@morlay/dsh-llm-openai-compatible 0.0.9-alpha.5 → 0.0.9-alpha.7

package/README.md

@@ -15,46 +15,54 @@ DeepSeek Harness 的 **OpenAI 兼容 LLM 适配器**插件。与内置 `llm-pi-a
 ## 配置
 
 `providers` 是 dict：**key 就是 provider 路由键**（选择器与
-`GenerateOptions.provider` 使用），值是 profile。
+`GenerateOptions.provider` 使用），值是 profile。它标了 `.volatile()`——**运行期可改**：
+值经 Loader 的引用读取，设置页（Models 页）改的就是它，不需要重挂这行插件（见
+[ADR-运行期改配置走volatile引用](./.agents/adrs/20260922-运行期改配置走volatile引用.md)）。
+
+配置写在**这行的 config** 里（bundle patch / profile patch / 设置页都落到这里）：
 
 ```yaml
-llm-openai-compatible:
-  providers:
-    ollama:
-      apiKeyEnv: OLLAMA_API_KEY
-      baseURL: https://ollama.com/v1
-      displayName: Ollama Gateway
-      # === 采样默认参数（请求级 temperature 优先）===
-      temperature: 1 # 0..2
-      topP: 0.95 # 0..1 → wire top_p
-      topK: 40 # 正整数 → wire top_k（非标准，仅网关支持时发送）
-      presencePenalty: 0 # -2..2 → wire presence_penalty
-      frequencyPenalty: 0 # -2..2 → wire frequency_penalty
-      seed: 42 # 正整数 → wire seed
-      # === 推理 ===
-      reasoning: high # 部署默认档位（省略 = 提供方默认）
-      # === 模型目录 ===
-      defaultContextWindow: 262144
-      defaultMaxTokens: 32768
-      models:
-        - id: deepseek-v4-flash:0731
-          name: DeepSeek V4 Flash
-          contextWindow: 1000000
-          maxTokens: 65535
-          inputModalities: [text, image]
-          reasoningEfforts:
-            off: # off 空值 = 不发送 reasoning_effort
-            high: high # 档位 → wire reasoning_effort 拼写
-            max: max
-      # === 传输 ===
-      maxRequestImageBytes: 20971520
-      streamIdleTimeoutMs: 300000
-      timeoutMs: 600000 # 整体请求超时；缺省不设
-      retryPolicy:
-        mode: normal
-        maxRetries: 5
+- id: llm-openai-compatible
+  name: "@morlay/dsh-llm-openai-compatible"
+  config:
+    providers:
+      ollama:
+        apiKeyEnv: OLLAMA_API_KEY
+        baseURL: https://ollama.com/v1
+        displayName: Ollama Gateway
+        # === 采样默认参数（请求级 temperature 优先）===
+        temperature: 1 # 0..2
+        topP: 0.95 # 0..1 → wire top_p
+        topK: 40 # 正整数 → wire top_k（非标准，仅网关支持时发送）
+        presencePenalty: 0 # -2..2 → wire presence_penalty
+        frequencyPenalty: 0 # -2..2 → wire frequency_penalty
+        seed: 42 # 正整数 → wire seed
+        # === 推理 ===
+        reasoning: high # 部署默认档位（省略 = 提供方默认）
+        # === 模型目录 ===
+        defaultContextWindow: 262144
+        defaultMaxTokens: 32768
+        models:
+          - id: deepseek-v4-flash:0731
+            name: DeepSeek V4 Flash
+            contextWindow: 1000000
+            maxTokens: 65535
+            inputModalities: [text, image]
+            reasoningEfforts:
+              off: # off 空值 = 不发送 reasoning_effort
+              high: high # 档位 → wire reasoning_effort 拼写
+              max: max
+        # === 传输 ===
+        maxRequestImageBytes: 20971520
+        streamIdleTimeoutMs: 300000
+        timeoutMs: 600000 # 整体请求超时；缺省不设
+        retryPolicy:
+          mode: normal
+          maxRetries: 5
 ```
 
+> 旧 `$DSH_HOME/settings.yaml` 的 `llm-openai-compatible` 段由上游启动时一次性导入到同 id 的行 config。
+
 ## 规则与取舍
 
