diff --git a/.env.example b/.env.example
index 89fec9c..2d44424 100644
--- a/.env.example
+++ b/.env.example
@@ -32,7 +32,7 @@ OPENAI_COMPATIBLE_API_KEY=
 OPENAI_COMPATIBLE_BASE_URL=
 OPENAI_COMPATIBLE_MODEL=
 
-# 默认模型 (可选，如 deepseek:deepseek-v4-flash)
+# 默认模型 (可选，如 deepseek:deepseek-flash)
 X_CODE_MODEL=
 
 # Tavily 搜索 (可选，免费 1000 次/月，https://tavily.com)
diff --git a/.github/workflows/pr-check.yml b/.github/workflows/pr-check.yml
index 824d00c..48768c4 100644
--- a/.github/workflows/pr-check.yml
+++ b/.github/workflows/pr-check.yml
@@ -23,8 +23,6 @@ jobs:
       - uses: actions/checkout@v7
 
       - uses: pnpm/action-setup@v4
-        with:
-          version: 10
 
       - uses: actions/setup-node@v7
         with:
@@ -43,8 +41,6 @@ jobs:
       - uses: actions/checkout@v7
 
       - uses: pnpm/action-setup@v4
-        with:
-          version: 10
 
       - uses: actions/setup-node@v7
         with:
@@ -67,8 +63,6 @@ jobs:
       - uses: actions/checkout@v7
 
       - uses: pnpm/action-setup@v4
-        with:
-          version: 10
 
       - uses: actions/setup-node@v7
         with:
@@ -85,8 +79,6 @@ jobs:
       - uses: actions/checkout@v7
 
       - uses: pnpm/action-setup@v4
-        with:
-          version: 10
 
       - uses: actions/setup-node@v7
         with:
@@ -121,8 +113,6 @@ jobs:
       - uses: actions/checkout@v7
 
       - uses: pnpm/action-setup@v4
-        with:
-          version: 10
 
       - uses: actions/setup-node@v7
         with:
@@ -141,8 +131,6 @@ jobs:
           fetch-depth: 0
 
       - uses: pnpm/action-setup@v4
-        with:
-          version: 10
 
       - uses: actions/setup-node@v7
         with:
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 0e3c3b2..58178c9 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -59,8 +59,6 @@ jobs:
       - uses: actions/checkout@v7
 
       - uses: pnpm/action-setup@v4
-        with:
-          version: 10
 
       - uses: actions/setup-node@v7
         with:
@@ -101,8 +99,6 @@ jobs:
           fetch-depth: 0
 
       - uses: pnpm/action-setup@v4
-        with:
-          version: 10
 
       - uses: actions/setup-node@v7
         with:
diff --git a/README.md b/README.md
index c1f7162..5a520be 100644
--- a/README.md
+++ b/README.md
@@ -60,6 +60,8 @@ The current release stores ChatGPT tokens in a plaintext credential file under `
 Alternatively, configure at least one provider API key:
 
 > **Recommended: [DeepSeek](https://platform.deepseek.com/)** — affordable and capable enough for everyday coding. Promotional credits and prices can change; check the official console for current terms.
+>
+> The default DeepSeek model is V4.1 Flash (`deepseek-flash`). It receives supported image attachments and tool screenshots natively; X-Code does not send them to a separate vision model.
 
 | Variable                       | Provider           | Sign up                                                                     |
 | ------------------------------ | ------------------ | --------------------------------------------------------------------------- |
@@ -72,6 +74,21 @@ Alternatively, configure at least one provider API key:
 | `ZHIPU_API_KEY`                | Zhipu (GLM)        | [open.bigmodel.cn](https://open.bigmodel.cn/usercenter/apikeys)             |
 | `MOONSHOT_API_KEY`             | Moonshot (Kimi)    | [Choose a service](#moonshot-kimi-endpoints)                                |
 
+The curated `/model` catalog tracks the current API families below. Older supported entries remain selectable.
+
+| Provider  | Current catalog highlights                                      | Smart default       |
+| --------- | --------------------------------------------------------------- | ------------------- |
+| Anthropic | Claude Fable 5.1, Opus 5, Sonnet 5, Haiku 4.5                   | Claude Sonnet 5     |
+| OpenAI    | GPT-6 Astra, GPT-5.6 Sol / Terra / Luna                         | GPT-5.6 Sol         |
+| DeepSeek  | DeepSeek V4.1 Flash (native vision), V4 Pro (text only)         | DeepSeek V4.1 Flash |
+| Alibaba   | Qwen3.8 Max / Flash (native vision), Qwen3.7 and Qwen3 variants | Qwen3.8 Max         |
+| Google    | Gemini 3.8 / 3.7 / 3.6 / 3.5 Flash, Gemini 2.5                  | Gemini 3.8 Flash    |
+| xAI       | Grok 4.6, Grok 4.20 reasoning / non-reasoning                   | Grok 4.6            |
+| Zhipu     | GLM-5.3, GLM-5.3 Flash (native vision), earlier GLM models      | GLM-5.3             |
+| Moonshot  | Kimi K3, K2.7 Code / Highspeed, K2.6                            | Kimi K3             |
+
+GPT-6 Astra is intentionally not the OpenAI smart default: GPT-5.6 Sol remains the safer default for ChatGPT subscription compatibility and lower accidental API cost. Models whose reasoning cannot be disabled automatically map `/thinking off` to their lowest supported effort.
+
 **OpenAI-compatible escape hatch** (vLLM / OpenRouter / internal gateways): set both `OPENAI_COMPATIBLE_API_KEY` and `OPENAI_COMPATIBLE_BASE_URL`, then address models as `custom:<your-model-id>`.
 
 <details>
@@ -134,7 +151,7 @@ To enable the `webSearch` tool, configure any one of the following:
 
 > Tavily is recommended for first-time setup: simpler signup, LLM-optimized responses. When several keys are set, the first in the table order above wins. Set `X_CODE_WEB_SEARCH_PROVIDER` to `tavily`, `brave`, `exa`, `perplexity`, `firecrawl`, or `deepseek` to force a specific provider.
 >
-> **DeepSeek users need no extra key**: when the active model is a DeepSeek model and `DEEPSEEK_API_KEY` is set, `webSearch` automatically uses DeepSeek's built-in server-side web search. Note that each search is billed as a model turn (default `deepseek-v4-flash`), not as a flat search request.
+> **DeepSeek users need no extra key**: when the active model is a DeepSeek model and `DEEPSEEK_API_KEY` is set, `webSearch` automatically uses DeepSeek's built-in server-side web search. Note that each search is billed as a model turn (default `deepseek-flash`), not as a flat search request.
 
 </details>
 
@@ -147,7 +164,7 @@ Moonshot/Kimi credentials come from three separate services. A key only works wi
 - China Open Platform: [platform.kimi.com](https://platform.kimi.com/console/api-keys) → `https://api.moonshot.cn/v1`
 - International Open Platform: [platform.kimi.ai](https://platform.kimi.ai/console/api-keys) → `https://api.moonshot.ai/v1`
 
-After selecting a Kimi model via `/model`, an endpoint picker appears automatically.
+After selecting a Kimi model via `/model`, an endpoint picker appears automatically. On Coding Plan, the stable `kimi-for-coding` wire id currently auto-routes to K2.8 Preview; the public Moonshot API model ids remain unchanged.
 
 </details>
 
@@ -312,6 +329,8 @@ Log path: `~/.x-code/logs/debug.log` (Windows: `%USERPROFILE%\.x-code\logs\debug
 
 Requires Node.js 22+ and pnpm 10.x.
 
+The repository pins pnpm 10.28.2 through `packageManager`. If pnpm 11 is installed globally, run `corepack enable` and use the repository-local `pnpm` (or `corepack pnpm`); running pnpm 11 directly is intentionally rejected by `engines.pnpm`.
+
 ```bash
 git clone https://github.com/woai3c/x-code-cli.git
 cd x-code-cli
diff --git a/README.zh-CN.md b/README.zh-CN.md
index 3915bc0..6bc2136 100644
--- a/README.zh-CN.md
+++ b/README.zh-CN.md
@@ -60,6 +60,8 @@ xc logout                   # 退出并删除本地 ChatGPT 凭据
 也可以配置至少一个厂商的 API Key：
 
 > **推荐 [DeepSeek](https://platform.deepseek.com/)**：价格低、国内访问稳定，适合首次试用。赠送额度与价格可能变化，请以官方控制台为准。
+>
+> DeepSeek 默认模型为 V4.1 Flash（`deepseek-flash`）。它会原生接收支持的图片附件和工具截图，X-Code 不会将它们发送给其他视觉模型。
 
 | 环境变量                       | 厂商                | 注册地址                                                                    |
 | ------------------------------ | ------------------- | --------------------------------------------------------------------------- |
@@ -72,6 +74,21 @@ xc logout                   # 退出并删除本地 ChatGPT 凭据
 | `ZHIPU_API_KEY`                | 智谱（GLM）         | [open.bigmodel.cn](https://open.bigmodel.cn/usercenter/apikeys)             |
 | `MOONSHOT_API_KEY`             | Moonshot（Kimi）    | [按服务选择](#moonshot-kimi-endpoints)                                      |
 
+内置 `/model` 目录已跟进以下当前 API 型号；仍可用的旧型号继续保留。
+
+| 厂商      | 当前目录重点型号                                       | 智能默认值          |
+| --------- | ------------------------------------------------------ | ------------------- |
+| Anthropic | Claude Fable 5.1、Opus 5、Sonnet 5、Haiku 4.5          | Claude Sonnet 5     |
+| OpenAI    | GPT-6 Astra、GPT-5.6 Sol / Terra / Luna                | GPT-5.6 Sol         |
+| DeepSeek  | DeepSeek V4.1 Flash（原生视觉）、V4 Pro（仅文本）      | DeepSeek V4.1 Flash |
+| 阿里通义  | Qwen3.8 Max / Flash（原生视觉）、Qwen3.7 与 Qwen3 系列 | Qwen3.8 Max         |
+| Google    | Gemini 3.8 / 3.7 / 3.6 / 3.5 Flash、Gemini 2.5         | Gemini 3.8 Flash    |
+| xAI       | Grok 4.6、Grok 4.20 推理版 / 非推理版                  | Grok 4.6            |
+| 智谱      | GLM-5.3、GLM-5.3 Flash（原生视觉）及较早 GLM 型号      | GLM-5.3             |
+| Moonshot  | Kimi K3、K2.7 Code / Highspeed、K2.6                   | Kimi K3             |
+
+GPT-6 Astra 不会自动成为 OpenAI 智能默认值：保留 GPT-5.6 Sol 可以兼顾 ChatGPT 订阅兼容性，并降低误用高价 API 的风险。对于不能关闭推理的型号，`/thinking off` 会自动降到该型号支持的最低推理档位。
+
 **OpenAI 兼容接入**（vLLM / OpenRouter / 代理网关等）：同时设置 `OPENAI_COMPATIBLE_API_KEY` 与 `OPENAI_COMPATIBLE_BASE_URL`，模型 ID 写成 `custom:<your-model-id>`。
 
 <details>
@@ -134,7 +151,7 @@ setx DEEPSEEK_API_KEY "sk-..."
 
 > 推荐首次配 Tavily：注册简便，返回格式针对 LLM 优化。配置多个 key 时按上表顺序取第一个；也可通过 `X_CODE_WEB_SEARCH_PROVIDER` 显式指定（可选值：`tavily`、`brave`、`exa`、`perplexity`、`firecrawl`、`deepseek`）。
 >
-> **DeepSeek 用户无需额外 key**：当前模型为 DeepSeek 且已配置 `DEEPSEEK_API_KEY` 时，`webSearch` 自动使用 DeepSeek 内置的服务端联网搜索。注意每次搜索按一次模型调用计费（默认 `deepseek-v4-flash`），而非按搜索次数计费。
+> **DeepSeek 用户无需额外 key**：当前模型为 DeepSeek 且已配置 `DEEPSEEK_API_KEY` 时，`webSearch` 自动使用 DeepSeek 内置的服务端联网搜索。注意每次搜索按一次模型调用计费（默认 `deepseek-flash`），而非按搜索次数计费。
 
 </details>
 
@@ -147,7 +164,7 @@ Moonshot/Kimi 提供三套独立的凭证与端点，API Key 只能用于签发
 - 国内开放平台：[platform.kimi.com](https://platform.kimi.com/console/api-keys) → `https://api.moonshot.cn/v1`
 - 国际开放平台：[platform.kimi.ai](https://platform.kimi.ai/console/api-keys) → `https://api.moonshot.ai/v1`
 
-通过 `/model` 选择 Kimi 模型后，X-Code CLI 会自动显示端点选择器。
+通过 `/model` 选择 Kimi 模型后，X-Code CLI 会自动显示端点选择器。Coding Plan 的稳定 wire id `kimi-for-coding` 当前会自动路由到 K2.8 Preview；Moonshot 开放平台的模型 ID 不变。
 
 </details>
 
@@ -312,6 +329,8 @@ set DEBUG_STDOUT=1 && xc
 
 需要 Node.js 22+ 和 pnpm 10.x。
 
+仓库通过 `packageManager` 固定使用 pnpm 10.28.2。如果全局安装的是 pnpm 11，请先运行 `corepack enable`，再在仓库内使用 `pnpm`（或直接使用 `corepack pnpm`）；`engines.pnpm` 会有意拒绝 pnpm 11 直接执行。
+
 ```bash
 git clone https://github.com/woai3c/x-code-cli.git
 cd x-code-cli
diff --git a/package.json b/package.json
index aa4fec6..34c6fdd 100644
--- a/package.json
+++ b/package.json
@@ -1,5 +1,6 @@
 {
   "private": true,
+  "packageManager": "pnpm@10.28.2",
   "type": "module",
   "engines": {
     "node": ">=22.0.0",
diff --git a/packages/cli/src/ui/app/App.tsx b/packages/cli/src/ui/app/App.tsx
index 83b8f91..aa1ab50 100644
--- a/packages/cli/src/ui/app/App.tsx
+++ b/packages/cli/src/ui/app/App.tsx
@@ -2039,7 +2039,7 @@ export function App({
       // automatically on /model switch since switchModel updates
       // state.modelId. When the model has a reasoning-effort tier configured
       // (via the /model tier picker), it appends next to the name — e.g.
-      // "deepseek-v4-flash · High".
+      // "deepseek-flash · High".
       modelLabel={
         reasoningTierLabel
           ? `${renderModelLabel(state.modelId)} · ${reasoningTierLabel}`
diff --git a/packages/cli/src/ui/utils.ts b/packages/cli/src/ui/utils.ts
index 898a13f..a02282e 100644
--- a/packages/cli/src/ui/utils.ts
+++ b/packages/cli/src/ui/utils.ts
@@ -108,7 +108,8 @@ const SHELL_LABELS: Record<string, string> = {
 export function getToolLabel(toolName: string): string {
   const n = normalizeToolName(toolName)
   if (n === 'shell' || n === 'bash') return SHELL_LABELS[getShellProvider().type] ?? 'Shell'
-  if (n === 'readfile' || n === 'read' || n === 'fileingest') return 'Read'
+  if (n === 'fileingest') return 'Attach'
+  if (n === 'readfile' || n === 'read') return 'Read'
   if (n === 'writefile' || n === 'write') return 'Write'
   if (n === 'edit' || n === 'update') return 'Update'
   if (n === 'glob') return 'Glob'
@@ -192,6 +193,7 @@ export interface ReadGroupSummary {
 const TABLE_OUTPUT_MAX_LINES = 30
 
 export function formatReadGroupSummary(tools: readonly DisplayToolCall[]): ReadGroupSummary {
+  let attachCount = 0
   let readCount = 0
   let grepCount = 0
   let globCount = 0
@@ -200,7 +202,11 @@ export function formatReadGroupSummary(tools: readonly DisplayToolCall[]): ReadG
 
   for (const tc of tools) {
     const n = normalizeToolName(tc.toolName)
-    if (n === 'read' || n === 'readfile' || n === 'fileingest') {
+    if (n === 'fileingest') {
+      attachCount++
+      const p = (tc.input.filePath as string) || (tc.input.file_path as string) || (tc.input.path as string) || ''
+      if (p) readPaths.push(basename(p))
+    } else if (n === 'read' || n === 'readfile') {
       readCount++
       const p = (tc.input.filePath as string) || (tc.input.file_path as string) || (tc.input.path as string) || ''
       if (p) readPaths.push(basename(p))
@@ -214,6 +220,7 @@ export function formatReadGroupSummary(tools: readonly DisplayToolCall[]): ReadG
   }
 
