From 4a2fbed9028b052f9b2b0a214e852e6d80876795 Mon Sep 17 00:00:00 2001 From: fylorn <249551762+fylorn@users.noreply.github.com> Date: Mon, 5 Oct 2026 01:15:54 +0800 Subject: [PATCH] Claude Desktop takeover: follow the model-name rule of the current Claude Desktop Claude Desktop 2.19675.0 decides which gateway model names it accepts with a denylist of other vendors' name fragments followed by an allowlist: a name is rejected if it contains any fragment such as deepseek, gpt, qwen or codex (even next to "claude"), and otherwise accepted if it is a bare tier name (sonnet, opus-4.8) or contains claude, anthropic or a tier name (sonnet, opus, haiku, fable, mythos). An invalid inferenceModels entry is now dropped on its own with a configuration error instead of rejecting the whole list, and "X[1m]" is accepted as shorthand for X with a 1M-context variant. looks_like_claude now mirrors that rule, with the denylist copied verbatim from the app bundle (JS \b and \d translated to their ASCII meaning). Names Desktop would drop are still left out of inferenceModels, since each one produces a configuration error, and the fallback model with its routing rule stays for a gateway that lists nothing Desktop accepts. [1m] spellings are no longer refused: a gateway listing both X and X[1m] gets one model with a 1M variant in Desktop's picker. Co-Authored-By: Claude Opus 5.5 --- src-tauri/Cargo.lock | 1 + src-tauri/crates/tw-adopt/Cargo.toml | 2 + src-tauri/crates/tw-adopt/src/desktop.rs | 125 ++++++++++++++---- .../crates/tw-adopt/tests/claude_desktop.rs | 24 +++- 4 files changed, 128 insertions(+), 24 deletions(-) diff --git a/src-tauri/Cargo.lock b/src-tauri/Cargo.lock index 12921133..6b53bc71 100644 --- a/src-tauri/Cargo.lock +++ b/src-tauri/Cargo.lock @@ -5254,6 +5254,7 @@ dependencies = [ name = "tw-adopt" version = "0.1.0" dependencies = [ + "regex", "serde", "serde_json", "tempfile", diff --git a/src-tauri/crates/tw-adopt/Cargo.toml b/src-tauri/crates/tw-adopt/Cargo.toml index 8991f927..2bd4ccc6 100644 --- a/src-tauri/crates/tw-adopt/Cargo.toml +++ b/src-tauri/crates/tw-adopt/Cargo.toml @@ -18,6 +18,8 @@ serde_json = "1" # 会把用户的注释和排版一起抹掉 toml_edit = "0.25" thiserror = "2" +# Claude Desktop 收哪些模型名是一条正则，照抄它的原文，见 desktop.rs +regex = "1" [target.'cfg(windows)'.dependencies] # 原子替换一个已存在的文件：`rename` 在这个平台上做不到，见 foreign.rs diff --git a/src-tauri/crates/tw-adopt/src/desktop.rs b/src-tauri/crates/tw-adopt/src/desktop.rs index 4fdf3de4..55293191 100644 --- a/src-tauri/crates/tw-adopt/src/desktop.rs +++ b/src-tauri/crates/tw-adopt/src/desktop.rs @@ -33,6 +33,7 @@ //! 替用户拆掉一道安全边界，和「指向网关」无关。 use std::path::{Path, PathBuf}; +use std::sync::LazyLock; use tw_types::{Msg, msg}; @@ -192,29 +193,56 @@ fn managed_in_registry() -> Option { // ---------------------------------------------------------------- 模型 -/// 这个模型名 Claude Desktop 收不收。 +/// Claude 的几档 +const TIERS: [&str; 5] = ["sonnet", "opus", "haiku", "fable", "mythos"]; + +/// 别家模型的名字片段。**照抄 Claude Desktop 2.19675.0 的原文**，它换了就整段换掉： +/// 解开 `Claude.app/Contents/Resources/app.asar`，在 `.vite/build/index.chunk-*.js` 里搜 +/// `ark-code|astron`，档位表和判断函数就在它旁边。 +const OTHER_VENDORS: &str = r"ark-code|astron|command-r|deepseek|doubao|gemini|gemma|glm|gpt|grok|hermes|hy3|kimi|lfm|\bling\b|llama|longcat|mimo|minimax|mistral|mixtral|moonshot|nemotron|openai|phi-|qianfan|qwen|tc-code|\bunic\b|yi-|stepfun|step-3|seed-|bytedance|hunyuan|granite|amazon\.nova|nova-|devstral|ministral|ernie|codex|arcee|trinity|abab|phi\d|\bk2\.|\bm2\.|jamba|arctic|solar|mercury|zamba|kat-coder|\bds-|dpsk"; + +/// JS 的正则写成 Rust 的：没有 `u` 标志的 JS 正则里，`\b` 和 `\d` 只认 ASCII，Rust 的 +/// 默认认 Unicode（`é` 在 JS 里不算单词字符，`claude-éling` 要被 `\bling\b` 拦下） +fn js_regex(src: &str) -> regex::Regex { + regex::Regex::new(&src.replace(r"\b", r"(?-u:\b)").replace(r"\d", "[0-9]")) + .expect("Claude Desktop's model-name pattern") +} + +static OTHER_VENDOR: LazyLock = LazyLock::new(|| js_regex(OTHER_VENDORS)); + +/// 光一个档位名、可以带版本：`sonnet`、`opus-4.8` +static BARE_TIER: LazyLock = + LazyLock::new(|| js_regex(&format!(r"^({})(-[\d.]+)?$", TIERS.join("|")))); + +/// 这个模型名 Claude Desktop 收不收。照它网关模式的规则，名字先转小写： +/// +/// 1. 含 `OTHER_VENDORS` 里任一片段的**不收** —— 名字里有 Claude 也一样， +/// `claude-deepseek-v3` 不收； +/// 2. 否则光一个档位名（`BARE_TIER`）的收，名字里有 `claude`、`anthropic` 或任一档位名 +/// 的也收。 /// -/// 它只认看起来像 Claude 的名字：自动发现时只留这样的，写进 `inferenceModels` -/// 的名字里只要有一个不像，**整张列表都会被拒掉**（社区报告，anthropics/claude-code -/// #56990；CC Switch 实测过的规则是 `claude-` 之后紧跟 sonnet、opus、haiku、fable -/// 之一，再跟一段版本）。所以这里取严不取宽。带 `[1m]` 的写法它也不认 —— -/// 1M 上下文在它那里是单独的一个字段。 +/// `inferenceModels` 里的名字它先去掉首尾空白，结尾的 `[1m]`（不分大小写）也先去掉， +/// 再按这条判断 —— `claude-sonnet-5[1m]` 是「`claude-sonnet-5`，另给一个 1M 上下文的 +/// 版本」的简写。判断不过的那一条**单独去掉**并报一条配置错误，别的照用；一条都不剩 +/// 就没有模型可选。网关自己列出模型（`GET /v1/models`）时它也用同一条规则筛。 pub fn looks_like_claude(model: &str) -> bool { - let m = model.trim().to_ascii_lowercase(); - if m.contains("[1m]") { + let m = model.trim().to_lowercase(); + let m = match m.strip_suffix("[1m]") { + Some(base) if !base.is_empty() => base.trim(), + _ => m.as_str(), + }; + if OTHER_VENDOR.is_match(m) { return false; } - let tail = m.strip_prefix("anthropic/").unwrap_or(&m); - let Some(tail) = tail.strip_prefix("claude-") else { - return false; - }; - ["sonnet-", "opus-", "haiku-", "fable-"] - .iter() - .any(|role| tail.strip_prefix(role).is_some_and(|rest| !rest.is_empty())) + BARE_TIER.is_match(m) + || ["claude", "anthropic"] + .iter() + .chain(TIERS.iter()) + .any(|w| m.contains(w)) } -/// 网关一个像 Claude 的模型都没列出来时写进去的那一个。路由里要有一条规则把它 -/// 改写到上游真正在用的模型 —— 接管说明里给出那条规则。 +/// 网关列出的模型它一个都不收时写进去的那一个（一条都不剩，它就没有模型可选）。 +/// 路由里要有一条规则把它改写到上游真正在用的模型 —— 接管说明里给出那条规则。 pub const FALLBACK_MODEL: &str = "claude-sonnet-5"; // ---------------------------------------------------------------- 接管 @@ -255,6 +283,8 @@ pub fn plan_adopt_in( let mut notes: Vec = plan::cost_notes(c).collect(); match models { Some(all) => { + // 它不收的名字它自己会去掉，但每一条都报一个配置错误：不写进去。带 `[1m]` 的 + // 照写 —— 网关同时列出 `X` 和 `X[1m]` 时，它在选择器里合成一个带 1M 版本的模型 let claude: Vec = all .iter() .filter(|m| looks_like_claude(m)) @@ -414,7 +444,7 @@ fn deployment_mode() -> Edit { } } -/// 网关没列出像 Claude 的模型：要在路由里加一条规则，把 [`FALLBACK_MODEL`] 改写到 +/// 网关列出的模型它一个都不收：要在路由里加一条规则，把 [`FALLBACK_MODEL`] 改写到 /// 上游真正在用的那个模型。