@siliconflow-official/dsh-llm-siliconflow 0.1.0-rc.7 → 0.2.0-rc.2

package/README.md

@@ -1,6 +1,6 @@
 # @siliconflow-official/dsh-llm-siliconflow
 
-面向 [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) LLM 接缝的 SiliconFlow chat-completions 适配器插件：用直接 `fetch` + SSE（由 `eventsource-parser` 分帧）把 SiliconFlow 的 OpenAI 兼容线上格式翻译成 `StreamChunk` 协议。SiliconFlow 托管着广泛的开源模型目录，其中包含 delta 携带 `reasoning_content` 的推理模型（DeepSeek-R1、QwQ、Kimi-K2-Thinking）——适配器把该通道翻译成 harness reasoning 块，并在工具调用轮次按这些模型的要求将其回传。
+面向 [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) LLM 接缝的 SiliconFlow chat-completions 适配器插件：用直接 `fetch` + SSE（由 `eventsource-parser` 分帧）把 SiliconFlow 的 OpenAI 兼容线上格式翻译成 `StreamChunk` 协议。SiliconFlow 托管着广泛的开源模型目录，其中包含 delta 携带 `reasoning_content` 的推理模型（DeepSeek-R1、QwQ、Kimi-K2-Thinking）——适配器把该通道翻译成 harness reasoning 块，并在工具调用轮次按这些模型的要求将其回传。同时支持视觉语言模型（VLM）：当 harness 挂载了附件存储服务（`ctx.attachments`）时，用户消息中的图片块会被解析为 base64 data URL 并序列化为 OpenAI 兼容的 `image_url` 多模态内容部分。
 
 本包拥有 `siliconflow` 提供方路由，因此部署只需提供一个 SiliconFlow API key 即可使用。其模型选择器从实时 `GET /models?sub_type=chat` 列表按端点顺序填充；配置的 `models` 列表是在没有 key 或发现失败时展示的回退目录。这是一个纯 OpenAI 兼容端点，没有 `thinking`/`reasoning_effort` 开关，因此适配器不暴露任何推理档位元数据、也不序列化任何推理档位字段：推理模型通过其目录 id 选择，请求上显式指定 `reasoningEffort` 会在网络 I/O 前以 `UNSUPPORTED_REASONING_EFFORT` 被拒绝。为 `siliconflow` 注册另一个适配器会抛出 `LlmError('DUPLICATE_ADAPTER')`。
 
@@ -34,6 +34,14 @@ dsh-siliconflow-setup
 export SILICONFLOW_API_KEY=sk-...   # 或写入 $DSH_HOME/.credentials.yaml
 ```
 
+写入 `.credentials.yaml` 时使用 dsh ≥ 0.1.2 的 version-1 布局（`version: 1` + `refs:` 嵌套）；向导会识别并就地升级本向导早期版本写入的 pre-release 顶层扁平布局，其他无法识别的文档会明确报错而不是被改写：
+
+```yaml
+version: 1
+refs:
+  SILICONFLOW_API_KEY: sk-...
+```
+
 在包发布到 npm 之前，先用 `pnpm install && pnpm build` 构建出 `lib/`，再从本地路径安装：
 
 ```sh
@@ -90,7 +98,9 @@ nohup npx @deepseek-ai/dsh --profile headless "任务" > ~/.dsh/task.log 2>&1 &
       - id: deepseek-ai/DeepSeek-V4-Flash
 ```
 
-插件把单个提供方路由 `siliconflow` 连同其已解析的 `retryPolicy` 一并注册。请求用 `provider: siliconflow` 选中它；其 `model` 原样作为线上 `model` 字符串透传，因此更换 SiliconFlow 模型不需要生命周期级重新注册。线上模型 id 是 SiliconFlow 的 `org/model` 写法（如 `deepseek-ai/DeepSeek-V4-Flash`），绝不是短别名。省略 `models` 时保留一份由六个当前托管对话模型组成的小回退目录；显式列表会替换这些默认值，而 `models: []` 则一个都不通告。目录条目通过 `ctx.llm.listModels('siliconflow')` 暴露给 ACP 编辑器与 Web 选择器这类客户端，但始终是建议性的：未列出的模型 id 依然原样透传。省略的条目名默认等于其 id。
+目录条目可携带 `inputModalities` 字段（`['text']` 或 `['text', 'image']`）以显式声明模型接受的输入模态；省略时适配器从内置的权威 VLM 模型集合（从模型广场 `vlm` 属性提取，静态维护，最后更新 2026-08-22）查找。
+
+插件把单个提供方路由 `siliconflow` 连同其已解析的 `retryPolicy` 一并注册。请求用 `provider: siliconflow` 选中它；其 `model` 原样作为线上 `model` 字符串透传，因此更换 SiliconFlow 模型不需要生命周期级重新注册。线上模型 id 是 SiliconFlow 的 `org/model` 写法（如 `deepseek-ai/DeepSeek-V4-Flash`），绝不是短别名。省略 `models` 时保留一份由当前托管对话模型（含 VLM）组成的回退目录；显式列表会替换这些默认值，而 `models: []` 则一个都不通告。目录条目通过 `ctx.llm.listModels('siliconflow')` 暴露给 ACP 编辑器与 Web 选择器这类客户端，但始终是建议性的：未列出的模型 id 依然原样透传。省略的条目名默认等于其 id。
 
 `contextWindow` 按模型可选。`ctx.llm.resolveModelInfo('siliconflow', model).context` 先返回精确值——来自配置条目或温热的发现缓存——再对未被任何来源定容的模型回退到 `defaultContextWindow`。适配器默认值是 32,768；SiliconFlow 目录大致横跨 8k 到 1M 上下文（GLM-5.2、DeepSeek-V4-Pro/Flash 均支持 1M），因此披露了上下文的发现列表是权威值，回退值仅在没有任何来源披露时使用。对压力敏感的插件由此获得部署自有的容量，而不把模型选择器当作权威。
 
@@ -161,10 +171,11 @@ nohup npx @deepseek-ai/dsh --profile headless "任务" > ~/.dsh/task.log 2>&1 &
 
 ## Known Limitations and Deferred Work
 
-- **回退 `models` 列表是手工维护的** —— 六个默认值是一份小快照，仅在发现无法运行时展示；实时列表才是权威目录。
+- **回退 `models` 列表是手工维护的** —— 默认值是一份快照，仅在发现无法运行时展示；实时列表才是权威目录。
 - **发现不跨 baseURL 变化缓存** —— 缓存按端点键控，因此改指路由会在下一次 `listModels` 重新询问。
 - **settings 的 `models` 列表整体替换组合列表** —— settings 层合并在字段粒度进行，数组是一个字段；按条目合并目录需要键控结构。
 - **未映射 `tool_choice`** —— 不属于核心词汇表（MVP 裁剪，与 pi-ai 和 DeepSeek 双胞胎相同）。
 - **请求使用原始 `fetch`，而非 `@cordisjs/plugin-http`** —— 没有共享代理/拦截配置；待有第二个直接 fetch 适配器需要时再采用（`TODO(http)`）。
 - **序列化把 user 与 tool-result 内容扁平化为文本块** —— 插件添加的块类型被跳过，空工具输出以字面 `(no output)` 上线。
-- **图片内容被拒绝** —— 这里的 chat-completions 线上路由仅文本；多模态 SiliconFlow 路由需要自己的内容序列化器。
+- **VLM 模型集与回退目录的 contextWindow 均为静态数据** —— SiliconFlow 的 OpenAI 兼容 `GET /models` API 不返回模型的 `vlm` 属性或 `contextLen` 字段，因此适配器内置了从模型广场（siliconflow.cn/models）SSR 数据提取的权威 VLM 模型 ID 集合（当前 23 个）和回退目录的上下文窗口配置。这些静态数据会随模型广场更新而定期同步，直至 OpenAPI 本身暴露相关属性（如 `sub_type=vision` 或模型列表中的 `vlm` 字段）后被运行时查询取代。目录条目可通过 `inputModalities` 字段显式声明以覆盖静态集合。
+- **图片附件依赖存储服务** —— 用户消息中的图片块通过 `ctx.attachments` 解析为 base64 data URL；未挂载该服务时图片被替换为 `[image omitted]` 占位文本，请求仍可继续。工具结果中的图片始终被替换为占位文本。

package/lib/bin.js

@@ -1,11 +1,11 @@
 #!/usr/bin/env node
-import { a as PUBLIC_BASE_URL, h as discoverChatModels, i as PROVIDER, n as DEFAULT_API_KEY_ENV, r as DEFAULT_MODELS } from "./types-CmhCs_oH.js";
+import { a as PUBLIC_BASE_URL, g as discoverChatModels, i as PROVIDER, n as DEFAULT_API_KEY_ENV, r as DEFAULT_MODELS } from "./types-iMqgaeX4.js";