   const clauses: string[] = []
+  if (attachCount > 0) clauses.push(`attached ${attachCount} file${attachCount === 1 ? '' : 's'}`)
   if (readCount > 0) clauses.push(`read ${readCount} file${readCount === 1 ? '' : 's'}`)
   if (grepCount > 0) clauses.push(`searched for ${grepCount} pattern${grepCount === 1 ? '' : 's'}`)
   if (globCount > 0) clauses.push(`globbed ${globCount} pattern${globCount === 1 ? '' : 's'}`)
diff --git a/packages/cli/tests/e2e/README.md b/packages/cli/tests/e2e/README.md
index aa46d6c..ce464cc 100644
--- a/packages/cli/tests/e2e/README.md
+++ b/packages/cli/tests/e2e/README.md
@@ -16,7 +16,7 @@ pnpm test:e2e
 ```
 
 The runner detects which `*_API_KEY` you have set, lists the matching models,
-and asks you to pick one. Default is `deepseek:deepseek-v4-flash` (cheap, fast).
+and asks you to pick one. Default is `deepseek:deepseek-flash` (V4.1 Flash, cheap, fast, and vision-capable).
 
 ## Flags
 
@@ -34,8 +34,8 @@ pnpm test:e2e --max-turns 8        # cap agent loop turns
 
 ## Cost
 
-`deepseek-v4-flash` runs the whole suite (26 scenarios, ~50-100K tokens total)
-for roughly **$0.10–0.18 per full run**. Each scenario takes 5–30 seconds.
+`deepseek-flash` runs the whole suite (26 scenarios, ~50-100K tokens total).
+Cost varies with input/output mix, cache hits, and DeepSeek peak/off-peak pricing. Each scenario takes 5–30 seconds.
 Full suite: 4–8 minutes typically.
 
 If you only want to verify changes near a specific area, use `--filter` and
@@ -85,7 +85,7 @@ export default scenario
 | Method                        | What it does                                                     |
 | ----------------------------- | ---------------------------------------------------------------- |
 | `ctx.tmpDir`                  | Absolute path to this scenario's temp directory. CLI runs there. |
-| `ctx.modelId`                 | Resolved model id (e.g. `deepseek:deepseek-v4-flash`).           |
+| `ctx.modelId`                 | Resolved model id (e.g. `deepseek:deepseek-flash`).              |
 | `ctx.writeFile(rel, content)` | Write a file inside tmpDir.                                      |
 | `ctx.readFile(rel)`           | Read a file.                                                     |
 | `ctx.fileExists(rel)`         | Returns `Promise<boolean>`.                                      |
diff --git a/packages/cli/tests/e2e/framework/models.ts b/packages/cli/tests/e2e/framework/models.ts
index 35910ad..3e0265d 100644
--- a/packages/cli/tests/e2e/framework/models.ts
+++ b/packages/cli/tests/e2e/framework/models.ts
@@ -7,39 +7,70 @@ import { fileURLToPath } from 'node:url'
 const __dirname = path.dirname(fileURLToPath(import.meta.url))
 const REPO_ROOT = path.resolve(__dirname, '..', '..', '..', '..', '..')
 
-export const DEFAULT_MODEL = 'deepseek:deepseek-v4-flash'
+export const DEFAULT_MODEL = 'deepseek:deepseek-flash'
 
 /** Provider env-var → list of model ids that can be selected when that key is set.
  *  Keep aligned with `packages/core/src/types/index.ts::PROVIDER_DETECTION_ORDER`. */
 const PROVIDER_MODELS: Record<string, string[]> = {
-  DEEPSEEK_API_KEY: ['deepseek:deepseek-v4-flash', 'deepseek:deepseek-v4-pro'],
+  DEEPSEEK_API_KEY: ['deepseek:deepseek-flash', 'deepseek:deepseek-v4-pro'],
   ANTHROPIC_API_KEY: [
+    'anthropic:claude-fable-5-1',
+    'anthropic:claude-opus-5',
     'anthropic:claude-fable-5',
     'anthropic:claude-opus-4-8',
     'anthropic:claude-sonnet-5',
     'anthropic:claude-haiku-4-5',
   ],
-  OPENAI_API_KEY: ['openai:gpt-5.6-sol', 'openai:gpt-5.6-terra', 'openai:gpt-5.6-luna', 'openai:gpt-5.4-mini'],
-  GOOGLE_GENERATIVE_AI_API_KEY: ['google:gemini-3.5-flash', 'google:gemini-2.5-pro', 'google:gemini-2.5-flash'],
-  XAI_API_KEY: ['xai:grok-4.3', 'xai:grok-4.5'],
-  ALIBABA_API_KEY: ['alibaba:qwen3.7-max', 'alibaba:qwen3.7-plus', 'alibaba:qwen3-coder-plus', 'alibaba:qwq-plus'],
-  ZHIPU_API_KEY: ['zhipu:glm-5.2', 'zhipu:glm-5', 'zhipu:glm-4.7'],
-  MOONSHOT_API_KEY: ['moonshotai:kimi-k2.6', 'moonshotai:kimi-k3'],
+  OPENAI_API_KEY: [
+    'openai:gpt-6-astra',
+    'openai:gpt-5.6-sol',
+    'openai:gpt-5.6-terra',
+    'openai:gpt-5.6-luna',
+    'openai:gpt-5.4-mini',
+  ],
+  GOOGLE_GENERATIVE_AI_API_KEY: [
+    'google:gemini-3.8-flash',
+    'google:gemini-3.7-flash',
+    'google:gemini-3.6-flash',
+    'google:gemini-3.5-flash',
+    'google:gemini-2.5-pro',
+    'google:gemini-2.5-flash',
+  ],
+  XAI_API_KEY: ['xai:grok-4.6', 'xai:grok-4.20', 'xai:grok-4.20-non-reasoning', 'xai:grok-4.5'],
+  ALIBABA_API_KEY: [
+    'alibaba:qwen3.8-max',
+    'alibaba:qwen3.8-flash',
+    'alibaba:qwen3.7-max',
+    'alibaba:qwen3.7-plus',
+    'alibaba:qwen3.7-flash',
+    'alibaba:qwen3-coder-plus',
+    'alibaba:qwq-plus',
+  ],
+  ZHIPU_API_KEY: ['zhipu:glm-5.3', 'zhipu:glm-5.3-flash', 'zhipu:glm-5.2', 'zhipu:glm-5', 'zhipu:glm-4.7'],
+  MOONSHOT_API_KEY: [
+    'moonshotai:kimi-k3',
+    'moonshotai:kimi-k2.7-code',
+    'moonshotai:kimi-k2.7-code-highspeed',
+    'moonshotai:kimi-k2.6',
+  ],
 }
 
 /** Short aliases — accepted on CLI `--model` flag. Aligns with product's
  *  `MODEL_ALIASES` table; keep them in sync. */
 export const ALIASES: Record<string, string> = {
-  fable: 'anthropic:claude-fable-5',
+  fable: 'anthropic:claude-fable-5-1',
   sonnet: 'anthropic:claude-sonnet-5',
-  opus: 'anthropic:claude-opus-4-8',
+  opus: 'anthropic:claude-opus-5',
   haiku: 'anthropic:claude-haiku-4-5',
+  gpt6: 'openai:gpt-6-astra',
+  astra: 'openai:gpt-6-astra',
   gpt5: 'openai:gpt-5.6-sol',
-  gemini: 'google:gemini-3.5-flash',
-  deepseek: 'deepseek:deepseek-v4-flash',
+  gemini: 'google:gemini-3.8-flash',
+  deepseek: 'deepseek:deepseek-flash',
   'deepseek-pro': 'deepseek:deepseek-v4-pro',
-  qwen: 'alibaba:qwen3.7-max',
-  glm: 'zhipu:glm-5.2',
+  qwen: 'alibaba:qwen3.8-max',
+  grok: 'xai:grok-4.6',
+  glm: 'zhipu:glm-5.3',
   kimi: 'moonshotai:kimi-k3',
 }
 
diff --git a/packages/cli/tests/e2e/scenarios/14-large-file.ts b/packages/cli/tests/e2e/scenarios/14-large-file.ts
index 30d2f23..07e9744 100644
--- a/packages/cli/tests/e2e/scenarios/14-large-file.ts
+++ b/packages/cli/tests/e2e/scenarios/14-large-file.ts
@@ -13,7 +13,7 @@ const scenario: Scenario = {
       lines.push(`line ${i}: lorem ipsum dolor sit amet`)
     }
     // 关键标记放在尾部。前 4999 行模式完全一致 — 一个偷懒的模型只看 head
-    // 就会"按模式外推"答 `lorem ipsum dolor sit amet`（deepseek-v4-flash 实测会）；
+    // 就会"按模式外推"答 `lorem ipsum dolor sit amet`（deepseek-flash 实测会）；
     // 只有真的二次调 readFile(offset≈5000) 拿到尾部才能引用出这个 token。
     lines[4999] = 'line 5000: FINAL_SENTINEL_TOKEN_XYZ'
     await ctx.writeFile('big.txt', lines.join('\n'))
diff --git a/packages/cli/tests/openai-auth-transition-state.test.ts b/packages/cli/tests/openai-auth-transition-state.test.ts
index 5388bf2..02ef5ce 100644
--- a/packages/cli/tests/openai-auth-transition-state.test.ts
+++ b/packages/cli/tests/openai-auth-transition-state.test.ts
@@ -69,7 +69,7 @@ describe('OpenAI auth transition state', () => {
 
   it('falls back or disables deterministically when the new account has no models', () => {
     expect(planOpenAIModelReconciliation('openai:old', [], ['deepseek'], 'blocked', (id) => id)).toMatchObject({
-      modelId: 'deepseek:deepseek-v4-flash',
+      modelId: 'deepseek:deepseek-flash',
     })
     expect(planOpenAIModelReconciliation('openai:old', [], [], 'blocked', (id) => id)).toEqual({
       blockedMessage: 'blocked',
diff --git a/packages/cli/tests/pty/tui-input.test.ts b/packages/cli/tests/pty/tui-input.test.ts
index d5f3330..6c7111f 100644
--- a/packages/cli/tests/pty/tui-input.test.ts
+++ b/packages/cli/tests/pty/tui-input.test.ts
@@ -115,10 +115,10 @@ describe('TUI input and lifecycle', () => {
         const raw = harness.raw()
         const oldReplyTail = raw.lastIndexOf('initial-finished')
         const queuedUser = raw.lastIndexOf('@"queued attachment.txt" analyze this file')
-        const readSummary = raw.lastIndexOf('Read')
+        const attachSummary = raw.lastIndexOf('Attach')
         expect(oldReplyTail).toBeGreaterThanOrEqual(0)
         expect(queuedUser).toBeGreaterThan(oldReplyTail)
-        expect(readSummary).toBeGreaterThan(queuedUser)
+        expect(attachSummary).toBeGreaterThan(queuedUser)
       },
       {
         beforeStart: async (workspace) => {
diff --git a/packages/cli/tests/stdout-writer-spacing.test.ts b/packages/cli/tests/stdout-writer-spacing.test.ts
index 94bbd60..d1aea91 100644
--- a/packages/cli/tests/stdout-writer-spacing.test.ts
+++ b/packages/cli/tests/stdout-writer-spacing.test.ts
@@ -152,7 +152,7 @@ describe('stdout writer spacing', () => {
     })
 
     const plain = output.replace(/\x1b\[[0-9;]*m/g, '')
-    expect(plain).toContain('Read 2 files(invoice.pdf, analysis.docx)')
+    expect(plain).toContain('Attached 2 files(invoice.pdf, analysis.docx)')
     expect(plain.match(/Prepared for analysis\./g)).toBeNull()
   })
 
diff --git a/packages/cli/tests/utils.test.ts b/packages/cli/tests/utils.test.ts
index 73ea176..2a787e6 100644
--- a/packages/cli/tests/utils.test.ts
+++ b/packages/cli/tests/utils.test.ts
@@ -50,8 +50,8 @@ describe('CLI tool result summaries', () => {
 })
 
 describe('local file ingestion tool display', () => {
-  it('renders the built-in preflight as Read with portable file-name previews', () => {
-    expect(getToolLabel('fileIngest')).toBe('Read')
+  it('renders the built-in preflight as Attach with portable file-name previews', () => {
+    expect(getToolLabel('fileIngest')).toBe('Attach')
     expect(getToolInputPreview('fileIngest', { filePath: 'C:\\reports\\invoice.pdf' })).toBe('C:\\reports\\invoice.pdf')
   })
 })
diff --git a/packages/core/package.json b/packages/core/package.json
index 2fb2d6b..2489d5b 100644
--- a/packages/core/package.json
+++ b/packages/core/package.json
@@ -28,7 +28,7 @@
   "dependencies": {
     "@ai-sdk/alibaba": "^2.0.0",
     "@ai-sdk/anthropic": "^4.0.0",
-    "@ai-sdk/deepseek": "^3.0.28",
+    "@ai-sdk/deepseek": "^3.0.45",
     "@ai-sdk/google": "^4.0.0",
     "@ai-sdk/moonshotai": "^3.0.0",
     "@ai-sdk/openai": "^4.0.0",
diff --git a/packages/core/src/agent/compression.ts b/packages/core/src/agent/compression.ts
index 796a970..3056f40 100644
--- a/packages/core/src/agent/compression.ts
+++ b/packages/core/src/agent/compression.ts
@@ -362,6 +362,9 @@ async function finishLightweightCompression(
 ): Promise<void> {
   await markBoundaryAndReflush(state, undefined, candidate)
   setTrackedTranscript(state, candidate)
+  // A compacted transcript may no longer contain the file bodies that made
+  // these de-dup entries valid. Let readFile deliver them again on demand.
+  state.readFileCache.clear()
   state.lastInputTokens = 0
   markExpectedCacheMiss(state, 'compaction')
   emitCompactionHook(hookCtx, {
@@ -389,6 +392,7 @@ async function summarizeLoopState(
   )
   await markBoundaryAndReflush(state, compressed.summary, compressed.trackedMessages)
   setTrackedTranscript(state, compressed.trackedMessages)
+  state.readFileCache.clear()
   await recordCompressionUsage(state, compressed, callbacks, hookCtx)
   state.lastInputTokens = 0
   markExpectedCacheMiss(state, 'compaction')
diff --git a/packages/core/src/agent/context-window.ts b/packages/core/src/agent/context-window.ts
index 85753bd..7fc8c09 100644
--- a/packages/core/src/agent/context-window.ts
+++ b/packages/core/src/agent/context-window.ts
@@ -49,41 +49,57 @@ export function setContextWindowOverride(value: unknown): number | undefined {
 /** Context window sizes per model (tokens). */
 const MODEL_CONTEXT_WINDOWS: ReadonlyMap<string, number> = new Map([
   // Anthropic
+  ['anthropic:claude-fable-5-1', 1000000],
+  ['anthropic:claude-opus-5', 1000000],
   ['anthropic:claude-fable-5', 1000000],
   ['anthropic:claude-opus-4-8', 1000000],
   ['anthropic:claude-sonnet-5', 1000000],
   ['anthropic:claude-haiku-4-5', 200000],
   // OpenAI
+  ['openai:gpt-6-astra', 1050000],
   ['openai:gpt-5.6-sol', 1047576],
   ['openai:gpt-5.6-terra', 1047576],
   ['openai:gpt-5.6-luna', 1047576],
   ['openai:gpt-5.4-mini', 1047576],
   ['openai:gpt-5.4-nano', 1047576],
   // Google
+  ['google:gemini-3.8-flash', 1048576],
+  ['google:gemini-3.7-flash', 1048576],
+  ['google:gemini-3.6-flash', 1048576],
   ['google:gemini-3.5-flash', 1000000],
   ['google:gemini-2.5-pro', 1000000],
   ['google:gemini-2.5-flash', 1000000],
   // DeepSeek
+  ['deepseek:deepseek-flash', 1000000],
   ['deepseek:deepseek-v4-flash', 1000000],
   ['deepseek:deepseek-v4-pro', 1000000],
-  // Alibaba — per DashScope docs: qwen3.7-max and qwen3-coder-plus extend to 1M;
+  // Alibaba — current Qwen3.8 models, qwen3.7-max/plus, and qwen3-coder-plus extend to 1M;
   // qwen-max still caps at 32k. Values verified against
   // https://help.aliyun.com/zh/model-studio/models.
+  ['alibaba:qwen3.8-max', 1000000],
+  ['alibaba:qwen3.8-flash', 1000000],
   ['alibaba:qwen3.7-max', 1000000],
-  ['alibaba:qwen3.7-plus', 131072],
+  ['alibaba:qwen3.7-plus', 1000000],
+  ['alibaba:qwen3.7-flash', 1000000],
   ['alibaba:qwen3-coder-plus', 1000000],
   ['alibaba:qwq-plus', 131072],
   ['alibaba:qwen-max', 32768],
-  // xAI — grok-4.5 has 500k window; grok-4.3 has 1M.
-  ['xai:grok-4.5', 512000],
+  // xAI — Grok 4.6/4.5 have 500k windows; Grok 4.20/4.3 have 1M.
+  ['xai:grok-4.6', 500000],
+  ['xai:grok-4.20', 1000000],
+  ['xai:grok-4.20-non-reasoning', 1000000],
+  ['xai:grok-4.5', 500000],
   ['xai:grok-4.3', 1000000],
   // Zhipu
+  ['zhipu:glm-5.3', 1000000],
+  ['zhipu:glm-5.3-flash', 1000000],
   ['zhipu:glm-5.2', 1000000],
   ['zhipu:glm-5', 200000],
   ['zhipu:glm-4.7', 128000],
   // Moonshot
   ['moonshotai:kimi-k3', 1000000],
   ['moonshotai:kimi-k2.7-code', 262144],
+  ['moonshotai:kimi-k2.7-code-highspeed', 262144],
   ['moonshotai:kimi-k2.6', 262144],
 ])
 
@@ -125,15 +141,31 @@ export function getCompressionThreshold(modelId: string): number {
  */
 const DEFAULT_MAX_OUTPUT_TOKENS = 16384
 const MODEL_MAX_OUTPUT_TOKENS: ReadonlyMap<string, number> = new Map([
-  // DeepSeek V4: both flash and pro advertise up to 384K output tokens.
+  // Current flagships with provider-documented extended output ceilings.
+  ['anthropic:claude-fable-5-1', 128000],
+  ['anthropic:claude-opus-5', 128000],
+  ['anthropic:claude-fable-5', 128000],
+  ['anthropic:claude-opus-4-8', 128000],
+  ['anthropic:claude-sonnet-5', 128000],
+  ['anthropic:claude-haiku-4-5', 64000],
+  ['openai:gpt-6-astra', 128000],
+  ['google:gemini-3.8-flash', 65536],
+  ['google:gemini-3.7-flash', 65536],
+  ['google:gemini-3.6-flash', 65536],
+  // DeepSeek V4/V4.1: flash and pro advertise up to 384K output tokens.
   // We cap at a generous but conservative 131072 to avoid edge-case 400s.
+  ['deepseek:deepseek-flash', 131072],
   ['deepseek:deepseek-v4-flash', 131072],
   ['deepseek:deepseek-v4-pro', 131072],
-  // Alibaba — Qwen3.7 models support 32768 (non-thinking) / 81920 (thinking).
+  // Alibaba — Qwen3.8 and multimodal Qwen3.7 Plus support 64k output.
+  ['alibaba:qwen3.8-max', 65536],
+  ['alibaba:qwen3.8-flash', 65536],
+  ['alibaba:qwen3.7-plus', 65536],
+  ['alibaba:qwen3.7-flash', 65536],
+  // Older Qwen3.7 models support 32768 (non-thinking) / 81920 (thinking).
   // We cap at the non-thinking ceiling so the request always succeeds.
   ['alibaba:qwen-max', 8192],
   ['alibaba:qwen3.7-max', 32000],
-  ['alibaba:qwen3.7-plus', 32000],
   ['alibaba:qwen3-coder-plus', 32000],
   ['alibaba:qwq-plus', 32000],
 ])
diff --git a/packages/core/src/agent/file-ingest.ts b/packages/core/src/agent/file-ingest.ts
index 56a1ebd..d989fd0 100644
--- a/packages/core/src/agent/file-ingest.ts
+++ b/packages/core/src/agent/file-ingest.ts
@@ -659,6 +659,7 @@ export async function ingestFile(
         parts.push({ type: 'text', text: buildCompressionCaption(compressed) })
         onNotice?.(`Normalized image: ${formatBytes(buffer.length)} → ${formatBytes(compressed.data.length)}`)
       }
+      parts.push({ type: 'text', text: BUILT_IN_MEDIA_ANALYSIS_NOTE })
       return parts
     }
 
@@ -732,7 +733,11 @@ function hasCompleteLocalFile(parts: IngestedPart[], filePath: string): boolean
     if (part.type !== 'text' || !part.text.startsWith('<<file ')) return false
     const openingEnd = part.text.indexOf('>>')
     if (openingEnd === -1 || !part.text.slice(0, openingEnd).includes(pathAttribute)) return false
-    return part.text.indexOf('<</file>>', openingEnd + 2) !== -1
+    if (part.text.indexOf('<</file>>', openingEnd + 2) !== -1) return true
+    return (
+      part.text.slice(0, openingEnd).includes('kind="image"') &&
+      parts.some((candidate) => candidate.type === 'file' && candidate.mediaType.startsWith('image/'))
+    )
   })
 }
 
diff --git a/packages/core/src/agent/local-media.ts b/packages/core/src/agent/local-media.ts
index 3c44ceb..1929a2b 100644
--- a/packages/core/src/agent/local-media.ts
+++ b/packages/core/src/agent/local-media.ts
@@ -16,7 +16,7 @@ export type ProcessedLocalPart =
 
 export const BUILT_IN_MEDIA_ANALYSIS_NOTE =
   '[Built-in local media processing succeeded. Analyze the supplied content directly. ' +
