codeep 3.4.0 → 3.5.0

package/dist/config/providers.js

@@ -17,11 +17,15 @@ export const PROVIDERS = {
                 supportsNativeTools: true,
             },
         },
+        // The plan accepts exactly these two ("Only the following two models can be
+        // called: GLM-5.3, GLM-5.3-Flash" — docs.z.ai/devpack/faq, 2026-09-23). It
+        // routes GLM-5.2/5.1 to 5.3, so a 5.2 entry labelled 5.3 as an older model,
+        // and Turbo is not on the plan at all: Z.AI warns other models risk
+        // "unexpected charges", which a flat-fee surface would never show in /cost.
+        // Stored ids migrate via RETIRED_MODEL_REPLACEMENTS.
         models: [
             { id: 'glm-5.3', name: 'GLM-5.3', description: 'Latest flagship for project-scale engineering (1M context)' },
             { id: 'glm-5.3-flash', name: 'GLM-5.3 Flash', description: 'Same 1M context at roughly a ninth of the price' },
-            { id: 'glm-5.2', name: 'GLM-5.2', description: 'Previous flagship for project-scale engineering (1M context)' },
-            { id: 'glm-5-turbo', name: 'GLM-5 Turbo', description: 'Fast GLM-5 variant, available to all users' },
         ],
         defaultModel: 'glm-5.3',
         defaultProtocol: 'openai',
@@ -48,14 +52,16 @@ export const PROVIDERS = {
             },
         },
         models: [
-            // GLM-5.3 reached the standalone pay-per-use API on 2026-08-19 and is
-            // priced on docs.z.ai/guides/overview/pricing. The China platform
-            // (`z.ai-cn*`) bills separately and is not covered by that page, so it is
-            // left alone until its own listing is checked.
+            // GLM-5.3 reached the standalone pay-per-use API on 2026-08-18 and is
+            // priced on docs.z.ai/guides/overview/pricing, as is FlashX (not on
+            // either Coding Plan). GLM-5-Turbo is no longer on that price list, the
+            // models overview or the OpenAPI model enum — it was last listed on
+            // 2026-07-31, with no deprecation notice — so it is not offered here;
+            // stored configs move to Flash. China (`z.ai-cn-api`) still sells it.
             { id: 'glm-5.3', name: 'GLM-5.3', description: 'Latest flagship for project-scale engineering (1M context)' },
             { id: 'glm-5.3-flash', name: 'GLM-5.3 Flash', description: 'Same 1M context at roughly a ninth of the price' },
+            { id: 'glm-5.3-flashx', name: 'GLM-5.3 FlashX', description: 'Faster GLM-5.3 Flash (about 200 tokens/s), 1M context' },
             { id: 'glm-5.2', name: 'GLM-5.2', description: 'Previous flagship for project-scale engineering (1M context)' },
-            { id: 'glm-5-turbo', name: 'GLM-5 Turbo', description: 'Fast GLM-5 variant' },
         ],
         defaultModel: 'glm-5.3',
         defaultProtocol: 'openai',
@@ -80,12 +86,13 @@ export const PROVIDERS = {
                 supportsNativeTools: true,
             },
         },
-        // GLM-5.3 is on the China gateway and the China Coding Plan (bigmodel.cn).
-        // Flash is listed only on China pay-per-use, so the plan does not offer it.
+        // The China plan now carries the same two models on every tier ("所有套餐均
+        // 支持 GLM-5.3、GLM-5.3-Flash" — docs.bigmodel.cn/cn/coding-plan/overview) and
+        // switches GLM-5.2/5.1 to 5.3 and GLM-5-Turbo to 5.3-Flash. FlashX is not on
+        // the plan. Stored ids migrate via RETIRED_MODEL_REPLACEMENTS.
         models: [
             { id: 'glm-5.3', name: 'GLM-5.3', description: 'Latest flagship for project-scale engineering (1M context)' },
-            { id: 'glm-5.2', name: 'GLM-5.2', description: 'Previous flagship for project-scale engineering (1M context)' },
-            { id: 'glm-5-turbo', name: 'GLM-5 Turbo', description: 'Fast GLM-5 variant, available to all users' },
+            { id: 'glm-5.3-flash', name: 'GLM-5.3 Flash', description: 'Same 1M context, much lighter on plan quota' },
         ],
         defaultModel: 'glm-5.3',
         defaultProtocol: 'openai',
@@ -114,7 +121,10 @@ export const PROVIDERS = {
         models: [
             { id: 'glm-5.3', name: 'GLM-5.3', description: 'Latest flagship for project-scale engineering (1M context)' },
             { id: 'glm-5.3-flash', name: 'GLM-5.3 Flash', description: 'Same 1M context at a tenth of the price' },
+            { id: 'glm-5.3-flashx', name: 'GLM-5.3 FlashX', description: 'Faster GLM-5.3 Flash (about 200 tokens/s), 1M context' },
             { id: 'glm-5.2', name: 'GLM-5.2', description: 'Previous flagship for project-scale engineering (1M context)' },
+            // Still listed and priced on BigModel pay-per-use (CNY 5/22 below 32K
+            // input); only the China Coding Plan reroutes it.
             { id: 'glm-5-turbo', name: 'GLM-5 Turbo', description: 'Fast GLM-5 variant' },
         ],
         defaultModel: 'glm-5.3',
@@ -212,17 +222,23 @@ export const PROVIDERS = {
                 supportsNativeTools: true,
             },
         },
-        // One model. DeepSeek's own pricing page names `deepseek-flash` (V4.1 Flash)
-        // as the model to use and says it outperforms V4 Pro. `deepseek-v4-flash`
-        // is already served by it, and from 2026-09-14 12:00 Beijing time every
-        // `deepseek-v4-pro` request is routed to it and billed at its price until
-        // a V4.1 Pro exists — so a Pro entry would be a label on a different model.
-        // Both old ids migrate via RETIRED_MODEL_REPLACEMENTS.
+        // DeepSeek's pricing page names `deepseek-flash` (V4.1 Flash) as the model
+        // to use, so it stays the default; the retired `deepseek-v4-flash` ids are
+        // served by it and migrate via RETIRED_MODEL_REPLACEMENTS.
