zall 0.4.0__py3-none-any.whl

zall/__init__.py

@@ -0,0 +1,7 @@
+"""zall — model-agnostic engineering agent.
+
+Model-agnostic engineering agent. This package MUST NOT import any model SDK;
+Model adapters reside in `zall.adapters`. See IMPL.md IPR-3.
+"""
+
+__version__ = "0.2.7"

zall/__main__.py

@@ -0,0 +1,3 @@
+from zall.cli import main
+
+raise SystemExit(main())

zall/_util/__init__.py

@@ -0,0 +1,15 @@
+"""zall._util — 内部共享toolfunction库。
+
+本包为 zall 内部使用的工具函数, 不构成 public API。
+各模块消除重复代码用 (B22/B23/B24 等)。
+"""
+
+from zall._util.file import is_binary, read_text_file
+from zall._util.path import NOISE_DIRS, is_noise, skip_noise_dirs
+from zall._util.string import unquote
+
+__all__ = [
+    "is_binary", "read_text_file",
+    "NOISE_DIRS", "is_noise", "skip_noise_dirs",
+    "unquote",
+]

zall/_util/file.py

@@ -0,0 +1,115 @@
+"""zall._util.file — File operations共享toolfunction。"""
+
+from __future__ import annotations
+
+import locale
+from pathlib import Path
+
+
+def _preferred_encoding() -> str:
+    """Get the system's preferred encoding (Windows Chinese is GBK/CP936, do not hardcode UTF-8)."""
+    try:
+        enc = locale.getpreferredencoding(False)
+        if enc:
+            return enc
+    except (ValueError, LookupError):
+        pass
+    return "utf-8"  # fallback
+
+
+def detect_text_encoding(path: Path) -> str:
+    """Detect text file encoding: try UTF-8 first, fallback to system preferred encoding.
+
+    Most modern code/text files are UTF-8. On Chinese Windows, the system default is GBK,
+    but opening a UTF-8 file with GBK produces mojibake. This function tries UTF-8 first
+    (strict), and verifies via round-trip to avoid GBK text accidentally passing UTF-8
+    validation (e.g., GBK bytes that happen to form valid UTF-8 sequences).
+
+    Note: the returned encoding should be used WITHOUT errors='replace', because
+    replacement characters (U+FFFD) cannot be encoded in the target encoding
+    (e.g., GBK) when the content is sent to the API pipeline.
+    """
+    try:
+        with open(path, "rb") as f:
+            raw = f.read(8192)
+        # Strict UTF-8 decode
+        decoded = raw.decode("utf-8")
+        # Round-trip verification: re-encode must match original bytes exactly
+        # This catches GBK text that happens to be valid UTF-8 (different characters)
+        if decoded.encode("utf-8") == raw:
+            return "utf-8"
+    except (UnicodeDecodeError, OSError):
+        pass
+
+    # Try system encoding (GBK/CP936 on Chinese Windows)
+    sys_enc = _preferred_encoding()
+    if sys_enc.lower() != "utf-8":
+        try:
+            raw.decode(sys_enc)
+            return sys_enc
+        except (UnicodeDecodeError, LookupError):
+            pass
+
+    return "utf-8"  # best-effort fallback
+
+
+def is_binary(path: Path) -> bool:
+    """检测二进制file: 读前 8KB, check空字节。
+
+    失败时返回 True (fail-safe: 二进制文件不读).
+    B23: 统一 grep.py 的 _is_binary 和 read_file.py 的内联检测。
+    """
+    try:
+        with open(path, "rb") as f:
+            chunk = f.read(8192)
+        return b"\0" in chunk
+    except OSError:
+        return True
+
+
+def read_text_file(path: Path, encoding: str | None = None) -> str:
+    """read文本file, 统一exceptionhandle。
+
+    B22: 统一 edit_file.py 和 batch_edit.py 的文件读取逻辑。
+    自动解析相对路径, 检查存在性/类型, 返回文件内容。
+    抛出 OSError 时调用方自行处理。
+    编码默认自动检测: 先尝试 UTF-8, 失败回退系统编码。
+    """
+    if encoding is None:
+        encoding = detect_text_encoding(path)
+        path = Path.cwd() / path
+    if not path.exists():
+        raise FileNotFoundError(f"file not found: {path}")
+    if not path.is_file():
+        raise IsADirectoryError(f"not a file: {path}")
+    return path.read_text(encoding=encoding)
+
+
+def atomic_write(path: Path, content: str, encoding: str | None = None) -> Path:
+    """原子writefile — 使用唯一临时file名, 避免 concurrent write竞态。
+
+    v2 fix: 旧实现 path.with_suffix(suffix + ".zall_tmp") 在两个 agent
+    实例同时编辑同一文件时会产生相同临时文件路径, 导致竞态。
+    新实现使用 uuid 生成唯一文件名, 保证并发安全。
+    编码默认 UTF-8 (现代标准, 与 detect_text_encoding 的优先尝试一致)。
+
+    Returns: 临时文件路径 (已被 os.replace 移走, 不再存在)
+    Raises: OSError on write/replace failure (临时文件已清理)
+    """
+    import os
+    import uuid
+    if encoding is None:
+        encoding = "utf-8"
+    tmp_name = f".zall_tmp_{uuid.uuid4().hex[:8]}"
+    tmp = path.parent / tmp_name
+    try:
+        tmp.write_text(content, encoding=encoding)
+        os.replace(str(tmp), str(path))
+    except OSError:
+        try:
+            if tmp.exists():
+                tmp.unlink()
+        except Exception:
+            pass
+        raise
+    return tmp