-  'Do not invoke shell, Node.js, Python, FFmpeg, or other external programs merely to re-read, parse, OCR, ' +
+  'Do not invoke readFile, shell, Node.js, Python, FFmpeg, or other external programs merely to re-read, parse, OCR, ' +
   'transcribe, or independently validate values from this attachment. Use external programs only if the built-in ' +
   'pipeline reports a failure, or the user explicitly asks for conversion, codec diagnostics, or independent validation.]'
 
diff --git a/packages/core/src/agent/loop.ts b/packages/core/src/agent/loop.ts
index 8110b96..5af3e9b 100644
--- a/packages/core/src/agent/loop.ts
+++ b/packages/core/src/agent/loop.ts
@@ -21,8 +21,13 @@ import { listMcpResources, readMcpResource } from '../mcp/resources.js'
 import { bridgeMcpTool, toSystemPromptEntries } from '../mcp/tool-bridge.js'
 import { listAgentsTool, sendMessageTool } from '../peers/tools.js'
 import { applyCacheControl, openAICacheComparisonTtlMs } from '../providers/cache-control.js'
-import { withZhipuReasoningHeader } from '../providers/registry.js'
-import { getReasoningLevel, getThinkingProviderOptions, mergeThinkingOptions } from '../providers/thinking.js'
+import { withXaiReasoningHeader, withZhipuReasoningHeader } from '../providers/registry.js'
+import {
+  getReasoningEffort,
+  getReasoningLevel,
+  getThinkingProviderOptions,
+  mergeThinkingOptions,
+} from '../providers/thinking.js'
 import { createActivateSkillTool } from '../tools/activate-skill.js'
 import { BROWSER_VISUAL_CHECK_TOOL_NAME, browserVisualCheck } from '../tools/browser-visual-check.js'
 import { createGetGoalTool } from '../tools/get-goal.js'
@@ -516,11 +521,14 @@ async function runTurnAttempt(
   // /thinking toggle. If the user explicitly chose a reasoning effort level
   // for this model (stored in config.modelReasoningEffort), we use it.
   const effort = userConfig.modelReasoningEffort?.[options.modelId]
+  const effectiveEffort = getReasoningEffort(options.modelId, options.thinking ?? false, effort)
   const reasoningLevel = getReasoningLevel(options.modelId, options.thinking ?? false, effort)
   const thinkingOptions = getThinkingProviderOptions(options.modelId, options.thinking ?? false, effort)
   const mergedProviderOptions = mergeThinkingOptions(cached.providerOptions, thinkingOptions)
-  const requestHeaders =
-    options.modelId.split(':')[0] === 'zhipu' ? withZhipuReasoningHeader(cached.headers, effort) : cached.headers
+  const provider = options.modelId.split(':')[0]
+  let requestHeaders = cached.headers
+  if (provider === 'zhipu') requestHeaders = withZhipuReasoningHeader(requestHeaders, effectiveEffort)
+  if (provider === 'xai') requestHeaders = withXaiReasoningHeader(requestHeaders, effectiveEffort)
 
   const requestTimestamp = new Date().toISOString()
   let result: StreamResult
@@ -600,7 +608,7 @@ async function runTurnAttempt(
       // notices and retry once. stripBinaryPartsFromMessages returns false
       // when nothing matched — then the bad part isn't in a shape we
       // recognize, so fall through and report instead of looping forever.
-      if (stripBinaryPartsFromMessages(state.messages)) {
+      if (stripBinaryPartsFromMessages(state.messages, state.readFileCache)) {
         recalculateContextSecurity(state)
         state.transcriptRequiresSnapshot = true
         await flushPendingMessages(state)
diff --git a/packages/core/src/agent/provider-compat.ts b/packages/core/src/agent/provider-compat.ts
index 7cd6f90..727c808 100644
--- a/packages/core/src/agent/provider-compat.ts
+++ b/packages/core/src/agent/provider-compat.ts
@@ -24,8 +24,8 @@ import type { ToolImage } from './messages.js'
 
 // ── Image/PDF downgrade for text-only providers ───────────────────────────
 //
-// If the active provider can't receive image/file parts (DeepSeek today,
-// plus `custom` unless the user opts in), walk every message that would be
+// If the active model can't receive image/file parts (for example DeepSeek V4
+// Pro, plus `custom` unless the user opts in), walk every message that would be
 // sent on the next turn and replace each binary part with something the
 // provider CAN accept.
 //
@@ -320,7 +320,10 @@ function imagePartToBuffer(part: { image: unknown; mediaType?: string }): Buffer
  *  retry the turn. Returns true only when something actually changed, so the
  *  caller can tell "retry is worthwhile" from "the bad part isn't in a shape
  *  we recognize — report the error instead of looping". */
-export function stripBinaryPartsFromMessages(messages: ModelMessage[]): boolean {
+export function stripBinaryPartsFromMessages(
+  messages: ModelMessage[],
+  deliveredFileCache?: { clear(): void },
+): boolean {
   let changed = false
   const fileNotice = (mediaType?: string) =>
     `[File attachment omitted (${mediaType ?? 'binary'}) — the provider rejected it; removed so the session can continue.]`
@@ -358,6 +361,7 @@ export function stripBinaryPartsFromMessages(messages: ModelMessage[]): boolean
       }
     }
   }
+  if (changed) deliveredFileCache?.clear()
   return changed
 }
 
diff --git a/packages/core/src/agent/vision-fallback.ts b/packages/core/src/agent/vision-fallback.ts
index c0a885b..d1b9908 100644
--- a/packages/core/src/agent/vision-fallback.ts
+++ b/packages/core/src/agent/vision-fallback.ts
@@ -47,6 +47,7 @@ export interface VisionUsageEvent {
 const VISION_MODELS: Record<string, { modelId: string; label: string }> = {
   google: { modelId: 'google:gemini-2.5-flash', label: 'Gemini 2.5 Flash' },
   zhipu: { modelId: 'zhipu:glm-4.6v', label: 'GLM-4.6V' },
+  deepseek: { modelId: 'deepseek:deepseek-flash', label: 'DeepSeek V4.1 Flash' },
   alibaba: { modelId: 'alibaba:qwen3-vl-flash', label: 'Qwen3-VL Flash' },
   openai: { modelId: 'openai:gpt-5.4-mini', label: 'GPT-5.4 Mini' },
   anthropic: { modelId: 'anthropic:claude-haiku-4-5', label: 'Claude Haiku 4.5' },
@@ -59,8 +60,9 @@ const VISION_MODELS: Record<string, { modelId: string; label: string }> = {
  *  last. Gemini 2.5 Flash leads because its free tier is the most
  *  generous (1500/day) and the model is also the strongest at the
  *  free price point. GLM-4.6V is second because it's cheap/free
- *  and reachable from China without a proxy. */
-const VISION_PRIORITY = ['google', 'zhipu', 'alibaba', 'openai', 'anthropic', 'moonshotai', 'xai']
+ *  and reachable from China without a proxy. DeepSeek V4.1 Flash follows as
+ *  the cheapest configured paid fallback with native vision. */
+const VISION_PRIORITY = ['google', 'zhipu', 'deepseek', 'alibaba', 'openai', 'anthropic', 'moonshotai', 'xai']
 
 /**
  * Pick the best available vision sub-agent given the keys the user has
diff --git a/packages/core/src/knowledge/memory/inference.ts b/packages/core/src/knowledge/memory/inference.ts
index 06a4816..4163460 100644
--- a/packages/core/src/knowledge/memory/inference.ts
+++ b/packages/core/src/knowledge/memory/inference.ts
@@ -6,7 +6,7 @@ import { z } from 'zod'
 import type { MemoryReasoningMode } from '../../config/index.js'
 import { providerOf } from '../../providers/capabilities.js'
 import { getOpenAIChatGPTReasoningTiers } from '../../providers/openai-chatgpt-models.js'
-import { getReasoningLevel, getThinkingProviderOptions } from '../../providers/thinking.js'
+import { acceptsReasoningControl, getReasoningLevel, getThinkingProviderOptions } from '../../providers/thinking.js'
 import { debugLog } from '../../utils.js'
 
 export type MemoryReasoningControl = 'off' | 'low' | 'provider-default'
@@ -174,6 +174,7 @@ function inferenceSettings(
   omitTemperature: boolean,
 ): MemoryInferenceSettings {
   if (!modelId || control === 'provider-default') return { maxOutputTokens }
+  if (!acceptsReasoningControl(modelId)) return { maxOutputTokens }
   if (control === 'off') {
     const reasoning = getReasoningLevel(modelId, false)
     const providerOptions = getThinkingProviderOptions(modelId, false)
@@ -186,7 +187,7 @@ function inferenceSettings(
   }
   const provider = providerOf(modelId)
   if (provider === 'alibaba' || provider === 'zhipu') {
-    const providerOptions = getThinkingProviderOptions(modelId, true)
+    const providerOptions = getThinkingProviderOptions(modelId, true, 'low')
     return { maxOutputTokens, ...(Object.keys(providerOptions).length ? { providerOptions } : {}) }
   }
   if (provider === 'custom') return { maxOutputTokens }
@@ -201,9 +202,11 @@ function requiresThinking(modelId: string): boolean {
   if (provider === 'openai') {
     const chatGPTTiers = getOpenAIChatGPTReasoningTiers(modelId)
     if (chatGPTTiers?.length && !chatGPTTiers.some((tier) => tier.value === 'none')) return true
-    return /(?:^|:)(?:gpt-5(?:$|[-.])|o[134](?:$|[-.]))/.test(normalized)
+    return /(?:^|:)(?:gpt-(?:5|6)(?:$|[-.])|o[134](?:$|[-.]))/.test(normalized)
   }
-  return provider === 'anthropic' && /claude-opus-4-5/.test(normalized)
+  if (provider === 'anthropic') return /claude-(?:fable-5|opus-4-5)/.test(normalized)
+  if (provider === 'zhipu') return /glm-5\.3(?:$|-)/.test(normalized)
+  return provider === 'xai' && /grok-(?:4\.6|4\.20)$/.test(normalized)
 }
 
 function rejectControl(modelId: string, control: MemoryReasoningControl): void {
diff --git a/packages/core/src/providers/capabilities.ts b/packages/core/src/providers/capabilities.ts
index b393909..c188544 100644
--- a/packages/core/src/providers/capabilities.ts
+++ b/packages/core/src/providers/capabilities.ts
@@ -37,7 +37,7 @@ const CAPS: Record<string, ProviderCapabilities> = {
   moonshotai: { image: true, pdf: true, audio: false, filesApi: true, toolImageTransport: 'user-message' },
   alibaba: { image: true, pdf: true, audio: false, filesApi: true, toolImageTransport: 'user-message' },
   zhipu: { image: true, pdf: true, audio: false, filesApi: true, toolImageTransport: 'user-message' },
-  deepseek: { image: false, pdf: false, audio: false, filesApi: false, toolImageTransport: 'unsupported' },
+  deepseek: { image: true, pdf: false, audio: false, filesApi: true, toolImageTransport: 'user-message' },
   custom: { image: false, pdf: false, audio: false, filesApi: false, toolImageTransport: 'unsupported' },
 }
 
@@ -141,8 +141,8 @@ export function capabilitiesOf(modelId: string): ProviderCapabilities {
 /** Can this specific MODEL natively see images? Unlike `capabilitiesOf` (which
  *  is provider-level — "does the API accept image parts"), this is per-model,
  *  because providers mix vision and text-only models under one id namespace
- *  (Qwen-VL vs Qwen-Max, GLM-4V vs GLM-5, kimi-k2.6 is multimodal but a plain
- *  DeepSeek is not). Used to gate the browser agent's `--caps vision` so a
+ *  (DeepSeek Flash vs V4 Pro, Qwen-VL vs Qwen-Max, GLM-4V vs GLM-5). Used to
+ *  gate the browser agent's `--caps vision` so a
  *  text-only model never gets screenshots it can't read.
  *
  *  Resolution: alias-expand, look the id up in the curated catalog and trust
diff --git a/packages/core/src/providers/catalog.ts b/packages/core/src/providers/catalog.ts
index 3d8d38f..5ee5c82 100644
--- a/packages/core/src/providers/catalog.ts
+++ b/packages/core/src/providers/catalog.ts
@@ -20,8 +20,8 @@ export interface ProviderModel {
    *  just whether its provider's API accepts image parts. Drives the browser
    *  agent's visual gating (modelSupportsVision) so a text-only model never
    *  gets `--caps vision` / screenshots. Set per-model because providers mix
-   *  vision and text-only models under one id namespace (e.g. Qwen-VL vs
-   *  Qwen-Max, GLM-4V vs GLM-5). */
+   *  vision and text-only models under one id namespace (e.g. DeepSeek Flash
+   *  vs V4 Pro, Qwen-VL vs Qwen-Max, GLM-4V vs GLM-5). */
   vision: boolean
 }
 
@@ -36,6 +36,14 @@ export interface ReasoningTierOption {
   description: string
 }
 
+export interface ReasoningTierProfile {
+  /** Restrict this profile to matching model ids. Omit for a provider-wide profile. */
+  modelPattern?: RegExp
+  options: readonly ReasoningTierOption[]
+  /** Lowest valid effort for models that reject disabling reasoning. */
+  offValue?: string
+}
+
 export interface ProviderInfo {
   /** Provider key used in `<provider>:<model>` ids and config maps. */
   name: string
@@ -48,9 +56,9 @@ export interface ProviderInfo {
   /** Hand-curated models shown in the interactive `/model` picker. Users can
    *  still type any full id into `/model <provider>:<model>` for variants
    *  not listed here. Vision flags reflect model FAMILY: Claude / GPT /
-   *  Gemini / Grok flagships and Kimi K2.x are multimodal; DeepSeek and the
-   *  Qwen-Max / GLM text flagships are text-only; the dedicated *-VL /
-   *  GLM-4V / *-vision-preview models see images. */
+   *  Gemini / Grok flagships, DeepSeek Flash, and Kimi K2.x are multimodal;
+   *  Qwen-Max / GLM text flagships and DeepSeek V4 Pro are text-only; the
+   *  dedicated *-VL / GLM-4V / *-vision-preview models see images. */
   models: readonly ProviderModel[]
   /** Providers that serve multiple endpoints for the same API (regional
    *  platforms, plan-specific gateways, etc.). When a user picks a model
@@ -67,12 +75,12 @@ export interface ProviderInfo {
    *  Providers with no entry here (alibaba) only support the binary
    *  /thinking toggle — skip the tier picker.
    *
-   *  `modelPattern` gates the tier to the models that actually honor it:
+   *  `modelPattern` gates each profile to the models that actually honor it:
    *  within a provider, only some model families expose the granular knob
    *  (e.g. thinkingLevel is Gemini 3-only, Kimi's reasoningEffort is
    *  K3-only). Models that don't match fall back to the binary /thinking
    *  toggle. */
-  reasoningTiers?: { modelPattern?: RegExp; options: readonly ReasoningTierOption[] }
+  reasoningTiers?: readonly ReasoningTierProfile[]
 }
 
 // ─── The table ───
@@ -88,6 +96,18 @@ export const PROVIDERS: readonly ProviderInfo[] = [
     defaultModel: 'anthropic:claude-sonnet-5',
     keyUrl: 'https://console.anthropic.com/',
     models: [
+      {
+        id: 'anthropic:claude-fable-5-1',
+        label: 'Fable 5.1',
+        description: 'Newest adaptive-reasoning flagship, 1M context',
+        vision: true,
+      },
+      {
+        id: 'anthropic:claude-opus-5',
+        label: 'Opus 5',
+        description: 'Latest Opus for complex reasoning and agentic coding, 1M context',
+        vision: true,
+      },
       {
         id: 'anthropic:claude-fable-5',
         label: 'Fable 5',
@@ -113,13 +133,29 @@ export const PROVIDERS: readonly ProviderInfo[] = [
         vision: true,
       },
     ],
-    reasoningTiers: {
-      options: [
-        { label: 'Low', value: 'low', description: 'Minimal reasoning, fastest' },
-        { label: 'Medium', value: 'medium', description: 'Balanced reasoning' },
-        { label: 'High', value: 'high', description: 'Thorough reasoning (default)' },
-      ],
-    },
+    reasoningTiers: [
+      {
+        modelPattern: /claude-fable-5(?:-1)?$/,
+        offValue: 'low',
+        options: [
+          { label: 'Low', value: 'low', description: 'Minimum adaptive reasoning' },
+          { label: 'Medium', value: 'medium', description: 'Balanced reasoning' },
+          { label: 'High', value: 'high', description: 'Thorough reasoning' },
+          { label: 'XHigh', value: 'xhigh', description: 'Very deep reasoning' },
+          { label: 'Max', value: 'max', description: 'Maximum reasoning depth' },
+        ],
+      },
+      {
+        modelPattern: /claude-(?:opus-(?:5|4-8)|sonnet-5)$/,
+        options: [
+          { label: 'Low', value: 'low', description: 'Minimal reasoning, fastest' },
+          { label: 'Medium', value: 'medium', description: 'Balanced reasoning' },
+          { label: 'High', value: 'high', description: 'Thorough reasoning' },
+          { label: 'XHigh', value: 'xhigh', description: 'Very deep reasoning' },
+          { label: 'Max', value: 'max', description: 'Maximum reasoning depth' },
+        ],
+      },
+    ],
   },
   {
     name: 'openai',
@@ -127,22 +163,28 @@ export const PROVIDERS: readonly ProviderInfo[] = [
     defaultModel: 'openai:gpt-5.6-sol',
     keyUrl: 'https://platform.openai.com/api-keys',
     models: [
+      {
+        id: 'openai:gpt-6-astra',
+        label: 'GPT-6 Astra',
+        description: 'Newest flagship, 1.05M context and 128k output',
+        vision: true,
+      },
       {
         id: 'openai:gpt-5.6-sol',
         label: 'GPT-5.6 Sol',
-        description: 'Flagship, top reasoning + coding, $5/$30, 1M context',
+        description: 'Flagship reasoning and coding tier, $4/$20, 1M context',
         vision: true,
       },
       {
         id: 'openai:gpt-5.6-terra',
         label: 'GPT-5.6 Terra',
-        description: 'Balanced tier, $2.50/$15, 1M context',
+        description: 'Balanced tier, $2/$12, 1M context',
         vision: true,
       },
       {
         id: 'openai:gpt-5.6-luna',
         label: 'GPT-5.6 Luna',
-        description: 'Budget tier, $1/$6, 1M context',
+        description: 'Budget tier, $0.20/$1.20, 1M context',