+        //
+        // V4 Pro was announced to route to Flash from 2026-09-14, and Codeep 3.3.0
+        // migrated it away. DeepSeek reversed that on 2026-09-11, before the
+        // cutover: the changelog's 09-10 entry now says V4 Pro service continues
+        // "with the billing method remaining unchanged", and the pricing page and
+        // API reference list `deepseek-v4-pro` (DeepSeek-V4-Pro-0813) beside Flash.
+        // The 09-10 news post still carries the old routing text; the changelog is
+        // the newer source. So Pro is a separate, separately billed model again.
         //
-        // Note the id differs on OpenRouter, which lists the same model as
-        // `deepseek/deepseek-v4.1-flash`.
+        // The ids differ on OpenRouter: `deepseek/deepseek-v4.1-flash`, and
+        // `deepseek/deepseek-v4-pro-0813` for this Pro (see the fallback list).
         models: [
             { id: 'deepseek-flash', name: 'DeepSeek V4.1 Flash', description: 'Current DeepSeek model — thinking on by default, vision, 1M context, 384K output' },
+            { id: 'deepseek-v4-pro', name: 'DeepSeek V4 Pro', description: 'Larger V4 model (V4-Pro-0813) — thinking on, 1M context, no vision, about 4x Flash\'s price' },
         ],
         defaultModel: 'deepseek-flash',
         defaultProtocol: 'openai',
@@ -243,11 +259,18 @@ export const PROVIDERS = {
         protocols: {
             openai: { baseUrl: 'https://api.kimi.com/coding/v1', authHeader: 'Bearer', supportsNativeTools: true },
         },
+        // Kimi Code upgraded `kimi-for-coding` in place to K2.8 Preview on
+        // 2026-09-11 (1,048,576 context on every tier, effort low/high/max). The
+        // plans were then renamed: Plus ≈ Moderato, Pro ≈ Allegretto, and the new Go
+        // tier has no coding quota. K3 reaches 1M context only on Pro/Allegretto; on
+        // Plus/Moderato it is capped at 256K, and going past that returns HTTP 401
+        // "Your current plan supports only kimi-k3 up to 256K context"
+        // (kimi.com/code/docs/en/kimi-code/models.html, error-reference.html).
         models: [
-            { id: 'kimi-for-coding', name: 'Kimi Code', description: 'Available on every Kimi Code plan — maps to K2.7 Code' },
-            { id: 'k3', name: 'Kimi K3', description: '1M-context flagship — Moderato plan or higher' },
-            { id: 'k3-256k', name: 'Kimi K3 (256K)', description: 'K3 with a smaller context window — Moderato plan or higher' },
-            { id: 'kimi-for-coding-highspeed', name: 'Kimi Code (High-Speed)', description: 'Low-latency K2.7 Code — Allegretto plan or higher' },
+            { id: 'kimi-for-coding', name: 'Kimi Code', description: 'K2.8 Preview, 1M context — every Kimi Code plan (Plus/Andante and above)' },
+            { id: 'k3', name: 'Kimi K3', description: 'Flagship — 1M context on Pro/Allegretto and above; Plus/Moderato is capped at 256K (use K3 256K there)' },
+            { id: 'k3-256k', name: 'Kimi K3 (256K)', description: 'K3 with a 256K context window — Plus/Moderato plan or higher' },
+            { id: 'kimi-for-coding-highspeed', name: 'Kimi Code (High-Speed)', description: 'K2.7 Code HighSpeed, 256K context — Pro/Allegretto plan or higher' },
         ],
         defaultModel: 'kimi-for-coding',
         defaultProtocol: 'openai',
@@ -309,15 +332,17 @@ export const PROVIDERS = {
             openai: { baseUrl: 'https://api.x.ai/v1', authHeader: 'Bearer', supportsNativeTools: true },
         },
         models: [
-            { id: 'grok-4.6', name: 'Grok 4.6', description: 'Flagship reasoning model — xAI recommends it for code, 500K context' },
-            { id: 'grok-4.5', name: 'Grok 4.5', description: 'Previous flagship reasoning model, 500K context' },
+            { id: 'grok-4.7', name: 'Grok 4.7', description: 'Flagship reasoning model — xAI recommends it for code, 500K context' },
+            { id: 'grok-4.6', name: 'Grok 4.6', description: 'Previous flagship reasoning model, 500K context' },
+            { id: 'grok-4.5', name: 'Grok 4.5', description: 'Older flagship reasoning model, 500K context' },
             { id: 'grok-build-0.1', name: 'Grok Build 0.1', description: 'Agentic coding model — fast, 256K context' },
             { id: 'grok-4.3', name: 'Grok 4.3', description: 'Older flagship, 1M context' },
         ],
-        // Stays on the agentic coder, not the new flagship. grok-4.6 is the better
-        // model and xAI recommends it for code, but it bills 2x input and 3x output
-        // against grok-build-0.1 — moving every unpinned user onto it silently is
-        // not ours to decide. It is one `/model` away for anyone who wants it.
+        // Stays on the agentic coder, not the new flagship. grok-4.7 is the better
+        // model — xAI's models page recommends only it, "including code", and it is
+        // the default of xAI's own Grok Build agent — but it bills 2x input and 3x
+        // output against grok-build-0.1, so moving every unpinned user onto it
+        // silently is not ours to decide. It is one `/model` away.
         defaultModel: 'grok-build-0.1',
         defaultProtocol: 'openai',
         useMaxCompletionTokens: true, // reasoning models reject max_tokens
@@ -358,8 +383,7 @@ export const PROVIDERS = {
             openai: { baseUrl: 'https://dashscope-intl.aliyuncs.com/compatible-mode/v1', authHeader: 'Bearer', supportsNativeTools: true },
         },
         // 3.8 Max and Flash are GA on Model Studio international and cost less
-        // than the 3.7 Max / 3.6 Flash they sit above. China (`qwen-cn-api`) is a
-        // separate listing and is not changed on this evidence.