 规则细节（合并表、wire 字段落点、端点与错误映射）在各自 ADR 里维护，这里只列结论：

package/dist/index.d.mts

@@ -1,7 +1,11 @@
 import { a as OpenAICompatibleAdapter, c as ResolvedModelProfile, i as DEFAULT_STREAM_IDLE_TIMEOUT_MS, l as ResolvedProviderProfile, n as DEFAULT_MAX_REQUEST_IMAGE_BYTES, o as OpenAICompatibleAdapterOptions, r as DEFAULT_MAX_TOKENS, s as ReasoningEffort, t as DEFAULT_CONTEXT_WINDOW } from "./adapter-CYd_pegB.mjs";
 import z from "@deepseek-ai/schemastery";
 import { ModelModality, RetryPolicyConfig } from "@deepseek-ai/dsh-llm";
-import { Context } from "@deepseek-ai/cordis";
+import { Context, Volatile } from "@deepseek-ai/cordis";
+//#region ../../../vendor/deepseek-harness/vendor/cosmokit/lib/types/misc.d.ts
+/** String/symbol keyed dictionary type. */
+type Dict<T = any, K extends string | symbol = string> = { [key in K]: T; };
+//#endregion
 //#region src/index.d.ts
 declare const name = "llm-openai-compatible";
 declare const inject: string[];
@@ -38,12 +42,23 @@ interface ProviderProfileSource {
   retryPolicy?: RetryPolicyConfig;
 }
 interface Config {
-  providers?: Record<string, ProviderProfileSource>;
+  /**
+   * provider 路由，key 是路由键。**volatile**：值经 Loader 的引用读取
+   * （`config.providers.get()`），设置页改它时不需要重挂这行插件——上游 0.1.7 起
+   * 「运行期可改」只有这一条路（旧的 settings namespace section 覆盖已取消）。
+   */
+  providers: Volatile<Record<string, ProviderProfileSource>>;
 }
-declare const Config: z<Config>;
+/** 解析成普通值之后的配置形状（校验与解析只认它）。 */
+type Options = { [K in keyof Config]?: Config[K] extends Volatile<infer T> ? T : never; };
+declare const Config: z<Schemastery.ObjectS<NoInfer<{
+  providers: z<NoInfer<Dict<ProviderProfileSource, string>>, NoInfer<Dict<ProviderProfileSource, string>>, "volatile-defined">;
+}>>, Schemastery.ObjectT<NoInfer<{
+  providers: z<NoInfer<Dict<ProviderProfileSource, string>>, NoInfer<Dict<ProviderProfileSource, string>>, "volatile-defined">;
+}>>, "plain">;
 declare function resolveAdapterOptions(provider: string, source: ProviderProfileSource): ResolvedProviderProfile;
 declare function resolveProfiles(providers: Readonly<Record<string, ProviderProfileSource>> | undefined): Map<string, ResolvedProviderProfile>;
-declare function assertServiceable(config: Config): void;
+declare function assertServiceable(config: Options): void;
 declare function apply(ctx: Context, config: Config): void;
 //#endregion
-export { Config, DEFAULT_CONTEXT_WINDOW, DEFAULT_MAX_REQUEST_IMAGE_BYTES, DEFAULT_MAX_TOKENS, DEFAULT_STREAM_IDLE_TIMEOUT_MS, MODEL_MODALITIES, ModelProfileSource, NS, OpenAICompatibleAdapter, type OpenAICompatibleAdapterOptions, ProviderProfileSource, REASONING_LEVELS, type ReasoningEffort, type ResolvedModelProfile, type ResolvedProviderProfile, apply, assertServiceable, inject, name, resolveAdapterOptions, resolveProfiles };
+export { Config, DEFAULT_CONTEXT_WINDOW, DEFAULT_MAX_REQUEST_IMAGE_BYTES, DEFAULT_MAX_TOKENS, DEFAULT_STREAM_IDLE_TIMEOUT_MS, MODEL_MODALITIES, ModelProfileSource, NS, OpenAICompatibleAdapter, type OpenAICompatibleAdapterOptions, Options, ProviderProfileSource, REASONING_LEVELS, type ReasoningEffort, type ResolvedModelProfile, type ResolvedProviderProfile, apply, assertServiceable, inject, name, resolveAdapterOptions, resolveProfiles };