         vision: true,
       },
       {
@@ -158,50 +200,97 @@ export const PROVIDERS: readonly ProviderInfo[] = [
         vision: true,
       },
     ],
-    reasoningTiers: {
-      options: [
-        { label: 'Minimal', value: 'minimal', description: 'Bare-minimum reasoning' },
-        { label: 'Low', value: 'low', description: 'Fast, concise reasoning' },
-        { label: 'Medium', value: 'medium', description: 'Balanced (default)' },
-        { label: 'High', value: 'high', description: 'Thorough reasoning' },
-      ],
-    },
+    reasoningTiers: [
+      {
+        modelPattern: /gpt-6-astra$/,
+        offValue: 'low',
+        options: [
+          { label: 'Low', value: 'low', description: 'Minimum supported reasoning' },
+          { label: 'Medium', value: 'medium', description: 'Balanced reasoning' },
+          { label: 'High', value: 'high', description: 'Thorough reasoning' },
+          { label: 'XHigh', value: 'xhigh', description: 'Very deep reasoning' },
+          { label: 'Max', value: 'max', description: 'Maximum reasoning depth' },
+        ],
+      },
+      {
+        modelPattern: /gpt-5\.6(?:-(?:sol|terra|luna))?$/,
+        options: [
+          { label: 'Low', value: 'low', description: 'Fast, concise reasoning' },
+          { label: 'Medium', value: 'medium', description: 'Balanced reasoning' },
+          { label: 'High', value: 'high', description: 'Thorough reasoning' },
+          { label: 'XHigh', value: 'xhigh', description: 'Very deep reasoning' },
+          { label: 'Max', value: 'max', description: 'Maximum reasoning depth' },
+        ],
+      },
+      {
+        modelPattern: /gpt-5\.5$/,
+        options: [
+          { label: 'Low', value: 'low', description: 'Fast, concise reasoning' },
+          { label: 'Medium', value: 'medium', description: 'Balanced reasoning' },
+          { label: 'High', value: 'high', description: 'Thorough reasoning' },
+          { label: 'XHigh', value: 'xhigh', description: 'Maximum reasoning depth' },
+        ],
+      },
+      {
+        modelPattern: /gpt-5\.4-(?:mini|nano)$/,
+        options: [
+          { label: 'Low', value: 'low', description: 'Fast, concise reasoning' },
+          { label: 'Medium', value: 'medium', description: 'Balanced (default)' },
+          { label: 'High', value: 'high', description: 'Thorough reasoning' },
+          { label: 'XHigh', value: 'xhigh', description: 'Maximum reasoning depth' },
+        ],
+      },
+    ],
   },
   {
     name: 'deepseek',
     envKey: 'DEEPSEEK_API_KEY',
-    defaultModel: 'deepseek:deepseek-v4-flash',
+    defaultModel: 'deepseek:deepseek-flash',
     keyUrl: 'https://platform.deepseek.com/api_keys',
     models: [
       {
-        id: 'deepseek:deepseek-v4-flash',
-        label: 'DeepSeek V4 Flash',
-        description: 'Fast, efficient general-purpose, $0.14/$0.28, 1M context (text-only)',
-        vision: false,
+        id: 'deepseek:deepseek-flash',
+        label: 'DeepSeek V4.1 Flash',
+        description: 'Latest fast, efficient model with native vision and 1M context',
+        vision: true,
       },
       {
         id: 'deepseek:deepseek-v4-pro',
         label: 'DeepSeek V4 Pro',
-        description: 'Flagship, stronger reasoning, $0.44/$0.87, 1M context (text-only)',
+        description: 'Flagship V4 model with 1M context and text-only API surface',
         vision: false,
       },
     ],
-    reasoningTiers: {
-      // V4 Flash and Pro both support low/high/max; medium/xhigh map to high server-side.
-      modelPattern: /deepseek-v4/,
-      options: [
-        { label: 'Low', value: 'low', description: 'Faster, less reasoning' },
-        { label: 'High', value: 'high', description: 'Standard reasoning (default)' },
-        { label: 'Max', value: 'max', description: 'Maximum reasoning depth' },
-      ],
-    },
+    reasoningTiers: [
+      {
+        // Flash and V4 Pro support low/high/max; medium/xhigh map to high server-side.
+        modelPattern: /deepseek-(?:flash|v4)/,
+        options: [
+          { label: 'Low', value: 'low', description: 'Faster, less reasoning' },
+          { label: 'High', value: 'high', description: 'Standard reasoning (default)' },
+          { label: 'Max', value: 'max', description: 'Maximum reasoning depth' },
+        ],
+      },
+    ],
   },
   {
     name: 'alibaba',
     envKey: 'ALIBABA_API_KEY',
-    defaultModel: 'alibaba:qwen3.7-max',
+    defaultModel: 'alibaba:qwen3.8-max',
     keyUrl: 'https://dashscope.console.aliyun.com/apiKey',
     models: [
+      {
+        id: 'alibaba:qwen3.8-max',
+        label: 'Qwen3.8 Max',
+        description: 'Newest flagship with native vision, 1M context',
+        vision: true,
+      },
+      {
+        id: 'alibaba:qwen3.8-flash',
+        label: 'Qwen3.8 Flash',
+        description: 'Fast multimodal model with 1M context',
+        vision: true,
+      },
       {
         id: 'alibaba:qwen3.7-max',
         label: 'Qwen3.7 Max',
@@ -211,8 +300,14 @@ export const PROVIDERS: readonly ProviderInfo[] = [
       {
         id: 'alibaba:qwen3.7-plus',
         label: 'Qwen3.7 Plus',
-        description: 'Mid-tier, balanced cost/quality',
-        vision: false,
+        description: 'Multimodal mid-tier with 1M context',
+        vision: true,
+      },
+      {
+        id: 'alibaba:qwen3.7-flash',
+        label: 'Qwen3.7 Flash',
+        description: 'Fast multimodal model with 1M context',
+        vision: true,
       },
       {
         id: 'alibaba:qwen3-coder-plus',
@@ -243,9 +338,27 @@ export const PROVIDERS: readonly ProviderInfo[] = [
   {
     name: 'google',
     envKey: 'GOOGLE_GENERATIVE_AI_API_KEY',
-    defaultModel: 'google:gemini-3.5-flash',
+    defaultModel: 'google:gemini-3.8-flash',
     keyUrl: 'https://aistudio.google.com/apikey',
     models: [
+      {
+        id: 'google:gemini-3.8-flash',
+        label: 'Gemini 3.8 Flash',
+        description: 'Newest multimodal flagship, 1M context and 64k output',
+        vision: true,
+      },
+      {
+        id: 'google:gemini-3.7-flash',
+        label: 'Gemini 3.7 Flash',
+        description: 'Previous-generation multimodal agentic model, 1M context',
+        vision: true,
+      },
+      {
+        id: 'google:gemini-3.6-flash',
+        label: 'Gemini 3.6 Flash',
+        description: 'Fast multimodal agentic model, 1M context',
+        vision: true,
+      },
       {
         id: 'google:gemini-3.5-flash',
         label: 'Gemini 3.5 Flash',
@@ -265,21 +378,52 @@ export const PROVIDERS: readonly ProviderInfo[] = [
         vision: true,
       },
     ],
-    reasoningTiers: {
-      // thinkingLevel is a Gemini 3 feature; Gemini 2.5 uses thinkingBudget.
-      modelPattern: /gemini-3/,
-      options: [
-        { label: 'Low', value: 'low', description: 'Lower latency, lower cost' },
-        { label: 'High', value: 'high', description: 'Deeper reasoning, higher quality' },
-      ],
-    },
+    reasoningTiers: [
+      {
+        modelPattern: /gemini-3\.[78]-flash$/,
+        offValue: 'low',
+        options: [
+          { label: 'Low', value: 'low', description: 'Minimum supported reasoning' },
+          { label: 'Medium', value: 'medium', description: 'Balanced reasoning' },
+          { label: 'High', value: 'high', description: 'Deeper reasoning, higher quality' },
+        ],
+      },
+      {
+        modelPattern: /gemini-3\.[56]-flash$/,
+        offValue: 'minimal',
+        options: [
+          { label: 'Minimal', value: 'minimal', description: 'Minimum supported reasoning' },
+          { label: 'Low', value: 'low', description: 'Lower latency, lower cost' },
+          { label: 'Medium', value: 'medium', description: 'Balanced reasoning' },
+          { label: 'High', value: 'high', description: 'Deeper reasoning, higher quality' },
+        ],
+      },
+    ],
   },
   {
     name: 'xai',
     envKey: 'XAI_API_KEY',
-    defaultModel: 'xai:grok-4.5',
+    defaultModel: 'xai:grok-4.6',
     keyUrl: 'https://console.x.ai/',
     models: [
+      {
+        id: 'xai:grok-4.6',
+        label: 'Grok 4.6',
+        description: 'Recommended flagship, native vision and 500k context',
+        vision: true,
+      },
+      {
+        id: 'xai:grok-4.20',
+        label: 'Grok 4.20',
+        description: 'Fixed-reasoning model with native vision and 1M context',
+        vision: true,
+      },
+      {
+        id: 'xai:grok-4.20-non-reasoning',
+        label: 'Grok 4.20 Non-Reasoning',
+        description: 'Low-latency non-reasoning variant with 1M context',
+        vision: true,
+      },
       {
         id: 'xai:grok-4.5',
         label: 'Grok 4.5',
@@ -293,19 +437,55 @@ export const PROVIDERS: readonly ProviderInfo[] = [
         vision: true,
       },
     ],
-    reasoningTiers: {
-      options: [
-        { label: 'Low', value: 'low', description: 'Faster, cheaper responses' },
-        { label: 'High', value: 'high', description: 'Deeper reasoning' },
-      ],
-    },
+    reasoningTiers: [
+      {
+        modelPattern: /grok-4\.6$/,
+        offValue: 'low',
+        options: [
+          { label: 'Low', value: 'low', description: 'Minimum supported reasoning' },
+          { label: 'Medium', value: 'medium', description: 'Balanced reasoning' },
+          { label: 'High', value: 'high', description: 'Deeper reasoning' },
+          { label: 'XHigh', value: 'xhigh', description: 'Maximum supported reasoning' },
+        ],
+      },
+      {
+        modelPattern: /grok-4\.5$/,
+        offValue: 'low',
+        options: [
+          { label: 'Low', value: 'low', description: 'Minimum supported reasoning' },
+          { label: 'Medium', value: 'medium', description: 'Balanced reasoning' },
+          { label: 'High', value: 'high', description: 'Maximum reasoning depth' },
+        ],
+      },
+      {
+        modelPattern: /grok-4\.3$/,
+        options: [
+          { label: 'Low', value: 'low', description: 'Faster, cheaper responses' },
+          { label: 'Medium', value: 'medium', description: 'Balanced reasoning' },
+          { label: 'High', value: 'high', description: 'Deeper reasoning' },
+          { label: 'XHigh', value: 'xhigh', description: 'Maximum supported reasoning' },
+        ],
+      },
+    ],
   },
   {
     name: 'zhipu',
     envKey: 'ZHIPU_API_KEY',
-    defaultModel: 'zhipu:glm-5.2',
+    defaultModel: 'zhipu:glm-5.3',
     keyUrl: 'https://open.bigmodel.cn/usercenter/apikeys',
     models: [
+      {
+        id: 'zhipu:glm-5.3',
+        label: 'GLM-5.3',
+        description: 'Newest reasoning flagship with 1M context',
+        vision: false,
+      },
+      {
+        id: 'zhipu:glm-5.3-flash',
+        label: 'GLM-5.3 Flash',
+        description: 'Fast native-multimodal model with 1M context',
+        vision: true,
+      },
       {
         id: 'zhipu:glm-5.2',
         label: 'GLM-5.2',
@@ -337,14 +517,24 @@ export const PROVIDERS: readonly ProviderInfo[] = [
         vision: true,
       },
     ],
-    reasoningTiers: {
-      // reasoning_effort is GLM-5.2+; earlier models use the binary thinking switch.
-      modelPattern: /glm-5\.2/,
-      options: [
-        { label: 'High', value: 'high', description: 'Enhanced reasoning' },
-        { label: 'Max', value: 'max', description: 'Deep reasoning (default)' },
-      ],
-    },
+    reasoningTiers: [
+      {
+        modelPattern: /glm-5\.3(?:-flash)?$/,
+        offValue: 'low',
+        options: [
+          { label: 'Low', value: 'low', description: 'Minimum supported reasoning' },
+          { label: 'High', value: 'high', description: 'Enhanced reasoning' },
+          { label: 'Max', value: 'max', description: 'Deep reasoning (default)' },
+        ],
+      },
+      {
+        modelPattern: /glm-5\.2$/,
+        options: [
+          { label: 'High', value: 'high', description: 'Enhanced reasoning' },
+          { label: 'Max', value: 'max', description: 'Deep reasoning (default)' },
+        ],
+      },
+    ],
   },
   {
     name: 'moonshotai',
@@ -364,6 +554,12 @@ export const PROVIDERS: readonly ProviderInfo[] = [
         description: 'Dedicated coding model, 256k context',
         vision: true,
       },
+      {
+        id: 'moonshotai:kimi-k2.7-code-highspeed',
+        label: 'Kimi K2.7 Code Highspeed',
+        description: 'High-speed coding variant, 256k context',
+        vision: true,
+      },
       {
         id: 'moonshotai:kimi-k2.6',
         label: 'Kimi K2.6',
@@ -376,31 +572,36 @@ export const PROVIDERS: readonly ProviderInfo[] = [
       { label: 'api.moonshot.cn (China)', url: 'https://api.moonshot.cn/v1' },
       { label: 'api.moonshot.ai (International)', url: 'https://api.moonshot.ai/v1' },
     ],
-    reasoningTiers: {
-      // reasoning_effort is K3-only; K2.x uses the binary thinking switch.
-      modelPattern: /kimi-k3/,
-      options: [
-        { label: 'Low', value: 'low', description: 'Faster, concise reasoning' },
-        { label: 'High', value: 'high', description: 'Deeper reasoning' },
-        { label: 'Max', value: 'max', description: 'Maximum reasoning (default)' },
-      ],
-    },
+    reasoningTiers: [
+      {
+        // reasoning_effort is K3-only; K2.x uses the binary thinking switch.
+        modelPattern: /kimi-k3/,
+        options: [
+          { label: 'Low', value: 'low', description: 'Faster, concise reasoning' },
+          { label: 'High', value: 'high', description: 'Deeper reasoning' },
+          { label: 'Max', value: 'max', description: 'Maximum reasoning (default)' },
+        ],
+      },
+    ],
   },
 ]
 
 // ─── Model aliases ───
 
 export const MODEL_ALIASES: Record<string, string> = {
-  fable: 'anthropic:claude-fable-5',
+  fable: 'anthropic:claude-fable-5-1',
   sonnet: 'anthropic:claude-sonnet-5',
-  opus: 'anthropic:claude-opus-4-8',
+  opus: 'anthropic:claude-opus-5',
   haiku: 'anthropic:claude-haiku-4-5',
+  gpt6: 'openai:gpt-6-astra',
+  astra: 'openai:gpt-6-astra',
   gpt5: 'openai:gpt-5.6-sol',
-  gemini: 'google:gemini-3.5-flash',
-  deepseek: 'deepseek:deepseek-v4-flash',
+  gemini: 'google:gemini-3.8-flash',
+  deepseek: 'deepseek:deepseek-flash',
   'deepseek-pro': 'deepseek:deepseek-v4-pro',
-  qwen: 'alibaba:qwen3.7-max',
-  glm: 'zhipu:glm-5.2',
+  qwen: 'alibaba:qwen3.8-max',
+  grok: 'xai:grok-4.6',
+  glm: 'zhipu:glm-5.3',
   kimi: 'moonshotai:kimi-k3',
 }
 
@@ -425,7 +626,24 @@ export const PROVIDER_BASE_URLS: Record<string, { options: readonly ProviderBase
   PROVIDERS.filter((p) => p.baseUrlOptions).map((p) => [p.name, { options: p.baseUrlOptions! }]),
 )
 
+export const PROVIDER_REASONING_PROFILES: Record<string, readonly ReasoningTierProfile[]> = Object.fromEntries(
+  PROVIDERS.filter((p) => p.reasoningTiers).map((p) => [p.name, p.reasoningTiers!]),
+)
+
+/** Legacy provider-wide view retained for public API compatibility. Internal
+ * model selection uses PROVIDER_REASONING_PROFILES so each family gets only
+ * the effort values it actually supports. */
 export const PROVIDER_REASONING_TIERS: Record<
   string,
   { modelPattern?: RegExp; options: readonly ReasoningTierOption[] }
-> = Object.fromEntries(PROVIDERS.filter((p) => p.reasoningTiers).map((p) => [p.name, p.reasoningTiers!]))
+> = Object.fromEntries(
+  Object.entries(PROVIDER_REASONING_PROFILES).map(([name, profiles]) => {
+    const options = Array.from(
+      new Map(profiles.flatMap((profile) => profile.options).map((option) => [option.value, option])).values(),
+    )
+    const patterns = profiles.map((profile) => profile.modelPattern).filter((pattern) => pattern !== undefined)
+    const modelPattern =
+      patterns.length === profiles.length ? new RegExp(patterns.map((pattern) => pattern.source).join('|')) : undefined
+    return [name, { ...(modelPattern ? { modelPattern } : {}), options }]
+  }),
+)
diff --git a/packages/core/src/providers/registry.ts b/packages/core/src/providers/registry.ts
index fb2bc0f..21a2210 100644
--- a/packages/core/src/providers/registry.ts
+++ b/packages/core/src/providers/registry.ts
@@ -18,6 +18,7 @@ import { createOpenAIChatGPTFetch } from './openai-chatgpt-fetch.js'
 
 const KIMI_CODING_MODEL_IDS = {
   'kimi-k3': 'k3',
+  // The stable Coding Plan id currently auto-routes to K2.8 Preview.
   'kimi-k2.7-code': 'kimi-for-coding',
   'kimi-k2.7-code-highspeed': 'kimi-for-coding-highspeed',
   // Coding Plan exposes K2.6 by disabling thinking on kimi-for-coding rather
@@ -194,6 +195,7 @@ const moonshotConvertUsage = (usage: any) => {
  */
 let _zhipuReasoningEffort: string | undefined
 const ZHIPU_REASONING_HEADER = 'x-x-code-zhipu-reasoning-effort'
+const XAI_REASONING_HEADER = 'x-x-code-xai-reasoning-effort'
 
 export function setZhipuReasoningEffort(effort: string | undefined): void {
   _zhipuReasoningEffort = effort
@@ -202,8 +204,28 @@ export function setZhipuReasoningEffort(effort: string | undefined): void {
 export function withZhipuReasoningHeader(
   headers: Record<string, string | undefined> | undefined,
   effort: string | undefined,
