fix(ai): preserve literal DSML tool arguments

## What Problem This Solves

DSML parameter recovery loses the literal `__proto__` argument, causing a call containing only that argument to remain unrecovered. The same provider pipeline also repeats endpoint parsing, model-family checks, and tool-result text cleanup.

## User Impact

Recovered DeepSeek tool calls retain declared argument names, including `__proto__`, with the same ordinary object shape and duplicate-name last-value behavior as JSON arguments. The helper cleanup preserves existing endpoint restrictions, request shaping, usage and failover behavior.

## Why This Change Was Made

- Define recovered DSML parameters as own data properties, preserving their normal enumerability, writability and configurability.
- Reuse the existing Gemini-family predicate and non-empty tool-result text sanitizer.
- Replace three parse-only URL exception paths with `URL.parse`, preserving their existing route predicates.

Measured production reduction: 23 lines (33 added, 56 removed). The regression adds 31 test lines; no dependency or public API is added.

## Fixes found along the way

An ordinary object assignment invokes the inherited `__proto__` setter. Because DSML values are strings, that argument disappears; if it is the only argument, recovery rejects the call. The JSON fallback already preserves the key. The regression runs through the completions stream reducer and checks reserved and ordinary names plus duplicate-name overwrite behavior.

## Evidence

- Original code: the `__proto__` regression failed (`stop` instead of `toolUse`); the other 14 tests passed. Fixed code: all 15 tests passed, preserving own enumerable properties, the ordinary prototype, and duplicate-name last-write behavior.
- Full AI suite: 2,904 tests in 149 files passed. Both cycle checks report zero. All seven local changed files match the remotely tested SHA-256 digests.
- Test cost: `pnpm test packages/ai/src/transports/openai-completions-dsml.test.ts --maxWorkers=1 --reporter=dot` took 1.94 seconds wall on Linux Testbox; Vitest reported 945 ms total. Original failing run took 2.88 seconds wall. CI timing is pending the hosted run.
- Independent Codex review found no actionable P0-P2 issues. Full `check-changed` passed, including all core test-type groups, lint and boundary guards. The canonical SDK API comparison reports no API changes.
- Tested candidate tree: `0eec1534eed2fab3567828267914e0c045e656cc`; published commit will contain that exact tree. Proof ran on Blacksmith Testbox `tbx_01m3tmabk0s6555s9b8k83rxkt`.

packages/ai/src/openai-completions-messages.ts

@@ -15,6 +15,7 @@ import {
   extractToolResultText,
 } from "./providers/tool-result-text.js";
 import type { ResolvedOpenAICompletionsCompat } from "./transports/openai-completions-compat.js";
+import { sanitizeNonEmptyTransportPayloadText } from "./transports/transport-stream-shared.js";
 import type { Context, Model, ThinkingContent, ToolCall } from "./types.js";
 import { sanitizeSurrogates } from "./utils/sanitize-unicode.js";
 import {
@@ -23,17 +24,11 @@ import {
   stripSystemPromptRelocatableBoundary,
 } from "./utils/system-prompt-cache-boundary.js";
 
-const EMPTY_TOOL_RESULT_TEXT = "(no output)";
 type ChatCompletionContentPartVideo = {
   type: "video_url";
   video_url: { url: string };
 };
 
-function sanitizeToolResultText(text: string, fallback: string): string {
-  const sanitized = sanitizeSurrogates(text);
-  return sanitized.trim().length > 0 ? sanitized : fallback;
-}
-
 /** Whether replayed messages require a tools marker for proxy compatibility. */
 export function hasToolCallHistory(messages: Context["messages"]): boolean {
   return messages.some(
@@ -252,10 +247,7 @@ export function convertMessages(
         const textResult = extractToolResultText(toolMsg.content);
         const mediaPlaceholder = describeToolResultMediaPlaceholder(toolMsg.content);
         const images = toolMsg.content.filter(isImageWithMediaPayload);
-        const content = sanitizeToolResultText(
-          textResult,
-          mediaPlaceholder ?? EMPTY_TOOL_RESULT_TEXT,
-        );
+        const content = sanitizeNonEmptyTransportPayloadText(textResult, mediaPlaceholder);
         const toolResultMsg: ChatCompletionToolMessageParam = {
           role: "tool",
           content,