package/dist/index.mjs

@@ -1,4 +1,4 @@
-import { a as serializeCallOptions, o as serializeCallOptionsWithImages, r as translate } from "./translate-CwniWJOm.mjs";
+import { a as serializeCallOptions, o as serializeCallOptionsWithImages, r as translate } from "./translate-D9PyJrw2.mjs";
 import z from "@deepseek-ai/schemastery";
 import { CONTEXT_WINDOW_EXCEEDED_CODE, LlmAdapter, LlmError, ProviderRequestId, QUOTA_EXCEEDED_CODE, ReasoningEffortId, RetryPolicySchema, assertUsableApiKey, attributionHeaders, contentHasImage, isContextWindowExceededError, isQuotaExceededError, resolveRetryPolicy } from "@deepseek-ai/dsh-llm";
 import { credentialRef } from "@deepseek-ai/dsh-credentials";
@@ -288,7 +288,7 @@ const providerSchema = z.object({
 	timeoutMs: z.number().min(Number.MIN_VALUE).max(MAX_TIMER_DELAY_MS),
 	retryPolicy: RetryPolicySchema
 });
-const Config = z.object({ providers: z.dict(providerSchema).default({}) });
+const Config = z.object({ providers: z.dict(providerSchema).default({}).volatile() });
 function isReasoningEffort(value) {
 	return REASONING_LEVELS.includes(value);
 }
@@ -397,25 +397,25 @@ function registrationFacts(profiles) {
 		retryPolicy: profile.retryPolicy
 	})).sort((left, right) => left.provider.localeCompare(right.provider));
 }
-function directoryEntries(profiles) {
+function directoryEntries(profiles, settingsNs) {
 	const entries = /* @__PURE__ */ new Map();
 	for (const [provider, profile] of profiles) entries.set(provider, {
 		provider,
 		displayName: profile.displayName,
-		settingsNs: NS,
+		settingsNs,
 		settingsPath: ["providers", provider],
 		declared: true
 	});
 	return [...entries.values()];
 }
 function apply(ctx, config) {
-	let current = () => config;
+	const settingsNs = ctx.fiber.entry?.options.id ?? "llm-openai-compatible";
 	let lastRaw;
 	let memoized;
 	const profiles = () => {
-		const raw = current();
+		const raw = config.providers.get();
 		if (raw === lastRaw && memoized !== void 0) return memoized;
-		const next = resolveProfiles(raw.providers);
+		const next = resolveProfiles(raw);
 		lastRaw = raw;
 		memoized = next;
 		return next;
@@ -445,7 +445,7 @@ function apply(ctx, config) {
 	let directory;
 	let directoryFacts;
 	const ensureDirectory = () => {
-		const entries = directoryEntries(profiles());
+		const entries = directoryEntries(profiles(), settingsNs);
 		if (deepEqualJson(entries, directoryFacts)) return;
 		if (directory === void 0) directory = ctx.llm.registerConfigurableProviders(entries);
 		else directory.replace(entries);
@@ -468,27 +468,27 @@ function apply(ctx, config) {
 		registeredFacts = facts;
 	};
 	ensureRegistrationFacts();
-	ctx.inject(["settings"], (settingsCtx) => {
-		settingsCtx.settings.installSection(ctx, NS, Config, config, {
-			validate: assertServiceable,
-			setSource: (source) => {
-				current = source;
-			},
-			onChange: () => {
-				try {
-					ensureRegistrationFacts();
-				} catch (error) {
-					ctx.logger.error("llm-openai-compatible: keeping the previously registered routes after a refused update");
-					ctx.logger.error(error);
-				}
-				try {
-					ensureDirectory();
-				} catch (error) {
-					ctx.logger.error("llm-openai-compatible: keeping the previous configurable-provider directory after a refused update");
-					ctx.logger.error(error);
-				}
-			}
-		});
+	const refresh = () => {
+		try {
+			ensureRegistrationFacts();
+		} catch (error) {
+			ctx.logger.error("llm-openai-compatible: keeping the previously registered routes after a refused update");
+			ctx.logger.error(error);
+		}
+		try {
+			ensureDirectory();
+		} catch (error) {
+			ctx.logger.error("llm-openai-compatible: keeping the previous configurable-provider directory after a refused update");
+			ctx.logger.error(error);
+		}
+	};
+	ctx.on("loader/volatile-update", refresh);
+	ctx.on("internal/config", function(_raw, next) {
+		const raw = next();
+		if (this !== ctx.fiber) return raw;
+		const candidate = Config(raw);
+		assertServiceable({ providers: structuredClone(candidate.providers.get()) });
+		return raw;
 	});
 }
 //#endregion

