dsh-connect-minimaxcode 0.5.2 → 0.5.4

package/README.en.md

@@ -50,6 +50,7 @@ Another implementation in the same direction is **[dingminhua/dsh-connect-workbu
 - **Read-mostly**: apart from "refreshing an expired desktop app login state" and the sign-in you start yourself, the plugin does not write to or modify any MiniMax file
 - **Follows the selected model**: the balance sits in the **composer tool row beside the model picker**, and is drawn **only while a MiniMax model is selected** — picking another connector's model hides it entirely instead of misleading you with MiniMax's balance
 - **An icon and two languages of its own**: a 48×48 SVG ships with the package, so the plugin page no longer shows the default placeholder; the title and the introduction come in Chinese and English, and the interface follows the host's language
+- **Image input**: models whose catalog entry declares the `image` input modality (`M3`, `M3.1-Flash`) take pictures directly; the host's attachment service reads the image into request bytes, and the plugin only has to connect the durable reference to it
 
 ## Prerequisites
 
@@ -60,6 +61,8 @@ Either:
 
 Plus DSH desktop (built-in core `>=0.2.1-rc.2 <0.3.0`, see `engines.dsh` in `package.json`; tested on `0.2.1-rc.2`) or the Web profile.
 
+> **Sending a picture needs the host's attachment service**: the plugin does not read images itself — the host's durable attachment service turns a stored reference into request bytes, and the plugin only connects the two. On a host without that service the plugin still loads and text still works; only image requests are refused (`pi-ai image input requires the durable attachment service`).
+
 ## Installation
 
 **Option 1: install from the Git repository** (recommended, gets the latest code)
@@ -293,10 +296,24 @@ The OAuth token is valid for about an hour and is refreshed automatically by the
 - **The automatic claim is off by default**, and it only asks when the server explicitly says the day can be claimed, at most once a day. Whether to turn it on, and whether that is within MiniMax's terms of service, is your call.
 - **No settings card is provided**. The plugin does not register a card on the DSH settings page; the interface it adds is a single pill in the composer tool row (**drawn only while this plugin's model is selected**) and a small card on the plugin page.
 - **The upstream APIs are non-public** and may stop working when MiniMax updates its client. If the balance API breaks, the pill shows "credits unavailable", which does not affect model traffic.
+- **Image input depends on the host's attachment service.** The catalog declares the `image` modality (`M3`, `M3.1-Flash`), but turning a durable attachment into request bytes is the host's job: without that service the plugin still loads and serves text, and only image requests are refused.
+- **The context window is declared as the gateway's real ceiling (1,000,000), not the 512,000 in the catalog.** Upstream's `limit.context` is the *default* window and the billing tier boundary (input above 512k costs more); the largest entry of `context_window_options` is the ceiling the gateway enforces. Measured on 2026-10-25: 946,206 tokens were served normally, 1,050,000 were refused (`invalid params, context window exceeds limit (2013)`). Declaring 512,000 meant every turn of a long conversation was reported as `pi-ai detected context overflow` — while the model had answered perfectly well and the reply was thrown away.
 - **The device code flow needs `account.minimax.cn` to be reachable**, and the browser that authorizes it has to be able to sign in to a MiniMax account. On some restricted networks the authorization page cannot be opened.
 