packages/ai/src/transports/openai-completions-compat.ts

@@ -94,28 +94,21 @@ export function isOpenAICodexResponsesModel(model: {
 
 function isNativeOpenAICodexResponsesBaseUrl(baseUrl?: string): boolean {
   const trimmed = typeof baseUrl === "string" ? baseUrl.trim() : "";
-  if (!trimmed) {
+  const url = URL.parse(trimmed);
+  if (!url || (url.protocol !== "http:" && url.protocol !== "https:")) {
     return false;
   }
-  try {
-    const url = new URL(trimmed);
-    if (url.protocol !== "http:" && url.protocol !== "https:") {
-      return false;
-    }
-    if (url.hostname.toLowerCase() !== "chatgpt.com") {
-      return false;
-    }
-    const pathname = url.pathname.replace(/\/+$/u, "").toLowerCase();
-    return [
-      "/backend-api",
-      "/backend-api/v1",
-      "/backend-api/codex",
-      "/backend-api/codex/v1",
-      "/backend-api/codex/responses",
-    ].includes(pathname);
-  } catch {
+  if (url.hostname.toLowerCase() !== "chatgpt.com") {
     return false;
   }
+  const pathname = url.pathname.replace(/\/+$/u, "").toLowerCase();
+  return [
+    "/backend-api",
+    "/backend-api/v1",
+    "/backend-api/codex",
+    "/backend-api/codex/v1",
+    "/backend-api/codex/responses",
+  ].includes(pathname);
 }
 
 export function usesNativeOpenAICodexResponsesBackend(model: {

packages/ai/src/transports/openai-completions-dsml.test.ts

@@ -273,6 +273,37 @@ describe("openai completions DSML", () => {
     expect(JSON.stringify(events)).not.toContain("DSML");
   });
 
+  it.each(["__proto__", "constructor", "text"])(
+    "preserves literal DSML parameter name %s as an own scalar argument",
+    async (name) => {
+      const model = createDeepSeekCompletionsModel();
+      const output = createAssistantOutput(model);
+      const content =
+        '<|DSML|tool_calls><|DSML|invoke name="echo">' +
+        `<|DSML|parameter name="${name}" string="true">first value</|DSML|parameter>` +
+        `<|DSML|parameter name="${name}" string="true">scalar value</|DSML|parameter>` +
+        "</|DSML|invoke></|DSML|tool_calls>";
+
+      await processCompletionsStream(
+        streamChunks([makeCompletionsChunk({ content }, "stop")]),
+        output,
+        model,
+        { push() {} },
+      );
+
+      expect(output.stopReason).toBe("toolUse");
+      expect(output.content).toHaveLength(1);
+      const call = output.content.find((block) => block.type === "toolCall");
+      expect(call).toMatchObject({ type: "toolCall", name: "echo" });
+      expect(call?.arguments).toStrictEqual({ [name]: "scalar value" });
+      expect(Object.getOwnPropertyDescriptor(call?.arguments, name)).toMatchObject({
+        value: "scalar value",
+        enumerable: true,
+      });
+      expect(Object.getPrototypeOf(call?.arguments)).toBe(Object.prototype);
+    },
+  );
+
   it("rejects an oversized DeepSeek DSML block when the crossing chunk contains its close", async () => {
     const model = createDeepSeekCompletionsModel();
     const output = createAssistantOutput(model);