zall/_util/file_state.py

@@ -0,0 +1,123 @@
+"""File state cache for tracking read file contents and modification state.
+
+缓存每个文件的最后读取时间和 mtime, 支持:
+  - is_file_unchanged(path) — 文件是否在上次读取后未被修改
+  - get_cached_mtime(path) — 获取缓存的 mtime
+  - mark_file_read(path) — 标记文件已读取
+  - invalidate(path) — 使缓存失效
+
+用途: 避免 agent 重复读取未修改的文件, 节省 API token。
+当 agent 调用 read_file 后, 文件状态被缓存; 下次读取前先检查 mtime,
+如果未变化则提示 agent "file unchanged, no need to re-read"。
+
+IPR constraints:
+  IPR-3: stdlib only, no model SDK
+"""
+
+from __future__ import annotations
+
+import threading
+import time
+from pathlib import Path
+from typing import Any
+
+
+class FileStateCache:
+    """filestatecache — 跟踪file mtime + read时间。
+
+    线程安全: 使用 RLock 保护内部 dict。
+    缓存上限: 1000 条 (LRU 淘汰), 防长会话无界增长。
+    """
+
+    __test__ = False
+
+    _MAX_ENTRIES = 1000
+
+    def __init__(self) -> None:
+        self._lock = threading.RLock()
+        # key: str(path), value: {"mtime": float, "read_at": float, "size": int}
+        self._cache: dict[str, dict[str, Any]] = {}
+
+    def mark_file_read(self, path: Path | str) -> None:
+        """标记file已read, cache其 mtime 和 size。"""
+        key = str(path)
+        try:
+            stat = Path(path).stat()
+            mtime = stat.st_mtime
+            size = stat.st_size
+        except OSError:
+            return
+        with self._lock:
+            # LRU 淘汰
+            if len(self._cache) >= self._MAX_ENTRIES:
+                # 淘汰最旧的条目
+                oldest_key = min(self._cache, key=lambda k: self._cache[k].get("read_at", 0))
+                self._cache.pop(oldest_key, None)
+            self._cache[key] = {"mtime": mtime, "read_at": time.time(), "size": size}
+
+    def is_file_unchanged(self, path: Path | str) -> bool:
+        """checkfile是否在上次read后未被修改。
+
+        Returns:
+            True — 文件未修改 (可以跳过重复读取)
+            False — 文件已修改或未被缓存过 (需要重新读取)
+        """
+        key = str(path)
+        with self._lock:
+            cached = self._cache.get(key)
+        if cached is None:
+            return False
+        try:
+            current_mtime = Path(path).stat().st_mtime
+        except OSError:
+            return False
+        return bool(cached["mtime"] == current_mtime)
+
+    def get_cached_mtime(self, path: Path | str) -> float | None:
+        """获取cache的 mtime (不读盘)。"""
+        key = str(path)
+        with self._lock:
+            entry = self._cache.get(key)
+            return entry["mtime"] if entry else None
+
+    def invalidate(self, path: Path | str | None = None) -> None:
+        """使cache失效。path=None 时清除全部cache。"""
+        with self._lock:
+            if path is None:
+                self._cache.clear()
+            else:
+                self._cache.pop(str(path), None)
+
+    def get_changed_files(self) -> list[str]:
+        """return自上次read以来被修改的filelist (供 /doctor 用)。"""
+        changed = []
+        with self._lock:
+            for key, entry in self._cache.items():
+                try:
+                    current_mtime = Path(key).stat().st_mtime
+                    if current_mtime != entry["mtime"]:
+                        changed.append(key)
+                except OSError:
+                    changed.append(key)  # 文件被删除也算变更
+        return changed
+
+    @property
+    def size(self) -> int:
+        """当前cache条目数。"""
+        with self._lock:
+            return len(self._cache)
+
+
+# 全局singleton (process级)
+_global_cache: FileStateCache | None = None
+_global_lock = threading.Lock()
+
+
+def get_file_state_cache() -> FileStateCache:
+    """获取全局 FileStateCache singleton。"""
+    global _global_cache
+    if _global_cache is None:
+        with _global_lock:
+            if _global_cache is None:
+                _global_cache = FileStateCache()
+    return _global_cache