-): Record<string, string | undefined> {
-  return { ...headers, [ZHIPU_REASONING_HEADER]: effort ?? '' }
+): Record<string, string> {
+  return withInternalReasoningHeader(headers, ZHIPU_REASONING_HEADER, effort)
+}
+
+export function withXaiReasoningHeader(
+  headers: Record<string, string | undefined> | undefined,
+  effort: string | undefined,
+): Record<string, string> {
+  return withInternalReasoningHeader(headers, XAI_REASONING_HEADER, effort)
+}
+
+function withInternalReasoningHeader(
+  headers: Record<string, string | undefined> | undefined,
+  name: string,
+  effort: string | undefined,
+): Record<string, string> {
+  const result: Record<string, string> = {}
+  for (const [key, value] of Object.entries(headers ?? {})) {
+    if (value !== undefined) result[key] = value
+  }
+  result[name] = effort ?? ''
+  return result
 }
 
 /** Move x-code's per-session affinity hint to the wire location required by
@@ -212,16 +234,18 @@ export function withZhipuReasoningHeader(
 const xaiPromptCacheFetch: typeof fetch = async (input, init) => {
   const headers = new Headers(init?.headers ?? (input instanceof Request ? input.headers : undefined))
   const promptCacheKey = headers.get(XAI_PROMPT_CACHE_KEY_HEADER)
+  const reasoningEffort = headers.get(XAI_REASONING_HEADER) || undefined
   headers.delete(XAI_PROMPT_CACHE_KEY_HEADER)
+  headers.delete(XAI_REASONING_HEADER)
   const sanitizedInit = { ...init, headers }
-  if (!promptCacheKey) return permanentErrorFetch(input, sanitizedInit)
+  if (!promptCacheKey && !reasoningEffort) return permanentErrorFetch(input, sanitizedInit)
 
   const url = input instanceof URL ? input : new URL(typeof input === 'string' ? input : input.url)
   if (url.pathname.endsWith('/chat/completions')) {
-    headers.set('x-grok-conv-id', promptCacheKey)
+    if (promptCacheKey) headers.set('x-grok-conv-id', promptCacheKey)
+  } else if (!url.pathname.endsWith('/responses')) {
     return permanentErrorFetch(input, sanitizedInit)
   }
-  if (!url.pathname.endsWith('/responses')) return permanentErrorFetch(input, sanitizedInit)
 
   const rawBody = init?.body ?? (input instanceof Request ? await input.clone().text() : undefined)
   if (typeof rawBody !== 'string') return permanentErrorFetch(input, sanitizedInit)
@@ -231,7 +255,18 @@ const xaiPromptCacheFetch: typeof fetch = async (input, init) => {
       return permanentErrorFetch(input, sanitizedInit)
     }
     const body = parsed as Record<string, unknown>
-    if (!body.prompt_cache_key) body.prompt_cache_key = promptCacheKey
+    if (url.pathname.endsWith('/chat/completions')) {
+      if (reasoningEffort) body.reasoning_effort = reasoningEffort
+    } else {
+      if (promptCacheKey && !body.prompt_cache_key) body.prompt_cache_key = promptCacheKey
+      if (reasoningEffort) {
+        const current =
+          body.reasoning && typeof body.reasoning === 'object' && !Array.isArray(body.reasoning)
+            ? (body.reasoning as Record<string, unknown>)
+            : {}
+        body.reasoning = { ...current, effort: reasoningEffort }
+      }
+    }
     return permanentErrorFetch(input, { ...sanitizedInit, body: JSON.stringify(body) })
   } catch {
     return permanentErrorFetch(input, sanitizedInit)
diff --git a/packages/core/src/providers/thinking.ts b/packages/core/src/providers/thinking.ts
index ce56a3d..35d7740 100644
--- a/packages/core/src/providers/thinking.ts
+++ b/packages/core/src/providers/thinking.ts
@@ -5,9 +5,13 @@
 // We use it as the primary mechanism for reasoning control.
 //
 // Exceptions that still need providerOptions / fetch shim injection:
+//   - deepseek: explicit tiers use native reasoningEffort so `max` does not
+//     round-trip through portable `xhigh` and emit a compatibility warning.
 //   - zhipu: goes through @ai-sdk/openai-compatible, SDK doesn't auto-translate
 //     `reasoning` for it. We inject `reasoning_effort` via fetch shim.
 //   - alibaba: uses `enableThinking` in providerOptions (no top-level support).
+//   - native `max` controls for newer Anthropic/OpenAI models also travel via
+//     providerOptions because the portable reasoning enum tops out at xhigh.
 //
 // The user-facing controls:
 //   /thinking on|off — binary toggle (maps to 'high' / 'none')
@@ -17,8 +21,8 @@
 // over the `enabled` flag. The /thinking toggle is only used as a fallback
 // for models without an explicit tier.
 import { providerOf } from './capabilities.js'
-import { PROVIDER_REASONING_TIERS } from './catalog.js'
-import type { ReasoningTierOption } from './catalog.js'
+import { PROVIDER_REASONING_PROFILES } from './catalog.js'
+import type { ReasoningTierOption, ReasoningTierProfile } from './catalog.js'
 import { getOpenAIChatGPTReasoningTiers, getOpenAIChatGPTRuntimeModel } from './openai-chatgpt-models.js'
 
 /** Whether the model exposes a granular reasoning-effort tier (vs. the
@@ -26,20 +30,30 @@ import { getOpenAIChatGPTReasoningTiers, getOpenAIChatGPTRuntimeModel } from './
  *  model families honor them — modelPattern in PROVIDER_REASONING_TIERS
  *  gates that. Drives both the /model tier picker and the effort branch
  *  in getReasoningLevel. */
+function getReasoningTierProfile(modelId: string): ReasoningTierProfile | undefined {
+  return PROVIDER_REASONING_PROFILES[providerOf(modelId)]?.find(
+    (profile) => !profile.modelPattern || profile.modelPattern.test(modelId),
+  )
+}
+
+/** Some fixed-reasoning xAI models reject the effort field entirely. The
+ * public grok-4.20 alias is not recognized by the SDK's equivalent guard. */
+export function acceptsReasoningControl(modelId: string): boolean {
+  if (providerOf(modelId) !== 'xai') return true
+  const providerModelId = modelId.slice(modelId.indexOf(':') + 1)
+  return !/^grok-4\.20(?!-multi-agent(?:$|-))/.test(providerModelId)
+}
+
 export function supportsReasoningTier(modelId: string): boolean {
   const chatGPTTiers = getOpenAIChatGPTReasoningTiers(modelId)
   if (chatGPTTiers !== undefined) return chatGPTTiers.length > 0
-  const config = PROVIDER_REASONING_TIERS[providerOf(modelId)]
-  if (!config) return false
-  return !config.modelPattern || config.modelPattern.test(modelId)
+  return getReasoningTierProfile(modelId) !== undefined
 }
 
 export function getReasoningTierOptions(modelId: string): readonly ReasoningTierOption[] | undefined {
   const chatGPTTiers = getOpenAIChatGPTReasoningTiers(modelId)
   if (chatGPTTiers !== undefined) return chatGPTTiers
-  const config = PROVIDER_REASONING_TIERS[providerOf(modelId)]
-  if (!config || (config.modelPattern && !config.modelPattern.test(modelId))) return undefined
-  return config.options
+  return getReasoningTierProfile(modelId)?.options
 }
 
 export type ReasoningLevel = 'provider-default' | 'none' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh'
@@ -55,6 +69,15 @@ const TIER_TO_REASONING: Record<string, ReasoningLevel> = {
   max: 'xhigh',
 }
 
+/** Resolve the provider-native effort after applying model support and the
+ * minimum valid effort for models that cannot turn reasoning off. */
+export function getReasoningEffort(modelId: string, enabled: boolean, effort?: string): string | undefined {
+  const profile = getReasoningTierProfile(modelId)
+  if (!profile) return undefined
+  if (effort && profile.options.some((option) => option.value === effort)) return effort
+  return !enabled ? profile.offValue : undefined
+}
+
 /**
  * Compute the top-level `reasoning` value for streamText/generateText.
  * Returns undefined when the model doesn't support reasoning control
@@ -69,7 +92,7 @@ export function getReasoningLevel(modelId: string, enabled: boolean, effort?: st
   const chatGPTTiers = getOpenAIChatGPTReasoningTiers(modelId)
 
   // These providers use separate mechanisms (providerOptions / fetch shim)
-  if (provider === 'alibaba' || provider === 'zhipu' || provider === 'custom') {
+  if (provider === 'alibaba' || provider === 'zhipu' || provider === 'custom' || !acceptsReasoningControl(modelId)) {
     return undefined
   }
 
@@ -83,20 +106,22 @@ export function getReasoningLevel(modelId: string, enabled: boolean, effort?: st
     return chatGPTTiers[Math.floor((chatGPTTiers.length - 1) / 2)]?.value as ReasoningLevel
   }
 
-  // Tiered reasoning — user picked an explicit effort level AND the model
-  // honors it.
-  if (effort && supportsReasoningTier(modelId)) {
-    return TIER_TO_REASONING[effort] ?? (effort as ReasoningLevel)
+  const effectiveEffort = getReasoningEffort(modelId, enabled, effort)
+  // DeepSeek exposes provider-native low/high/max values. Sending the
+  // portable max equivalent (`xhigh`) makes the SDK map it back to `max`
+  // and emit a compatibility warning, so explicit tiers travel only through
+  // providerOptions.deepseek.reasoningEffort.
+  if (provider === 'deepseek' && effectiveEffort) return undefined
+  if (effectiveEffort) {
+    return TIER_TO_REASONING[effectiveEffort] ?? (effectiveEffort as ReasoningLevel)
   }
 
   return enabled ? 'high' : 'none'
 }
 
 /**
- * Build providerOptions for providers that can't use the top-level
- * `reasoning` parameter: Alibaba (enableThinking) and Zhipu (thinking toggle).
- *
- * Returns an empty object for providers that use top-level `reasoning`.
+ * Build provider-native reasoning options where the portable top-level
+ * `reasoning` value cannot express the exact provider control.
  */
 export function getThinkingProviderOptions(
   modelId: string,
@@ -106,17 +131,32 @@ export function getThinkingProviderOptions(
   const provider = providerOf(modelId)
 
   switch (provider) {
+    case 'deepseek':
+      const deepseekEffort = getReasoningEffort(modelId, enabled, effort)
+      return deepseekEffort ? { deepseek: { reasoningEffort: deepseekEffort } } : {}
+
     case 'alibaba':
       return { alibaba: { enableThinking: enabled } }
 
     case 'zhipu':
       // Binary toggle via providerOptions for models that don't use tiers.
       // Tiered models get reasoning_effort injected by the fetch shim.
-      if (effort && supportsReasoningTier(modelId)) {
-        return { zhipu: { thinking: { type: 'enabled' } } }
+      const zhipuEffort = getReasoningEffort(modelId, enabled, effort)
+      if (zhipuEffort) {
+        return { zhipu: { thinking: { type: 'enabled' }, reasoningEffort: zhipuEffort } }
       }
       return enabled ? { zhipu: { thinking: { type: 'enabled' } } } : { zhipu: { thinking: { type: 'disabled' } } }
 
+    case 'anthropic':
+      return effort === 'max' && getReasoningTierOptions(modelId)?.some((option) => option.value === 'max')
+        ? { anthropic: { thinking: { type: 'adaptive' }, effort: 'max' } }
+        : {}
+
+    case 'openai':
+      return effort === 'max' && getReasoningTierOptions(modelId)?.some((option) => option.value === 'max')
+        ? { openai: { reasoningEffort: 'max' } }
+        : {}
+
     default:
       return {}
   }
diff --git a/packages/core/src/tools/read-file.ts b/packages/core/src/tools/read-file.ts
index aa80ea1..f7b5370 100644
--- a/packages/core/src/tools/read-file.ts
+++ b/packages/core/src/tools/read-file.ts
@@ -247,7 +247,8 @@ async function checkReadCache(
       stub:
         `[readFile: ${filePath} is unchanged since its full content was added to this conversation ` +
         `(same mtime and size); its full content is already in the conversation above. ` +
-        `Re-read with an explicit offset/limit to revisit a specific range, or use grep to search within it.]`,
+        `Analyze that content directly instead of reading it again. For text files, use an explicit offset/limit ` +
+        `only to revisit a specific range, or use grep to search within it.]`,
     }
   }
   return { hit: false, entry: { mtimeMs: stat.mtimeMs, size: stat.size } }
@@ -337,6 +338,8 @@ Usage:
         }
 
         if (kind === 'image') {
+          const verdict = await checkReadCache(cache, filePath, false)
+          if (verdict && verdict.hit) return verdict.stub
           const stats = await fs.stat(filePath)
           if (stats.size > MAX_IMAGE_SOURCE_BYTES) {
             return `[Image ${filePath} is too large to process safely (${(stats.size / (1024 * 1024)).toFixed(1)} MB, cap ${MAX_IMAGE_SOURCE_BYTES / (1024 * 1024)} MB).]`
@@ -380,6 +383,7 @@ Usage:
           const header = compressed.changed
             ? `Loaded image: ${filePath} (compressed from ${buffer.length} to ${compressed.data.length} bytes)`
             : `Loaded image: ${filePath}`
+          if (verdict && !verdict.hit) cache?.set(filePath, verdict.entry)
           return {
             type: 'content',
             value: [
@@ -390,6 +394,7 @@ Usage:
                 mediaType: finalMime,
                 filename: path.basename(filePath),
               },
+              { type: 'text', text: BUILT_IN_MEDIA_ANALYSIS_NOTE },
             ],
           }
         }
diff --git a/packages/core/src/tools/web-search.ts b/packages/core/src/tools/web-search.ts
index ecad6f7..6cd2251 100644
--- a/packages/core/src/tools/web-search.ts
+++ b/packages/core/src/tools/web-search.ts
@@ -236,7 +236,7 @@ async function searchWithDeepseek(query: string, maxResults: number, signal?: Ab
     body: JSON.stringify({
       // A full (cheap) model turn powers each search; override if DeepSeek
       // retires the id or a cheaper search-capable model appears.
-      model: process.env.DEEPSEEK_SEARCH_MODEL || 'deepseek-v4-flash',
+      model: process.env.DEEPSEEK_SEARCH_MODEL || 'deepseek-flash',
       max_tokens: 4096,
       messages: [{ role: 'user', content: [{ type: 'text', text: `Perform a web search for the query: ${query}` }] }],
       // Each use is billed server-side, so don't grant more searches than the
diff --git a/packages/core/tests/agent-loop.test.ts b/packages/core/tests/agent-loop.test.ts
index 248cf3b..cd8d303 100644
--- a/packages/core/tests/agent-loop.test.ts
+++ b/packages/core/tests/agent-loop.test.ts
@@ -999,7 +999,7 @@ describe('agent loop', () => {
     await agentLoop(
       'Try this task',
       {} as any,
-      { modelId: 'deepseek:deepseek-v4-flash', trustMode: false, maxTurns: 10, printMode: false },
+      { modelId: 'deepseek:deepseek-flash', trustMode: false, maxTurns: 10, printMode: false },
       mockCallbacks,
     )
 
@@ -1025,7 +1025,7 @@ describe('agent loop', () => {
     const { state } = await agentLoop(
       'Give a complete answer',
       {} as any,
-      { modelId: 'deepseek:deepseek-v4-flash', trustMode: false, maxTurns: 10, printMode: false },
+      { modelId: 'deepseek:deepseek-flash', trustMode: false, maxTurns: 10, printMode: false },
       mockCallbacks,
     )
 
diff --git a/packages/core/tests/capabilities.test.ts b/packages/core/tests/capabilities.test.ts
index ffe0e17..c109452 100644
--- a/packages/core/tests/capabilities.test.ts
+++ b/packages/core/tests/capabilities.test.ts
@@ -14,7 +14,18 @@ describe('modelSupportsVision', () => {
     expect(modelSupportsVision('anthropic:claude-haiku-4-5')).toBe(true)
     expect(modelSupportsVision('moonshotai:kimi-k3')).toBe(true)
     expect(modelSupportsVision('moonshotai:kimi-k2.6')).toBe(true)
+    expect(modelSupportsVision('deepseek:deepseek-flash')).toBe(true)
+    expect(modelSupportsVision('deepseek:deepseek-v4-flash')).toBe(true)
+    expect(modelSupportsVision('alibaba:qwen3.8-max')).toBe(true)
+    expect(modelSupportsVision('alibaba:qwen3.8-flash')).toBe(true)
+    expect(modelSupportsVision('alibaba:qwen3.7-plus')).toBe(true)
+    expect(modelSupportsVision('alibaba:qwen3.7-flash')).toBe(true)
     expect(modelSupportsVision('alibaba:qwen3-vl-flash')).toBe(true)
+    expect(modelSupportsVision('google:gemini-3.8-flash')).toBe(true)
+    expect(modelSupportsVision('google:gemini-3.7-flash')).toBe(true)
+    expect(modelSupportsVision('xai:grok-4.6')).toBe(true)
+    expect(modelSupportsVision('zhipu:glm-5.3-flash')).toBe(true)
+    expect(modelSupportsVision('moonshotai:kimi-k2.7-code-highspeed')).toBe(true)
     expect(modelSupportsVision('zhipu:glm-4.6v')).toBe(true)
   })
 
@@ -23,18 +34,18 @@ describe('modelSupportsVision', () => {
     // specific models are text-only — the per-model flag must win.
     expect(modelSupportsVision('alibaba:qwen3.7-max')).toBe(false)
     expect(modelSupportsVision('zhipu:glm-5.2')).toBe(false)
-    expect(modelSupportsVision('deepseek:deepseek-v4-flash')).toBe(false)
+    expect(modelSupportsVision('deepseek:deepseek-v4-pro')).toBe(false)
   })
 
   it('expands aliases before lookup', () => {
-    expect(modelSupportsVision('opus')).toBe(true) // → anthropic:claude-opus-4-8
-    expect(modelSupportsVision('deepseek')).toBe(false) // → deepseek:deepseek-v4-flash
+    expect(modelSupportsVision('opus')).toBe(true) // → anthropic:claude-opus-5
+    expect(modelSupportsVision('deepseek')).toBe(true) // → deepseek:deepseek-flash
   })
 
   it('falls back to provider-level capability for unlisted ids', () => {
     // Not in the catalog → defer to the provider's image capability.
     expect(modelSupportsVision('anthropic:claude-some-future-model')).toBe(true)