packages/ai/src/transports/openai-completions-dsml.ts

@@ -222,7 +222,12 @@ function parseDeepSeekDsmlInvokeArguments(body: string): Record<string, unknown>
     if (rawValue.length === 0) {
       continue;
     }
-    args[name] = decodeDeepSeekDsmlText(rawValue);
+    Object.defineProperty(args, name, {
+      value: decodeDeepSeekDsmlText(rawValue),
+      enumerable: true,
+      configurable: true,
+      writable: true,
+    });
   }
   if (Object.keys(args).length > 0) {
     return args;

packages/ai/src/transports/openai-completions-params.ts

@@ -70,11 +70,8 @@ function isKnownOpenAICompletionsEndpoint(model: Pick<Model, "baseUrl">): boolea
   if (endpointClass === "openai-public" || endpointClass === "azure-openai") {
     return true;
   }
-  try {
-    return isAzureOpenAICompatibleHost(new URL(model.baseUrl).hostname.toLowerCase());
-  } catch {
-    return false;
-  }
+  const endpoint = URL.parse(model.baseUrl);
+  return endpoint !== null && isAzureOpenAICompatibleHost(endpoint.hostname.toLowerCase());
 }
 
 function resolveOpenAICompletionsMaxTokens(

packages/ai/src/transports/openai-completions-replay.ts

@@ -1,9 +1,6 @@
 import type { Context, Model } from "@openclaw/llm-core";
 import { uniqueStrings } from "@openclaw/normalization-core/string-normalization";
-import {
-  isGoogleGemini3FlashModel,
-  isGoogleGemini3ProModel,
-} from "../internal/google-model-family.js";
+import { isGoogleGemini3ThinkingLevelModel } from "../internal/google-model-family.js";
 import { detectOpenAICompletionsCompat } from "./openai-completions-compat.js";
 import {
   GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP,
@@ -20,10 +17,6 @@ function isGoogleOpenAICompatModel(model: OpenAIModeModel): boolean {
   );
 }
 
-function requiresGoogleCompatToolCallThoughtSignature(model: OpenAIModeModel): boolean {
-  return isGoogleGemini3ProModel(model.id) || isGoogleGemini3FlashModel(model.id);
-}
-
 const GOOGLE_COMPAT_THOUGHT_SIGNATURE_ELLIPSIS_RE = /[\u2026]|\.\.\./;
 const GOOGLE_COMPAT_THOUGHT_SIGNATURE_BASE64_RE = /^[A-Za-z0-9+/=]+$/;
 
@@ -43,7 +36,7 @@ function injectToolCallThoughtSignatures(
     return;
   }
   const sigById = new Map<string, string>();
-  const fallbackSig = requiresGoogleCompatToolCallThoughtSignature(model)
+  const fallbackSig = isGoogleGemini3ThinkingLevelModel(model.id)
     ? GEMINI_THOUGHT_SIGNATURE_VALIDATOR_SKIP
     : undefined;
   for (const msg of context.messages ?? []) {

packages/ai/src/transports/openai-responses-websocket.ts

@@ -80,19 +80,16 @@ function isOfficialOpenAIResponsesBaseUrl(baseUrl: string | undefined): boolean
   if (!baseUrl) {
     return false;
   }
-  try {
-    const url = new URL(baseUrl);
-    return (
-      url.origin === "https://api.openai.com" &&
-      url.username === "" &&
-      url.password === "" &&
-      url.search === "" &&
-      url.hash === "" &&
-      url.pathname.replace(/\/+$/, "") === "/v1"
-    );
-  } catch {
-    return false;
-  }
+  const url = URL.parse(baseUrl);
+  return (
+    url !== null &&
+    url.origin === "https://api.openai.com" &&
+    url.username === "" &&
+    url.password === "" &&
+    url.search === "" &&
+    url.hash === "" &&
+    url.pathname.replace(/\/+$/, "") === "/v1"
+  );
 }
 export function supportsNativeOpenAIResponsesEndpoint(params: {
   provider: string;