zall/_util/logging.py

@@ -0,0 +1,130 @@
+"""zall._util.logging — Unified logging configuration (IPR-0: self-falsifiable errors).
+
+Inspired by Grok Build's xai-tracing crate: structured, contextual logging that
+makes non-fatal errors observable without changing control flow.
+
+Design:
+  - Single entry point: get_zall_logger(__name__) → standard library logger
+  - ZALL_LOG_LEVEL env var controls verbosity (default: WARNING)
+  - IPR-0 invariant: observer/renderer errors are swallowed but LOGGED
+    (silent pass → violates falsifiability; logged pass → observable)
+
+Usage:
+    from zall._util.logging import get_zall_logger
+    logger = get_zall_logger(__name__)
+    ...
+    except Exception as e:
+        logger.warning("operation failed: %s", e)  # observable, non-fatal
+
+Corresponds to:
+  §0     PR-0: self-falsifiability — silent errors are unfalsifiable
+  IPR-0  Self-falsifiability code form — errors must be observable
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import sys
+from typing import TextIO
+
+# ── Defaults ──
+
+_DEFAULT_LEVEL = logging.WARNING
+_ENV_VAR = "ZALL_LOG_LEVEL"
+
+# Level name → int mapping for env var parsing
+_LEVEL_NAMES: dict[str, int] = {
+    "DEBUG": logging.DEBUG,
+    "INFO": logging.INFO,
+    "WARNING": logging.WARNING,
+    "ERROR": logging.ERROR,
+    "CRITICAL": logging.CRITICAL,
+}
+
+
+def _resolve_level() -> int:
+    """Resolve log level from environment or default."""
+    raw = os.environ.get(_ENV_VAR, "").strip().upper()
+    if raw in _LEVEL_NAMES:
+        return _LEVEL_NAMES[raw]
+    if raw and raw.isdigit():
+        return int(raw)
+    return _DEFAULT_LEVEL
+
+
+# ── Formatter ──
+
+_FORMAT = "%(asctime)s [%(levelname)s] %(name)s: %(message)s"
+_FORMAT_DATE = "%H:%M:%S"
+
+
+def _make_handler(stream: TextIO = sys.stderr) -> logging.Handler:
+    """Create a stderr handler with the standard format."""
+    handler = logging.StreamHandler(stream)
+    handler.setFormatter(logging.Formatter(_FORMAT, _FORMAT_DATE))
+    return handler
+
+
+# ── Global state (lazy-initialized) ──
+
+_initialized = False
+
+
+def setup_logging(
+    *,
+    level: int | None = None,
+    stream: TextIO | None = None,
+    force: bool = False,
+) -> None:
+    """Configure the root zall logger.
+
+    Called once at process start (from main/REPL). Safe to call multiple
+    times — subsequent calls are no-ops unless *force=True*.
+
+    Args:
+        level: Log level (default: from ZALL_LOG_LEVEL env, or WARNING).
+        stream: Output stream (default: sys.stderr).
+        force: Reconfigure even if already initialized.
+    """
+    global _initialized
+    if _initialized and not force:
+        return
+
+    root = logging.getLogger("zall")
+    root.setLevel(level if level is not None else _resolve_level())
+
+    # Remove existing handlers to avoid duplicates on force re-init
+    if force:
+        root.handlers.clear()
+
+    if not root.handlers:
+        root.addHandler(_make_handler(stream or sys.stderr))
+
+    _initialized = True
+
+
+def get_zall_logger(name: str) -> logging.Logger:
+    """Get a zall-namespaced logger.
+
+    Usage:
+        logger = get_zall_logger(__name__)
+        logger.warning("non-fatal issue: %s", detail)
+
+    The logger name is prefixed with 'zall.' to keep it under the
+    zall root logger's level configuration.
+    """
+    # Strip leading 'zall.' if already present to avoid 'zall.zall...'
+    if name.startswith("zall."):
+        logger_name = name
+    elif name.startswith("zall"):
+        logger_name = f"zall.{name}"
+    else:
+        logger_name = f"zall.{name}"
+
+    return logging.getLogger(logger_name)
+
+
+def get_log_level() -> int:
+    """Return the current effective log level for the zall root logger."""
+    return logging.getLogger("zall").level