`upstream` 是网关列出的第一个模型，拿来填示例。 fn no_claude_model(upstream: Option<&String>) -> Msg { let target = upstream.map_or("", |s| s.as_str()); @@ -731,6 +761,7 @@ mod tests { assert!(p.starts_with(THIRD_PARTY_DIR.resolve(home))); } + /// 每一个都拿 Claude Desktop 2.19675.0 自己的判断（从 app.asar 里原样取出来跑）核对过 #[test] fn only_names_claude_desktop_accepts_make_the_list() { for ok in [ @@ -738,19 +769,67 @@ mod tests { "claude-opus-4-8", "claude-haiku-4-5-20251001", "claude-fable-5", + "claude-mythos-1", "anthropic/claude-sonnet-5", "Claude-Sonnet-5", + "us.anthropic.claude-sonnet-4-5-20250929-v1:0", + " claude-sonnet-5 ", + // 1M 上下文的写法 + "claude-sonnet-5[1m]", + "claude-opus-4-8[1M]", + // 光一个档位名 + "sonnet", + "Sonnet", + "opus-4.8", + "haiku-4.5", + "fable", + "mythos-1", + // 名字里有 claude、anthropic 或档位名就收，在哪儿都行 + "claude", + "claude-sonnet-", + "my-claude-proxy", + "claude-3-5-sonnet-latest", + "anthropic-router", + "opus-proxy", + // `\bling\b`、`\bunic\b` 要整个词：sibling、unicorn 里的不算 + "claude-sibling", + "claude-unicorn", + // `\d` 只认 ASCII 数字：阿拉伯-印度数字的 3 不算 + "claude-phi\u{663}", ] { assert!(looks_like_claude(ok), "{ok}"); } for no in [ "deepseek-chat", "gpt-5", - "claude-sonnet-", - "claude-sonnet-5[1m]", - "claude", - "my-claude-sonnet-5", - "claude-3-5-sonnet-latest", + "gemini-2.5-pro", + "qwen3-coder-plus", + "kimi-k2", + "glm-4.6", + // 有别家的片段，名字里有 Claude 也不收 + "claude-deepseek-v3", + "DeepSeek-Claude", + "claude-gpt-proxy", + "gpt-5-codex", + "claude-codex", + "anthropic/kimi-k2", + "sonnet-qwen", + "claude-sonnet-5-gpt[1m]", + "claude-ling", + "claude-unic", + // JS 的 `\b` 只认 ASCII：é 不是单词字符，ling 前面就是词的边界 + "claude-\u{e9}ling", + "claude-phi4", + "claude-k2.5", + "claude-m2.1", + "claude-yi-34b", + "claude-seed-2", + "opus-ds-1", + "claude-dpsk", + // 什么都不像 + "[1m]", + "", + " ", ] { assert!(!looks_like_claude(no), "{no}"); } diff --git a/src-tauri/crates/tw-adopt/tests/claude_desktop.rs b/src-tauri/crates/tw-adopt/tests/claude_desktop.rs index c97288a9..590ed3c4 100644 --- a/src-tauri/crates/tw-adopt/tests/claude_desktop.rs +++ b/src-tauri/crates/tw-adopt/tests/claude_desktop.rs @@ -177,6 +177,28 @@ fn restoring_puts_every_file_back_byte_for_byte() { ); } +/// 带 `[1m]` 的照写：网关同时列出 `X` 和 `X[1m]` 时，Claude Desktop 在选择器里把它们 +/// 合成一个带 1M 版本的模型。名字里有别家片段的不写，名字里有 Claude 也一样 —— 它会 +/// 去掉这一条，并报一个配置错误 +#[test] +fn one_m_spellings_are_written_and_other_vendors_are_left_out() { + let b = bed(); + adopt( + &b, + Some(&names(&[ + "claude-sonnet-5", + "claude-sonnet-5[1m]", + "claude-deepseek-v3", + "gpt-5-codex", + "claude-opus-5", + ])), + ); + assert_eq!( + json(&desktop::profile_path(&b.home))["inferenceModels"], + serde_json::json!(["claude-sonnet-5", "claude-sonnet-5[1m]", "claude-opus-5"]) + ); +} + #[test] fn a_deployment_mode_that_was_there_goes_back_to_its_old_value() { let b = bed(); @@ -376,7 +398,7 @@ fn without_a_claude_model_the_fallback_is_written_and_the_rule_is_spelled_out() &c, &b.home, &gw(), - Some(&names(&["deepseek-chat", "gpt-5"])), + Some(&names(&["deepseek-chat", "claude-deepseek-v3", "gpt-5"])), None, &Around::default(), )