package/dist/{translate-CwniWJOm.mjs → translate-D9PyJrw2.mjs}

@@ -22,7 +22,7 @@ function assertTextOnly(blocks) {
 	if (contentHasImage(blocks)) throw new LlmError("The OpenAI-compatible chat-completions adapter does not support image content in this message.", "UNSUPPORTED_CONTENT");
 }
 function assertSupportedImageRoles(messages) {
-	for (const message of messages) if (message.role !== "user" && contentHasImage(message.content)) throw new LlmError(`The OpenAI-compatible chat-completions adapter cannot represent image content in a ${message.role} message.`, "UNSUPPORTED_CONTENT");
+	for (const message of messages) if (message.role !== "user" && message.role !== "tool" && contentHasImage(message.content)) throw new LlmError(`The OpenAI-compatible chat-completions adapter cannot represent image content in a ${message.role} message.`, "UNSUPPORTED_CONTENT");
 }
 function assertRetainedImagesFit(messages, maxRequestImageBytes) {
 	const offloadImages = requiredImageOffload(messages, {
@@ -94,8 +94,6 @@ async function userParts(blocks, resolveImage, signal) {
 		case "image":
 			if (resolveImage === void 0) throw new LlmError("The OpenAI-compatible chat-completions adapter does not support image content in this message.", "UNSUPPORTED_CONTENT");
 			parts.push(await resolveImage(block, signal));
-			break;
-		case "tool-result": parts.push(...await userParts(block.content, resolveImage, signal));
 	}
 	return parts;
 }
@@ -117,6 +115,8 @@ async function serializePrompt(messages, resolveImage, signal) {
 		pendingToolImages = [];
 	};
 	for (const message of messages) {
+		if (message.role === "developer") throw new LlmError("The OpenAI-compatible chat-completions adapter cannot serialize developer messages.", "UNSUPPORTED_CONTENT");
+		if (message.content.some((block) => block.type === "tool-addition" || block.type === "tool-removal")) throw new LlmError("The OpenAI-compatible chat-completions adapter cannot serialize tool-change blocks outside developer messages.", "UNSUPPORTED_CONTENT");
 		if (message.role === "system") {
 			flushToolImages();
 			prompt.push({
@@ -134,34 +134,34 @@ async function serializePrompt(messages, resolveImage, signal) {
 			});
 			continue;
 		}
-		const regular = message.content.filter((block) => block.type !== "tool-result");
-		const toolResults = message.content.filter((block) => block.type === "tool-result");
-		const content = await userParts(regular, resolveImage, signal);
-		if (content.length > 0 || toolResults.length === 0) {
-			flushToolImages();
-			prompt.push({
-				role: "user",
-				content
-			});
-		}
-		for (const result of toolResults) {
+		if (message.role === "tool") {
 			const images = [];
 			if (resolveImage !== void 0) {
-				for (const block of result.content) if (block.type === "image") images.push(await resolveImage(block, signal));
+				for (const block of message.content) if (block.type === "image") images.push(await resolveImage(block, signal));
 			}
+			flushToolImages();
 			prompt.push({
 				role: "tool",
 				content: [{
 					type: "tool-result",
-					toolCallId: result.toolCallId,
-					toolName: toolNames.get(result.toolCallId) ?? "",
+					toolCallId: message.toolCallId,
+					toolName: toolNames.get(message.toolCallId) ?? "",
 					output: {
 						type: "text",
-						value: flattenText(result.content) || (images.length > 0 ? "(see attached image)" : "(no output)")
+						value: flattenText(message.content) || (images.length > 0 ? "(see attached image)" : "(no output)")
 					}
 				}]
 			});
 			pendingToolImages.push(...images);
+			continue;
+		}
+		const content = await userParts(message.content, resolveImage, signal);
+		if (content.length > 0) {
+			flushToolImages();
+			prompt.push({
+				role: "user",
+				content
+			});
 		}
 	}
 	flushToolImages();

