feat(providers): add a Tsubasa descriptor on the compatible transport (#6695)

Tsubasa gets one bundled row in provider_descriptors.json on the existing
OpenAI-compatible Chat Completions path, as @cenab proposed and the founder
approved: base https://api.tsubasa.sh/v1, key env TSUBASA_API_KEY, bootstrap
model tsubasa-pro, docs https://tsubasa.sh/docs. No new ProviderKind, no
transport, no prices, capabilities or underlying-model claims.

Discovery stays `models_endpoint`, like every neighbouring row: Tsubasa's
published API contract (tsubasa.sh/docs) names GET /v1/models as the catalog
authority, and the live 404 is not specific to discovery — on 2026-09-28
POST /v1/chat/completions also returned 404 without a key, and the public
model page lists both models as Disabled.

The descriptor schema has no context or output fields, so the 32,768-token
window and the second public id (tsubasa-fast) reach the user through the
row's guidance and docs/PROVIDERS.md, with the exact
`codewhale config set providers.tsubasa.context_window 32768` command
(an unknown custom model otherwise budgets 128,000 tokens).

Pricing default-coverage ledger records the route as unknown billing basis
and the compatible-default breadth receipt moves from 7 to 8.

Tests (focused):
- cargo test -p codewhale-config --lib -- descriptors route::export:
  9 passed; 0 failed; 1 ignored (new tsubasa_is_a_compatible_row_with_its_own_key;
  golden_providers_export_matches_registry passes, golden unchanged)
- cargo test -p codewhale-tui --lib pricing::default_coverage_tests:
  1 passed; 0 failed
- cargo test -p codewhale-tui --lib provider_picker: 148 passed; 0 failed
- scripts/check-provider-registry.py: passed
- cargo fmt --all -- --check: clean

Closes #6695

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_014ZwqatxgVFxHvovngywnks

crates/config/assets/provider_descriptors.json

@@ -102,6 +102,17 @@
         "meituan-sensenova",
         "meituan-sensenova-cn"
       ]
+    },
+    {
+      "id": "tsubasa",
+      "label": "Tsubasa",
+      "wire": "openai-compatible",
+      "base_url": "https://api.tsubasa.sh/v1",
+      "api_key_env": "TSUBASA_API_KEY",
+      "default_model": "tsubasa-pro",
+      "discovery": "models_endpoint",
+      "docs_url": "https://tsubasa.sh/docs",
+      "guidance": "OpenAI-compatible host (Chat Completions + GET /v1/models only). Public model ids: tsubasa-pro and tsubasa-fast, both with a 32,768-token context, so set context_window = 32768 on [providers.tsubasa]. Store TSUBASA_API_KEY, not a raw key."
     }
   ]
 }

crates/config/src/descriptors.rs

@@ -153,6 +153,26 @@ mod tests {
         }
     }
 
+    /// #6695: Tsubasa is a data row on the existing compatible transport with
+    /// its own key env. The row carries no context field, so the guidance is
+    /// where the 32K window and the second public model id reach the user.
+    #[test]
+    fn tsubasa_is_a_compatible_row_with_its_own_key() {
+        let tsubasa = provider_descriptor("tsubasa").expect("tsubasa");
+        assert_eq!(tsubasa.wire, DescriptorWire::OpenaiCompatible);
+        assert_eq!(tsubasa.base_url, "https://api.tsubasa.sh/v1");
+        assert_eq!(tsubasa.api_key_env, "TSUBASA_API_KEY");
+        assert_eq!(tsubasa.default_model, "tsubasa-pro");
+        let guidance = tsubasa.guidance.as_deref().expect("tsubasa guidance");
+        for needle in [
+            "tsubasa-fast",
+            "context_window = 32768",
+            "Store TSUBASA_API_KEY",
+        ] {
+            assert!(guidance.contains(needle), "{needle}: {guidance}");
+        }
+    }
+
     #[test]
     fn descriptors_do_not_embed_model_rosters() {
         let raw = DESCRIPTORS_JSON;

crates/tui/src/pricing/default_coverage_tests.rs

@@ -123,7 +123,7 @@ fn shipped_default_routes_have_reviewed_pricing_coverage() {
     assert_eq!(built_in_count, 51, "review changed shipped-default breadth");
     assert_eq!(
         defaults.len() - built_in_count,
-        7,
+        8,
         "review compatible-default breadth"
     );
     let recorded_at = Utc.with_ymd_and_hms(2026, 9, 8, 12, 0, 0).unwrap();

