@qwen-code/arena-agent 0.0.10-dev.8-test.0 → 0.0.10-test.0

package/dist/code-tools.js

@@ -14,6 +14,7 @@ import { executeBraveWebSearch } from "./brave-search.js";
 import { FetchPageReader } from "./fetch-page.js";
 import { replaceFirstText, replaceUniqueText } from "./file-edit.js";
 import { detectedImageMimeType, generatedImageBytesForExtension } from "./image.js";
+import { codeArenaImageModel } from "./image-model.js";
 import { supportsImageInput } from "./model-capabilities.js";
 import { generateImageWithModelRouter, } from "./model-router-media.js";
 import { executeRoutifyWebFetch, routifyWebFetchEnabled, } from "./routify-web-fetch.js";
@@ -224,18 +225,18 @@ function hasReadableImageStructure(data, mediaType) {
     }
     return false;
 }
-// All 39 successful, unmutated list_files captures use this exact order,
-// independently of whether the requested path is the project root or src/.
-// It is the seed manifest order, not an alphabetical/filesystem order.
+// Unmutated React captures on 2026-10-02, 10-13 and 10-21 share this order.
+// Older September captures used a different seed order. Preserve the current
+// manifest order for both root and scoped listings, not filesystem ordering.
 const capturedReactSeedFileOrder = [
-    "src/utils/cn.ts",
     "src/index.css",
     "src/main.tsx",
+    "src/utils/cn.ts",
     "src/App.tsx",
-    "vite.config.ts",
+    "index.html",
     "tsconfig.json",
+    "vite.config.ts",
     "package.json",
-    "index.html",
 ];
 const capturedFullStackSeedFileOrder = [
     "postcss.config.mjs",
@@ -441,6 +442,17 @@ function stringValue(value) {
 }
 function contentTypeForFile(filePath) {
     const extension = path.extname(filePath).toLowerCase();
+    const capturedTypes = {
+        ".py": "text/x-python",
+        ".sh": "text/x-shellscript",
+        ".sql": "application/sql",
+        ".yaml": "application/x-yaml",
+        ".yml": "application/x-yaml",
+        ".mjs": "text/plain",
+        ".c": "text/plain",
+    };
+    if (capturedTypes[extension])
+        return capturedTypes[extension];
     if ([".js", ".jsx", ".mjs", ".cjs", ".ts", ".tsx", ".mts", ".cts"].includes(extension)) {
         return "text/javascript";
     }
@@ -578,6 +590,17 @@ function validationIssues(toolName, value, errors, definitions) {
             continue;
         }
         const actual = missing ? undefined : valueAtPath(value, error.instancePath);
+        if (error.keyword === "enum") {
+            const options = record(error.params).allowedValues;
+            issues.push({
+                received: actual,
+                code: "invalid_enum_value",
+                options,
+                path: pathParts,
+                message: `Invalid enum value. Expected ${options.map((option) => `'${option}'`).join(" | ")}, received '${actual}'`,
+            });
+            continue;
+        }
         const expected = error.keyword === "type"
             ? String(record(error.params).type || "string")
             : requiredPropertyType(toolName, key, definitions);
@@ -794,12 +817,36 @@ export class CodeToolExecutor {
         this.seedFileRank = capturedSeedFileRank(config.mode);
         addFormats(this.ajv);
         for (const definition of this.definitions) {
-            this.validators.set(definition.function.name, this.ajv.compile(definition.function.parameters));
+            const name = definition.function.name;
+            // The current execution docs add an optional millisecond timeout. Keep
+            // the outgoing captured contract intact until a new capture is approved.
+            const parameters = structuredClone(definition.function.parameters);
+            if (name === "bash") {
+                parameters.properties = {
+                    ...record(parameters.properties),
+                    timeout: { type: "integer", minimum: 1, maximum: 600_000 },
+                };
+            }
+            this.validators.set(name, this.ajv.compile(parameters));
+        }
+        // React Codex uses the same runtime under a different name. Like the
+        // existing hidden Codex tools, accept calls without inventing a public
+        // contract or switching the standard React prompt/tool set.