package/dist/wire.mjs

@@ -1,2 +1,2 @@
-import { a as serializeCallOptions, i as resolveReasoningWire, n as mapUsage, o as serializeCallOptionsWithImages, r as translate, t as mapFinishReason } from "./translate-CwniWJOm.mjs";
+import { a as serializeCallOptions, i as resolveReasoningWire, n as mapUsage, o as serializeCallOptionsWithImages, r as translate, t as mapFinishReason } from "./translate-D9PyJrw2.mjs";
 export { mapFinishReason, mapUsage, resolveReasoningWire, serializeCallOptions, serializeCallOptionsWithImages, translate };

package/locale/en.json

@@ -0,0 +1,6 @@
+{
+  "meta": {
+    "title": "OpenAI-Compatible LLM",
+    "description": "OpenAI-compatible LLM adapter with configurable default sampling parameters over a providers dict."
+  }
+}

package/locale/zh.json

@@ -0,0 +1,6 @@
+{
+  "meta": {
+    "title": "OpenAI 兼容 LLM 适配",
+    "description": "OpenAI 兼容端点的 LLM 适配器：可在 providers 上配置默认采样参数。"
+  }
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@morlay/dsh-llm-openai-compatible",
-  "version": "0.0.9-alpha.5",
+  "version": "0.0.9-alpha.7",
   "description": "OpenAI-compatible LLM adapter plugin for DeepSeek Harness with configurable default sampling parameters (temperature / topP / topK / penalties / seed) over a providers dict.",
   "keywords": [
     "dsh",
@@ -16,6 +16,7 @@
     "url": "https://github.com/morlay/dsh-plugin.git"
   },
   "files": [
+    "locale/*.json",
     "dist",
     "src",
     "cordis.patch.yml",
@@ -27,24 +28,25 @@
     "./invariant": "./dist/invariant.mjs",
     "./wire": "./dist/wire.mjs",
     "./package.json": "./package.json",
-    "./cordis.patch.yml": "./cordis.patch.yml"
+    "./cordis.patch.yml": "./cordis.patch.yml",
+    "./locale/*.json": "./locale/*.json"
   },
   "dependencies": {
     "@ai-sdk/openai-compatible": "^3.0.32",
     "@ai-sdk/provider": "^4.0.7",
-    "@deepseek-ai/dsh-util-values": "0.1.6-alpha.2"
+    "@deepseek-ai/dsh-util-values": "0.1.7-alpha.2"
   },
   "peerDependencies": {
-    "@deepseek-ai/cordis": "4.0.2",
-    "@deepseek-ai/dsh-anonymous-user-id": "0.1.6-alpha.2",
-    "@deepseek-ai/dsh-attachment": "0.1.6-alpha.2",
-    "@deepseek-ai/dsh-credentials": "0.1.6-alpha.2",
-    "@deepseek-ai/dsh-invariants": "0.1.6-alpha.2",
-    "@deepseek-ai/dsh-launch-environment": "0.1.6-alpha.2",
-    "@deepseek-ai/dsh-llm": "0.1.6-alpha.2",
-    "@deepseek-ai/dsh-settings": "0.1.6-alpha.2",
-    "@deepseek-ai/dsh-timeout": "0.1.6-alpha.2",
-    "@deepseek-ai/schemastery": "3.18.2"
+    "@deepseek-ai/cordis": "4.0.4",
+    "@deepseek-ai/dsh-anonymous-user-id": "0.1.7-alpha.2",
+    "@deepseek-ai/dsh-attachment": "0.1.7-alpha.2",
+    "@deepseek-ai/dsh-credentials": "0.1.7-alpha.2",
+    "@deepseek-ai/dsh-invariants": "0.1.7-alpha.2",
+    "@deepseek-ai/dsh-launch-environment": "0.1.7-alpha.2",
+    "@deepseek-ai/dsh-llm": "0.1.7-alpha.2",
+    "@deepseek-ai/dsh-settings": "0.1.7-alpha.2",
+    "@deepseek-ai/dsh-timeout": "0.1.7-alpha.2",
+    "@deepseek-ai/schemastery": "3.18.4"
   },
   "dsh": {
     "bundle": {

package/src/index.ts

@@ -1,4 +1,4 @@
-import type { Context } from "@deepseek-ai/cordis";
+import type { Context, Volatile } from "@deepseek-ai/cordis";
 import z from "@deepseek-ai/schemastery";
 import {
   LlmError,
@@ -82,9 +82,17 @@ export interface ProviderProfileSource {
 }
 
 export interface Config {
-  providers?: Record<string, ProviderProfileSource>;
+  /**
+   * provider 路由，key 是路由键。**volatile**：值经 Loader 的引用读取
+   * （`config.providers.get()`），设置页改它时不需要重挂这行插件——上游 0.1.7 起
+   * 「运行期可改」只有这一条路（旧的 settings namespace section 覆盖已取消）。
+   */
+  providers: Volatile<Record<string, ProviderProfileSource>>;
 }
 