 import { createInterface } from "node:readline/promises";
 import { stdin, stdout } from "node:process";
 import { mkdir, readFile, writeFile } from "node:fs/promises";
 import { homedir } from "node:os";
 import { dirname, join, resolve } from "node:path";
-import { parseDocument } from "yaml";
+import { Pair, YAMLMap, parseDocument } from "yaml";
 //#region lib/types/setup.js
 /**
 * Interactive setup wizard core for {@link @siliconflow-official/dsh-llm-siliconflow}:
@@ -45,11 +45,81 @@ function settingsPath(home) {
 function isEnoent(error) {
 	return error.code === "ENOENT";
 }
+/** The credentials document layout version this wizard reads and writes. */
+const CREDENTIALS_LAYOUT_VERSION = 1;
+/** POSIX identifier rule the credentials seam addresses references by. */
+const REF_NAME_PATTERN = /^[A-Za-z_][A-Za-z0-9_]*$/;
 /**
-* Read one credential reference from a comment-preserving document.
+* Recognize one parsed credentials document.
+* @param document - the parsed document; parse errors make it unrecognizable.
+* @returns the admitted layout with its references — plus, under `extraKeys`,
+*   any top-level key a pre-fix release of this wizard left beside `version`,
+*   held out of `entries` so reads treat them as absent while the write path
+*   folds them back under `refs` — or `undefined` when the document does not
+*   parse, holds an unknown `version`, carries a `refs` section that is not a
+*   mapping of reference names to non-empty strings, or holds any other
+*   top-level entry this build cannot attribute to its own predecessor.
+*/
+function admitCredentialsDocument(document) {
+	if (document.errors.length > 0) return void 0;
+	const root = document.toJS() ?? {};
+	if (typeof root !== "object" || root === null) return void 0;
+	const fields = root;
+	const entries = /* @__PURE__ */ new Map();
+	if (!("version" in fields)) {
+		for (const [key, value] of Object.entries(fields)) {
+			if (!REF_NAME_PATTERN.test(key)) return void 0;
+			if (typeof value !== "string" || value.length === 0) return void 0;
+			entries.set(key, value);
+		}
+		return {
+			kind: "flat",
+			entries
+		};
+	}
+	if (fields.version !== CREDENTIALS_LAYOUT_VERSION) return void 0;
+	const refs = fields["refs"];
+	if (refs !== void 0) {
+		if (typeof refs !== "object" || refs === null) return void 0;
+		for (const [key, value] of Object.entries(refs)) {
+			if (!REF_NAME_PATTERN.test(key)) return void 0;
+			if (typeof value !== "string" || value.length === 0) return void 0;
+			entries.set(key, value);
+		}
+	}
+	const extraKeys = /* @__PURE__ */ new Map();
+	for (const [key, value] of Object.entries(fields)) {
+		if (key === "version" || key === "refs" || key === "records") continue;
+		if (!REF_NAME_PATTERN.test(key) || typeof value !== "string" || value.length === 0) return void 0;
+		extraKeys.set(key, value);
+	}
+	return {
+		kind: "versioned",
+		entries,
+		...extraKeys.size > 0 ? { extraKeys } : {}
+	};
+}
+/**
+* Nest a pre-release flat document under `refs:` with a `version` stamp. The
+* whole current root — every flat entry, comments included — becomes the
+* `refs` section verbatim, so every other provider's key migrates in the same
+* write; the new root carries `version` and `refs` in that order, matching how
+* `dsh-credentials-local` writes one from scratch.
+* @param document - the admitted flat document to upgrade in place.
+*/
+function upgradeFlatDocument(document) {
+	const refs = document.contents instanceof YAMLMap ? document.contents : new YAMLMap();
+	const root = new YAMLMap();
+	root.add(new Pair("version", CREDENTIALS_LAYOUT_VERSION));
+	root.add(new Pair("refs", refs));
+	document.contents = root;
+}
+/**
+* Read one credential reference from a versioned or pre-release flat document.
 * @param path - the credentials document path.
-* @param keyEnv - the top-level reference name (e.g. `SILICONFLOW_API_KEY`).
-* @returns the stored value, or `undefined` when absent or the file is missing.
+* @param keyEnv - the reference name to read (e.g. `SILICONFLOW_API_KEY`).
+* @returns the stored value, or `undefined` when the reference is absent, the
+*   file is missing, or the document is not a recognizable layout.
 */
 async function readCredential(path, keyEnv) {
 	let text;
@@ -59,18 +129,31 @@ async function readCredential(path, keyEnv) {
 		if (isEnoent(error)) return void 0;
 		throw error;
 	}
-	const value = parseDocument(text).toJS()?.[keyEnv];
-	return typeof value === "string" && value.length > 0 ? value : void 0;
+	const admitted = admitCredentialsDocument(parseDocument(text));
+	if (admitted === void 0) return void 0;
+	return admitted.entries.get(keyEnv);
 }
 /**
-* Set one credential reference, preserving every other entry and comment.