 ## Changelog
 
+### 0.5.4 (2026-10-29)
+
+- **Fixed long conversations being reported as context overflow**: upstream's `limit.context` (512,000) is the *default* window and the billing tier boundary, while the ceiling the gateway enforces is the 1,000,000 in `context_window_options`. Declaring 512,000 meant that once a conversation crossed that line, *every* turn was reported as `pi-ai detected context overflow` — even though the model answered normally and the reply was discarded. The window is now declared as the ceiling the gateway actually enforces.
+- The built-in fallback roster (used when the catalog cannot be fetched) now declares 1,000,000 as well.
+- Four new cases pin the window parsing: the largest of `context_window_options` wins, upstream's `limit.context` is the fallback, a malformed option list is ignored, and the fallback roster does not claim 512k.
+
+### 0.5.3 (2026-10-25)
+
+- **Fixed image input**: an upload used to fail every time — `pi-ai image input requires the durable attachment service` (`UNSUPPORTED_CONTENT`). The catalog always declared `image` among a model's input modalities, but the plugin built its `PiAiAdapter` without `resolveAttachments` or `resolveImageAccess`, so pi-ai had no way to turn a durable attachment reference into request bytes and refused the call before it reached the gateway. The plugin now reads the host's `attachments` service on demand and maps an attachment's host path into the current execution world with `fs.processPathFromHostPath`; a host that offers no attachment service gets `undefined` for both hooks, and text-only calls are unchanged.
+- New `tests/image-input.spec.ts` pins how both hooks resolve, the path mapping, and the "no attachment service still loads" case; `peerDependencies` gains `@deepseek-ai/dsh-attachment`.
+- Measured against the live gateway: a 2×2 PNG sent to `M3.1-Flash` came back answered correctly (`inputTokens: 316`).
+
 ### 0.5.2 (2026-10-21)
 
 - **An icon for the plugin page**: a new `icon.svg` (48×48, under 800 bytes) ships with the package, declared as `"icon": "./icon.svg"` in `package.json`. The host reads it as a data URL and draws it on the plugin page instead of the default placeholder.
@@ -442,6 +459,12 @@ dsh plugin --profile web exec dsh-workbuddy-connect status
 
 > This plugin's version must match your DSH core exactly (each core line uses a different version); see the version table in its README before installing.
 
+## Contributors
+
+- **[eghrhegpe](https://github.com/eghrhegpe)** — raised the request for a plugin-page icon and Chinese/English copy ([#1](https://github.com/JeffTsuiCUG/dsh-connect-minimaxcode/issues/1)). It became 0.5.2: the plugin page has its own SVG artwork, the title and the introduction come in both languages, and the interface follows the host's language.
+- **[qzz19970504](https://github.com/qzz19970504)** — found and fixed the wiring gap that made every image upload fail ([#3](https://github.com/JeffTsuiCUG/dsh-connect-minimaxcode/pull/3)): the catalog declared the `image` modality, but the adapter never connected the host's durable attachment service, so pi-ai refused image requests before they were sent. The fix became 0.5.3.
+- Plugin author: **[JeffTsuiCUG](https://github.com/JeffTsuiCUG)**
+
 ## Acknowledgements
 
 - **[corrinehu/dsh-workbuddy-connect](https://github.com/corrinehu/dsh-workbuddy-connect)** — **the inspiration for this plugin**. The overall idea of "reuse the desktop app's local login state and bring the already-signed-in model into DSH" was established first by that project, and this plugin follows the same approach; the read-only reuse of credentials was designed with it in mind.

package/README.md

@@ -50,6 +50,7 @@
 - **只读为主**：除「续期已过期的桌面 App 登录态」和「你自己发起的登录」外，不写入、不修改 MiniMax 的任何文件
 - **跟随所选模型**：余额显示在**输入框工具行（模型选择器旁）**，且**只在选中 MiniMax 模型时出现** —— 切到其他反代插件的模型就整体隐藏，不会拿着 MiniMax 的余额误导你
 - **插件页有自己的图标与中英文**：随包发布一张 48×48 的 SVG，插件页不再显示默认占位图；标题与简介按宿主语言给中英文两版，界面文案也跟着宿主语言走
+- **能看图**：模型目录里声明了 `image` 输入模态的模型（`M3`、`M3.1-Flash`）可以直接发图；图片由宿主的附件服务读成请求字节，插件只负责把持久化的附件引用接上去
 
 ## 前置条件
 
@@ -60,6 +61,8 @@
 
 另外需要 DSH 桌面版（内置内核 `>=0.2.1-rc.2 <0.3.0`，见 `package.json` 的 `engines.dsh`；实测 `0.2.1-rc.2`）或 Web profile。
 