+/** 解析成普通值之后的配置形状（校验与解析只认它）。 */
+export type Options = { [K in keyof Config]?: Config[K] extends Volatile<infer T> ? T : never };
+
 const modelSchema = z.object({
   id: z.string().required(),
   name: z.string(),
@@ -95,7 +103,7 @@ const modelSchema = z.object({
   reasoningEfforts: z.union([z.const(false), z.dict(z.union([z.string(), z.const(null)]))]),
 });
 
-const providerSchema = z.object({
+const providerSchema: z<ProviderProfileSource> = z.object({
   apiKeyEnv: z.string().role("credential-ref"),
   displayName: z.string(),
   baseURL: z.string().required(),
@@ -120,8 +128,8 @@ const providerSchema = z.object({
   retryPolicy: RetryPolicySchema,
 });
 
-export const Config: z<Config> = z.object({
-  providers: z.dict(providerSchema).default({}),
+export const Config = z.object({
+  providers: z.dict(providerSchema).default({}).volatile(),
 });
 
 function isReasoningEffort(value: string): value is ReasoningEffort {
@@ -355,7 +363,7 @@ export function resolveProfiles(
   return resolved;
 }
 
-export function assertServiceable(config: Config): void {
+export function assertServiceable(config: Options): void {
   resolveProfiles(config.providers);
 }
 
@@ -369,10 +377,13 @@ function registrationFacts(profiles: ReadonlyMap<string, ResolvedProviderProfile
     .sort((left, right) => left.provider.localeCompare(right.provider));
 }
 
-function directoryEntries(profiles: ReadonlyMap<string, ResolvedProviderProfile>): {
+function directoryEntries(
+  profiles: ReadonlyMap<string, ResolvedProviderProfile>,
+  settingsNs: string,
+): {
   provider: string;
   displayName: string;
-  settingsNs: typeof NS;
+  settingsNs: string;
   settingsPath: readonly string[];
   declared: boolean;
 }[] {
@@ -381,7 +392,7 @@ function directoryEntries(profiles: ReadonlyMap<string, ResolvedProviderProfile>
     {
       provider: string;
       displayName: string;
-      settingsNs: typeof NS;
+      settingsNs: string;
       settingsPath: readonly string[];
       declared: boolean;
     }
@@ -390,7 +401,7 @@ function directoryEntries(profiles: ReadonlyMap<string, ResolvedProviderProfile>
     entries.set(provider, {
       provider,
       displayName: profile.displayName,
-      settingsNs: NS,
+      settingsNs,
       settingsPath: ["providers", provider],
       declared: true,
     });
@@ -399,14 +410,18 @@ function directoryEntries(profiles: ReadonlyMap<string, ResolvedProviderProfile>
 }
 
 export function apply(ctx: Context, config: Config): void {
-  let current = () => config;
-  let lastRaw: Config | undefined;
+  // `settingsNs` 是**行 id**（不是包短名）：设置页按 entry 定位这份配置。
+  const settingsNs = ctx.fiber.entry?.options.id ?? NS;
+  // 引用读回的是 schema 解析后的形状（只读、可选），不是手写的 `Options`。
+  let lastRaw: ReturnType<Config["providers"]["get"]> | undefined;
   let memoized: Map<string, ResolvedProviderProfile> | undefined;
 