+* Write one credential reference into the versioned layout, preserving every
+* other entry and comment. A pre-release flat document is upgraded in place,
+* and a top-level key a pre-fix release of this wizard left beside `version`
+* is folded back under `refs`, so one run repairs the file its predecessor
+* corrupted. An unrecognized document fails loud instead of being rewritten —
+* a silent rewrite would hide why the running harness rejects it.
 * @param path - the credentials document path; created when absent.
-* @param keyEnv - the top-level reference name to write.
+* @param keyEnv - the reference name to write.
 * @param key - the value.
+* @throws when the document exists but is not a recognizable credentials layout.
 */
 async function writeCredential(path, keyEnv, key) {
 	const doc = await loadDocument(path);
-	doc.set(keyEnv, key);
+	const admitted = admitCredentialsDocument(doc);
+	if (admitted === void 0) throw new Error(`setup: ${path} is not a recognizable credentials document (expected version 1 with a refs section, or the pre-release flat layout); fix it before running setup`);
+	if (admitted.kind === "flat") upgradeFlatDocument(doc);
+	for (const extra of admitted.extraKeys?.keys() ?? []) doc.deleteIn([extra]);
+	for (const [extra, value] of admitted.extraKeys ?? /* @__PURE__ */ new Map()) doc.setIn(["refs", extra], value);
+	doc.setIn(["version"], CREDENTIALS_LAYOUT_VERSION);
+	doc.setIn(["refs", keyEnv], key);
 	await persistDocument(path, doc);
 }
 /**

package/lib/index.js

@@ -1,2 +1,2 @@
-import { _ as readListing, a as PUBLIC_BASE_URL, c as name, d as DEFAULT_MAX_TOKENS, f as DEFAULT_STREAM_IDLE_TIMEOUT_MS, g as listingUrl, h as discoverChatModels, i as PROVIDER, l as resolveAdapterOptions, m as SiliconFlowAdapter, n as DEFAULT_API_KEY_ENV, o as apply, p as DISCOVERY_TTL_MS, r as DEFAULT_MODELS, s as inject, t as Config, u as DEFAULT_CONTEXT_WINDOW } from "./types-CmhCs_oH.js";
-export { Config, DEFAULT_API_KEY_ENV, DEFAULT_CONTEXT_WINDOW, DEFAULT_MAX_TOKENS, DEFAULT_MODELS, DEFAULT_STREAM_IDLE_TIMEOUT_MS, DISCOVERY_TTL_MS, PROVIDER, PUBLIC_BASE_URL, SiliconFlowAdapter, apply, discoverChatModels, inject, listingUrl, name, readListing, resolveAdapterOptions };
+import { _ as listingUrl, a as PUBLIC_BASE_URL, c as name, d as DEFAULT_MAX_TOKENS, f as DEFAULT_STREAM_IDLE_TIMEOUT_MS, g as discoverChatModels, h as inferInputModalities, i as PROVIDER, l as resolveAdapterOptions, m as SiliconFlowAdapter, n as DEFAULT_API_KEY_ENV, o as apply, p as DISCOVERY_TTL_MS, r as DEFAULT_MODELS, s as inject, t as Config, u as DEFAULT_CONTEXT_WINDOW, v as readListing } from "./types-iMqgaeX4.js";
+export { Config, DEFAULT_API_KEY_ENV, DEFAULT_CONTEXT_WINDOW, DEFAULT_MAX_TOKENS, DEFAULT_MODELS, DEFAULT_STREAM_IDLE_TIMEOUT_MS, DISCOVERY_TTL_MS, PROVIDER, PUBLIC_BASE_URL, SiliconFlowAdapter, apply, discoverChatModels, inferInputModalities, inject, listingUrl, name, readListing, resolveAdapterOptions };

package/lib/types/adapter.d.ts

@@ -8,8 +8,9 @@
  * @module dsh-llm-siliconflow/adapter
  */
 import { LlmAdapter } from '@deepseek-ai/dsh-llm';
-import type { GenerateOptions, LlmModelInfo, LlmProviderInfo, LlmResolvedModelInfo, ResolvedRetryPolicy, StreamChunk } from '@deepseek-ai/dsh-llm';
+import type { GenerateOptions, LlmModelInfo, LlmProviderInfo, LlmResolvedModelInfo, ModelModality, ResolvedRetryPolicy, StreamChunk } from '@deepseek-ai/dsh-llm';
 import type { CredentialRef } from '@deepseek-ai/dsh-credentials';
+import type { ImageAttachmentRef } from '@deepseek-ai/dsh-attachment';
 import type { AnonymousUserId } from '@deepseek-ai/dsh-anonymous-user-id';
 import type { WireError } from './types.ts';
 /** One optional model entry advertised by the direct-fetch adapter. */
@@ -24,6 +25,8 @@ export interface SiliconFlowCatalogModel {
     contextWindow?: number;
     /** Per-request output cap for this model; omission falls back to the profile's {@link SiliconFlowConnectionOptions.maxTokens}. */
     maxTokens?: number;
+    /** Accepted input modalities; omitted infers from the built-in VLM model set. */
+    inputModalities?: ModelModality[];
 }
 /**
  * Validated connection facts for one operation. The plugin's
@@ -65,6 +68,12 @@ export interface SiliconFlowAdapterOptions {
     resolveApiKey: (connection: SiliconFlowConnectionOptions) => Promise<string>;
     /** Resolve the harness-home anonymous id shared with telemetry and feedback. */
     resolveUserId: () => AnonymousUserId;
+    /**
+     * Resolve one image attachment to a base64 data URL for wire serialization.
+     * Returns `undefined` when the attachment store is unavailable or the read
+     * fails; the serializer substitutes the `OFFLOADED_IMAGE_TEXT` sentinel.
+     */
+    resolveImage?: (ref: ImageAttachmentRef) => Promise<string | undefined>;
 }
 /** Default maximum idle interval while an adapter stream read is outstanding. */
 export declare const DEFAULT_STREAM_IDLE_TIMEOUT_MS = 300000;
@@ -74,6 +83,16 @@ export declare const DEFAULT_CONTEXT_WINDOW = 32768;
 export declare const DEFAULT_MAX_TOKENS = 8192;
 /** How long a cached model-listing discovery stays fresh before the next `listModels` re-interrogates. */
 export declare const DISCOVERY_TTL_MS: number;
+/**
+ * Determine the input modalities for a SiliconFlow model id.
+ *
+ * Uses the authoritative VLM model set from the SiliconFlow marketplace rather
+ * than naming-pattern heuristics. Catalog entries can declare explicit
+ * `inputModalities` to override this lookup.
+ * @param id - the wire model id (e.g. `Qwen/Qwen3-VL-8B-Instruct`).