+        // than the 3.7 Max / 3.6 Flash they sit above.
         models: [
             { id: 'qwen3.8-max', name: 'Qwen3.8-Max', description: 'Latest flagship for complex coding and reasoning (1M context)' },
             { id: 'qwen3.8-flash', name: 'Qwen3.8-Flash', description: 'Low-latency, low-cost multimodal model (1M context)' },
@@ -382,14 +406,22 @@ export const PROVIDERS = {
         protocols: {
             openai: { baseUrl: 'https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1', authHeader: 'Bearer', supportsNativeTools: true },
         },
+        // qwen3.8-max-preview is retired here: Alibaba routes it to qwen3.8-max,
+        // bills it at that rate and says to update the id (token-plan-personal-
+        // overview, 2026-09-23); stored configs migrate. The plan now has Personal
+        // and Team editions on this same URL and key format, with different
+        // allowlists — qwen3.6-plus is Team-only, and a Personal key gets
+        // "403 AccessDenied.Unpurchased" for it. Nothing here can tell the two
+        // editions apart, so the entry says so rather than disappearing on Team.
         models: [
-            { id: 'qwen3.8-max-preview', name: 'Qwen3.8-Max Preview', description: 'Newest Token Plan flagship for complex agentic work' },
-            { id: 'qwen3.7-max', name: 'Qwen3.7-Max', description: 'Production flagship for complex coding and reasoning' },
+            { id: 'qwen3.8-max', name: 'Qwen3.8-Max', description: 'Token Plan flagship for complex agentic work (1M context)' },
+            { id: 'qwen3.8-flash', name: 'Qwen3.8-Flash', description: 'Low-latency, credit-efficient model (1M context)' },
+            { id: 'qwen3.7-max', name: 'Qwen3.7-Max', description: 'Previous flagship for complex coding and reasoning' },
             { id: 'qwen3.7-plus', name: 'Qwen3.7-Plus', description: 'Balanced quality and throughput' },
-            { id: 'qwen3.6-plus', name: 'Qwen3.6-Plus', description: 'Fast multimodal model with a 1M context window' },
-            { id: 'qwen3.6-flash', name: 'Qwen3.6-Flash', description: 'Low-latency, credit-efficient model' },
+            { id: 'qwen3.6-plus', name: 'Qwen3.6-Plus', description: 'Team edition only — a Personal plan returns 403 for it' },
+            { id: 'qwen3.6-flash', name: 'Qwen3.6-Flash', description: 'Previous low-latency, credit-efficient model' },
         ],
-        defaultModel: 'qwen3.8-max-preview',
+        defaultModel: 'qwen3.8-max',
         defaultProtocol: 'openai',
         maxOutputTokens: 131_072,
         noStreamWithTools: true,
@@ -426,10 +458,18 @@ export const PROVIDERS = {
         protocols: {
             openai: { baseUrl: 'https://dashscope.aliyuncs.com/compatible-mode/v1', authHeader: 'Bearer', supportsNativeTools: true },
         },
+        // 3.8 Max and Flash have been GA in China (Beijing) since 2026-08-02 and
+        // 2026-08-26, at CNY 12/36 and 0.8/2.7 (help.aliyun.com/zh/model-studio/
+        // qwen3-8-max, qwen3-8-flash). Pricing is keyed by id, so they reuse the
+        // international USD rows — an over-estimate, the same policy as z.ai-cn.
+        // The default is unchanged: 3.8 Max costs the same as 3.7 Max in Beijing,
+        // and moving unpinned users was not part of adding it.
         models: [
-            { id: 'qwen3.7-max', name: 'Qwen3.7-Max', description: 'Latest flagship for complex coding and reasoning' },
+            { id: 'qwen3.8-max', name: 'Qwen3.8-Max', description: 'Latest flagship for complex coding and reasoning (1M context)' },
+            { id: 'qwen3.8-flash', name: 'Qwen3.8-Flash', description: 'Low-latency, low-cost multimodal model (1M context)' },
+            { id: 'qwen3.7-max', name: 'Qwen3.7-Max', description: 'Previous flagship for complex coding and reasoning' },
             { id: 'qwen3.7-plus', name: 'Qwen3.7-Plus', description: 'Balanced quality, speed, and price (1M context)' },
-            { id: 'qwen3.6-flash', name: 'Qwen3.6-Flash', description: 'Low-latency, low-cost multimodal model' },
+            { id: 'qwen3.6-flash', name: 'Qwen3.6-Flash', description: 'Previous low-latency, low-cost multimodal model' },
         ],
         defaultModel: 'qwen3.7-max',
         defaultProtocol: 'openai',
@@ -446,10 +486,16 @@ export const PROVIDERS = {
         protocols: {
             openai: { baseUrl: 'https://api-inference.modelscope.cn/v1', authHeader: 'Bearer', supportsNativeTools: true },
         },
+        // Shown only until the live catalogue loads. The previous fallback,
+        // Qwen/Qwen3-Coder-480B-A35B-Instruct, is no longer served: it is absent from
+        // api-inference.modelscope.cn/v1/models (2026-09-23) and its hub entry has
+        // SupportApiInference false. ModelScope names no successor, so this pick is
+        // Codeep's: the largest Qwen in the live list, whose model card documents
+        // tool calling ("Qwen3.5 excels in tool calling capabilities").
         models: [
-            { id: 'Qwen/Qwen3-Coder-480B-A35B-Instruct', name: 'Qwen3-Coder 480B', description: 'Fallback model shown until the live catalog loads' },
+            { id: 'Qwen/Qwen3.5-397B-A17B', name: 'Qwen3.5 397B', description: 'Fallback model shown until the live catalog loads' },
         ],
-        defaultModel: 'Qwen/Qwen3-Coder-480B-A35B-Instruct',
+        defaultModel: 'Qwen/Qwen3.5-397B-A17B',
         defaultProtocol: 'openai',
         maxOutputTokens: 65_536,
         noStreamWithTools: true,
@@ -462,7 +508,7 @@ export const PROVIDERS = {
     },
     'openai': {
         name: 'OpenAI',
-        description: 'GPT and o-series models',
+        description: 'GPT models',
         protocols: {
             openai: {
                 baseUrl: 'https://api.openai.com/v1',
@@ -470,9 +516,21 @@ export const PROVIDERS = {
                 supportsNativeTools: true,
             },
         },
+        // Chat Completions is the only OpenAI endpoint Codeep calls, and GPT-6 has a
+        // documented tool restriction there (developers.openai.com guides/
+        // latest-model and reasoning): Sol and Luna call tools only with
+        // reasoning_effort "none", and "Chat Completions does not support function
+        // calling with GPT-6 Astra" at all. So Astra is not offered until a
+        // Responses API transport exists — stored configs move to Sol — and Sol and
+        // Luna run agent turns with reasoning off (see reasoningParamsFor).