+> **发图需要宿主提供附件服务**：图片不是插件自己读的，而是由宿主的持久化附件服务读成请求字节，插件只负责把附件引用接上去。宿主没有附件服务时，插件照常加载、文本照常对话，只是发图会被拒绝（`pi-ai image input requires the durable attachment service`）。
+
 ## 安装
 
 **方式一：从 Git 仓库安装**（推荐，拿到最新代码）
@@ -299,10 +302,24 @@ OAuth 令牌有效期约 1 小时，由客户端自动续期。插件**只在自
 - **自动签到默认关闭**，且只在服务端明确说「当天可领」时才发请求、每天至多一次。是否开启、以及这样做是否符合 MiniMax 的服务条款，请自行判断。
 - **不提供设置界面卡片**。插件不在 DSH 设置页注册卡片；界面只在输入框工具行增加一枚胶囊（**仅选中本插件的模型时显示**），以及插件页面里的一张小卡片。
 - **上游接口非公开**，MiniMax 客户端更新后可能失效。余额接口失效时胶囊会显示「积分不可用」，不会影响模型对话。
+- **发图依赖宿主的附件服务**。模型目录声明了 `image` 模态（`M3`、`M3.1-Flash`），但把持久化附件读成请求字节是宿主的活儿：宿主没有附件服务时插件仍能加载并服务文本，只是带图的请求会被拒绝。
+- **上下文窗口按网关的真实上限声明（1,000,000），不是目录里的 512,000**。上游 `limit.context` 是**默认档位与计费分界**（输入超过 512k 的部分按更高倍率计费），`context_window_options` 里最大的一档才是网关执行的硬上限。实测（2026-10-25）：946,206 tokens 正常受理，1,050,000 被拒（`invalid params, context window exceeds limit (2013)`）。此前按 512,000 声明，长会话一旦过线就**每一轮**都被判成「上下文溢出」，模型其实答得好好的，回复却被丢掉（`pi-ai detected context overflow`）。
 - **设备码流需要能访问 `account.minimax.cn`**，且授权时要在浏览器里能登录 MiniMax 账号。某些内网或受限网络下打不开授权页。
 
 ## 更新记录
 
+### 0.5.4（2026-10-29）
+
+- **修复长会话被误判为「上下文溢出」**：上游 `limit.context`（512,000）是**默认档位与计费分界**，网关实际执行的上限是 `context_window_options` 里的 1,000,000。此前按 512,000 声明，会话一旦超过这条线，**每一轮**都会被判定为 `pi-ai detected context overflow` —— 而模型其实正常答了，回复却被丢弃。现在按网关的真实上限声明窗口。
+- 兜底名录（离线时的内置猜测）同步改为 1,000,000。
+- 新增 4 个用例钉住窗口解析：取 `context_window_options` 的最大值、上游没给就回落 `limit.context`、非法列表被忽略、兜底名录的窗口不是 512k。
+
+### 0.5.3（2026-10-25）
+
+- **修复图片输入**：此前发图必然失败 —— `本轮运行失败 pi-ai image input requires the durable attachment service`（`UNSUPPORTED_CONTENT`）。模型目录一直把 `image` 写进模型的输入模态，但插件构造 `PiAiAdapter` 时没有给 `resolveAttachments` 与 `resolveImageAccess` 两个钩子，pi-ai 因此无法把持久化的附件引用变成请求字节，在发往网关之前就拒绝了调用。现在插件按需读取宿主的 `attachments` 服务，并用 `fs.processPathFromHostPath` 把附件的宿主路径映射进当前执行世界；宿主没有提供附件服务时这两个钩子为 `undefined`，纯文本调用不受影响。
+- 新增 `tests/image-input.spec.ts`，钉住两个钩子的解析、路径映射，以及「宿主没有附件服务仍能加载」；`peerDependencies` 增加 `@deepseek-ai/dsh-attachment`。
+- 实测：真实网关 + `M3.1-Flash` 发一张 2×2 PNG，回答正确（`inputTokens: 316`）。
+
 ### 0.5.2（2026-10-21）
 