+ * @returns `['text', 'image']` when the id is a known VLM, `['text']` otherwise.
+ */
+export declare function inferInputModalities(id: string): readonly ModelModality[];
 /**
  * Map an HTTP status to a stable LlmError code.
  * @param status - status of a non-2xx provider response.

package/lib/types/index.d.ts

@@ -8,6 +8,10 @@
  * restarting anything, while an in-flight stream keeps the facts it started
  * with. The one registration-captured fact — the retry policy — re-registers
  * the route in place when it changes.
+ *
+ * When the optional `ctx.attachments` service is mounted, image blocks in user
+ * messages are resolved to base64 data URLs and serialized as OpenAI-compatible
+ * `image_url` content parts, enabling VLM models (Qwen3-VL, GLM-4.5V, etc.).
  * @module @siliconflow-official/dsh-llm-siliconflow
  */
 import type { Context } from '@deepseek-ai/cordis';
@@ -16,6 +20,7 @@ import type { RetryPolicyConfig } from '@deepseek-ai/dsh-llm';
 import { type LaunchEnvironmentSnapshot } from '@deepseek-ai/dsh-launch-environment';
 import type { SiliconFlowCatalogModel, SiliconFlowConnectionOptions } from './adapter.ts';
 export { DEFAULT_CONTEXT_WINDOW, DEFAULT_MAX_TOKENS, DEFAULT_STREAM_IDLE_TIMEOUT_MS, DISCOVERY_TTL_MS, SiliconFlowAdapter, } from './adapter.ts';
+export { inferInputModalities } from './adapter.ts';
 export { discoverChatModels, listingUrl, readListing } from './discovery.ts';
 export type { SiliconFlowListingEntry } from './discovery.ts';
 export type { SiliconFlowAdapterOptions, SiliconFlowCatalogModel, SiliconFlowConnectionOptions } from './adapter.ts';
@@ -26,7 +31,18 @@ export declare const inject: string[];
 export declare const DEFAULT_API_KEY_ENV = "SILICONFLOW_API_KEY";
 /** The single provider route this plugin owns. */
 export declare const PROVIDER = "siliconflow";
-/** Fallback advisory catalog: six widely hosted chat models, also the setup CLI's discovery fallback. */
+/**
+ * Fallback advisory catalog: widely hosted chat models including VLMs, also the
+ * setup CLI's discovery fallback.
+ *
+ * **Static data** — the `contextWindow` values and VLM flags are extracted from
+ * the SiliconFlow model marketplace (siliconflow.cn/models). The live
+ * `GET /models?sub_type=chat` API does not return context window or VLM metadata,
+ * so these fields are maintained by hand. This catalog will be kept in sync with
+ * the marketplace until the API exposes these attributes natively.
+ *
+ * Last updated: 2026-08-22. VLM entries declare `inputModalities` explicitly.
+ */
 export declare const DEFAULT_MODELS: SiliconFlowCatalogModel[];
 /**
  * Plugin config, validated by the same-named schemastery schema and doubling
@@ -44,7 +60,7 @@ export interface Config {
     maxTokens?: number;
     /** Positive context capacity used when the selected model has no exact value (default 32,768). */
     defaultContextWindow?: number;
-    /** Advisory models shown by discovery consumers; defaults to six widely hosted models. */
+    /** Advisory models shown by discovery consumers; defaults to widely hosted models including VLMs. */
     models?: SiliconFlowCatalogModel[];
     /** Maximum provider idle time while one stream read is outstanding (default five minutes). */
     streamIdleTimeoutMs?: number;

package/lib/types/serialize.d.ts

@@ -3,28 +3,43 @@
  * joined; assistant text becomes `content`, tool calls become `tool_calls`,
  * and tool results become separate tool messages. Assistant reasoning is
  * replayed as `reasoning_content` only on tool-call turns, as hosted reasoning
- * models (DeepSeek-R1 and siblings) require. Core image blocks are rejected
- * explicitly because this wire route is text-only; unknown declaration-merged
- * block types retain the adapter's documented extension fallback.
+ * models (DeepSeek-R1 and siblings) require.
+ *
+ * **Multimodal**: image blocks in user messages are serialized as
+ * OpenAI-compatible `image_url` content parts with base64 data URLs. The
+ * serializer resolves image attachments through the optional attachment-store
+ * resolver; when no resolver is available, images are replaced with the
+ * `OFFLOADED_IMAGE_TEXT` sentinel. Tool-result content is flattened to text
+ * (images within tool results are also replaced with the sentinel). Unknown
+ * declaration-merged block types retain the adapter's documented extension
+ * fallback.
+ *
  * @module dsh-llm-siliconflow/serialize
  */
 import type { GenerateOptions, Message } from '@deepseek-ai/dsh-llm';
+import type { ImageAttachmentRef } from '@deepseek-ai/dsh-attachment';
 import type { WireMessage, WireRequest } from './types.ts';
 /**
  * Serialize the conversation. `tool-result` blocks become standalone
  * `{role: 'tool'}` messages; the harness puts each tool result in its own
  * user-role message, so a mixed user message contributes its text first and
  * its tool results as separate wire messages after.
+ *
+ * Image blocks within user messages are resolved through the `resolveImage`
+ * callback; images within tool results are flattened to the sentinel text
+ * (tool results carry structured data, not multimodal content).
  * @param messages - the harness conversation, in order.
+ * @param resolveImage - resolves an image attachment ref to a base64 data URL; returns `undefined` when unavailable.
  * @returns the wire messages; order preserved, each tool result expanded into its own entry.
  */
-export declare function serializeMessages(messages: Message[]): WireMessage[];
+export declare function serializeMessages(messages: Message[], resolveImage?: (ref: ImageAttachmentRef) => Promise<string | undefined>): Promise<WireMessage[]>;
 /**
  * Build the full wire request. Always streaming (`stream: true`, usage
  * reporting on); optional fields are omitted rather than sent as null, so
  * provider defaults apply.
  * @param options - the harness request (model, history, system, tools, sampling).
+ * @param resolveImage - resolves an image attachment ref to a base64 data URL; returns `undefined` when unavailable.
  * @returns the chat-completions request body.
  */
-export declare function serializeRequest(options: GenerateOptions): WireRequest;
+export declare function serializeRequest(options: GenerateOptions, resolveImage?: (ref: ImageAttachmentRef) => Promise<string | undefined>): Promise<WireRequest>;
 //# sourceMappingURL=serialize.d.ts.map

package/lib/types/setup.d.ts

@@ -56,17 +56,24 @@ export declare function credentialsPath(home: string): string;
 /** The settings document path under a harness home. */
 export declare function settingsPath(home: string): string;
 /**
- * Read one credential reference from a comment-preserving document.
+ * Read one credential reference from a versioned or pre-release flat document.
  * @param path - the credentials document path.
- * @param keyEnv - the top-level reference name (e.g. `SILICONFLOW_API_KEY`).