+        //
+        // The default stays 5.6 Sol: it is the newest model OpenAI documents with
+        // reasoning AND tool calls on Chat Completions ("The Chat Completions
+        // examples use GPT-5.6 for compatibility" — function-calling guide).
         models: [
-            { id: 'gpt-6-astra', name: 'GPT-6 Astra', description: 'Frontier GPT — 1M context, 128K output. Rolling out by organization, so a key may not have it yet. 2x the price of 5.6 Sol' },
-            { id: 'gpt-5.6-sol', name: 'GPT-5.6 Sol', description: 'Most capable GPT — best for coding & agentic work' },
+            { id: 'gpt-5.6-sol', name: 'GPT-5.6 Sol', description: 'Most capable GPT for coding & agentic work on Codeep\'s transport' },
+            { id: 'gpt-6-sol', name: 'GPT-6 Sol', description: 'GPT-6 at $2/$10, 1M context — agent turns run with reasoning off (tools need effort "none" here); /thinking applies to plain chat' },
+            { id: 'gpt-6-luna', name: 'GPT-6 Luna', description: 'Cheapest GPT-6 ($0.10/$0.50), 1M context — agent turns run with reasoning off, as for Sol' },
             { id: 'gpt-5.6-terra', name: 'GPT-5.6 Terra', description: 'Balanced — GPT-5.5 quality at about half the price' },
             { id: 'gpt-5.6-luna', name: 'GPT-5.6 Luna', description: 'Fast and cheap — high-volume workloads' },
         ],
@@ -497,12 +555,20 @@ export const PROVIDERS = {
         },
         models: [
             { id: 'claude-fable-5-1', name: 'Claude Fable 5.1', description: 'Most capable — hardest reasoning & long-horizon agentic work' },
+            { id: 'claude-opus-5-5', name: 'Claude Opus 5.5', description: 'Complex agentic coding & deep reasoning — $4/$20, 1M context' },
             { id: 'claude-fable-5', name: 'Claude Fable 5', description: 'Superseded by 5.1 — same price, kept for pinned configs' },
-            { id: 'claude-opus-5', name: 'Claude Opus 5', description: 'Complex agentic coding & deep reasoning — the Opus workhorse' },
+            { id: 'claude-opus-5', name: 'Claude Opus 5', description: 'Legacy since Opus 5.5 — kept for pinned configs' },
             { id: 'claude-sonnet-5', name: 'Claude Sonnet 5', description: 'Best balance of speed and intelligence' },
             { id: 'claude-haiku-4-5-20251001', name: 'Claude Haiku', description: 'Fastest and most affordable' },
         ],
-        defaultModel: 'claude-opus-5',
+        // Anthropic's models overview now says to "start with Claude Opus 5.5 for
+        // most workloads", and it is 20% cheaper than Opus 5 (now "Active
+        // (legacy)", retiring no sooner than 2027-07-24 — so no migration). One
+        // behaviour change comes with it: Opus 5.5 defaults to MEDIUM effort where
+        // Opus 5 defaulted to high, and /thinking auto sends no effort, so an
+        // unpinned user runs one level lower than before. It also thinks more per
+        // turn at a given effort — see minResponseTokensFor.
+        defaultModel: 'claude-opus-5-5',
         defaultProtocol: 'anthropic',
         envKey: 'ANTHROPIC_API_KEY',
         groupLabel: 'Anthropic',
@@ -543,28 +609,38 @@ export const PROVIDERS = {
                 supportsNativeTools: true,
             },
         },
-        // Top 12 — the full catalog (100+) is fetched lazily via
+        // A short fallback — the full catalog (100+) is fetched lazily via
         // fetchOpenRouterModels() because dynamicModels is true.
         // We keep these hardcoded so first-time users without network
         // get a working dropdown.
         models: [
             { id: 'openrouter/auto', name: 'Auto-route', description: 'OpenRouter picks the best model for the task' },
-            { id: 'anthropic/claude-fable-5-1', name: 'Claude Fable 5.1', description: 'Anthropic — most capable' },
+            // Dotted, as OpenRouter's /api/v1/models lists them. The hyphenated
+            // `anthropic/claude-fable-5-1` this used to carry resolves on OpenRouter's
+            // metadata lookup but is absent from the models list, so a pick made
+            // before the live catalogue loaded matched nothing in it.
+            { id: 'anthropic/claude-fable-5.1', name: 'Claude Fable 5.1', description: 'Anthropic — most capable' },
+            { id: 'anthropic/claude-opus-5.5', name: 'Claude Opus 5.5', description: 'Anthropic — current Opus' },
             { id: 'anthropic/claude-fable-5', name: 'Claude Fable 5', description: 'Anthropic — superseded by 5.1' },
-            { id: 'anthropic/claude-opus-5', name: 'Claude Opus 5', description: 'Anthropic — flagship Opus tier' },
+            { id: 'anthropic/claude-opus-5', name: 'Claude Opus 5', description: 'Anthropic — legacy Opus' },
             { id: 'anthropic/claude-sonnet-5', name: 'Claude Sonnet 5', description: 'Anthropic — balanced' },
             { id: 'openai/gpt-6-astra', name: 'GPT-6 Astra', description: 'OpenAI — frontier' },
+            { id: 'openai/gpt-6-sol', name: 'GPT-6 Sol', description: 'OpenAI — GPT-6, balanced' },
+            { id: 'openai/gpt-6-luna', name: 'GPT-6 Luna', description: 'OpenAI — GPT-6, fast/cheap' },
             { id: 'openai/gpt-5.6-sol', name: 'GPT-5.6 Sol', description: 'OpenAI — flagship' },
             { id: 'openai/gpt-5.6-luna', name: 'GPT-5.6 Luna', description: 'OpenAI — fast/efficient' },
             { id: 'google/gemini-3.8-flash', name: 'Gemini 3.8 Flash', description: 'Google — latest production Flash' },
             { id: 'google/gemini-3.7-flash', name: 'Gemini 3.7 Flash', description: 'Google — previous production Flash' },
             { id: 'deepseek/deepseek-v4.1-flash', name: 'DeepSeek V4.1 Flash', description: 'DeepSeek — current model' },
+            // The GA Pro, with a DeepSeek-hosted endpoint. `deepseek/deepseek-v4-pro`
+            // is the April 0423 preview, served only by third parties.