   const profiles = (): ReadonlyMap<string, ResolvedProviderProfile> => {
-    const raw = current();
+    const raw = config.providers.get();
     if (raw === lastRaw && memoized !== void 0) return memoized;
-    const next = resolveProfiles(raw.providers);
+    // 引用读回的是 schema 解析后的只读形状（`exactOptionalPropertyTypes` 下与手写的可选字段形态不同），
+    // 这里按上游 llm-pi-ai 的做法转换一次再解析。
+    const next = resolveProfiles(raw as Record<string, ProviderProfileSource> | undefined);
     lastRaw = raw;
     memoized = next;
     return next;
@@ -443,7 +458,7 @@ export function apply(ctx: Context, config: Config): void {
   let directory: ReturnType<typeof ctx.llm.registerConfigurableProviders> | undefined;
   let directoryFacts: unknown[] | undefined;
   const ensureDirectory = () => {
-    const entries = directoryEntries(profiles());
+    const entries = directoryEntries(profiles(), settingsNs);
     if (deepEqualJson(entries, directoryFacts)) return;
     if (directory === void 0) directory = ctx.llm.registerConfigurableProviders(entries);
     else directory.replace(entries);
@@ -468,30 +483,35 @@ export function apply(ctx: Context, config: Config): void {
     registeredFacts = facts;
   };
   ensureRegistrationFacts();
-  ctx.inject(["settings"], (settingsCtx) => {
-    settingsCtx.settings.installSection(ctx, NS, Config, config, {
-      validate: assertServiceable,
-      setSource: (source) => {
-        current = source;
-      },
-      onChange: () => {
-        try {
-          ensureRegistrationFacts();
-        } catch (error) {
-          ctx.logger.error(
-            "llm-openai-compatible: keeping the previously registered routes after a refused update",
-          );
-          ctx.logger.error(error);
-        }
-        try {
-          ensureDirectory();
-        } catch (error) {
-          ctx.logger.error(
-            "llm-openai-compatible: keeping the previous configurable-provider directory after a refused update",
-          );
-          ctx.logger.error(error);
-        }
-      },
+  const refresh = (): void => {
+    try {
+      ensureRegistrationFacts();
+    } catch (error) {
+      ctx.logger.error(
+        "llm-openai-compatible: keeping the previously registered routes after a refused update",
+      );
+      ctx.logger.error(error);
+    }
+    try {
+      ensureDirectory();
+    } catch (error) {
+      ctx.logger.error(
+        "llm-openai-compatible: keeping the previous configurable-provider directory after a refused update",
+      );
+      ctx.logger.error(error);
+    }
+  };
+  // volatile 更新（设置页保存）只把新值提交进运行引用并广播，不重挂这一行：这里重算路由注册与
+  // 可配置 provider 目录（上游 0.1.7 的运行期改配置机制；`profiles()` 读的就是引用里的新值）。
+  ctx.on("loader/volatile-update", refresh);
+  // 候选配置在提交前先校验：无效更新被拒绝，运行引用保持原值（loader 只记录这次拒绝）。
+  ctx.on("internal/config", function (this: Context["fiber"], _raw, next) {
+    const raw: unknown = next();
+    if (this !== ctx.fiber) return raw;
+    const candidate = Config(raw as Options);
+    assertServiceable({
+      providers: structuredClone(candidate.providers.get()) as Record<string, ProviderProfileSource>,
     });
+    return raw;
   });
 }

package/src/serialize.ts

@@ -6,7 +6,7 @@ import {
   projectOffloadedImages,
   requiredImageOffload,
 } from "@deepseek-ai/dsh-llm";
-import type { ContentBlock, GenerateOptions, Message } from "@deepseek-ai/dsh-llm";
+import type { ContentBlock, GenerateOptions, RequestMessage } from "@deepseek-ai/dsh-llm";
 import { AttachmentError } from "@deepseek-ai/dsh-attachment";
 import type { AttachmentStore } from "@deepseek-ai/dsh-attachment";
 import type {
@@ -83,9 +83,14 @@ function assertTextOnly(blocks: readonly ContentBlock[]): void {
   }
 }
 
-function assertSupportedImageRoles(messages: readonly Message[]): void {
+function assertSupportedImageRoles(messages: readonly RequestMessage[]): void {
   for (const message of messages) {
-    if (message.role !== "user" && contentHasImage(message.content)) {
+    // 图片只出现在 user 与 tool 消息里；tool 结果的图片由 serializePrompt 聚成一条 user 消息发出。
+    if (
+      message.role !== "user" &&
+      message.role !== "tool" &&
+      contentHasImage(message.content)
+    ) {
       throw new LlmError(
         `The OpenAI-compatible chat-completions adapter cannot represent image content in a ${message.role} message.`,
         "UNSUPPORTED_CONTENT",
@@ -94,7 +99,10 @@ function assertSupportedImageRoles(messages: readonly Message[]): void {
   }
 }
 