- * @returns the stored value, or `undefined` when absent or the file is missing.
+ * @param keyEnv - the reference name to read (e.g. `SILICONFLOW_API_KEY`).
+ * @returns the stored value, or `undefined` when the reference is absent, the
+ *   file is missing, or the document is not a recognizable layout.
  */
 export declare function readCredential(path: string, keyEnv: string): Promise<string | undefined>;
 /**
- * Set one credential reference, preserving every other entry and comment.
+ * Write one credential reference into the versioned layout, preserving every
+ * other entry and comment. A pre-release flat document is upgraded in place,
+ * and a top-level key a pre-fix release of this wizard left beside `version`
+ * is folded back under `refs`, so one run repairs the file its predecessor
+ * corrupted. An unrecognized document fails loud instead of being rewritten —
+ * a silent rewrite would hide why the running harness rejects it.
  * @param path - the credentials document path; created when absent.
- * @param keyEnv - the top-level reference name to write.
+ * @param keyEnv - the reference name to write.
  * @param key - the value.
+ * @throws when the document exists but is not a recognizable credentials layout.
  */
 export declare function writeCredential(path: string, keyEnv: string, key: string): Promise<void>;
 /**

package/lib/types/types.d.ts

@@ -31,10 +31,22 @@ export interface WireSystemMessage {
     role: 'system';
     content: string;
 }
-/** User-role message: a single string of user input. */
+/** One text part in a multipart user message. */
+export interface WireTextPart {
+    type: 'text';
+    text: string;
+}
+/** One image part in a multipart user message (OpenAI-compatible `image_url`). */
+export interface WireImagePart {
+    type: 'image_url';
+    image_url: {
+        url: string;
+    };
+}
+/** A user-role message: a plain string for text-only, or an array of content parts for multimodal. */
 export interface WireUserMessage {
     role: 'user';
-    content: string;
+    content: string | (WireTextPart | WireImagePart)[];
 }
 /** Tool-role message: the result of one tool call, keyed by its call id. */
 export interface WireToolMessage {

package/lib/{types-CmhCs_oH.js → types-iMqgaeX4.js}

@@ -1,8 +1,8 @@
 import z from "@deepseek-ai/schemastery";
-import { CONTEXT_WINDOW_EXCEEDED_CODE, CallId, EMPTY_RESPONSE_CODE, LlmAdapter, LlmError, ProviderRequestId, QUOTA_EXCEEDED_CODE, RetryPolicySchema, assertUsableApiKey, attributionHeaders, contentHasImage, isContextWindowExceededError, isQuotaExceededError, resolveRetryPolicy } from "@deepseek-ai/dsh-llm";
+import { CONTEXT_WINDOW_EXCEEDED_CODE, EMPTY_RESPONSE_CODE, LlmAdapter, LlmError, ProviderRequestId, QUOTA_EXCEEDED_CODE, RetryPolicySchema, ToolCallId, assertUsableApiKey, attributionHeaders, isContextWindowExceededError, isQuotaExceededError, resolveRetryPolicy } from "@deepseek-ai/dsh-llm";
 import { credentialRef } from "@deepseek-ai/dsh-credentials";
 import { launchEnvironmentOf } from "@deepseek-ai/dsh-launch-environment";
-import { deepEqualJson, installSettingsSection, settingsNamespace } from "@deepseek-ai/dsh-settings";
+import { deepEqualJson } from "@deepseek-ai/dsh-util-values";
 import { MAX_TIMER_DELAY_MS, idleWatchdog, timeoutOf } from "@deepseek-ai/dsh-timeout";
 import { getOrCreateAnonymousUserId } from "@deepseek-ai/dsh-anonymous-user-id";
 import { EventSourceParserStream } from "eventsource-parser/stream";
@@ -158,20 +158,79 @@ async function discoverChatModels(baseURL, apiKey, signal) {
 * joined; assistant text becomes `content`, tool calls become `tool_calls`,
 * and tool results become separate tool messages. Assistant reasoning is
 * replayed as `reasoning_content` only on tool-call turns, as hosted reasoning
-* models (DeepSeek-R1 and siblings) require. Core image blocks are rejected
-* explicitly because this wire route is text-only; unknown declaration-merged
-* block types retain the adapter's documented extension fallback.
+* models (DeepSeek-R1 and siblings) require.
+*
+* **Multimodal**: image blocks in user messages are serialized as
+* OpenAI-compatible `image_url` content parts with base64 data URLs. The
+* serializer resolves image attachments through the optional attachment-store
+* resolver; when no resolver is available, images are replaced with the
+* `OFFLOADED_IMAGE_TEXT` sentinel. Tool-result content is flattened to text
+* (images within tool results are also replaced with the sentinel). Unknown
+* declaration-merged block types retain the adapter's documented extension
+* fallback.
+*
 * @module dsh-llm-siliconflow/serialize
 */
+/**
+* Model-facing stand-in for an image that could not be resolved to a data URL.
+* Mirrors the upstream `OFFLOADED_IMAGE_TEXT` sentinel from `@deepseek-ai/dsh-llm`
+* (available since 0.1.1-rc.1); defined locally so the plugin works against
+* the 0.1.0-rc.x line pinned in CI as well.
+*/
+const OFFLOADED_IMAGE_TEXT = "[image omitted to keep the request within its image limit; older images are omitted first. If this image is still needed, read its file again when a path is available; otherwise ask the user to attach it again.]";
+/** Default image resolver: returns undefined (no data URL available). */
+const noopResolveImage = () => Promise.resolve(void 0);
 /** Join the text blocks of a message (used for user/tool-result content). */
 function flattenText(blocks) {
 	return blocks.filter((block) => block.type === "text").map((block) => block.text).join("");
 }
-/** Reject core image content before any text-flattening path can silently erase it. */
-function assertTextOnly(blocks) {
-	if (contentHasImage(blocks)) throw new LlmError("The SiliconFlow chat-completions adapter does not support image content.", "UNSUPPORTED_CONTENT");
+/**
+* Replace image blocks (including nested ones in tool results) with the
+* `OFFLOADED_IMAGE_TEXT` sentinel so the provider sees a coherent text
+* placeholder instead of silently dropped bytes. This is the no-store fallback;
+* the multimodal path resolves real base64 data URLs instead.
+*/
+function replaceImagesWithSentinel(blocks) {
+	return blocks.map((block) => block.type === "image" ? {
+		type: "text",
+		text: OFFLOADED_IMAGE_TEXT
+	} : block);
+}
+/** Whether this message's content has any image blocks (user side). */
+function hasImages(blocks) {
+	return blocks.some((block) => block.type === "image");
+}
+/**
+* Build the wire content parts for a user message: text blocks become `text`
+* parts, image blocks become `image_url` parts. Non-text/non-image blocks are
+* skipped (merge-extensible fallback). An image block whose attachment cannot
+* be resolved becomes the `OFFLOADED_IMAGE_TEXT` sentinel text part.