+            { id: 'deepseek/deepseek-v4-pro-0813', name: 'DeepSeek V4 Pro', description: 'DeepSeek — larger V4 model' },
             { id: 'moonshotai/kimi-k3', name: 'Kimi K3', description: 'Moonshot — long-horizon coding' },
             // Dated on purpose: OpenRouter lists only the snapshot. The undated
             // `qwen/qwen3.8-max` this used to carry does not exist there, so picking
             // it before the live catalogue loaded was an error.
             { id: 'qwen/qwen3.8-max-0902', name: 'Qwen 3.8 Max', description: 'Alibaba — latest flagship' },
-            { id: 'x-ai/grok-4.6', name: 'Grok 4.6', description: 'xAI — flagship reasoning' },
+            { id: 'x-ai/grok-4.7', name: 'Grok 4.7', description: 'xAI — flagship reasoning' },
         ],
         defaultModel: 'openrouter/auto',
         defaultProtocol: 'openai',
@@ -620,22 +696,36 @@ export function getProvider(id) {
  * Exact migrations for curated model ids that vendors replaced.
  *
  * Keep this deliberately narrower than the provider catalogue. Dynamic
- * OpenRouter/Ollama/custom ids are user-controlled and must never be rewritten.
+ * OpenRouter/Ollama/custom ids are user-controlled and must never be rewritten;
+ * ModelScope's one entry is the exception, explained where it sits.
  */
 const RETIRED_MODEL_REPLACEMENTS = {
-    'z.ai': { 'glm-5.1': 'glm-5.2', 'glm-5': 'glm-5.2' },
-    'z.ai-api': { 'glm-5.1': 'glm-5.2', 'glm-5': 'glm-5.2' },
-    'z.ai-cn': { 'glm-5.1': 'glm-5.2', 'glm-5': 'glm-5.2' },
+    // Both GLM Coding Plans accept exactly GLM-5.3 and GLM-5.3-Flash. Z.AI routes
+    // 5.2 and 5.1 to 5.3 on both, and the China plan routes Turbo to 5.3-Flash.
+    // Plain glm-5, and Turbo on the international plan, are Codeep's extension of
+    // the same routing: neither plan's text names them.
+    'z.ai': { 'glm-5.2': 'glm-5.3', 'glm-5.1': 'glm-5.3', 'glm-5': 'glm-5.3', 'glm-5-turbo': 'glm-5.3-flash' },
+    // Pay-per-use still sells 5.2 as its own id. Turbo left the international
+    // price list and model enum with no notice and no named successor, so it
+    // follows the China plan's routing to Flash.
+    'z.ai-api': { 'glm-5.1': 'glm-5.2', 'glm-5': 'glm-5.2', 'glm-5-turbo': 'glm-5.3-flash' },
+    'z.ai-cn': { 'glm-5.2': 'glm-5.3', 'glm-5.1': 'glm-5.3', 'glm-5': 'glm-5.3', 'glm-5-turbo': 'glm-5.3-flash' },
     'z.ai-cn-api': { 'glm-5.1': 'glm-5.2', 'glm-5': 'glm-5.2' },
     google: {
         'gemini-3.1-flash-lite': 'gemini-3.5-flash-lite',
+        // Google's deprecations table names 3.6 Flash (no shutdown date yet). Its
+        // /whats-new-gemini-3.5 guide suggests 3.5 Flash instead; the lifecycle
+        // table is the dedicated source, and 3.5 Flash would cost twice as much.
+        'gemini-3-flash-preview': 'gemini-3.6-flash',
+        // Shut down 2026-03-09; the server already aliases the id to 3.1 Pro.
+        'gemini-3-pro-preview': 'gemini-3.1-pro-preview',
     },
-    // DeepSeek named V4.1 Flash the model to use: the V4 Flash ids are retired
-    // and served by it, and V4 Pro is routed to it from 2026-09-14.
+    // The V4 Flash ids are retired and served by V4.1 Flash. There is no V4 Pro
+    // entry: DeepSeek cancelled its routing to Flash (see the catalogue above),
+    // and rewriting it moved users off a live, separately billed model.
     deepseek: {
         'deepseek-v4-flash': 'deepseek-flash',
         'deepseek-v4-flash-vision-exp': 'deepseek-flash',
-        'deepseek-v4-pro': 'deepseek-flash',
     },
     grok: {
         'grok-code-fast-1': 'grok-build-0.1',
@@ -645,6 +735,13 @@ const RETIRED_MODEL_REPLACEMENTS = {
         'gpt-5.5': 'gpt-5.6-sol',
         'gpt-5.4': 'gpt-5.6-terra',
         'gpt-5.4-mini': 'gpt-5.6-luna',
+        // Not retired by OpenAI. Astra cannot call tools on Chat Completions at all
+        // ("Chat Completions does not support function calling with GPT-6 Astra"),
+        // and agentChat turns a 400 on a tools request into the text-tool fallback,
+        // so every agent turn on it silently ran without native tools. Sol is the
+        // GPT-6 that can call tools here. Drop this entry once a Responses API
+        // transport exists and Astra is offered again.