 - **插件页图标**：新增随包发布的 `icon.svg`（48×48，800 字节内），并在 `package.json` 里声明 `"icon": "./icon.svg"`。宿主会把它读成 data URL 画在插件页卡片上，不再是默认占位图。
@@ -448,6 +465,12 @@ dsh plugin --profile web exec dsh-workbuddy-connect status
 
 > 该插件的版本需与你的 DSH 内核严格对应（不同内核线用不同版本），安装前请见其 README 的版本对应表。
 
+## 贡献者
+
+- **[eghrhegpe](https://github.com/eghrhegpe)** —— 提出了插件页图标与中英文文案的建议（[#1](https://github.com/JeffTsuiCUG/dsh-connect-minimaxcode/issues/1)）。这条建议促成了 0.5.2：插件页有了自己的 SVG 图标，标题与简介有了中英两版，界面文案也开始跟随宿主语言。
+- **[qzz19970504](https://github.com/qzz19970504)** —— 定位并修复了「发图必然失败」的接线缺口（[#3](https://github.com/JeffTsuiCUG/dsh-connect-minimaxcode/pull/3)）：模型目录声明了 `image` 模态，但适配器没有把宿主的持久化附件服务接上，导致带图请求被 pi-ai 在发出前拒掉。该修复成为 0.5.3。
+- 插件作者：**[JeffTsuiCUG](https://github.com/JeffTsuiCUG)**
+
 ## 致谢
 
 - **[corrinehu/dsh-workbuddy-connect](https://github.com/corrinehu/dsh-workbuddy-connect)** — **本插件的灵感来源**。「复用桌面 App 本机登录态、把已登录模型接进 DSH」的整体思路由该项目最早确立，本插件沿用同一做法；凭据只读复用的设计也参照于此。

package/lib/index.d.ts

@@ -1,5 +1,6 @@
 import "@earendil-works/pi-ai";
 import { LlmAdapter } from "@deepseek-ai/dsh-llm";
+import { PiAiAdapterOptions } from "@deepseek-ai/dsh-llm-pi-ai";
 import { Context } from "@deepseek-ai/cordis";
 //#region src/auth.d.ts
 /** File the desktop app writes its runtime auth context into. */
@@ -144,6 +145,11 @@ declare const MINIMAXCODE_PROVIDER_ID = "minimaxcode";
  * when upstream retired `MiniMax-M2.7` and `MiniMax-M2.7-highspeed` and
  * promoted the model previously served anonymously (internally "space bunny")
  * as `MiniMax-M3.1-Flash`.
+ *
+ * The window is 1,000,000 rather than the 512,000 the upstream `limit` block
+ * reports: 512k is the billing tier boundary, and the gateway itself was
+ * measured serving 946,206 tokens before refusing ~1,050,000 with
+ * `invalid params, context window exceeds limit (2013)`.
  */
 declare const FALLBACK_MODELS: readonly CatalogModel[];
 /** Where the currently served list came from. */
@@ -911,6 +917,17 @@ interface AdapterOptions {
    * transcript, so what is reported is what was actually billed.
    */
   onUsage?: (model: string, usage: ModelUsage) => void;
+  /**
+   * Resolve the host's durable attachment service at request time.
+   *
+   * Required for image input: without it the pi-ai adapter rejects any message
+   * carrying an image with `UNSUPPORTED_CONTENT`, because it cannot turn a
+   * durable reference into request bytes. Optional here only so a host that
+   * provides no such service can still load this plugin and serve text.
+   */
+  resolveAttachments?: PiAiAdapterOptions['resolveAttachments'];
+  /** Bridge one attachment into the current tool execution world. */
+  resolveImageAccess?: PiAiAdapterOptions['resolveImageAccess'];
 }
 /** What {@link createMinimaxCodeAdapter} hands back. */
 interface AdapterBundle {