+*/
+async function serializeUserContent(blocks, resolveImage) {
+	if (!hasImages(blocks)) return flattenText(blocks);
+	const parts = [];
+	for (const block of blocks) if (block.type === "text") parts.push({
+		type: "text",
+		text: block.text
+	});
+	else if (block.type === "image") {
+		const url = await resolveImage(block.attachment);
+		parts.push(url !== void 0 ? {
+			type: "image_url",
+			image_url: { url }
+		} : {
+			type: "text",
+			text: OFFLOADED_IMAGE_TEXT
+		});
+	}
+	return parts;
 }
-/** Serialize one assistant message (text + reasoning + tool calls). */
+/**
+* Serialize one assistant message (text + reasoning + tool calls). Image
+* blocks in assistant content (forward compatibility) are replaced with the
+* sentinel text — the chat-completions wire route carries images only in user
+* content.
+*/
 function serializeAssistant(message) {
 	const text = flattenText(message.content);
 	const reasoning = message.content.filter((block) => block.type === "reasoning").map((block) => block.text).join("");
@@ -195,13 +254,17 @@ function serializeAssistant(message) {
 * `{role: 'tool'}` messages; the harness puts each tool result in its own
 * user-role message, so a mixed user message contributes its text first and
 * its tool results as separate wire messages after.
+*
+* Image blocks within user messages are resolved through the `resolveImage`
+* callback; images within tool results are flattened to the sentinel text
+* (tool results carry structured data, not multimodal content).
 * @param messages - the harness conversation, in order.
+* @param resolveImage - resolves an image attachment ref to a base64 data URL; returns `undefined` when unavailable.
 * @returns the wire messages; order preserved, each tool result expanded into its own entry.
 */
-function serializeMessages(messages) {
+async function serializeMessages(messages, resolveImage = noopResolveImage) {
 	const wire = [];
 	for (const message of messages) {
-		assertTextOnly(message.content);
 		if (message.role === "system") {
 			wire.push({
 				role: "system",
@@ -214,16 +277,22 @@ function serializeMessages(messages) {
 			continue;
 		}
 		const toolResults = message.content.filter((block) => block.type === "tool-result");
-		const text = flattenText(message.content);
-		if (text.length > 0 || toolResults.length === 0) wire.push({
-			role: "user",
-			content: text
-		});
-		for (const result of toolResults) wire.push({
-			role: "tool",
-			tool_call_id: result.toolCallId,
-			content: flattenText(result.content) || "(no output)"
-		});
+		const userBlocks = message.content.filter((block) => block.type !== "tool-result");
+		if (userBlocks.length > 0 || toolResults.length === 0) {
+			const content = await serializeUserContent(userBlocks, resolveImage);
+			wire.push({
+				role: "user",
+				content
+			});
+		}
+		for (const result of toolResults) {
+			const flatContent = flattenText(replaceImagesWithSentinel(result.content)) || "(no output)";
+			wire.push({
+				role: "tool",
+				tool_call_id: result.toolCallId,
+				content: flatContent
+			});
+		}
 	}
 	return wire;
 }
@@ -232,15 +301,16 @@ function serializeMessages(messages) {
 * reporting on); optional fields are omitted rather than sent as null, so
 * provider defaults apply.
 * @param options - the harness request (model, history, system, tools, sampling).
+* @param resolveImage - resolves an image attachment ref to a base64 data URL; returns `undefined` when unavailable.
 * @returns the chat-completions request body.
 */
-function serializeRequest(options) {
+async function serializeRequest(options, resolveImage = noopResolveImage) {
 	const messages = [];
 	if (options.system !== void 0) messages.push({
 		role: "system",
 		content: options.system
 	});
-	messages.push(...serializeMessages(options.messages));
+	messages.push(...await serializeMessages(options.messages, resolveImage));
 	const tools = options.tools?.map((tool) => ({
 		type: "function",
 		function: {
@@ -338,7 +408,7 @@ function closeBlock(block) {
 		};
 		case "tool-call": return {
 			type: "tool-call",
-			id: CallId(block.callId ?? ""),
+			id: ToolCallId(block.callId ?? ""),
 			name: block.name ?? "",
 			arguments: block.text
 		};
@@ -453,7 +523,7 @@ async function* translate(payloads) {
 				yield {
 					type: "tool-call-delta",
 					index: block.index,
-					id: CallId(block.callId ?? ""),
+					id: ToolCallId(block.callId ?? ""),
 					...block.name !== void 0 ? { name: block.name } : {},
 					argumentsDelta: fragment
 				};
@@ -542,13 +612,71 @@ const DEFAULT_MAX_TOKENS = 8192;
 /** How long a cached model-listing discovery stays fresh before the next `listModels` re-interrogates. */
 const DISCOVERY_TTL_MS = 3e5;
 const STREAM_IDLE_TIMEOUT_CODE = "LLM_STREAM_IDLE_TIMEOUT";
+/**
+* Authoritative set of SiliconFlow model ids that accept image input.
+*
+* **Static data — see maintenance note below.**
+*
+* SiliconFlow's model marketplace (siliconflow.cn/models) tags every chat model
+* with a `vlm` boolean. The OpenAI-compatible `GET /models` API does not expose
+* this attribute, so the adapter ships a set of known VLM model ids extracted
+* from the marketplace's SSR data. Catalog entries can still declare explicit
+* `inputModalities` to override this set.
+*
+* **Maintenance**: this set is hand-maintained and will drift from the live
+* marketplace as new VLM models are added or existing ones are deprecated. It is
+* refreshed periodically. Once the OpenAI-compatible API exposes model
+* capabilities (e.g. a `vision` sub_type or a `vlm` field in the model listing),
+* this static set will be replaced by a runtime query.
+*
+* Last updated: 2026-08-22 (23 VLM models out of 88 chat models).
+*/
+const KNOWN_VLM_MODELS = /* @__PURE__ */ new Set([
+	"PaddlePaddle/PaddleOCR-VL-1.5",
+	"Pro/moonshotai/Kimi-K2.6",
+	"Qwen/Qwen3-Omni-30B-A3B-Captioner",
+	"Qwen/Qwen3-Omni-30B-A3B-Instruct",
+	"Qwen/Qwen3-Omni-30B-A3B-Thinking",
+	"Qwen/Qwen3-VL-30B-A3B-Instruct",
+	"Qwen/Qwen3-VL-30B-A3B-Thinking",
+	"Qwen/Qwen3-VL-32B-Instruct",
+	"Qwen/Qwen3-VL-32B-Thinking",
+	"Qwen/Qwen3-VL-8B-Instruct",
+	"Qwen/Qwen3-VL-8B-Thinking",
+	"Qwen/Qwen3.5-122B-A10B",
+	"Qwen/Qwen3.5-27B",
+	"Qwen/Qwen3.5-35B-A3B",
+	"Qwen/Qwen3.5-397B-A17B",
+	"Qwen/Qwen3.5-4B",
+	"Qwen/Qwen3.5-9B",
+	"Qwen/Qwen3.6-27B",
+	"Qwen/Qwen3.6-35B-A3B",
+	"deepseek-ai/DeepSeek-OCR",
+	"moonshotai/Kimi-K2.7-Code",
+	"nex-agi/Nex-N2-Pro",
+	"zai-org/GLM-4.5V"
+]);
+/**
+* Determine the input modalities for a SiliconFlow model id.