+        const bashValidator = this.validators.get("bash");
+        if (config.mode === "react" &&
+            bashValidator &&
+            !disallowed.has("shell_command")) {
+            this.validators.set("shell_command", bashValidator);
         }
         // Hidden mode: advertise nothing, but keep the call executable.
         if (!advertiseApplyPatch && !disallowed.has("apply_patch")) {
             this.validators.set("apply_patch", this.ajv.compile(APPLY_PATCH_PARAMETERS));
         }
+        // Captured React/HTML histories call this hidden tool. No public schema was
+        // captured: accept its object payload without adding model-visible tools.
+        if (config.mode !== "full-stack" && !disallowed.has("update_plan")) {
+            this.validators.set("update_plan", this.ajv.compile({ type: "object" }));
+        }
     }
     // The tools offered to the model for this session, with captured schemas
     // and descriptions adjusted for current capabilities and workspace runtime.
@@ -821,10 +868,9 @@ export class CodeToolExecutor {
         const argumentsValue = (this.config.mode === "react" && typeof parsedArguments === "string"
             ? {}
             : structuredClone(parsedArguments));
-        // The captured provider advertises these fields as required, but its
-        // runtime still applies the documented defaults when a model omits
-        // them. Preserve the byte-identical outgoing schema while matching the
-        // observed execution behavior.
+        // Captured schemas advertise these fields as required. Keep those public
+        // fixtures intact while applying runtime defaults. Since 2026-10-19,
+        // official Code Arena accepts optional nulls as omission for all providers.
         if (name === "read_file") {
             argumentsValue.offset ??= 1;
             argumentsValue.limit ??= 2000;
@@ -832,10 +878,13 @@ export class CodeToolExecutor {
         else if (name === "list_files" || name === "glob_files") {
             argumentsValue.path ??= "";
         }
-        else if (this.config.mode === "full-stack" && name === "bash" &&
+        else if ((name === "bash" || name === "shell_command") &&
             argumentsValue && typeof argumentsValue === "object" &&
-            !Array.isArray(argumentsValue) && argumentsValue.workdir === undefined) {
-            argumentsValue.workdir = "";
+            !Array.isArray(argumentsValue)) {
+            argumentsValue.workdir ??= "";
+            argumentsValue.description ??= "";
+            if (argumentsValue.timeout === null)
+                delete argumentsValue.timeout;
         }
         else if (name === "grep_files") {
             argumentsValue.path ??= "";
@@ -847,6 +896,21 @@ export class CodeToolExecutor {
             argumentsValue["-C"] ??= 0;
             argumentsValue.context ??= 0;
         }
+        if (argumentsValue && typeof argumentsValue === "object" && !Array.isArray(argumentsValue)) {
+            if (name === "web_fetch") {
+                argumentsValue.format ??= "markdown";
+            }
+            else if (name === "fetch_media") {
+                argumentsValue.media_type ??= "both";
+                argumentsValue.count ??= 6;
+                argumentsValue.orientation ??= "any";
+                argumentsValue.size ??= "any";
+                argumentsValue.locale ??= "en-US";
+            }
+            else if (name === "build_and_start") {
+                argumentsValue.description ??= "";
+            }
+        }
         return argumentsValue;
     }
     validationMessage(name, value, validate) {
@@ -1007,6 +1071,8 @@ export class CodeToolExecutor {
                 return this.editCodeFile(String(values.path), String(values.context), String(values.replacement));
             case "apply_patch":
                 return this.applyCodePatch(String(values.input));
+            case "update_plan":
+                return { status: "success" };
             case "delete_file":
                 return this.deleteCodeFile(String(values.path));
             case "list_files":
@@ -1020,7 +1086,8 @@ export class CodeToolExecutor {
             case "build_project":
                 return this.buildProject();
             case "bash":
-                return this.fullStackBash(String(values.command), String(values.workdir));
+            case "shell_command":
+                return this.codeBash(String(values.command), String(values.workdir), values.timeout);
             case "build_and_start":
                 return this.buildAndStart();
             case "web_search":
@@ -1531,6 +1598,11 @@ export class CodeToolExecutor {
         this.forgetListedPath(this.relativePath(target));
         return { status: "success" };
     }
+    discoverableFile(filePath) {
+        // Full Stack captures expose the lock to shell/read_file, but omit it
+        // from both project discovery tools, even before any package installation.