-function assertRetainedImagesFit(messages: readonly Message[], maxRequestImageBytes: number): void {
+function assertRetainedImagesFit(
+  messages: readonly RequestMessage[],
+  maxRequestImageBytes: number,
+): void {
   const offloadImages = requiredImageOffload(
     messages,
     { representation: "base64", maxBytes: maxRequestImageBytes },
@@ -133,7 +141,7 @@ async function imagePart(
 }
 
 function assistantParts(
-  message: Message,
+  message: Extract<RequestMessage, { role: "assistant" }>,
   toolNames: Map<string, string>,
 ): Extract<LanguageModelV4Prompt[number], { role: "assistant" }>["content"] {
   const parts: Extract<LanguageModelV4Prompt[number], { role: "assistant" }>["content"] = [];
@@ -190,9 +198,6 @@ async function userParts(
           );
         parts.push(await resolveImage(block, signal));
         break;
-      case "tool-result":
-        parts.push(...(await userParts(block.content, resolveImage, signal)));
-        break;
       default:
         break;
     }
@@ -201,7 +206,7 @@ async function userParts(
 }
 
 async function serializePrompt(
-  messages: readonly Message[],
+  messages: readonly RequestMessage[],
   resolveImage:
     | ((
         block: Extract<ContentBlock, { type: "image" }>,
@@ -227,6 +232,20 @@ async function serializePrompt(
     pendingToolImages = [];
   };
   for (const message of messages) {
+    // V4 把 developer 消息与 tool-change 块持久化下来，但 provider 侧序列化明确不支持它们
+    // （与上游 `llm-deepseek` 同口径：先失败，等生产者与消费方一起实现）。
+    if (message.role === "developer") {
+      throw new LlmError(
+        "The OpenAI-compatible chat-completions adapter cannot serialize developer messages.",
+        "UNSUPPORTED_CONTENT",
+      );
+    }
+    if (message.content.some((block) => block.type === "tool-addition" || block.type === "tool-removal")) {
+      throw new LlmError(
+        "The OpenAI-compatible chat-completions adapter cannot serialize tool-change blocks outside developer messages.",
+        "UNSUPPORTED_CONTENT",
+      );
+    }
     if (message.role === "system") {
       flushToolImages();
       prompt.push({ role: "system", content: flattenText(message.content) });
@@ -238,37 +257,38 @@ async function serializePrompt(
       if (parts.length > 0) prompt.push({ role: "assistant", content: parts });
       continue;
     }
-    const regular = message.content.filter((block) => block.type !== "tool-result");
-    const toolResults = message.content.filter((block) => block.type === "tool-result");
-    const content = await userParts(regular, resolveImage, signal);
-    if (content.length > 0 || toolResults.length === 0) {
-      flushToolImages();
-      prompt.push({ role: "user", content });
-    }
-    for (const result of toolResults) {
+    if (message.role === "tool") {
+      // 结果消息自己带 `toolCallId` 与结果块（V4 起结果不再是 user 消息里的一个块）。
       const images: UserContentPart[] = [];
       if (resolveImage !== void 0) {
-        for (const block of result.content) {
+        for (const block of message.content) {
           if (block.type === "image") images.push(await resolveImage(block, signal));
         }
       }
+      flushToolImages();
       prompt.push({
         role: "tool",
         content: [
           {
             type: "tool-result",
-            toolCallId: result.toolCallId,
-            toolName: toolNames.get(result.toolCallId) ?? "",
+            toolCallId: message.toolCallId,
+            toolName: toolNames.get(message.toolCallId) ?? "",
             output: {
               type: "text",
               value:
-                flattenText(result.content) ||
+                flattenText(message.content) ||
                 (images.length > 0 ? "(see attached image)" : "(no output)"),
             },
           },
         ],
       });
       pendingToolImages.push(...images);
+      continue;
+    }
+    const content = await userParts(message.content, resolveImage, signal);
+    if (content.length > 0) {
+      flushToolImages();
+      prompt.push({ role: "user", content });
     }
   }
   flushToolImages();