+*
+* Uses the authoritative VLM model set from the SiliconFlow marketplace rather
+* than naming-pattern heuristics. Catalog entries can declare explicit
+* `inputModalities` to override this lookup.
+* @param id - the wire model id (e.g. `Qwen/Qwen3-VL-8B-Instruct`).
+* @returns `['text', 'image']` when the id is a known VLM, `['text']` otherwise.
+*/
+function inferInputModalities(id) {
+	return KNOWN_VLM_MODELS.has(id) ? ["text", "image"] : ["text"];
+}
 function modelInfo(provider, model) {
+	const explicit = model.inputModalities;
+	const modalities = explicit !== void 0 && explicit.length > 0 ? explicit : inferInputModalities(model.id);
 	return {
 		provider,
 		id: model.id,
 		name: model.name ?? model.id,
 		...model.description === void 0 ? {} : { description: model.description },
-		inputModalities: ["text"]
+		inputModalities: modalities
 	};
 }
 function providerRetryAfterMs(value) {
@@ -655,7 +783,7 @@ var SiliconFlowAdapter = class extends LlmAdapter {
 				provider,
 				id: model,
 				name: model,
-				inputModalities: ["text"]
+				inputModalities: inferInputModalities(model)
 			} : modelInfo(provider, entry),
 			context: { contextWindow: entry?.contextWindow ?? connection.defaultContextWindow },
 			defaultMaxTokens: entry?.maxTokens ?? connection.maxTokens
@@ -706,7 +834,7 @@ var SiliconFlowAdapter = class extends LlmAdapter {
 		}
 	}
 	async *request(options, signal, connection, apiKey, userId, onComment) {
-		const body = serializeRequest(options);
+		const body = await serializeRequest(options, this.config.resolveImage);
 		const payload = JSON.stringify(body);
 		const headers = {
 			"authorization": `Bearer ${apiKey}`,
@@ -760,25 +888,36 @@ var SiliconFlowAdapter = class extends LlmAdapter {
 * restarting anything, while an in-flight stream keeps the facts it started
 * with. The one registration-captured fact — the retry policy — re-registers
 * the route in place when it changes.
+*
+* When the optional `ctx.attachments` service is mounted, image blocks in user
+* messages are resolved to base64 data URLs and serialized as OpenAI-compatible
+* `image_url` content parts, enabling VLM models (Qwen3-VL, GLM-4.5V, etc.).
 * @module @siliconflow-official/dsh-llm-siliconflow
 */
 const name = "llm-siliconflow";
 const inject = ["llm"];
-const NS = settingsNamespace("llm-siliconflow");
+const NS = "llm-siliconflow";
 /** Credential reference this plugin reads by default, also used by the setup CLI. */
 const DEFAULT_API_KEY_ENV = "SILICONFLOW_API_KEY";
 /** The single provider route this plugin owns. */
 const PROVIDER = "siliconflow";
-/** Fallback advisory catalog: six widely hosted chat models, also the setup CLI's discovery fallback. */
+/**
+* Fallback advisory catalog: widely hosted chat models including VLMs, also the
+* setup CLI's discovery fallback.
+*
+* **Static data** — the `contextWindow` values and VLM flags are extracted from
+* the SiliconFlow model marketplace (siliconflow.cn/models). The live
+* `GET /models?sub_type=chat` API does not return context window or VLM metadata,
+* so these fields are maintained by hand. This catalog will be kept in sync with
+* the marketplace until the API exposes these attributes natively.
+*
+* Last updated: 2026-08-22. VLM entries declare `inputModalities` explicitly.
+*/
 const DEFAULT_MODELS = [
 	{
 		id: "zai-org/GLM-5.2",
 		contextWindow: 1e6
 	},
-	{
-		id: "moonshotai/Kimi-K2.7-Code",
-		contextWindow: 256e3
-	},
 	{
 		id: "deepseek-ai/DeepSeek-V4-Pro",
 		contextWindow: 1e6
@@ -787,13 +926,54 @@ const DEFAULT_MODELS = [
 		id: "deepseek-ai/DeepSeek-V4-Flash",
 		contextWindow: 1e6
 	},
+	{
+		id: "Pro/zai-org/GLM-5.1",
+		contextWindow: 202752
+	},
+	{
+		id: "moonshotai/Kimi-K2.7-Code",
+		contextWindow: 262144,
+		inputModalities: ["text", "image"]
+	},
 	{
 		id: "Pro/moonshotai/Kimi-K2.6",
-		contextWindow: 256e3
+		contextWindow: 262144,
+		inputModalities: ["text", "image"]
 	},
 	{
 		id: "Qwen/Qwen3.5-397B-A17B",
-		contextWindow: 256e3
+		contextWindow: 262144,
+		inputModalities: ["text", "image"]
+	},
+	{
+		id: "zai-org/GLM-4.5V",
+		contextWindow: 65536,
+		inputModalities: ["text", "image"]
+	},
+	{
+		id: "Qwen/Qwen3-VL-32B-Instruct",
+		contextWindow: 262144,
+		inputModalities: ["text", "image"]
+	},
+	{
+		id: "Qwen/Qwen3-VL-8B-Instruct",
+		contextWindow: 262144,
+		inputModalities: ["text", "image"]
+	},
+	{
+		id: "Qwen/Qwen3-VL-32B-Thinking",
+		contextWindow: 262144,
+		inputModalities: ["text", "image"]
+	},
+	{
+		id: "Qwen/Qwen3-VL-8B-Thinking",
+		contextWindow: 262144,
+		inputModalities: ["text", "image"]
+	},
+	{
+		id: "deepseek-ai/DeepSeek-OCR",
+		contextWindow: 8192,
+		inputModalities: ["text", "image"]
 	}
 ];
 const catalogModel = z.object({
@@ -801,7 +981,8 @@ const catalogModel = z.object({
 	name: z.string(),
 	description: z.string(),
 	contextWindow: z.number().step(1).min(1),
-	maxTokens: z.number().step(1).min(1)
+	maxTokens: z.number().step(1).min(1),
+	inputModalities: z.array(z.union(["text", "image"]))
 });
 const Config = z.object({
 	apiKeyEnv: z.string().role("credential-ref").default(DEFAULT_API_KEY_ENV),
@@ -831,7 +1012,8 @@ function resolveModels(models) {
 			...model.name === void 0 ? {} : { name: model.name },
 			...model.description === void 0 ? {} : { description: model.description },
 			...model.contextWindow === void 0 ? {} : { contextWindow: model.contextWindow },
-			...model.maxTokens === void 0 ? {} : { maxTokens: model.maxTokens }
+			...model.maxTokens === void 0 ? {} : { maxTokens: model.maxTokens },
+			...model.inputModalities === void 0 || model.inputModalities.length === 0 ? {} : { inputModalities: model.inputModalities }
 		};
 	});
 }
@@ -895,6 +1077,22 @@ function apply(ctx, config) {
 		}
 		throw new LlmError(`llm-siliconflow: no API key for provider route "${PROVIDER}"; store ${ref} through the credentials service (the web Models page writes it), or export ${ref} in the launching environment`, "MISSING_CREDENTIAL");
 	};
+	/**
+	* Resolve one image attachment to a base64 data URL for wire serialization.