zall/_util/model_registry.py

@@ -0,0 +1,225 @@
+"""zall._util.model_registry — 统一model元数据 (window大小 + 价格 + 别名 + provider register表).
+
+C1: 消除模型元数据分散在两处的问题:
+  - compactor.py 的 _KNOWN_WINDOWS (窗口大小)
+  - app.py _cmd_cost 的 _PRICES (价格表)
+  - cli/config.py 的 _MODEL_PRESETS (别名/预设)
+
+新增模型时只需更新本文件一处。
+
+重命名自 model_meta.py (v0.2.2): 更准确的命名反映其包含 provider 注册表。"""
+
+from __future__ import annotations
+
+
+# ──────────────────────────────────────────────────────────────────────────
+# modelwindow大小 (token)
+# ──────────────────────────────────────────────────────────────────────────
+
+_KNOWN_WINDOWS: dict[str, int] = {
+    # OpenAI
+    "gpt-4o": 128000,
+    "gpt-4o-mini": 128000,
+    "gpt-4-turbo": 128000,
+    "gpt-4.1": 128000,
+    "gpt-4.1-mini": 128000,
+    "gpt-4.1-nano": 128000,
+    "o1": 200000,
+    "o3-mini": 200000,
+    "o4-mini": 200000,
+    # Anthropic
+    "claude-3-5-sonnet": 200000,
+    "claude-3-5-haiku": 200000,
+    "claude-3-opus": 200000,
+    "claude-3-haiku": 200000,
+    "claude-sonnet-4": 200000,
+    "claude-sonnet-4-20250514": 200000,
+    # Google
+    "gemini-2.5-pro": 1_000_000,
+    "gemini-2.5-flash": 1_000_000,
+    "gemini-2.0-flash": 1_000_000,
+    # DeepSeek
+    "deepseek-chat": 128000,
+    "deepseek-reasoner": 128000,
+    "deepseek-v3": 128000,
+    "deepseek-r1": 128000,
+    # Meta
+    "llama3.1": 128000,
+    "llama3": 8192,
+    "llama-3": 8192,
+    # Qwen (all modern variants 128K)
+    "qwen2.5": 128000,
+    "qwen2.5-coder": 128000,
+    "qwen-plus": 131072,
+    "qwen3": 131072,
+    # 其他
+    "agnes-1.5-flash": 128000,
+    "agnes-2.0-flash": 128000,
+    "agnes-2.5-flash": 128000,
+    "glm-4-flash": 128000,
+    "glm-4": 128000,
+    # llama.cpp (window大小由启动parameter决定, 写常见值)
+    "llama.cpp-local": 8192,
+}
+
+# defaultwindow大小 (未知model)
+_DEFAULT_WINDOW: int = 32000
+
+
+# ──────────────────────────────────────────────────────────────────────────
+# model价格表 ($/1M tokens)
+# ──────────────────────────────────────────────────────────────────────────
+
+_KNOWN_PRICES: dict[str, tuple[float, float]] = {
+    # OpenAI
+    "gpt-4o": (2.50, 10.00),
+    "gpt-4o-mini": (0.15, 0.60),
+    "gpt-4-turbo": (10.00, 30.00),
+    "gpt-4.1": (2.00, 8.00),
+    "gpt-4.1-mini": (0.40, 1.60),
+    "gpt-4.1-nano": (0.10, 0.40),
+    "o1": (15.00, 60.00),
+    "o3-mini": (1.10, 4.40),
+    "o4-mini": (0.15, 0.60),
+    # Anthropic
+    "claude-3-5-sonnet": (3.00, 15.00),
+    "claude-3-5-haiku": (0.80, 4.00),
+    "claude-3-opus": (15.00, 75.00),
+    "claude-3-haiku": (0.25, 1.25),
+    "claude-sonnet-4": (3.00, 15.00),
+    "claude-sonnet-4-20250514": (3.00, 15.00),
+    # Google
+    "gemini-2.5-pro": (1.25, 5.00),
+    "gemini-2.5-flash": (0.15, 0.60),
+    "gemini-2.0-flash": (0.10, 0.40),
+    # DeepSeek
+    "deepseek-chat": (0.14, 0.28),
+    "deepseek-reasoner": (0.55, 2.19),
+    "deepseek-v3": (0.27, 1.10),
+    "deepseek-r1": (0.55, 2.19),
+    # Qwen
+    "qwen2.5": (0.35, 1.20),
+    "qwen2.5-coder": (0.35, 1.20),
+    "qwen-plus": (0.80, 2.00),
+    "qwen3": (0.80, 2.00),
+    # Zhipu GLM
+    "glm-4-flash": (0.10, 0.10),
+    "glm-4": (0.50, 1.50),
+    # 本地model / 其他
+    "agnes-1.5-flash": (0.15, 0.60),
