirrors the SDK's