+        'gpt-6-astra': 'gpt-6-sol',
     },
     'kimi-api': {
         'kimi-k3-code': 'kimi-k3',
@@ -658,30 +755,66 @@ const RETIRED_MODEL_REPLACEMENTS = {
         'kimi-k3-thinking': 'kimi-k3',
         'kimi-k2.5': 'kimi-k2.6',
     },
+    // Alibaba retires qwen3-coder-plus, qwen3-coder-next and bare qwen3-max on
+    // 2026-10-10, on the plans as well (notices 1949 and 1950). It names
+    // qwen3.7-plus for both coders and qwen3.7-max for qwen3-max — but 3.7 Max is
+    // not on the Coding Plan allowlist, so the two Coding Plan surfaces land
+    // qwen3-max on 3.7 Plus, exactly as their own qwen3.7-max entry does.
     qwen: {
         'qwen3-coder-plus': 'qwen3.7-plus',
         'qwen3-coder-next': 'qwen3.7-plus',
         'qwen3.7-max': 'qwen3.7-plus',
+        'qwen3-max': 'qwen3.7-plus',
     },
+    // The coders used to go to qwen3.7-max here, which bills 6.25x the input rate
+    // of the qwen3.7-plus Alibaba names. Configs already moved stay where they
+    // are: 3.7 Max is a valid model.
     'qwen-api': {
-        'qwen3-coder-plus': 'qwen3.7-max',
-        'qwen3-coder-next': 'qwen3.7-max',
+        'qwen3-coder-plus': 'qwen3.7-plus',
+        'qwen3-coder-next': 'qwen3.7-plus',
         'qwen3-coder-flash': 'qwen3.6-flash',
+        'qwen3-max': 'qwen3.7-max',
+    },
+    // The preview is retired and routed to qwen3.8-max, which Alibaba says to use.
+    'qwen-token-plan': {
+        'qwen3.8-max-preview': 'qwen3.8-max',
+        'qwen3-coder-plus': 'qwen3.7-plus',
+        'qwen3-coder-next': 'qwen3.7-plus',
+        'qwen3-max': 'qwen3.7-max',
     },
     'qwen-cn': {
         'qwen3-coder-plus': 'qwen3.7-plus',
         'qwen3-coder-next': 'qwen3.7-plus',
         'qwen3.7-max': 'qwen3.7-plus',
+        'qwen3-max': 'qwen3.7-plus',
     },
     'qwen-cn-api': {
-        'qwen3-coder-plus': 'qwen3.7-max',
-        'qwen3-coder-next': 'qwen3.7-max',
+        'qwen3-coder-plus': 'qwen3.7-plus',
+        'qwen3-coder-next': 'qwen3.7-plus',
         'qwen3-coder-flash': 'qwen3.6-flash',
+        'qwen3-max': 'qwen3.7-max',
+    },
+    // The one entry on a dynamic catalogue, and the exception to the rule above:
+    // the 480B coder was Codeep's own fallback default, so most configs holding
+    // it never chose it, and ModelScope stopped serving it (see the catalogue
+    // entry). Exact id only; every other ModelScope id stays the user's. Drop
+    // this if ModelScope serves the 480B again.
+    modelscope: {
+        'Qwen/Qwen3-Coder-480B-A35B-Instruct': 'Qwen/Qwen3.5-397B-A17B',
     },
 };
+/**
+ * One exact lookup, never followed further: every target must already be a
+ * model its own provider offers (providers.test.ts holds the map to that), so a
+ * chain can never be needed and the macOS mirror stays a flat table too.
+ */
 export function replacementModelFor(providerId, modelId) {
     return RETIRED_MODEL_REPLACEMENTS[providerId]?.[modelId];
 }
+/** The whole map, read-only — for the invariant test and nothing else. */
+export function retiredModelReplacements() {
+    return RETIRED_MODEL_REPLACEMENTS;
+}
 /**
  * Curated display order for the first-run login flow + `/provider` /
  * `/login` pickers. Headline / popular providers float to the top so
@@ -811,11 +944,16 @@ export function providerNoStreamWithTools(providerId) {
  * omission as default. Kimi K2.x code/thinking models fix temperature
  * internally and 400 on any custom value, so they're here too. Google removed
  * the deprecated sampling parameters outright in the Gemini 3.7 generation.
+ * GPT-6 takes temperature/top_p only at reasoning_effort "none" (Astra never):
+ * the direct `openai` provider already omits them for every model through
+ * requiresDefaultTemperature, so this entry is for `openai/gpt-6-*` on
+ * OpenRouter, the one path where a GPT-6 id could still be sent one.
  */
 const SAMPLING_PARAMS_REJECTED = [
     'claude-fable-5', 'claude-opus-5', 'claude-opus-4-8', 'claude-opus-4-7', 'claude-sonnet-5',
     'kimi-k3', 'kimi-k2.7-code', 'kimi-for-coding', 'k3',
     'gemini-3.7-flash', 'gemini-3.8-flash',
+    'gpt-6',
 ];
 export function modelRejectsSamplingParams(model) {
     // Canonicalize both sides. OpenRouter routes these as `google/gemini-3.7-flash`
@@ -837,6 +975,25 @@ export function getEffectiveMaxTokens(providerId, requested) {
         return requested;
     return Math.min(requested, provider.maxOutputTokens);
 }
+/**
+ * The smallest response budget (`max_tokens`) worth sending this model, or 0
+ * for no floor. Callers take the larger of this and whatever they would have
+ * sent, then apply getEffectiveMaxTokens as usual.
+ *
+ * Claude Opus 5.5 thinks on every request — adaptive thinking cannot be turned
+ * off — and "tends to think more per turn than Claude Opus 5" at the same
+ * effort; Anthropic's notes say to "leave room in max_tokens for the thinking",
+ * which spends the same limit as the answer. The task planner's 2048, or a
+ * maxTokens lowered in /settings, could go on thinking alone and cut the reply
+ * off. 32K is Codeep's own default; the Max tier gets 64K, where Anthropic's
+ * advice for max effort on Opus 5 is "starting at 64k tokens". Matched on the
+ * canonical id, so `anthropic/claude-opus-5.5` on OpenRouter is covered too.