crates/tui/src/pricing/default_coverage_unpriced.json

@@ -36,7 +36,7 @@
   },
   {
     "reason": "unknown_billing_basis",
-    "review": "No mode/credential product is supplied. Both MiniMax dialects share a PAYG/plan endpoint that cannot prove product (route_billing.rs:253,402; pricing.rs:534); Ollama Cloud is unclassified possible spend, not local/free (pricing.rs:411). Generic Custom and six non-Baseten compatible defaults are transport descriptors without a reviewed product; production envelope refuses Unknown at cost_status.rs:239. AICraft is the sixth compatible default: dc170add3 restored its row from the closed PR #6171 with endpoint, credential env and bootstrap model only, and no billing product or pricing row was reviewed for it. Its default is claude-4.6-sonnet, which has no bundled/hand row either; that omission is recorded here on the shape of the descriptor, and a real source review of AICraft's billing basis is still owed.",
+    "review": "No mode/credential product is supplied. Both MiniMax dialects share a PAYG/plan endpoint that cannot prove product (route_billing.rs:253,402; pricing.rs:534); Ollama Cloud is unclassified possible spend, not local/free (pricing.rs:411). Generic Custom and seven non-Baseten compatible defaults are transport descriptors without a reviewed product; production envelope refuses Unknown at cost_status.rs:239. AICraft is the sixth compatible default: dc170add3 restored its row from the closed PR #6171 with endpoint, credential env and bootstrap model only, and no billing product or pricing row was reviewed for it. Its default is claude-4.6-sonnet, which has no bundled/hand row either; that omission is recorded here on the shape of the descriptor, and a real source review of AICraft's billing basis is still owed. Tsubasa (#6695) is the seventh: endpoint, credential env and bootstrap model tsubasa-pro only; its public site lists prepaid-credit per-token rates, but no pricing row was reviewed or frozen for it, so it stays Unknown here too.",
     "routes": [
       ["aicraft", "claude-4.6-sonnet", "https://aicraftapi.com/v1", "chat_completions", "unknown"],
       ["cerebras", "llama-3.3-70b", "https://api.cerebras.ai/v1", "chat_completions", "unknown"],
@@ -47,7 +47,8 @@
       ["minimax", "MiniMax-M3", "https://api.minimax.io/v1", "chat_completions", "unknown"],
       ["minimax-anthropic", "MiniMax-M3", "https://api.minimax.io/anthropic", "anthropic_messages", "unknown"],
       ["ollama-cloud", "gpt-oss:120b", "https://ollama.com/v1", "chat_completions", "unknown"],
-      ["sensenova", "deepseek-v4-flash", "https://token.sensenova.cn/v1", "chat_completions", "unknown"]
+      ["sensenova", "deepseek-v4-flash", "https://token.sensenova.cn/v1", "chat_completions", "unknown"],
+      ["tsubasa", "tsubasa-pro", "https://api.tsubasa.sh/v1", "chat_completions", "unknown"]
     ]
   },
   {

docs/PROVIDERS.md

@@ -42,11 +42,16 @@ against the vendor's own docs before trusting any value here:
 | Command Code | `https://api.commandcode.ai/provider/v1` | `deepseek/deepseek-v4-flash` | `COMMAND_CODE_API_KEY` |
 | Alibaba Model Studio (DashScope) | `https://dashscope-intl.aliyuncs.com/compatible-mode/v1` | `qwen3.8-flash` | `DASHSCOPE_API_KEY` |
 | AICraft | `https://aicraftapi.com/v1` | `claude-4.6-sonnet`; DeepSeek / Claude / Gemini / Qwen / GLM / MiniMax / Doubao families | `AICRAFT_API_KEY` |
+| Tsubasa | `https://api.tsubasa.sh/v1` | `tsubasa-pro`, `tsubasa-fast` (32,768-token context) | `TSUBASA_API_KEY` |
 