package/lib/index.js

@@ -4,7 +4,7 @@ import { homedir } from "node:os";
 import { dirname, join } from "node:path";
 import { createProvider } from "@earendil-works/pi-ai";
 import { anthropicMessagesApi } from "@earendil-works/pi-ai/api/anthropic-messages.lazy";
-import { LlmAdapter, resolveRetryPolicy } from "@deepseek-ai/dsh-llm";
+import { LlmAdapter, resolveImageAttachmentAccess, resolveRetryPolicy } from "@deepseek-ai/dsh-llm";
 import { PiAiAdapter } from "@deepseek-ai/dsh-llm-pi-ai";
 import { createHash, randomBytes } from "node:crypto";
 //#region src/auth.ts
@@ -238,11 +238,16 @@ const MINIMAXCODE_PROVIDER_ID = "minimaxcode";
 * when upstream retired `MiniMax-M2.7` and `MiniMax-M2.7-highspeed` and
 * promoted the model previously served anonymously (internally "space bunny")
 * as `MiniMax-M3.1-Flash`.
+*
+* The window is 1,000,000 rather than the 512,000 the upstream `limit` block
+* reports: 512k is the billing tier boundary, and the gateway itself was
+* measured serving 946,206 tokens before refusing ~1,050,000 with
+* `invalid params, context window exceeds limit (2013)`.
 */
 const FALLBACK_MODELS = [{
 	id: "MiniMax-M3",
 	name: "M3",
-	contextWindow: 512e3,
+	contextWindow: 1e6,
 	maxOutputTokens: 128e3,
 	toolCall: true,
 	reasoning: true,
@@ -255,7 +260,7 @@ const FALLBACK_MODELS = [{
 }, {
 	id: "MiniMax-M3.1-Flash",
 	name: "M3.1-Flash",
-	contextWindow: 512e3,
+	contextWindow: 1e6,
 	maxOutputTokens: 128e3,
 	toolCall: true,
 	reasoning: true,
@@ -278,12 +283,42 @@ function sameCatalog(a, b) {
 function usable(id, raw) {
 	return id.length > 0 && id.startsWith("MiniMax-") && typeof raw === "object" && raw !== null;
 }
+/**
+* The window the gateway actually enforces.
+*
+* Upstream reports two different numbers, and they mean different things:
+* `limit.context` is the *default* window — the billing tier boundary, above
+* which input costs more — while `context_window_options` lists every window
+* the gateway will accept, the largest being the hard ceiling.
+*
+* Measured against the live gateway on 2026-10-25: a prompt of 946,206 tokens
+* was served normally, and one of ~1,050,000 was refused with
+* `invalid params, context window exceeds limit (2013)`. The ceiling is
+* therefore the largest option, not `limit.context`.
+*
+* Declaring the default instead of the ceiling is not a cosmetic difference:
+* the adapter compares the gateway's own token count against this number, so a
+* conversation that runs past 512k is reported as a context overflow on every
+* turn even though the gateway answers it perfectly well.
+*
+* @param raw - The upstream model entry.
+* @param limit - The upstream `limit` block, already read.
+* @returns The window to declare, in tokens.
+*/
+function readContextWindow(raw, limit) {
+	const options = raw.context_window_options;
+	if (Array.isArray(options)) {
+		const sizes = options.filter((value) => typeof value === "number" && Number.isInteger(value) && value > 0);
+		if (sizes.length > 0) return Math.max(...sizes);
+	}
+	return typeof limit.context === "number" ? limit.context : 2e5;
+}
 /** Read the fields the upstream sends, tolerating their absence. */
 function readModel(id, raw) {
 	const limit = raw.limit ?? {};
 	const modalities = raw.modalities ?? {};
 	const input = Array.isArray(modalities.input) ? modalities.input.filter((m) => typeof m === "string") : ["text"];
-	const contextWindow = typeof limit.context === "number" ? limit.context : 2e5;
+	const contextWindow = readContextWindow(raw, limit);
 	return {
 		id,
 		name: typeof raw.name === "string" ? raw.name : id,
@@ -613,7 +648,9 @@ function createMinimaxCodeAdapter(options) {
 	const inner = new PiAiAdapter({
 		profiles: () => profiles,
 		resolveApiKey: options.resolveApiKey,
-		auth: INERT_AUTH
+		auth: INERT_AUTH,
+		...options.resolveAttachments === void 0 ? {} : { resolveAttachments: options.resolveAttachments },
+		...options.resolveImageAccess === void 0 ? {} : { resolveImageAccess: options.resolveImageAccess }
 	});
 	return {
 		adapter: options.onUsage === void 0 ? inner : new MeteredAdapter(inner, options.onUsage),
@@ -623,6 +660,30 @@ function createMinimaxCodeAdapter(options) {
 	};
 }
 //#endregion
+//#region src/image-input.ts
+/**
+* Build the two image-input hooks from the services the host provides.
+*
+* Both services are read per request through `ctx.get` rather than declared as
+* startup injections: they may be provided after this plugin loads, and a host
+* without them must still start and serve text. A host that provides no
+* attachment store simply keeps refusing images, which is the old behaviour
+* rather than a new failure.
+*
+* `resolveImageAccess` is what lets a model read the normalized image from a
+* tool as well as see it inline. It degrades to `undefined` when the host
+* exposes no host-path mapping, and the adapter then reports no local copy.
+*
+* @param ctx - the plugin fiber whose services are read.
+* @returns The optional {@link AdapterOptions} hooks, ready to spread.
+*/
+function imageInputHooks(ctx) {
+	return {
+		resolveAttachments: () => ctx.get("attachments"),
+		resolveImageAccess: (attachments, ref) => resolveImageAttachmentAccess(attachments, (hostPath) => ctx.get("fs")?.processPathFromHostPath(hostPath), ref)
+	};
+}
+//#endregion
 //#region src/status-route.ts
 /** Route this plugin mounts, matching its settings card. */
 const MINIMAX_STATUS_PATH = "/plugins/dsh-connect-minimaxcode/status";
@@ -2159,6 +2220,7 @@ function apply(ctx) {
 		region: () => region,
 		resolveApiKey: async () => (await readNow())?.token,
 		bearerToken: () => modelToken,
+		...imageInputHooks(ctx),
 		onUsage: (model, usage) => {
 			meter.record(model, usage);
 			refreshCredits();