+ */
+export function minResponseTokensFor(model, tier) {
+    if (!idMatches(canonicalModelId(model), 'claude-opus-5-5'))
+        return 0;
+    return tier === 'max' ? 65_536 : 32_768;
+}
 export const REASONING_TIERS = ['auto', 'low', 'medium', 'high', 'max'];
 /**
  * Canonicalize a model id for capability matching: lowercase, drop any
@@ -855,6 +1012,64 @@ export function canonicalModelId(model) {
 function idMatches(id, prefix) {
     return id === prefix || id.startsWith(`${prefix}-`);
 }
+/**
+ * GPT-6 Sol and Luna on Chat Completions support function calling "only with
+ * `reasoning_effort` set to `none`" (developers.openai.com models/gpt-6-sol,
+ * guides/latest-model). So a request that carries tools to them must send
+ * "none", whatever /thinking says. Direct `openai` only: OpenRouter may reach
+ * OpenAI through the Responses API, where the rule does not apply, and Astra
+ * rejects "none" with a 400 (it is not offered on `openai` at all).
+ */
+export function toolsForceReasoningOff(providerId, model) {
+    if (providerId !== 'openai')
+        return false;
+    const id = canonicalModelId(model);
+    return idMatches(id, 'gpt-6-sol') || idMatches(id, 'gpt-6-luna');
+}
+/**
+ * What /thinking must tell the user when the tier does not reach every request
+ * for this model, or null. Without it the setting would look applied while
+ * agent turns quietly ran at "none".
+ */
+export function agentTurnReasoningNote(providerId, model) {
+    if (!toolsForceReasoningOff(providerId, model))
+        return null;
+    return `Agent turns on ${model} send reasoning_effort "none" whatever the tier — OpenAI's Chat Completions API lets GPT-6 Sol and Luna call tools only with reasoning off. The tier applies to plain chat.`;
+}
+/**
+ * What our Max tier sends as OpenRouter's unified `reasoning.effort`.
+ *
+ * OpenRouter accepts "xhigh" and "max", but only where the model does. Its
+ * /api/v1/models `reasoning.supported_efforts` (read 2026-09-23) lists "max"
+ * for GPT-5.6 and GPT-6, the Claude 5 family and Opus 4.7/4.8, DeepSeek V4.1
+ * Flash and V4 Pro 0813, and Kimi K3; "xhigh" is the ceiling for GPT-5.4/5.5,
+ * Grok 4.6/4.7 and Qwen 3.8 Max. Everything else keeps the old "high" cap —
+ * including the 0423 `deepseek/deepseek-v4-pro` preview (xhigh|high only) and
+ * Grok 4.5/4.3, which list no xhigh — because OpenRouter does not document what
+ * happens to a level a model does not list.
+ */
+function openRouterMaxEffort(model) {
+    const id = canonicalModelId(model);
+    const max = ['gpt-5-6', 'gpt-6', 'claude-opus-5', 'claude-fable-5', 'claude-sonnet-5',
+        'claude-opus-4-7', 'claude-opus-4-8', 'deepseek-v4-1-flash', 'deepseek-v4-pro-0813', 'kimi-k3'];
+    if (max.some(prefix => idMatches(id, prefix)))
+        return 'max';
+    const xhigh = ['gpt-5-5', 'gpt-5-4', 'grok-4-6', 'grok-4-7', 'qwen3-8-max'];
+    if (xhigh.some(prefix => idMatches(id, prefix)))
+        return 'xhigh';
+    return 'high';
+}
+/**
+ * Grok models whose ceiling is "xhigh". xAI's reasoning guide: "`xhigh` is
+ * available on `grok-4.6` and later", and grok-4.7 lists it. grok-4.5 and
+ * grok-4.3 are disputed — their model pages list xhigh, while the guide says 4.5
+ * treats it as "high" and the May-15 page gives 4.3 four levels ending at high —
+ * so they keep the high ceiling until a live request settles it.
+ */
+function grokHasXhigh(model) {
+    const id = canonicalModelId(model);
+    return idMatches(id, 'grok-4-6') || idMatches(id, 'grok-4-7');
+}
 /**
  * Does this provider+model expose a GRADED thinking-effort control we can drive?
  * Used to gate the `/thinking` UI — hidden entirely for models without one.
@@ -890,16 +1105,21 @@ export function modelSupportsReasoningEffort(providerId, model) {
             // Turbo is a plain toggle.
             return idMatches(id, 'glm-5-2') || idMatches(id, 'glm-5-3');
         case 'kimi':
-            return idMatches(id, 'k3');
+            // `kimi-for-coding` is K2.8 Preview since 2026-09-11 and takes low/high/max
+            // (default max). Matched EXACTLY: idMatches would also take in
+            // `kimi-for-coding-highspeed`, which is K2.7 Code HighSpeed — thinking
+            // always on, no effort ladder.
+            return idMatches(id, 'k3') || id === 'kimi-for-coding';
         case 'kimi-api':
         case 'kimi-cn':
             return idMatches(id, 'kimi-k3');
         case 'grok':
-            // Grok reasoning models accept reasoning_effort (none/low/medium/high).
-            // The coders (grok-code-fast, grok-build — the default) are NON-reasoning
-            // and 400 on reasoning_effort; a 400 here silently drops the whole turn
-            // into the weaker text-tool fallback (agentChat.ts), so exclude them
-            // alongside the explicit *-non-reasoning variants.
+            // Grok reasoning models accept reasoning_effort (low/medium/high, plus
+            // xhigh from grok-4.6; 4.3 also takes none, which no tier sends). The
+            // coders (grok-code-fast, grok-build — the default) reason internally but
+            // have no effort control and 400 on the parameter; a 400 here silently
+            // drops the whole turn into the weaker text-tool fallback (agentChat.ts),
+            // so exclude them alongside the explicit *-non-reasoning variants.
             if (id.startsWith('grok-build') || id.startsWith('grok-code'))
                 return false;
             return id.startsWith('grok') && !id.includes('non-reasoning');
@@ -919,7 +1139,14 @@ export function modelSupportsReasoningEffort(providerId, model) {
  * models, or providers without a graded knob — so callers can spread it
  * unconditionally. Keep in lockstep with macOS `ModelTuning.reasoningParams`.
  */
-export function reasoningParamsFor(providerId, model, tier) {
+export function reasoningParamsFor(providerId, model, tier, 
+/** The request carries a non-empty `tools` array (native tool calling). */
+opts = {}) {
+    // Before the 'auto' return on purpose: auto sends nothing, GPT-6 Sol/Luna then
+    // run at their default "medium", and that is exactly the combination OpenAI
+    // documents as unable to call tools on Chat Completions.