 AICraft's roster spans DeepSeek, Anthropic Claude, Google Gemini, Qwen, GLM,
 MiniMax and Doubao ids on its OpenAI-compatible endpoint. The authority is
 `GET https://aicraftapi.com/v1/models` with your key — pick a model from that
 list, not from this table.
+Tsubasa implements only `GET /v1/models` and Chat Completions. Its two public
+model ids share a 32,768-token context, smaller than the 128,000 tokens Codewhale assumes
+for an unknown model, so set it on the route after saving:
+`codewhale config set providers.tsubasa.context_window 32768`.
 OpenCode Zen and OpenCode Go are first-class provider routes, configured like
 any other provider below; they are not part of this table. In `/provider`,
 type to filter the list (letters not bound to a row action); `Ctrl+T` probes the

docs/zh_hans/PROVIDERS.md

@@ -6,7 +6,7 @@
 
 DeepSeek 仍是默认提供商，但 `ProviderKind::ALL` 中的每个条目都是一等公民、可选的提供商路由。`ALL` 是目录/选择器表面——每个厂商一个身份。双线协议方言种类（`*Anthropic`，例如 `deepseek-anthropic`）和 Model Studio 套餐变体保留在枚举中用于 serde 和 `provider_for_kind`，但刻意**不**作为目录行：套餐是主提供商配置（`crates/config/src/provider_kind.rs:221-226`）上的 `mode`/`base_url`，方言则是 `wire = openai|anthropic`。托管路由、通用 OpenAI 兼容端点、OpenAI Codex/ChatGPT 路由、原生 Anthropic 以及本地运行时，都在所选提供商/模型/base URL 上运行同一个终端 harness。
 
-经普通 Chat Completions 访问的主机是普通的具名 provider（`[providers.<name>]` 表：base URL、模型、密钥环境变量），而不是 `ProviderKind`；`/provider` 与 `/setup` 保留「粘贴 Base URL 和密钥」路径。英文版中的「已知可用主机」表列出 SenseNova、Baseten、Groq、Cerebras、Command Code、阿里云百炼（DashScope）与 AICraft 的 URL 和密钥变量；这些主机作为内置描述符行随附于 `crates/config/assets/provider_descriptors.json`（仅描述如何连接主机，模型 ID 以实时 `GET /v1/models` 与 Codewhale 目录为准），仅供参考，请以各厂商文档为准。AICraft 的模型列表涵盖 DeepSeek、Anthropic Claude、Google Gemini、Qwen、GLM、MiniMax 与 Doubao（例如 `claude-4.6-sonnet`），以带密钥请求 `GET https://aicraftapi.com/v1/models` 的结果为准。OpenCode Zen 与 OpenCode Go 是下方的一等路由。在 `/provider` 中直接输入即可筛选列表（已绑定行操作的字母除外）；`Ctrl+T` 探测所选行的 `/models`，只记录可达性（2xx 并不代表模型可用）。
+经普通 Chat Completions 访问的主机是普通的具名 provider（`[providers.<name>]` 表：base URL、模型、密钥环境变量），而不是 `ProviderKind`；`/provider` 与 `/setup` 保留「粘贴 Base URL 和密钥」路径。英文版中的「已知可用主机」表列出 SenseNova、Baseten、Groq、Cerebras、Command Code、阿里云百炼（DashScope）、AICraft 与 Tsubasa 的 URL 和密钥变量；这些主机作为内置描述符行随附于 `crates/config/assets/provider_descriptors.json`（仅描述如何连接主机，模型 ID 以实时 `GET /v1/models` 与 Codewhale 目录为准），仅供参考，请以各厂商文档为准。AICraft 的模型列表涵盖 DeepSeek、Anthropic Claude、Google Gemini、Qwen、GLM、MiniMax 与 Doubao（例如 `claude-4.6-sonnet`），以带密钥请求 `GET https://aicraftapi.com/v1/models` 的结果为准。Tsubasa（`https://api.tsubasa.sh/v1`，`TSUBASA_API_KEY`）仅实现 `GET /v1/models` 与 Chat Completions，公开模型 `tsubasa-pro` 与 `tsubasa-fast` 的上下文均为 32,768 tokens；保存后请执行 `codewhale config set providers.tsubasa.context_window 32768`。OpenCode Zen 与 OpenCode Go 是下方的一等路由。在 `/provider` 中直接输入即可筛选列表（已绑定行操作的字母除外）；`Ctrl+T` 探测所选行的 `/models`，只记录可达性（2xx 并不代表模型可用）。
 
 需要保持同步的来源：