+    "agnes-2.0-flash": (0.15, 0.60),
+    "agnes-2.5-flash": (0.15, 0.60),
+    "llama3.1": (0.0, 0.0),  # local
+    "llama.cpp-local": (0.0, 0.0),  # local (llama.cpp server)
+}
+
+# default价格 (未知model用保守值)
+_DEFAULT_PRICE_IN: float = 3.0   # $3/M input
+_DEFAULT_PRICE_OUT: float = 15.0  # $15/M output
+
+# O8: 模块级预sort (按名称长度降序)，避免每次调 get_window_size/get_price 都 sorted()
+_SORTED_WINDOWS: list[tuple[str, int]] = sorted(
+    _KNOWN_WINDOWS.items(), key=lambda x: -len(x[0])
+)
+_SORTED_PRICES: list[tuple[str, tuple[float, float]]] = sorted(
+    _KNOWN_PRICES.items(), key=lambda x: -len(x[0])
+)
+
+
+def get_window_size(model_name: str) -> int:
+    """查modelwindow大小。已知modelreturn精确值, 未知return保守default值。"""
+    if not model_name:
+        return _DEFAULT_WINDOW
+    # B1 fix: 先精确匹配完整名称
+    if model_name in _KNOWN_WINDOWS:
+        return _KNOWN_WINDOWS[model_name]
+    # 再按前缀匹配, 按名称长度降序 (长前缀优先, 防 gpt-4o-mini 误配 gpt-4o)
+    for known, size in _SORTED_WINDOWS:
+        if model_name.startswith(known):
+            return size
+    return _DEFAULT_WINDOW
+
+
+def get_price(model_name: str) -> tuple[float, float]:
+    """获取model价格 (input_price, output_price) $/1M tokens。
+
+    未知模型返回保守默认值 ($3/$15 per 1M)。
+    B1 fix: 先精确匹配完整名称, 再按长前缀降序匹配。
+    """
+    if not model_name:
+        return (_DEFAULT_PRICE_IN, _DEFAULT_PRICE_OUT)
+    # 先精确匹配
+    if model_name in _KNOWN_PRICES:
+        return _KNOWN_PRICES[model_name]
+    # 再按前缀长度降序匹配
+    for prefix, prices in _SORTED_PRICES:
+        if model_name.startswith(prefix):
+            return prices
+    return (_DEFAULT_PRICE_IN, _DEFAULT_PRICE_OUT)
+
+
+# ──────────────────────────────────────────────────────────────────────────
+# Item B: Provider register表 (统一元数据, 消除 if/elif 链)
+# ──────────────────────────────────────────────────────────────────────────
+
+# Item D: 6-tuple: (display, env_var, default_api_base, get_key_url, model_prefixes, adapter_import_path)
+# adapter_import_path 格式: "module.path:ClassName" — 供 _build_adapter 用 importlib dynamicload
+_ProviderMeta = dict[str, tuple[str, str, str, str, tuple[str, ...], str]]
+
+_PROVIDER_REGISTRY: _ProviderMeta = {
+    "openai":    ("OpenAI-compatible",    "ZALL_API_KEY",       "https://api.openai.com/v1",                     "https://platform.openai.com/api-keys",             ("gpt-", "o1", "o3", "o4", "glm-", "qwen"), "zall.adapters.openai_compat:OpenAICompatAdapter"),
+    "anthropic": ("Anthropic Claude",     "ANTHROPIC_API_KEY",  "https://api.anthropic.com",                     "https://console.anthropic.com/",                   ("claude-", "claude"),                                "zall.adapters.anthropic:AnthropicAdapter"),
+    "gemini":    ("Google Gemini",        "GOOGLE_API_KEY",     "https://generativelanguage.googleapis.com",      "https://aistudio.google.com/app/apikey",           ("gemini-", "gemini"),                                "zall.adapters.gemini:GeminiAdapter"),
+    "deepseek":  ("DeepSeek",             "DEEPSEEK_API_KEY",   "https://api.deepseek.com/v1",                   "https://platform.deepseek.com/api_keys",           ("deepseek-", "deepseek"),                            "zall.adapters.openai_compat:OpenAICompatAdapter"),