+	* Uses the optional `ctx.attachments` service; returns `undefined` when the
+	* store is not mounted or the read fails, so the serializer substitutes the
+	* `OFFLOADED_IMAGE_TEXT` sentinel instead of blocking the request.
+	*/
+	const resolveImage = async (ref) => {
+		const attachments = ctx.get("attachments");
+		if (attachments === void 0) return void 0;
+		try {
+			const { data } = await attachments.readImage(ref);
+			return `data:${ref.mediaType};base64,${Buffer.from(data).toString("base64")}`;
+		} catch {
+			return;
+		}
+	};
 	let userId;
 	const resolveUserId = () => userId ??= getOrCreateAnonymousUserId();
 	const storedApiKey = async () => {
@@ -907,7 +1105,8 @@ function apply(ctx, config) {
 	const adapter = new SiliconFlowAdapter({
 		options,
 		resolveApiKey,
-		resolveUserId
+		resolveUserId,
+		resolveImage
 	});
 	ctx.llm.registerConfigurableProviders([{
 		provider: PROVIDER,
@@ -926,12 +1125,14 @@ function apply(ctx, config) {
 		registration.replace([PROVIDER]);
 		registeredPolicy = policy;
 	};
-	installSettingsSection(ctx, NS, Config, config, {
-		setSource: (source) => {
-			current = source;
-		},
-		onChange: ensureRegistrationFacts
+	ctx.inject(["settings"], (settingsCtx) => {
+		settingsCtx.settings.installSection(ctx, NS, Config, config, {
+			setSource: (source) => {
+				current = source;
+			},
+			onChange: ensureRegistrationFacts
+		});
 	});
 }
 //#endregion
-export { readListing as _, PUBLIC_BASE_URL as a, name as c, DEFAULT_MAX_TOKENS as d, DEFAULT_STREAM_IDLE_TIMEOUT_MS as f, listingUrl as g, discoverChatModels as h, PROVIDER as i, resolveAdapterOptions as l, SiliconFlowAdapter as m, DEFAULT_API_KEY_ENV as n, apply as o, DISCOVERY_TTL_MS as p, DEFAULT_MODELS as r, inject as s, Config as t, DEFAULT_CONTEXT_WINDOW as u };
+export { listingUrl as _, PUBLIC_BASE_URL as a, name as c, DEFAULT_MAX_TOKENS as d, DEFAULT_STREAM_IDLE_TIMEOUT_MS as f, discoverChatModels as g, inferInputModalities as h, PROVIDER as i, resolveAdapterOptions as l, SiliconFlowAdapter as m, DEFAULT_API_KEY_ENV as n, apply as o, DISCOVERY_TTL_MS as p, DEFAULT_MODELS as r, inject as s, Config as t, DEFAULT_CONTEXT_WINDOW as u, readListing as v };

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@siliconflow-official/dsh-llm-siliconflow",
   "description": "SiliconFlow (OpenAI-compatible) chat-completions adapter plugin for the DeepSeek Harness LLM seam",
-  "version": "0.1.0-rc.7",
+  "version": "0.2.0-rc.2",
   "publishConfig": {
     "access": "public"
   },
@@ -45,12 +45,14 @@
   "peerDependencies": {
     "@deepseek-ai/cordis": "^4.0.1",
     "@deepseek-ai/dsh-anonymous-user-id": ">=0.0.1-rc.0",
+    "@deepseek-ai/dsh-attachment": ">=0.0.1-rc.0",
     "@deepseek-ai/dsh-credentials": ">=0.0.1-rc.0",
     "@deepseek-ai/dsh-invariants": ">=0.0.1-rc.0",
     "@deepseek-ai/dsh-launch-environment": ">=0.0.1-rc.0",
     "@deepseek-ai/dsh-llm": ">=0.0.1-rc.0",
     "@deepseek-ai/dsh-settings": ">=0.0.1-rc.0",
-    "@deepseek-ai/dsh-timeout": ">=0.0.1-rc.0"
+    "@deepseek-ai/dsh-timeout": ">=0.0.1-rc.0",
+    "@deepseek-ai/dsh-util-values": "*"
   },
   "dependencies": {
     "@deepseek-ai/schemastery": "^3.18.1",
@@ -58,14 +60,16 @@
     "yaml": "^2.9.0"
   },
   "devDependencies": {
-    "@deepseek-ai/cordis": "^4.0.1",
-    "@deepseek-ai/dsh-anonymous-user-id": ">=0.0.1-rc.0",
-    "@deepseek-ai/dsh-credentials": ">=0.0.1-rc.0",
-    "@deepseek-ai/dsh-invariants": ">=0.0.1-rc.0",
-    "@deepseek-ai/dsh-launch-environment": ">=0.0.1-rc.0",
-    "@deepseek-ai/dsh-llm": ">=0.0.1-rc.0",
-    "@deepseek-ai/dsh-settings": ">=0.0.1-rc.0",
-    "@deepseek-ai/dsh-timeout": ">=0.0.1-rc.0",
+    "@deepseek-ai/cordis": "^4.0.2",
+    "@deepseek-ai/dsh-anonymous-user-id": "0.1.5-rc.2",
+    "@deepseek-ai/dsh-attachment": "0.1.5-rc.2",
+    "@deepseek-ai/dsh-credentials": "0.1.5-rc.2",
+    "@deepseek-ai/dsh-invariants": "0.1.5-rc.2",
+    "@deepseek-ai/dsh-launch-environment": "0.1.5-rc.2",
+    "@deepseek-ai/dsh-llm": "0.1.5-rc.2",
+    "@deepseek-ai/dsh-settings": "0.1.5-rc.2",
+    "@deepseek-ai/dsh-timeout": "0.1.5-rc.2",
+    "@deepseek-ai/dsh-util-values": "0.1.5-rc.2",
     "@types/node": "^22.0.0",
     "tsdown": "^0.22.0",
     "typescript": "^6.0.0"