-    expect(modelSupportsVision('deepseek:some-future-model')).toBe(false)
+    expect(modelSupportsVision('deepseek:some-future-model')).toBe(true)
     expect(modelSupportsVision('unknownprovider:whatever')).toBe(false)
   })
 
@@ -71,13 +82,19 @@ describe('toolImageTransport capability', () => {
   })
 
   it('reattaches media in a following user message for Chat Completions providers', () => {
-    for (const id of ['moonshotai:kimi-k2.6', 'alibaba:qwen3-vl-flash', 'zhipu:glm-4.6v', 'xai:grok-4.3']) {
+    for (const id of [
+      'deepseek:deepseek-flash',
+      'moonshotai:kimi-k2.6',
+      'alibaba:qwen3-vl-flash',
+      'zhipu:glm-4.6v',
+      'xai:grok-4.3',
+    ]) {
       expect(capabilitiesOf(id).toolImageTransport, id).toBe('user-message')
     }
   })
 
   it('marks text-only and unknown providers unsupported', () => {
-    for (const id of ['deepseek:deepseek-v4', 'custom:whatever', 'unknownprovider:whatever']) {
+    for (const id of ['custom:whatever', 'unknownprovider:whatever']) {
       expect(capabilitiesOf(id).toolImageTransport, id).toBe('unsupported')
     }
   })
diff --git a/packages/core/tests/compression.test.ts b/packages/core/tests/compression.test.ts
index 62b8420..d9f5403 100644
--- a/packages/core/tests/compression.test.ts
+++ b/packages/core/tests/compression.test.ts
@@ -269,11 +269,46 @@ describe('checkAndCompressContext', () => {
     const state = createLoopState()
     state.messages = padMessages(10)
     state.lastInputTokens = 999_999
+    state.readFileCache.set('/tmp/previously-read.png', { mtimeMs: 1, size: 1 })
 
     await checkAndCompressContext(state, fakeModel, 1, makeCallbacks())
 
     expect(state.lastInputTokens).toBe(0)
     expect(state.expectCacheMiss).toBe(true)
+    expect(state.readFileCache.size).toBe(0)
+  })
+
+  it('clears delivered-file de-dup state after lightweight compression rewrites history', async () => {
+    const state = createLoopState()
+    state.messages = [
+      { role: 'user', content: 'start' },
+      {
+        role: 'assistant',
+        content: [{ type: 'tool-call', toolCallId: 'looped', toolName: 'readFile', input: { filePath: '/tmp/a' } }],
+      },
+      {
+        role: 'tool',
+        content: [
+          {
+            type: 'tool-result',
+            toolCallId: 'looped',
+            toolName: 'readFile',
+            output: { type: 'text', value: `[loop-guard] ${'x'.repeat(10_000)}` },
+          },
+        ],
+      },
+      { role: 'user', content: 'one' },
+      { role: 'assistant', content: 'two' },
+      { role: 'user', content: 'three' },
+      { role: 'assistant', content: 'four' },
+    ] as ModelMessage[]
+    state.lastInputTokens = 999_999
+    state.readFileCache.set('/tmp/previously-read.png', { mtimeMs: 1, size: 1 })
+
+    await checkAndCompressContext(state, fakeModel, 100, makeCallbacks())
+
+    expect(generateText).not.toHaveBeenCalled()
+    expect(state.readFileCache.size).toBe(0)
   })
 
   it('does not return until the compact boundary and recall-window reset are durable', async () => {
diff --git a/packages/core/tests/config.test.ts b/packages/core/tests/config.test.ts
index 784e134..d4aa087 100644
--- a/packages/core/tests/config.test.ts
+++ b/packages/core/tests/config.test.ts
@@ -101,8 +101,15 @@ describe('resolveModelId', () => {
 
   it('resolves alias from CLI argument', () => {
     expect(resolveModelId('sonnet')).toBe('anthropic:claude-sonnet-5')
-    expect(resolveModelId('opus')).toBe('anthropic:claude-opus-4-8')
-    expect(resolveModelId('deepseek')).toBe('deepseek:deepseek-v4-flash')
+    expect(resolveModelId('fable')).toBe('anthropic:claude-fable-5-1')
+    expect(resolveModelId('opus')).toBe('anthropic:claude-opus-5')
+    expect(resolveModelId('astra')).toBe('openai:gpt-6-astra')
+    expect(resolveModelId('gpt6')).toBe('openai:gpt-6-astra')
+    expect(resolveModelId('gemini')).toBe('google:gemini-3.8-flash')
+    expect(resolveModelId('qwen')).toBe('alibaba:qwen3.8-max')
+    expect(resolveModelId('grok')).toBe('xai:grok-4.6')
+    expect(resolveModelId('glm')).toBe('zhipu:glm-5.3')
+    expect(resolveModelId('deepseek')).toBe('deepseek:deepseek-flash')
   })
 
   it('falls back to env var X_CODE_MODEL', () => {
@@ -130,12 +137,26 @@ describe('resolveModelId', () => {
     expect(resolveModelId()).toBe('openai:gpt-5.6-sol')
   })
 
+  it('uses the updated provider smart defaults', () => {
+    const cases: Array<[string, string]> = [
+      ['GOOGLE_GENERATIVE_AI_API_KEY', 'google:gemini-3.8-flash'],
+      ['ALIBABA_API_KEY', 'alibaba:qwen3.8-max'],
+      ['XAI_API_KEY', 'xai:grok-4.6'],
+      ['ZHIPU_API_KEY', 'zhipu:glm-5.3'],
+    ]
+    for (const [envKey, expected] of cases) {
+      process.env[envKey] = 'test-key'
+      expect(resolveModelId()).toBe(expected)
+      delete process.env[envKey]
+    }
+  })
+
   it('returns null when no providers configured', () => {
     expect(resolveModelId()).toBeNull()
   })
 
   it('returns model even if provider key missing when explicitly requested', () => {
-    expect(resolveModelId('deepseek')).toBe('deepseek:deepseek-v4-flash')
+    expect(resolveModelId('deepseek')).toBe('deepseek:deepseek-flash')
   })
 
   it('uses ChatGPT authentication as the OpenAI smart default without an API key', async () => {
diff --git a/packages/core/tests/context-usage.test.ts b/packages/core/tests/context-usage.test.ts
index 9e81a54..7e4b621 100644
--- a/packages/core/tests/context-usage.test.ts
+++ b/packages/core/tests/context-usage.test.ts
@@ -235,7 +235,7 @@ describe('calibrateContextBreakdown', () => {
 describe('buildContextBreakdownInput', () => {
   it('returns null before the system prompt has been built', () => {
     const state = createLoopState()
-    expect(buildContextBreakdownInput({ modelId: 'deepseek:deepseek-v4-flash' } as any, state)).toBeNull()
+    expect(buildContextBreakdownInput({ modelId: 'deepseek:deepseek-flash' } as any, state)).toBeNull()
   })
 
   it('derives the deferred-tools block from the catalog', () => {
@@ -252,7 +252,7 @@ describe('buildContextBreakdownInput', () => {
         def: makeTool('Search the web'),
       },
     ]
-    const options = { modelId: 'deepseek:deepseek-v4-flash' } as any
+    const options = { modelId: 'deepseek:deepseek-flash' } as any
     const input = buildContextBreakdownInput(options, state)!
     expect(input.systemPrompt).toBe('prompt')
     expect(input.mcpDeferredBlock).toBe(formatDeferredCapabilities([{ name: 'webSearch', source: 'builtin' }]))
@@ -264,7 +264,7 @@ describe('buildContextBreakdownInput', () => {
     const state = createLoopState()
     state.systemPromptCache = 'prompt with embedded blocks'
     state.systemPromptBlocks = { knowledge: 'knowledge', skill: 'skills block', mcpDeferred: 'mcp block' }
-    const options = { modelId: 'deepseek:deepseek-v4-flash' } as any
+    const options = { modelId: 'deepseek:deepseek-flash' } as any
     const input = buildContextBreakdownInput(options, state)!
     expect(input.knowledgeContext).toBe('knowledge')
     expect(input.skillBlock).toBe('skills block')
@@ -286,7 +286,7 @@ describe('buildContextBreakdownInput', () => {
     state.deferredCatalog = catalog
     const executor = async () => 'ok'
     state.manualToolExecutors.set('existing', executor)
-    buildContextBreakdownInput({ modelId: 'deepseek:deepseek-v4-flash' } as any, state)
+    buildContextBreakdownInput({ modelId: 'deepseek:deepseek-flash' } as any, state)
     expect(state.deferredCatalog).toBe(catalog)
     expect([...state.manualToolExecutors]).toEqual([['existing', executor]])
   })
diff --git a/packages/core/tests/context-window.test.ts b/packages/core/tests/context-window.test.ts
index d59bc1a..00b1667 100644
--- a/packages/core/tests/context-window.test.ts
+++ b/packages/core/tests/context-window.test.ts
@@ -25,10 +25,18 @@ afterEach(() => {
 
 describe('getContextWindow', () => {
   it('returns exact value for known models', () => {
+    expect(getContextWindow('anthropic:claude-fable-5-1')).toBe(1000000)
     expect(getContextWindow('anthropic:claude-opus-4-8')).toBe(1000000)
+    expect(getContextWindow('openai:gpt-6-astra')).toBe(1050000)
     expect(getContextWindow('openai:gpt-5.6-sol')).toBe(1047576)
+    expect(getContextWindow('google:gemini-3.8-flash')).toBe(1048576)
+    expect(getContextWindow('google:gemini-3.7-flash')).toBe(1048576)
     expect(getContextWindow('google:gemini-2.5-flash')).toBe(1000000)
-    expect(getContextWindow('deepseek:deepseek-v4-flash')).toBe(1000000)
+    expect(getContextWindow('deepseek:deepseek-flash')).toBe(1000000)
+    expect(getContextWindow('alibaba:qwen3.8-max')).toBe(1000000)
+    expect(getContextWindow('alibaba:qwen3.7-plus')).toBe(1000000)
+    expect(getContextWindow('xai:grok-4.20')).toBe(1000000)
+    expect(getContextWindow('zhipu:glm-5.3-flash')).toBe(1000000)
     expect(getContextWindow('alibaba:qwen-max')).toBe(32768)
   })
 
@@ -72,8 +80,14 @@ describe('getCompressionThreshold', () => {
 
 describe('getMaxOutputTokens', () => {
   it('returns specific ceiling for known models', () => {
-    expect(getMaxOutputTokens('deepseek:deepseek-v4-flash')).toBe(131072)
-    expect(getMaxOutputTokens('alibaba:qwen3.7-plus')).toBe(32000)
+    expect(getMaxOutputTokens('deepseek:deepseek-flash')).toBe(131072)
+    expect(getMaxOutputTokens('anthropic:claude-sonnet-5')).toBe(128000)
+    expect(getMaxOutputTokens('anthropic:claude-haiku-4-5')).toBe(64000)
+    expect(getMaxOutputTokens('openai:gpt-6-astra')).toBe(128000)
+    expect(getMaxOutputTokens('google:gemini-3.8-flash')).toBe(65536)
+    expect(getMaxOutputTokens('google:gemini-3.7-flash')).toBe(65536)
+    expect(getMaxOutputTokens('alibaba:qwen3.8-flash')).toBe(65536)
+    expect(getMaxOutputTokens('alibaba:qwen3.7-plus')).toBe(65536)
     expect(getMaxOutputTokens('alibaba:qwen-max')).toBe(8192)
   })
 
diff --git a/packages/core/tests/file-ingest.test.ts b/packages/core/tests/file-ingest.test.ts
index 89db111..f67c0ca 100644
--- a/packages/core/tests/file-ingest.test.ts
+++ b/packages/core/tests/file-ingest.test.ts
@@ -584,6 +584,24 @@ describe('ingestFile', () => {
       expect(imagePart.data).toEqual({ type: 'data', data: source.toString('base64') })
       expect(JSON.parse(JSON.stringify(imagePart))).toEqual(imagePart)
     }
+    expect(JSON.stringify(parts)).toContain('Analyze the supplied content directly')
+    expect(JSON.stringify(parts)).toContain('Do not invoke readFile')
+  })
+
+  it('marks a natively attached image as already delivered for readFile de-duplication', async () => {
+    const cache = new Map()
+
+    await buildUserContent(
+      `'${imageFile}' describe this image`,
+      multimodalCaps,
+      undefined,
+      undefined,
+      undefined,
+      'deepseek:deepseek-flash',
+      cache,
+    )
+
+    expect(cache.has(imageFile)).toBe(true)
   })
 
   it('rejects GIF before session insertion for an xAI model', async () => {
diff --git a/packages/core/tests/memory-inference.test.ts b/packages/core/tests/memory-inference.test.ts
index a4f62e2..14124b1 100644
--- a/packages/core/tests/memory-inference.test.ts
+++ b/packages/core/tests/memory-inference.test.ts
@@ -43,6 +43,52 @@ describe('memory inference policy', () => {
     expect(generate).toHaveBeenCalledWith({ maxOutputTokens: 1500, reasoning: 'low' })
   })
 
+  it('keeps current always-thinking families at their minimum effort', async () => {
+    const generate = vi.fn().mockResolvedValue({ output: {} })
+
+    await runMemoryInference({
+      modelId: 'anthropic:claude-fable-5-1',
+      maxOutputTokens: 1500,
+      maxTotalOutputTokens: 8192,
+      generate,
+    })
+
+    expect(generate).toHaveBeenCalledWith({ maxOutputTokens: 1500, reasoning: 'low' })
+  })
+
+  it('uses the minimum GLM-5.3 effort for memory inference', async () => {
+    for (const modelId of ['zhipu:glm-5.3', 'zhipu:glm-5.3-flash']) {
+      const generate = vi.fn().mockResolvedValue({ output: {} })
+
+      await runMemoryInference({
+        modelId,
+        maxOutputTokens: 1500,
+        maxTotalOutputTokens: 8192,
+        generate,
+      })
+
+      expect(generate).toHaveBeenCalledWith({
+        maxOutputTokens: 1500,
+        providerOptions: {
+          zhipu: { thinking: { type: 'enabled' }, reasoningEffort: 'low' },
+        },
+      })
+    }
+  })
+
+  it('omits unsupported reasoning controls for fixed-reasoning Grok 4.20', async () => {
+    const generate = vi.fn().mockResolvedValue({ output: {} })
+
+    await runMemoryInference({
+      modelId: 'xai:grok-4.20',
+      maxOutputTokens: 1500,
+      maxTotalOutputTokens: 8192,
+      generate,
+    })
+
+    expect(generate).toHaveBeenCalledWith({ maxOutputTokens: 1500 })
+  })
+
   it('starts OpenAI models without an off tier at low effort without temperature', async () => {
     const generate = vi.fn().mockResolvedValue({ output: {} })
 
diff --git a/packages/core/tests/openai-chatgpt-models.test.ts b/packages/core/tests/openai-chatgpt-models.test.ts
index b3dc4be..9b886ec 100644
--- a/packages/core/tests/openai-chatgpt-models.test.ts
+++ b/packages/core/tests/openai-chatgpt-models.test.ts
@@ -26,7 +26,7 @@ import {
   refreshOpenAIChatGPTModelsAfterNotFound,
   resetOpenAIChatGPTModelsForTesting,
 } from '../src/providers/openai-chatgpt-models.js'
-import { getReasoningLevel } from '../src/providers/thinking.js'
+import { getReasoningLevel, getThinkingProviderOptions } from '../src/providers/thinking.js'
 
 describe('OpenAI ChatGPT model catalog', () => {
   let testHome: string
@@ -107,6 +107,28 @@ describe('OpenAI ChatGPT model catalog', () => {
     expect(String(fetchMock.mock.calls[0]?.[0])).not.toContain('client_version=test')
   })
 
+  it('does not reintroduce a stale max effort after the server catalog removes it', async () => {
+    const fetchMock = vi.fn<typeof fetch>(async () =>
+      Response.json({
+        models: [
+          {
+            slug: 'gpt-5.6-sol',
+            display_name: 'GPT-5.6 Sol',
+            input_modalities: ['text', 'image'],
+            default_reasoning_level: 'low',
+            supported_reasoning_levels: [{ effort: 'low' }],
+            visibility: 'list',
+          },
+        ],
+      }),
+    )
+
+    await refreshOpenAIChatGPTModels('test', { fetch: fetchMock, force: true })
+
+    expect(getReasoningLevel('openai:gpt-5.6-sol', false, 'max')).toBe('low')
+    expect(getThinkingProviderOptions('openai:gpt-5.6-sol', false, 'max')).toEqual({})
+  })
+
   it('uses the raw ChatGPT context window ahead of the static OpenAI model table', async () => {
     const fetchMock = vi.fn<typeof fetch>(async () =>
       Response.json({
diff --git a/packages/core/tests/provider-compat.test.ts b/packages/core/tests/provider-compat.test.ts
index 23630de..1bb25d0 100644
--- a/packages/core/tests/provider-compat.test.ts
+++ b/packages/core/tests/provider-compat.test.ts
@@ -1,5 +1,6 @@
 import { describe, expect, it, vi } from 'vitest'
 
+import { createDeepSeek } from '@ai-sdk/deepseek'
 import { createOpenAICompatible } from '@ai-sdk/openai-compatible'
 import { generateText } from 'ai'
 import type { ModelMessage } from 'ai'
@@ -38,6 +39,7 @@ function imageToolResult(toolCallId: string, data: string, toolName = 'readFile'
 
 describe('stripBinaryPartsFromMessages', () => {
   it('replaces user image/file parts and tool-result media with text notices', () => {
+    const deliveredFileCache = new Map([['/tmp/image.png', { mtimeMs: 1, size: 1 }]])
     const messages = [
       {
         role: 'user',
@@ -65,7 +67,8 @@ describe('stripBinaryPartsFromMessages', () => {
       },
     ] as unknown as ModelMessage[]
 
-    expect(stripBinaryPartsFromMessages(messages)).toBe(true)
+    expect(stripBinaryPartsFromMessages(messages, deliveredFileCache)).toBe(true)
+    expect(deliveredFileCache.size).toBe(0)
 
     const userContent = messages[0]!.content as Array<{ type: string; text?: string }>
     expect(userContent.map((p) => p.type)).toEqual(['text', 'text'])
@@ -87,7 +90,7 @@ describe('stripBinaryPartsFromMessages', () => {
 })
 
 describe('reattachToolResultImagesForProvider', () => {
-  it('moves a contiguous Kimi tool-image group into one following user message', () => {
+  it('moves a contiguous DeepSeek tool-image group into one following user message', () => {
     const messages: ModelMessage[] = [
       { role: 'assistant', content: [] },
       imageToolResult('tc-1', 'AAAA1'),
@@ -95,7 +98,7 @@ describe('reattachToolResultImagesForProvider', () => {
       { role: 'assistant', content: 'continued' },
     ]
 
-    const requestMessages = reattachToolResultImagesForProvider(messages, 'moonshotai:kimi-k3')
+    const requestMessages = reattachToolResultImagesForProvider(messages, 'deepseek:deepseek-flash')
 
     expect(requestMessages.map((message) => message.role)).toEqual(['assistant', 'tool', 'tool', 'user', 'assistant'])
     for (const message of requestMessages.slice(1, 3)) {
@@ -169,7 +172,7 @@ describe('reattachToolResultImagesForProvider', () => {
   })
 
   it('leaves native and text-only transports unchanged', () => {