+        return this.config.mode !== "full-stack" || this.relativePath(filePath) !== "package-lock.json";
+    }
     async listCodeFiles(rawPath) {
         if (this.config.mode === "react" && rawPath === "/")
             rawPath = ".";
@@ -1542,7 +1614,7 @@ export class CodeToolExecutor {
             !existsSync(target)) {
             return { status: "success", files: [] };
         }
-        const files = this.orderListedFiles(await this.walkFiles(target));
+        const files = this.orderListedFiles((await this.walkFiles(target)).filter(file => this.discoverableFile(file)));
         return {
             status: "success",
             files: (this.config.mode === "react" ? files.slice(0, 500) : files)
@@ -1744,6 +1816,7 @@ export class CodeToolExecutor {
             return { status: "success", paths: [], truncated: false };
         }
         const allPaths = files
+            .filter(file => this.discoverableFile(file))
             .filter((filePath) => matchesGlob(path.relative(target, filePath).split(path.sep).join("/"), pattern) || (this.config.mode === "react" && matchesGlob(this.relativePath(filePath), pattern)))
             .map((filePath) => this.relativePath(filePath));
         return {
@@ -1859,16 +1932,23 @@ export class CodeToolExecutor {
             stderr: this.virtualizeCodeOutput(result.stderr),
         };
     }
-    async fullStackBash(command, workdir) {
+    async codeBash(command, workdir, timeoutMs) {
+        if (!command.trim()) {
+            return { status: "error", message: "Command cannot be empty" };
+        }
         const cwd = workdir.trim() || ".";
         if (path.isAbsolute(cwd) || cwd.includes("..")) {
             throw new Error(`Invalid workdir: ${cwd}. Commands run from the project root by default. Only use workdir for relative subdirectories, never absolute paths.`);
         }
-        const configuredTimeout = Number(process.env.CODE_HARNESS_BASH_TIMEOUT_SECONDS || 300);
-        const timeoutSeconds = Number.isFinite(configuredTimeout) && configuredTimeout > 0
-            ? configuredTimeout
-            : 300;
+        const configuredTimeout = Number(process.env.CODE_HARNESS_BASH_TIMEOUT_SECONDS || 180);
+        const timeoutSeconds = timeoutMs !== undefined
+            ? timeoutMs / 1000
+            : Number.isFinite(configuredTimeout) && configuredTimeout > 0
+                ? configuredTimeout
+                : 180;
         const result = await this.bash(command, cwd, timeoutSeconds);
+        if (result.timed_out)
+            return { status: "error", message: "Command timed out" };
         const exitCode = result.exit_code;
         const stdout = String(result.stdout || "");
         const stderr = String(result.stderr || "");
@@ -2068,12 +2148,6 @@ export class CodeToolExecutor {
                     reject(this.signal.reason || new Error("operation cancelled"));
                     return;
                 }
-                // 生产契约：status 恒为 "completed"（包括非零退出码），超时无参考样本，
-                // 按最小猜测同样返回 completed + exit_code null，stderr 附超时说明。
-                if (timedOut && this.config.mode !== "full-stack") {
-                    stderr +=
-                        (stderr ? "\n" : "") + `command exceeded ${timeoutSeconds}s`;
-                }
                 const out = capBashOutput(stdout);
                 const err = capBashOutput(stderr);
                 resolve({
@@ -2081,6 +2155,7 @@ export class CodeToolExecutor {
                     stderr: err.text,
                     stdout: out.text,
                     exit_code: timedOut ? null : code,
+                    timed_out: timedOut,
                     duration_ms: Date.now() - startedAt,
                     stderr_truncated: err.truncated,
                     stdout_truncated: out.truncated,
@@ -2299,9 +2374,6 @@ export class CodeToolExecutor {
             throw new Error("file_path must end in .jpg, .jpeg, or .png");
         }
         mkdirSync(path.dirname(target), { recursive: true });
-        const model = process.env.CODE_HARNESS_IMAGE_MODEL ||
-            process.env.ARENA_IMAGE_MODEL ||
-            "gemini-3.1-flash-image";
         const references = [];
         for (const imagePath of images || []) {
             const source = await this.confinedPath(imagePath);
@@ -2322,7 +2394,7 @@ export class CodeToolExecutor {
         }
         const generated = await generateImageWithModelRouter({
             images: references,
-            model,
+            model: codeArenaImageModel(),
             prompt,
             signal: this.signal,
             size: process.env.CODE_HARNESS_IMAGE_SIZE ||