+    "ollama":    ("Ollama (local)",       "",                    "http://localhost:11434",                        "https://ollama.ai",                                ("ollama-", "llama"),                                 "zall.adapters.ollama:OllamaAdapter"),
+    "agnes":     ("Agnes AI",             "ZALL_API_KEY",       "https://apihub.agnes-ai.com/v1",                "https://apihub.agnes-ai.com",                      ("agnes-",),                                          "zall.adapters.openai_compat:OpenAICompatAdapter"),
+}
+
+# compatible性: preserve旧 dict reference (代码中仍reference _PROVIDER_DISPLAY 等)
+_provider_display = {k: v[0] for k, v in _PROVIDER_REGISTRY.items()}
+_provider_env_vars = {k: v[1] for k, v in _PROVIDER_REGISTRY.items()}
+_provider_api_bases = {k: v[2] for k, v in _PROVIDER_REGISTRY.items()}
+_provider_key_urls = {k: v[3] for k, v in _PROVIDER_REGISTRY.items()}
+
+# Item D: 已知 adapter type → import path mapping (供 TOML 自定义 provider 用)
+_ADAPTER_TYPE_MAP: dict[str, str] = {
+    "openai-compat": "zall.adapters.openai_compat:OpenAICompatAdapter",
+    "anthropic":     "zall.adapters.anthropic:AnthropicAdapter",
+    "gemini":        "zall.adapters.gemini:GeminiAdapter",
+    "ollama":        "zall.adapters.ollama:OllamaAdapter",
+}
+
+
+def get_model_provider(model_name: str) -> str:
+    """根据model名推断 provider type (Item B: 基于register表)。"""
+    if not model_name:
+        return "openai"
+    model_lower = model_name.lower()
+    for provider, (_display, _env, _base, _url, prefixes, _adapter) in _PROVIDER_REGISTRY.items():
+        for prefix in prefixes:
+            if model_lower.startswith(prefix):
+                return provider
+    return "openai"  # default
+
+
+# ──────────────────────────────────────────────────────────────────────────
+# model预设list (供交互式 picker 使用, 从 config.py 迁移至此)
+# ──────────────────────────────────────────────────────────────────────────
+
+_MODEL_PRESETS: list[tuple[str, str, str, str]] = [
+    ("agnes-2.0-flash", "agnes-2.0-flash", "fast / cheap (default + latest)", "agnes"),
+    ("agnes-2.5-flash", "agnes-2.5-flash", "fast / cheap (newer)", "agnes"),
+    ("agnes-1.5-flash", "agnes-1.5-flash", "fast / cheap (legacy)", "agnes"),
+    ("gpt-4o-mini",     "gpt-4o-mini",     "OpenAI, cheap", "openai"),
+    ("gpt-4o",          "gpt-4o",          "OpenAI, capable", "openai"),
+    ("claude-3-5-sonnet", "claude-3-5-sonnet", "Anthropic", "anthropic"),
+    ("claude-sonnet-4", "claude-sonnet-4-20250514", "Anthropic Claude Sonnet 4", "anthropic"),
+    ("deepseek-chat",   "deepseek-chat",   "DeepSeek", "deepseek"),
+    ("glm-4-flash",     "glm-4-flash",     "Zhipu GLM", "openai"),
+    ("qwen-plus",       "qwen-plus",       "Alibaba Qwen", "openai"),
+    ("gemini-2.5-pro",  "gemini-2.5-pro-exp-03-25", "Google Gemini 2.5 Pro", "gemini"),
+    ("gemini-2.5-flash", "gemini-2.5-flash-001", "Google Gemini 2.5 Flash", "gemini"),
+    ("ollama-llama3",   "llama3.1",        "Ollama local (llama3.1)", "ollama"),
+    ("ollama-qwen25",   "qwen2.5",         "Ollama local (qwen2.5)", "ollama"),
+    ("llama.cpp",       "llama.cpp-local", "llama.cpp local server (OpenAI-compat)", "openai"),
+]