-    for (const modelId of ['openai:gpt-5.6-sol', 'anthropic:claude-sonnet-5', 'deepseek:deepseek-v4-flash']) {
+    for (const modelId of ['openai:gpt-5.6-sol', 'anthropic:claude-sonnet-5', 'deepseek:deepseek-v4-pro']) {
       const messages = [imageToolResult('tc-1', 'AAAA1')]
       const original = structuredClone(messages)
       const requestMessages = reattachToolResultImagesForProvider(messages, modelId)
@@ -178,13 +181,13 @@ describe('reattachToolResultImagesForProvider', () => {
     }
   })
 
-  it('serializes reattached base64 as image_url rather than tool text', async () => {
+  it('serializes DeepSeek Flash direct and tool images natively as image_url parts', async () => {
     let requestBody: {
-      messages?: Array<{ role?: string; content?: unknown }>
+      model?: string
+      messages?: Array<{ role?: string; content?: unknown; reasoning_content?: string }>
     } = {}
-    const provider = createOpenAICompatible({
-      name: 'moonshotai',
-      baseURL: 'https://example.test/v1',
+    const provider = createDeepSeek({
+      baseURL: 'https://example.test',
       apiKey: 'test-key',
       fetch: async (_input, init) => {
         requestBody = JSON.parse(String(init?.body))
@@ -193,7 +196,7 @@ describe('reattachToolResultImagesForProvider', () => {
             id: 'response-1',
             object: 'chat.completion',
             created: 0,
-            model: 'k3',
+            model: 'deepseek-flash',
             choices: [
               {
                 index: 0,
@@ -222,10 +225,12 @@ describe('reattachToolResultImagesForProvider', () => {
       },
       imageToolResult('tc-1', 'QUFBQQ=='),
     ]
-    const requestMessages = reattachToolResultImagesForProvider(messages, 'moonshotai:kimi-k3')
+    const requestMessages = reattachToolResultImagesForProvider(messages, 'deepseek:deepseek-flash')
 
-    await generateText({ model: provider('k3'), messages: requestMessages })
+    await generateText({ model: provider('deepseek-flash'), messages: requestMessages })
 
+    expect(requestBody.model).toBe('deepseek-flash')
+    expect(requestBody.messages?.find((message) => message.role === 'assistant')?.reasoning_content).toBe('')
     const toolMessage = requestBody.messages?.find((message) => message.role === 'tool')
     expect(toolMessage?.content).toBe('Loaded tc-1')
     const imageMessage = requestBody.messages?.find(
@@ -243,6 +248,34 @@ describe('reattachToolResultImagesForProvider', () => {
         image_url: { url: 'data:image/png;base64,QUFBQQ==' },
       },
     ])
+
+    await generateText({
+      model: provider('deepseek-flash'),
+      messages: [
+        {
+          role: 'user',
+          content: [
+            { type: 'text', text: 'Describe this image directly.' },
+            {
+              type: 'file',
+              data: { type: 'data', data: 'QUFBQQ==' },
+              mediaType: 'image/png',
+              filename: 'image.png',
+            },
+            { type: 'text', text: 'Analyze the supplied content directly; do not re-read the local path.' },
+          ],
+        },
+      ],
+    })
+
+    expect(requestBody.messages?.[0]?.content).toEqual([
+      { type: 'text', text: 'Describe this image directly.' },
+      {
+        type: 'image_url',
+        image_url: { url: 'data:image/png;base64,QUFBQQ==' },
+      },
+      { type: 'text', text: 'Analyze the supplied content directly; do not re-read the local path.' },
+    ])
   })
 })
 
@@ -280,9 +313,12 @@ describe('downgradeBinaryPartsForProvider', () => {
     const messages = [{ role: 'user', content: [{ type: 'text', text: 'look' }, imagePart] }] as ModelMessage[]
     const canonical = structuredClone(messages)
 
-    const vision = await downgradeBinaryPartsForProvider(messages, 'moonshotai:kimi-k3')
-    expect(vision[0]).toEqual(messages[0])
-    const text = await downgradeBinaryPartsForProvider(messages, 'deepseek:deepseek-v4-flash')
+    for (const modelId of ['moonshotai:kimi-k3', 'deepseek:deepseek-flash']) {
+      const vision = await downgradeBinaryPartsForProvider(messages, modelId)
+      expect(vision[0], modelId).toEqual(messages[0])
+      expect(JSON.stringify(vision), modelId).not.toContain('mock compatibility OCR')
+    }
+    const text = await downgradeBinaryPartsForProvider(messages, 'deepseek:deepseek-v4-pro')
     expect(JSON.stringify(text)).toContain('mock compatibility OCR')
     expect(JSON.stringify(text)).not.toContain('"type":"file"')
     expect(messages).toEqual(canonical)
@@ -309,8 +345,8 @@ describe('downgradeBinaryPartsForProvider', () => {
       ] as ModelMessage[]
     const callsBefore = vi.mocked(ocrImage).mock.calls.length
 
-    await downgradeBinaryPartsForProvider(asMessages(first), 'deepseek:deepseek-v4-flash')
-    await downgradeBinaryPartsForProvider(asMessages(second), 'deepseek:deepseek-v4-flash')
+    await downgradeBinaryPartsForProvider(asMessages(first), 'deepseek:deepseek-v4-pro')
+    await downgradeBinaryPartsForProvider(asMessages(second), 'deepseek:deepseek-v4-pro')
 
     expect(vi.mocked(ocrImage).mock.calls.length - callsBefore).toBe(2)
   })
diff --git a/packages/core/tests/provider-registry.test.ts b/packages/core/tests/provider-registry.test.ts
index bf40fc6..d3bc16b 100644
--- a/packages/core/tests/provider-registry.test.ts
+++ b/packages/core/tests/provider-registry.test.ts
@@ -22,7 +22,12 @@ import {
   XAI_PROMPT_CACHE_KEY_HEADER,
   applyCacheControl,
 } from '../src/providers/cache-control.js'
-import { createModelRegistry, kimiCodingModelId } from '../src/providers/registry.js'
+import {
+  createModelRegistry,
+  kimiCodingModelId,
+  withXaiReasoningHeader,
+  withZhipuReasoningHeader,
+} from '../src/providers/registry.js'
 
 function sseResponse(events: unknown[]): Response {
   return new Response(`${events.map((event) => `data: ${JSON.stringify(event)}\n\n`).join('')}data: [DONE]\n\n`, {
@@ -50,6 +55,7 @@ describe('Kimi endpoint model ids', () => {
     delete process.env.MOONSHOT_API_KEY
     delete process.env.OPENAI_API_KEY
     delete process.env.XAI_API_KEY
+    delete process.env.ZHIPU_API_KEY
     fs.rmSync(testHome, { recursive: true, force: true })
   })
 
@@ -74,7 +80,7 @@ describe('Kimi endpoint model ids', () => {
     expect(registry.languageModel('moonshotai:kimi-k2.7-code').modelId).toBe('kimi-k2.7-code')
   })
 
-  it('moves xAI session affinity into the Responses prompt_cache_key body field', async () => {
+  it('moves xAI session affinity and xhigh reasoning into the Responses body', async () => {
     const fetchMock = vi.fn<typeof fetch>(
       async () =>
         new Response(JSON.stringify({ error: { message: 'test stop' } }), {
@@ -87,14 +93,15 @@ describe('Kimi endpoint model ids', () => {
     const cache = applyCacheControl({
       instructions: 'stable instructions',
       messages: [{ role: 'user', content: 'hello' }],
-      modelId: 'xai:grok-4.5',
+      modelId: 'xai:grok-4.6',
       sessionId: 'session-1',
     })
     const result = streamText({
-      model: createModelRegistry().languageModel('xai:grok-4.5'),
+      model: createModelRegistry().languageModel('xai:grok-4.6'),
       instructions: cache.instructions,
       messages: cache.messages,
-      headers: cache.headers,
+      headers: withXaiReasoningHeader(cache.headers, 'xhigh'),
+      reasoning: 'xhigh',
       abortSignal: controller.signal,
       onError: () => undefined,
     })
@@ -105,13 +112,84 @@ describe('Kimi endpoint model ids', () => {
     expect(fetchMock).toHaveBeenCalledOnce()
     const [url, init] = fetchMock.mock.calls[0]!
     expect(String(url)).toBe('https://api.x.ai/v1/responses')
-    expect(JSON.parse(String(init?.body))).toMatchObject({ prompt_cache_key: 'session-1' })
+    expect(JSON.parse(String(init?.body))).toMatchObject({
+      prompt_cache_key: 'session-1',
+      reasoning: { effort: 'xhigh' },
+    })
     const headers = new Headers(init?.headers)
     expect(headers.get(XAI_PROMPT_CACHE_KEY_HEADER)).toBeNull()
     expect(headers.get('x-grok-conv-id')).toBeNull()
     expect(init?.signal).toBe(controller.signal)
   })
 
+  it('omits unsupported reasoning parameters for the Grok 4.20 alias', async () => {
+    const fetchMock = vi.fn<typeof fetch>(async () =>
+      Response.json({ error: { message: 'test stop' } }, { status: 400 }),
+    )
+    vi.stubGlobal('fetch', fetchMock)
+    const result = streamText({
+      model: createModelRegistry().languageModel('xai:grok-4.20'),
+      messages: [{ role: 'user', content: 'hello' }],
+      headers: withXaiReasoningHeader(undefined, undefined),
+      onError: () => undefined,
+    })
+    for await (const _chunk of result.textStream) {
+      // The mock returns a deliberate error after the outbound request is captured.
+    }
+
+    const [, init] = fetchMock.mock.calls[0]!
+    const body = JSON.parse(String(init?.body))
+    expect(body).not.toHaveProperty('reasoning')
+    expect(body).not.toHaveProperty('reasoning_effort')
+    expect(new Headers(init?.headers).get('x-x-code-xai-reasoning-effort')).toBeNull()
+  })
+
+  it('serializes max Anthropic effort with adaptive thinking enabled', async () => {
+    process.env.ANTHROPIC_API_KEY = 'test-key'
+    const fetchMock = vi.fn<typeof fetch>(async () =>
+      Response.json({ error: { message: 'test stop' } }, { status: 400 }),
+    )
+    vi.stubGlobal('fetch', fetchMock)
+    const result = streamText({
+      model: createModelRegistry().languageModel('anthropic:claude-opus-5'),
+      messages: [{ role: 'user', content: 'hello' }],
+      reasoning: 'xhigh',
+      providerOptions: {
+        anthropic: { thinking: { type: 'adaptive' }, effort: 'max' },
+      },
+      onError: () => undefined,
+    })
+    for await (const _chunk of result.textStream) {
+      // The mock returns a deliberate error after the outbound request is captured.
+    }
+
+    const body = JSON.parse(String(fetchMock.mock.calls[0]?.[1]?.body))
+    expect(body.thinking).toEqual({ type: 'adaptive' })
+    expect(body.output_config).toMatchObject({ effort: 'max' })
+  })
+
+  it('serializes the minimum Zhipu reasoning effort for always-thinking models', async () => {
+    process.env.ZHIPU_API_KEY = 'test-key'
+    const fetchMock = vi.fn<typeof fetch>(async () =>
+      Response.json({ error: { message: 'test stop' } }, { status: 400 }),
+    )
+    vi.stubGlobal('fetch', fetchMock)
+    const result = streamText({
+      model: createModelRegistry().languageModel('zhipu:glm-5.3-flash'),
+      messages: [{ role: 'user', content: 'hello' }],
+      headers: withZhipuReasoningHeader(undefined, 'low'),
+      onError: () => undefined,
+    })
+    for await (const _chunk of result.textStream) {
+      // The mock returns a deliberate error after the outbound request is captured.
+    }
+
+    const [url, init] = fetchMock.mock.calls[0]!
+    expect(String(url)).toBe('https://open.bigmodel.cn/api/paas/v4/chat/completions')
+    expect(JSON.parse(String(init?.body))).toMatchObject({ model: 'glm-5.3-flash', reasoning_effort: 'low' })
+    expect(new Headers(init?.headers).get('x-x-code-zhipu-reasoning-effort')).toBeNull()
+  })
+
   it('serializes Anthropic cache breakpoints without putting a system role in messages', async () => {
     process.env.ANTHROPIC_API_KEY = 'test-key'
     vi.stubEnv('ANTHROPIC_BASE_URL', 'https://api.anthropic.com')
diff --git a/packages/core/tests/read-file.test.ts b/packages/core/tests/read-file.test.ts
index 96b8f1a..0f116b2 100644
--- a/packages/core/tests/read-file.test.ts
+++ b/packages/core/tests/read-file.test.ts
@@ -228,7 +228,7 @@ describe('readFile tool', () => {
     const filePath = path.join(tmpDir, 'image.png')
     const { Jimp } = await import('jimp')
     await fs.writeFile(filePath, await new Jimp({ width: 3, height: 2, color: 0xffffffff }).getBuffer('image/png'))
-    const tool = createReadFileTool(undefined, { modelId: 'deepseek:deepseek-v4-flash' })
+    const tool = createReadFileTool(undefined, { modelId: 'deepseek:deepseek-v4-pro' })
     const result = await tool.execute!({ filePath }, {
       toolCallId: 'image-ocr-test',
       messages: [],
@@ -239,6 +239,45 @@ describe('readFile tool', () => {
     expect(JSON.stringify(result)).not.toContain('image-data')
     await fs.rm(tmpDir, { recursive: true })
   })
+
+  it('returns native image content to DeepSeek Flash without local OCR', async () => {
+    const tmpDir = await fs.mkdtemp(path.join(os.tmpdir(), 'xc-rf-deepseek-image-'))
+    const filePath = path.join(tmpDir, 'image.png')
+    const { Jimp } = await import('jimp')
+    await fs.writeFile(filePath, await new Jimp({ width: 3, height: 2, color: 0xffffffff }).getBuffer('image/png'))
+    const tool = createReadFileTool(undefined, { modelId: 'deepseek:deepseek-flash' })
+    const result = await tool.execute!({ filePath }, {
+      toolCallId: 'deepseek-image-test',
+      messages: [],
+      abortSignal: undefined,
+    } as never)
+
+    expect(result).toMatchObject({
+      type: 'content',
+      value: expect.arrayContaining([
+        expect.objectContaining({ type: 'file', mediaType: 'image/png' }),
+        expect.objectContaining({ type: 'text', text: expect.stringContaining('Do not invoke readFile') }),
+      ]),
+    })
+    expect(JSON.stringify(result)).not.toContain('mock readFile OCR')
+    await fs.rm(tmpDir, { recursive: true })
+  })
+
+  it('does not deliver an unchanged image twice through readFile', async () => {
+    const tmpDir = await fs.mkdtemp(path.join(os.tmpdir(), 'xc-rf-image-cache-'))
+    const filePath = path.join(tmpDir, 'image.png')
+    const { Jimp } = await import('jimp')
+    await fs.writeFile(filePath, await new Jimp({ width: 3, height: 2, color: 0xffffffff }).getBuffer('image/png'))
+    const tool = createReadFileTool(new Map(), { modelId: 'deepseek:deepseek-flash' })
+    const options = { toolCallId: 'deepseek-image-cache-test', messages: [], abortSignal: undefined } as never
+
+    const first = await tool.execute!({ filePath }, options)
+    const second = await tool.execute!({ filePath }, options)
+
+    expect(first).toMatchObject({ type: 'content' })
+    expect(second).toContain('is unchanged since its full content was added')
+    await fs.rm(tmpDir, { recursive: true })
+  })
 })
 
 describe('parsePdfPageRange', () => {
diff --git a/packages/core/tests/thinking.test.ts b/packages/core/tests/thinking.test.ts
index 5e25a9f..728bdb5 100644
--- a/packages/core/tests/thinking.test.ts
+++ b/packages/core/tests/thinking.test.ts
@@ -1,4 +1,13 @@
-import { getReasoningLevel, getThinkingProviderOptions, supportsReasoningTier } from '../src/providers/thinking.js'
+import { createDeepSeek } from '@ai-sdk/deepseek'
+import { generateText } from 'ai'
+
+import {
+  getReasoningEffort,
+  getReasoningLevel,
+  getReasoningTierOptions,
+  getThinkingProviderOptions,
+  supportsReasoningTier,
+} from '../src/providers/thinking.js'
 import { isolateOpenAIAuth } from './provider-env.js'
 
 let restoreOpenAIAuth: () => void
@@ -12,10 +21,18 @@ afterEach(() => restoreOpenAIAuth())
 describe('supportsReasoningTier', () => {
   it('is true for tier-capable models', () => {
     expect(supportsReasoningTier('moonshotai:kimi-k3')).toBe(true)
+    expect(supportsReasoningTier('google:gemini-3.8-flash')).toBe(true)
+    expect(supportsReasoningTier('google:gemini-3.7-flash')).toBe(true)
+    expect(supportsReasoningTier('google:gemini-3.6-flash')).toBe(true)
     expect(supportsReasoningTier('google:gemini-3.5-flash')).toBe(true)
+    expect(supportsReasoningTier('openai:gpt-5.6')).toBe(true)
     expect(supportsReasoningTier('openai:gpt-5.6-sol')).toBe(true)
+    expect(supportsReasoningTier('openai:gpt-5.5')).toBe(true)
     expect(supportsReasoningTier('anthropic:claude-sonnet-5')).toBe(true)
     expect(supportsReasoningTier('xai:grok-4.5')).toBe(true)
+    expect(supportsReasoningTier('xai:grok-4.6')).toBe(true)
+    expect(supportsReasoningTier('anthropic:claude-fable-5-1')).toBe(true)
+    expect(supportsReasoningTier('openai:gpt-6-astra')).toBe(true)
   })
 
   it('is false for models whose provider has tiers but the model family does not', () => {
@@ -23,15 +40,19 @@ describe('supportsReasoningTier', () => {
     expect(supportsReasoningTier('google:gemini-2.5-flash')).toBe(false)
     expect(supportsReasoningTier('moonshotai:kimi-k2.6')).toBe(false)
     expect(supportsReasoningTier('moonshotai:kimi-k2.7-code')).toBe(false)
+    expect(supportsReasoningTier('xai:grok-4.20')).toBe(false)
+    expect(supportsReasoningTier('xai:grok-4.20-non-reasoning')).toBe(false)
   })
 
-  it('is true for DeepSeek V4 models', () => {
-    expect(supportsReasoningTier('deepseek:deepseek-v4-flash')).toBe(true)
+  it('is true for current DeepSeek Flash and V4 models', () => {
+    expect(supportsReasoningTier('deepseek:deepseek-flash')).toBe(true)
     expect(supportsReasoningTier('deepseek:deepseek-v4-pro')).toBe(true)
   })
 
   it('is true for Zhipu GLM-5.2', () => {
     expect(supportsReasoningTier('zhipu:glm-5.2')).toBe(true)