package/package.json

@@ -2,8 +2,22 @@
   "name": "dsh-connect-minimaxcode",
   "displayName": "DSH Connect MiniMax Code",
   "description": "把本机已登录的 MiniMax Code 桌面 App 模型接入 DeepSeek Harness：零配置、Anthropic Messages 兼容、动态模型目录，并显示积分余额与到期 — bring MiniMax Code desktop app models into DSH with a dynamic model catalog and a credit balance with expiry details.",
-  "version": "0.5.2",
+  "version": "0.5.4",
   "author": "JeffTsuiCUG",
+  "contributors": [
+    {
+      "name": "JeffTsuiCUG",
+      "url": "https://github.com/JeffTsuiCUG"
+    },
+    {
+      "name": "eghrhegpe",
+      "url": "https://github.com/eghrhegpe"
+    },
+    {
+      "name": "qzz19970504",
+      "url": "https://github.com/qzz19970504"
+    }
+  ],
   "license": "MIT",
   "private": false,
   "publishConfig": {
@@ -65,6 +79,7 @@
   "peerDependencies": {
     "@deepseek-ai/cordis": ">=4.0.4 <5.0.0",
     "@deepseek-ai/dsh-atomic-write": ">=0.1.7-rc.1 <0.3.0-0",
+    "@deepseek-ai/dsh-attachment": ">=0.1.7-rc.1 <0.3.0-0",
     "@deepseek-ai/dsh-home-paths": ">=0.1.7-rc.1 <0.3.0-0",
     "@deepseek-ai/dsh-host-webserver": ">=0.1.7-rc.1 <0.3.0-0",
     "@deepseek-ai/dsh-llm": ">=0.1.7-rc.1 <0.3.0-0",