+    if (opts.tools && toolsForceReasoningOff(providerId, model))
+        return { reasoning_effort: 'none' };
     // 'auto', or any unexpected value from an older/garbled config, → no param.
     // (Guards against ever emitting e.g. `effort: undefined`, which could 400.)
     if (tier === 'auto' || !REASONING_TIERS.includes(tier))
@@ -930,9 +1157,14 @@ export function reasoningParamsFor(providerId, model, tier) {
         case 'anthropic':
             // low / medium / high / max — all valid on the capable Claude models.
             return { output_config: { effort: tier } };
-        case 'openai':
-            // none/low/medium/high/xhigh — no "max"; map our Max → xhigh (the ceiling).
-            return { reasoning_effort: tier === 'max' ? 'xhigh' : tier };
+        case 'openai': {
+            // Chat Completions takes none/minimal/low/medium/high/xhigh/max, but "max"
+            // is per model: it arrived with GPT-5.6 (changelog 2026-07-09) and every
+            // GPT-6 page lists it. GPT-5.5 and earlier top out at xhigh.
+            const id = canonicalModelId(model);
+            const hasMax = idMatches(id, 'gpt-5-6') || idMatches(id, 'gpt-6');
+            return { reasoning_effort: tier === 'max' ? (hasMax ? 'max' : 'xhigh') : tier };
+        }
         case 'google':
             // Gemini's OpenAI-compat layer maps reasoning_effort onto thinking_level
             // and documents low | medium | high. (Medium 400'd on Gemini 3 Preview,
@@ -941,13 +1173,12 @@ export function reasoningParamsFor(providerId, model, tier) {
             // 'minimal' is deliberately not emitted: 3.7 Flash rejects it outright.
             return { reasoning_effort: tier === 'max' ? 'high' : tier };
         case 'deepseek':
-            // V4.1 Flash distinguishes low / high / max (DeepSeek's mapping table:
-            // minimal+low→low, medium+high+xhigh→high, max+ultra→max). The retired V4
-            // ids graded only high|max. Always an effort, never a disabled block.
-            if (idMatches(canonicalModelId(model), 'deepseek-flash')) {
-                return { reasoning_effort: tier === 'low' ? 'low' : tier === 'max' ? 'max' : 'high' };
-            }
-            return { reasoning_effort: tier === 'max' ? 'max' : 'high' };
+            // V4.1 Flash and V4 Pro both distinguish low / high / max — "The thinking
+            // modes of V4-Pro and V4-Flash now support three thinking effort levels"
+            // (changelog 2026-08-13; mapping table: minimal+low→low,
+            // medium+high+xhigh→high, max+ultra→max). Always an effort, never a
+            // disabled block.
+            return { reasoning_effort: tier === 'low' ? 'low' : tier === 'max' ? 'max' : 'high' };
         case 'z.ai':
         case 'z.ai-api':
         case 'z.ai-cn':
@@ -965,11 +1196,12 @@ export function reasoningParamsFor(providerId, model, tier) {
             // Kimi K3 accepts low/high/max; collapse our medium tier to high.
             return { reasoning_effort: tier === 'low' ? 'low' : tier === 'max' ? 'max' : 'high' };
         case 'grok':
-            // none/low/medium/high — no "max"; map our Max → high (the ceiling).
-            return { reasoning_effort: tier === 'max' ? 'high' : tier };
+            // No "max" on xAI; our Max maps to the model's own ceiling — xhigh from
+            // grok-4.6, high below it (see grokHasXhigh).
+            return { reasoning_effort: tier === 'max' ? (grokHasXhigh(model) ? 'xhigh' : 'high') : tier };
         case 'openrouter':
-            // Unified reasoning object; no "max" effort → cap at high.
-            return { reasoning: { effort: tier === 'max' ? 'high' : tier } };
+            // Unified reasoning object; Max goes as high as the model lists.
+            return { reasoning: { effort: tier === 'max' ? openRouterMaxEffort(model) : tier } };
         default:
             return {};
     }
@@ -977,8 +1209,11 @@ export function reasoningParamsFor(providerId, model, tier) {
 /**
  * The DISTINCT tiers a given provider+model actually exposes — used to build a
  * per-model picker that only offers levels the model can tell apart (e.g.
- * GLM-5.2/DeepSeek grade only high|max; Gemini via the OpenAI-compat layer only
- * low|high). Always leads with 'auto'. `[]` for models with no graded knob.
+ * GLM-5.2 grades only high|max; Gemini via the OpenAI-compat layer has no max).
+ * Always leads with 'auto'. `[]` for models with no graded knob.
+ *
+ * GPT-6 Sol/Luna list their full set: it is what they run on plain chat. Agent
+ * turns send "none" regardless (toolsForceReasoningOff), and /thinking says so.
  *
  * Kept in lockstep with `reasoningParamsFor` (the providers-test asserts every
  * listed tier yields a DISTINCT param, so this can't silently drift). Mirrors
@@ -997,9 +1232,7 @@ export function availableReasoningTiers(providerId, model) {
             // coding, so collapsing it hid the setting most users want.
             return ['auto', 'low', 'medium', 'high'];
         case 'deepseek':
-            return idMatches(canonicalModelId(model), 'deepseek-flash')
-                ? ['auto', 'low', 'high', 'max']
-                : ['auto', 'high', 'max'];
+            return ['auto', 'low', 'high', 'max'];
         case 'z.ai':
         case 'z.ai-api':
         case 'z.ai-cn':
@@ -1013,9 +1246,13 @@ export function availableReasoningTiers(providerId, model) {
         case 'kimi-cn':
             return ['auto', 'low', 'high', 'max'];
         case 'grok':
-            return ['auto', 'low', 'medium', 'high'];
+            return grokHasXhigh(model)
+                ? ['auto', 'low', 'medium', 'high', 'max']
+                : ['auto', 'low', 'medium', 'high'];
         case 'openrouter':
-            return ['auto', 'low', 'medium', 'high'];
+            return openRouterMaxEffort(model) === 'high'
+                ? ['auto', 'low', 'medium', 'high']
+                : ['auto', 'low', 'medium', 'high', 'max'];
         default:
             return [];
     }