+    expect(supportsReasoningTier('zhipu:glm-5.3')).toBe(true)
+    expect(supportsReasoningTier('zhipu:glm-5.3-flash')).toBe(true)
   })
 
   it('is false for providers without any tier support', () => {
@@ -46,25 +67,85 @@ describe('supportsReasoningTier', () => {
   })
 })
 
+describe('model-specific reasoning profiles', () => {
+  it('returns only tiers supported by the selected model family', () => {
+    expect(getReasoningTierOptions('google:gemini-3.8-flash')?.map((tier) => tier.value)).toEqual([
+      'low',
+      'medium',
+      'high',
+    ])
+    expect(getReasoningTierOptions('google:gemini-3.6-flash')?.map((tier) => tier.value)).toEqual([
+      'minimal',
+      'low',
+      'medium',
+      'high',
+    ])
+    expect(getReasoningTierOptions('google:gemini-3.5-flash')?.map((tier) => tier.value)).toEqual([
+      'minimal',
+      'low',
+      'medium',
+      'high',
+    ])
+    expect(getReasoningTierOptions('openai:gpt-5.4-mini')?.map((tier) => tier.value)).toEqual([
+      'low',
+      'medium',
+      'high',
+      'xhigh',
+    ])
+    expect(getReasoningTierOptions('openai:gpt-5.6')?.map((tier) => tier.value)).toEqual([
+      'low',
+      'medium',
+      'high',
+      'xhigh',
+      'max',
+    ])
+    expect(getReasoningTierOptions('openai:gpt-5.5')?.map((tier) => tier.value)).toEqual([
+      'low',
+      'medium',
+      'high',
+      'xhigh',
+    ])
+  })
+
+  it('floors disabled reasoning for models that reject off', () => {
+    for (const id of [
+      'anthropic:claude-fable-5-1',
+      'openai:gpt-6-astra',
+      'google:gemini-3.8-flash',
+      'google:gemini-3.7-flash',
+      'xai:grok-4.6',
+      'xai:grok-4.5',
+      'zhipu:glm-5.3',
+    ]) {
+      expect(getReasoningEffort(id, false), id).toBe('low')
+    }
+    expect(getReasoningLevel('google:gemini-3.8-flash', false)).toBe('low')
+    expect(getReasoningEffort('google:gemini-3.6-flash', false)).toBe('minimal')
+    expect(getReasoningLevel('openai:gpt-6-astra', false)).toBe('low')
+  })
+})
+
 describe('getReasoningLevel', () => {
   it('returns the effort tier for tier-capable models', () => {
-    expect(getReasoningLevel('deepseek:deepseek-v4-flash', false, 'high')).toBe('high')
+    expect(getReasoningLevel('deepseek:deepseek-flash', false, 'high')).toBeUndefined()
+    expect(getReasoningLevel('openai:gpt-5.6', false, 'low')).toBe('low')
     expect(getReasoningLevel('openai:gpt-5.6-sol', false, 'medium')).toBe('medium')
+    expect(getReasoningLevel('openai:gpt-5.5', false, 'xhigh')).toBe('xhigh')
     expect(getReasoningLevel('anthropic:claude-sonnet-5', true, 'low')).toBe('low')
   })
 
-  it('maps "max" tier to "xhigh" reasoning level', () => {
-    expect(getReasoningLevel('deepseek:deepseek-v4-flash', false, 'max')).toBe('xhigh')
+  it('maps portable "max" tiers to "xhigh" while leaving DeepSeek on its native option', () => {
+    expect(getReasoningLevel('deepseek:deepseek-flash', false, 'max')).toBeUndefined()
     expect(getReasoningLevel('moonshotai:kimi-k3', false, 'max')).toBe('xhigh')
   })
 
   it('returns high when enabled and no effort specified', () => {
-    expect(getReasoningLevel('deepseek:deepseek-v4-flash', true)).toBe('high')
+    expect(getReasoningLevel('deepseek:deepseek-flash', true)).toBe('high')
     expect(getReasoningLevel('openai:gpt-5.6-sol', true)).toBe('high')
   })
 
   it('returns none when disabled and no effort specified', () => {
-    expect(getReasoningLevel('deepseek:deepseek-v4-flash', false)).toBe('none')
+    expect(getReasoningLevel('deepseek:deepseek-flash', false)).toBe('none')
     expect(getReasoningLevel('openai:gpt-5.6-sol', false)).toBe('none')
   })
 
@@ -74,6 +155,12 @@ describe('getReasoningLevel', () => {
     expect(getReasoningLevel('custom:my-model', true)).toBeUndefined()
   })
 
+  it('omits reasoning control for fixed-reasoning Grok 4.20 variants', () => {
+    expect(getReasoningLevel('xai:grok-4.20', false)).toBeUndefined()
+    expect(getReasoningLevel('xai:grok-4.20', true, 'low')).toBeUndefined()
+    expect(getReasoningLevel('xai:grok-4.20-non-reasoning', true, 'high')).toBeUndefined()
+  })
+
   it('ignores effort for models that do not support tiers', () => {
     expect(getReasoningLevel('deepseek:deepseek-chat', false, 'high')).toBe('none')
     expect(getReasoningLevel('deepseek:deepseek-chat', true, 'high')).toBe('high')
@@ -101,14 +188,73 @@ describe('getThinkingProviderOptions', () => {
 
   it('returns zhipu thinking enabled when tier is set', () => {
     expect(getThinkingProviderOptions('zhipu:glm-5.2', false, 'high')).toEqual({
-      zhipu: { thinking: { type: 'enabled' } },
+      zhipu: { thinking: { type: 'enabled' }, reasoningEffort: 'high' },
+    })
+  })
+
+  it('keeps always-thinking Zhipu models enabled at low effort when switched off', () => {
+    expect(getThinkingProviderOptions('zhipu:glm-5.3-flash', false)).toEqual({
+      zhipu: { thinking: { type: 'enabled' }, reasoningEffort: 'low' },
     })
   })
 
+  it.each(['low', 'high', 'max'])('uses the provider-native DeepSeek %s tier', (effort) => {
+    expect(getThinkingProviderOptions('deepseek:deepseek-flash', true, effort)).toEqual({
+      deepseek: { reasoningEffort: effort },
+    })
+  })
+
+  it('uses provider-native max where the portable reasoning enum stops at xhigh', () => {
+    expect(getThinkingProviderOptions('anthropic:claude-opus-5', true, 'max')).toEqual({
+      anthropic: { thinking: { type: 'adaptive' }, effort: 'max' },
+    })
+    expect(getThinkingProviderOptions('openai:gpt-6-astra', true, 'max')).toEqual({
+      openai: { reasoningEffort: 'max' },
+    })
+    expect(getThinkingProviderOptions('openai:gpt-5.4-mini', true, 'max')).toEqual({})
+  })
+
   it('returns empty object for providers using top-level reasoning', () => {
-    expect(getThinkingProviderOptions('deepseek:deepseek-v4-flash', true)).toEqual({})
+    expect(getThinkingProviderOptions('deepseek:deepseek-flash', true)).toEqual({})
     expect(getThinkingProviderOptions('openai:gpt-5.6-sol', true, 'high')).toEqual({})
     expect(getThinkingProviderOptions('anthropic:claude-sonnet-5', true)).toEqual({})
     expect(getThinkingProviderOptions('google:gemini-3.5-flash', true, 'low')).toEqual({})
   })
+
+  it('sends DeepSeek max natively without an xhigh compatibility warning', async () => {
+    let requestBody: { reasoning_effort?: string } = {}
+    const provider = createDeepSeek({
+      baseURL: 'https://example.test',
+      apiKey: 'test-key',
+      fetch: async (_input, init) => {
+        requestBody = JSON.parse(String(init?.body))
+        return new Response(
+          JSON.stringify({
+            id: 'response-1',
+            object: 'chat.completion',
+            created: 0,
+            model: 'deepseek-flash',
+            choices: [{ index: 0, message: { role: 'assistant', content: 'done' }, finish_reason: 'stop' }],
+            usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
+          }),
+          { status: 200, headers: { 'content-type': 'application/json' } },
+        )
+      },
+    })
+    const modelId = 'deepseek:deepseek-flash'
+    const reasoning = getReasoningLevel(modelId, true, 'max')
+    const providerOptions = getThinkingProviderOptions(modelId, true, 'max')
+
+    const result = await generateText({
+      model: provider('deepseek-flash'),
+      messages: [{ role: 'user', content: 'hello' }],
+      ...(reasoning ? { reasoning } : {}),
+      providerOptions: providerOptions as Parameters<typeof generateText>[0]['providerOptions'],
+    })
+
+    expect(requestBody.reasoning_effort).toBe('max')
+    expect(result.warnings ?? []).not.toEqual(
+      expect.arrayContaining([expect.objectContaining({ feature: 'reasoning' })]),
+    )
+  })
 })
diff --git a/packages/core/tests/tool-result-images.test.ts b/packages/core/tests/tool-result-images.test.ts
index 4b08aca..5def177 100644
--- a/packages/core/tests/tool-result-images.test.ts
+++ b/packages/core/tests/tool-result-images.test.ts
@@ -45,21 +45,25 @@ describe('deliverToolImages', () => {
     expect(captionImageBuffer).not.toHaveBeenCalled()
   })
 
-  it('keeps Kimi images in canonical tool history for request-time reattachment', async () => {
-    const r = await deliverToolImages(ctx('moonshotai:kimi-k3'), 'shot taken', IMG)
-    expect(r.images).toEqual(IMG)
-    expect(r.text).toBe('shot taken')
+  it('keeps Chat Completions vision images in canonical tool history for request-time reattachment', async () => {
+    for (const modelId of ['moonshotai:kimi-k3', 'deepseek:deepseek-flash']) {
+      const r = await deliverToolImages(ctx(modelId), 'shot taken', IMG)
+      expect(r.images, modelId).toEqual(IMG)
+      expect(r.text, modelId).toBe('shot taken')
+    }
     expect(captionImageBuffer).not.toHaveBeenCalled()
   })
 
-  it('does not borrow a separate vision provider when the active Kimi model can view the image', async () => {
+  it('does not borrow a separate vision provider when the active model can view the image', async () => {
     vi.mocked(pickVisionProvider).mockReturnValue({
       provider: 'google',
       modelId: 'google:gemini-2.5-flash',
       label: 'Gemini 2.5 Flash',
     })
-    const r = await deliverToolImages(ctx('moonshotai:kimi-k3'), 'shot taken', IMG)
-    expect(r.images).toEqual(IMG)
+    for (const modelId of ['moonshotai:kimi-k3', 'deepseek:deepseek-flash']) {
+      const r = await deliverToolImages(ctx(modelId), 'shot taken', IMG)
+      expect(r.images, modelId).toEqual(IMG)
+    }
     expect(captionImageBuffer).not.toHaveBeenCalled()
   })
 
@@ -69,7 +73,7 @@ describe('deliverToolImages', () => {
       modelId: 'google:gemini-2.5-flash',
       label: 'Gemini 2.5 Flash',
     })
-    const r = await deliverToolImages(ctx('deepseek:deepseek-v4-flash'), 'shot taken', IMG)
+    const r = await deliverToolImages(ctx('deepseek:deepseek-v4-pro'), 'shot taken', IMG)
     expect(r.images).toBeUndefined()
     expect(r.text).toContain('A MAP OF BERLIN')
     expect(r.text).toContain('Privacy notice')
@@ -83,7 +87,7 @@ describe('deliverToolImages', () => {
       modelId: 'google:gemini-2.5-flash',
       label: 'Gemini 2.5 Flash',
     })
-    await deliverToolImages(ctx('deepseek:deepseek-v4-flash'), 'shot taken', IMG, {
+    await deliverToolImages(ctx('deepseek:deepseek-v4-pro'), 'shot taken', IMG, {
       captionPrompt: 'Report only visible UI defects.',
       maxOutputTokens: 400,
     })
@@ -115,7 +119,7 @@ describe('deliverToolImages', () => {
           releaseUsageWrite = resolve
         }),
     )
-    const context = ctx('deepseek:deepseek-v4-flash')
+    const context = ctx('deepseek:deepseek-v4-pro')
     let completed = false
     const resultPromise = deliverToolImages(context, 'shot taken', IMG).then((result) => {
       completed = true
@@ -134,14 +138,14 @@ describe('deliverToolImages', () => {
   })
 
   it('drops the image with a clear note when no vision model is available', async () => {
-    const r = await deliverToolImages(ctx('deepseek:deepseek-v4-flash'), 'shot taken', IMG)
+    const r = await deliverToolImages(ctx('deepseek:deepseek-v4-pro'), 'shot taken', IMG)
     expect(r.images).toBeUndefined()
     expect(r.text).toContain('no vision model is available')
     expect(captionImageBuffer).not.toHaveBeenCalled()
   })
 
   it('uses a caller-specific fallback when no accessibility snapshot exists', async () => {
-    const r = await deliverToolImages(ctx('deepseek:deepseek-v4-flash'), 'shot taken', IMG, {
+    const r = await deliverToolImages(ctx('deepseek:deepseek-v4-pro'), 'shot taken', IMG, {
       unavailableFallback: 'No accessibility snapshot was returned for this check.',
     })
 
@@ -157,7 +161,7 @@ describe('deliverToolImages', () => {
       label: 'Gemini 2.5 Flash',
     })
     const two = [...IMG, { data: Buffer.from('second').toString('base64'), mediaType: 'image/png' }]
-    const r = await deliverToolImages(ctx('deepseek:deepseek-v4-flash'), 'shots', two)
+    const r = await deliverToolImages(ctx('deepseek:deepseek-v4-pro'), 'shots', two)
     expect(captionImageBuffer).toHaveBeenCalledTimes(2)
     expect(r.text).toContain('Screenshot 1')
     expect(r.text).toContain('Screenshot 2')
diff --git a/packages/core/tests/vision-fallback.test.ts b/packages/core/tests/vision-fallback.test.ts
index b274c1c..294cbeb 100644
--- a/packages/core/tests/vision-fallback.test.ts
+++ b/packages/core/tests/vision-fallback.test.ts
@@ -89,10 +89,19 @@ describe('pickVisionProvider', () => {
     expect(pickVisionProvider()?.provider).toBe('xai')
   })
 
-  it('ignores DeepSeek key when picking — still selects vision provider if present', () => {
+  it('uses DeepSeek Flash when it is the only configured vision provider', () => {
+    process.env.DEEPSEEK_API_KEY = 'test'
+    expect(pickVisionProvider()).toEqual({
+      provider: 'deepseek',
+      modelId: 'deepseek:deepseek-flash',
+      label: 'DeepSeek V4.1 Flash',
+    })
+  })
+
+  it('prefers DeepSeek Flash over Anthropic for borrowed vision', () => {
     process.env.DEEPSEEK_API_KEY = 'test'
     process.env.ANTHROPIC_API_KEY = 'test'
-    expect(pickVisionProvider()?.provider).toBe('anthropic')
+    expect(pickVisionProvider()?.provider).toBe('deepseek')
   })
 })
 
diff --git a/packages/core/tests/web-search.test.ts b/packages/core/tests/web-search.test.ts
index ca05493..e014bf9 100644
--- a/packages/core/tests/web-search.test.ts
+++ b/packages/core/tests/web-search.test.ts
@@ -200,6 +200,7 @@ describe('webSearch', () => {
       }),
     })
     const body = JSON.parse(init!.body as string)
+    expect(body.model).toBe('deepseek-flash')
     expect(body.tools).toEqual([{ type: 'web_search_20250305', name: 'web_search', max_uses: 5 }])
     // Deduped by URL, snippet stitched from the text-block citation.
     expect(result).toContain('cited snippet')
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index b3600ab..13d1c33 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -138,8 +138,8 @@ importers:
         specifier: ^4.0.0
         version: 4.0.25(zod@3.25.76)
       '@ai-sdk/deepseek':
-        specifier: ^3.0.28
-        version: 3.0.28(zod@3.25.76)
+        specifier: ^3.0.45
+        version: 3.0.45(zod@3.25.76)
       '@ai-sdk/google':
         specifier: ^4.0.0
         version: 4.0.29(zod@3.25.76)
@@ -261,8 +261,8 @@ packages:
     peerDependencies:
       zod: ^3.25.76 || ^4.1.8
 
-  '@ai-sdk/deepseek@3.0.28':
-    resolution: {integrity: sha512-ySiafZL/1KQq8NF49eGWmVRYP4ezBjxAwAlMK+hhTV+9ee37LcCxlSIbURe4C4KFq8kYaTFKsTGI2poLFLOu0w==}
+  '@ai-sdk/deepseek@3.0.45':
+    resolution: {integrity: sha512-b2rEm2IdG1/JfLpry0rNi8A7dav9cvXZviFUo0g0XRRYAkNPkOgsO2XMEYfCQvKKymt2uWiqmDijxcZh17oYXg==}
     engines: {node: '>=22'}
     peerDependencies:
       zod: ^3.25.76 || ^4.1.8
@@ -303,18 +303,18 @@ packages:
     peerDependencies:
       zod: ^3.25.76 || ^4.1.8
 
-  '@ai-sdk/provider-utils@5.0.27':
-    resolution: {integrity: sha512-EzAn4pdgG5g0xXtH6lE2zyNmfjDQIDjATkfqzuidEI35g++hh4+07vnjzkT/RmGmIClPZiRj/Q2GMPV2V7mkHw==}
+  '@ai-sdk/provider-utils@5.0.41':
+    resolution: {integrity: sha512-hSosMII5k33z0buaZ6dugH/g8yIJjPQBILuQvr4ThdiMJxrkiZxFWTgIkvKsT6Dk493JbsnGz5W/q88Y5bR7WA==}
     engines: {node: '>=22'}
     peerDependencies:
       zod: ^3.25.76 || ^4.1.8
 
-  '@ai-sdk/provider@4.0.4':
-    resolution: {integrity: sha512-tbHKNLirllUNF3ZlkCsXnwab2ZV1Sl4b1H/Cp9ruCce15IBmskE8Gwkk0yo9xDWY+jho2of7lVXtwSsyrq7cwQ==}
+  '@ai-sdk/provider@4.0.15':
+    resolution: {integrity: sha512-LIDCrT8xgWVMiVUV13+v9FXPJ8B5uX/5ByEMXxp9Sspr2SyB1PwMSCPshZoI/t+4AgxyZCFEVutdaigCnqje3w==}
     engines: {node: '>=22'}
 
-  '@ai-sdk/provider@4.0.7':
-    resolution: {integrity: sha512-6or44XprPzKbr8zkmzosowSE0pxkvJcoojBL+mCZvPUt3kvXp3XSNqeVun9golb1acEfSo6yaEBRT18h2VU+1Q==}
+  '@ai-sdk/provider@4.0.4':
+    resolution: {integrity: sha512-tbHKNLirllUNF3ZlkCsXnwab2ZV1Sl4b1H/Cp9ruCce15IBmskE8Gwkk0yo9xDWY+jho2of7lVXtwSsyrq7cwQ==}
     engines: {node: '>=22'}
 
   '@ai-sdk/xai@4.0.23':
@@ -3363,10 +3363,10 @@ snapshots:
       '@ai-sdk/provider-utils': 5.0.16(zod@3.25.76)
       zod: 3.25.76
 
-  '@ai-sdk/deepseek@3.0.28(zod@3.25.76)':
+  '@ai-sdk/deepseek@3.0.45(zod@3.25.76)':
     dependencies:
-      '@ai-sdk/provider': 4.0.7
-      '@ai-sdk/provider-utils': 5.0.27(zod@3.25.76)
+      '@ai-sdk/provider': 4.0.15
+      '@ai-sdk/provider-utils': 5.0.41(zod@3.25.76)
       zod: 3.25.76
 
   '@ai-sdk/gateway@4.0.33(zod@3.25.76)':
@@ -3410,20 +3410,20 @@ snapshots:
       undici: 7.29.0
       zod: 3.25.76
 
-  '@ai-sdk/provider-utils@5.0.27(zod@3.25.76)':
+  '@ai-sdk/provider-utils@5.0.41(zod@3.25.76)':
     dependencies:
-      '@ai-sdk/provider': 4.0.7
+      '@ai-sdk/provider': 4.0.15
       '@standard-schema/spec': 1.1.0
       '@workflow/serde': 4.1.0
       eventsource-parser: 3.1.1
       undici: 7.29.0
       zod: 3.25.76
 
-  '@ai-sdk/provider@4.0.4':
+  '@ai-sdk/provider@4.0.15':
     dependencies:
       json-schema: 0.4.0
 
-  '@ai-sdk/provider@4.0.7':
+  '@ai-sdk/provider@4.0.4':
     dependencies:
       json-schema: 0.4.0