@botbotgo/agent-harness 0.0.320 → 0.0.322

package/dist/cli/chat-rendering.js

@@ -1,9 +1,10 @@
+import { markdownToConsole } from "../projections/presentation.js";
 import { renderChatRuntimeFailure } from "./chat-workspace.js";
 export function countRenderedLines(text) {
     return text.replace(/\n$/, "").split("\n").length;
 }
 export function renderChatTextChunk(text, modelInfo) {
-    return renderChatRuntimeFailure(text, modelInfo);
+    return markdownToConsole(renderChatRuntimeFailure(text, modelInfo));
 }
 export function renderChatRequestRunning(input) {
     const parts = [

package/dist/package-version.d.ts

@@ -1 +1 @@
-export declare const AGENT_HARNESS_VERSION = "0.0.319";
+export declare const AGENT_HARNESS_VERSION = "0.0.321";

package/dist/package-version.js

@@ -1 +1 @@
-export const AGENT_HARNESS_VERSION = "0.0.319";
+export const AGENT_HARNESS_VERSION = "0.0.321";

package/dist/projections/presentation.js

@@ -94,7 +94,7 @@ export function markdownToHtml(markdown) {
 }
 export function markdownToConsole(markdown) {
     if (!markdown.includes("\n")) {
-        return renderInlineConsoleMarkdown(markdown);
+        return renderInlineConsoleMarkdown(markdown.trim());
     }
     const normalized = markdown.replace(/\r\n/g, "\n");
     const blocks = normalized.split(/\n\n+/);
@@ -114,11 +114,12 @@ export function markdownToConsole(markdown) {
             const language = lines[0]?.slice(3).trim();
             const code = lines.slice(1, -1);
             const label = language ? ` ${language} ` : " code ";
+            const borderWidth = Math.max(label.length, 6);
             rendered.push([
-                `\u001b[2m┌${"─".repeat(Math.max(label.length, 6))}\u001b[0m`,
+                `\u001b[2m┌${"─".repeat(borderWidth)}\u001b[0m`,
                 `\u001b[2m│\u001b[0m${label}`,
                 ...code.map((line) => `  ${line}`),
-                "\u001b[2m└──────\u001b[0m",
+                `\u001b[2m└${"─".repeat(borderWidth)}\u001b[0m`,
             ].join("\n"));
             continue;
         }
@@ -130,25 +131,30 @@ export function markdownToConsole(markdown) {
             rendered.push(`\u001b[1m${title}\u001b[0m\n\u001b[2m${underline.repeat(Math.max(3, title.replace(/\u001b\[[0-9;]*m/g, "").length))}\u001b[0m`);
             continue;
         }
-        if (trimmed.split("\n").every((line) => /^[-*]\s+/.test(line))) {
+        if (trimmed.split("\n").every((line) => /^\s*[-*]\s+/.test(line))) {
             rendered.push(trimmed
                 .split("\n")
-                .map((line) => line.replace(/^[-*]\s+/, "• "))
+                .map((line) => {
+                const match = line.match(/^(\s*)[-*]\s+(.*)$/);
+                const indent = match?.[1] ?? "";
+                const content = match?.[2] ?? line;
+                return `${indent}• ${content}`;
+            })
                 .map((line) => renderInlineConsoleMarkdown(line))
                 .join("\n"));
             continue;
         }
-        if (trimmed.split("\n").every((line) => /^\d+\.\s+/.test(line))) {
+        if (trimmed.split("\n").every((line) => /^\s*\d+\.\s+/.test(line))) {
             rendered.push(trimmed
                 .split("\n")
                 .map((line) => renderInlineConsoleMarkdown(line))
                 .join("\n"));
             continue;
         }
-        if (trimmed.split("\n").every((line) => /^>\s?/.test(line))) {
+        if (trimmed.split("\n").every((line) => /^\s*>\s?/.test(line))) {
             rendered.push(trimmed
                 .split("\n")
-                .map((line) => line.replace(/^>\s?/, ""))
+                .map((line) => line.replace(/^\s*>\s?/, ""))
                 .map((line) => `\u001b[2m│\u001b[0m ${renderInlineConsoleMarkdown(line)}`)
                 .join("\n"));
             continue;

package/dist/resources/prompts/runtime/autonomous-investigation-recovery.md

@@ -0,0 +1 @@
+Do not stop at a plan or ask the user to choose the next obvious diagnostic step when the request is for deep investigation, root-cause analysis, or step-by-step execution and tool evidence already exists in the conversation. Continue the investigation yourself with the next concrete tool call. If the task is non-trivial, call write_todos, then keep executing the next diagnostic steps until you can explain the likely causes, impact, and recommended next actions.

package/dist/runtime/parsing/output-recovery.d.ts

@@ -11,4 +11,4 @@ export declare function resolveExecutionWithoutToolEvidenceTextInstruction(reque
 export declare function resolveToolCallRecoveryInstruction(error: unknown): string | null;
 export declare function appendToolRecoveryInstruction(input: unknown, instruction: string): unknown;
 export declare function wrapResolvedModel<T>(value: T): T;
-export { BROWSER_CAPABILITY_DISCLAIMER_RECOVERY_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_RETRY_INSTRUCTION, INVALID_TOOL_SELECTION_RECOVERY_INSTRUCTION, STRICT_TOOL_JSON_INSTRUCTION, WORKSPACE_RELATIVE_PATH_INSTRUCTION, WRITE_TODOS_DESCRIPTIVE_CONTENT_INSTRUCTION, WRITE_TODOS_FULL_ENTRY_INSTRUCTION, WRITE_TODOS_NON_EMPTY_INITIAL_LIST_INSTRUCTION } from "../prompts/runtime-prompts.js";
+export { AUTONOMOUS_INVESTIGATION_RECOVERY_INSTRUCTION, BROWSER_CAPABILITY_DISCLAIMER_RECOVERY_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_RETRY_INSTRUCTION, INVALID_TOOL_SELECTION_RECOVERY_INSTRUCTION, STRICT_TOOL_JSON_INSTRUCTION, WORKSPACE_RELATIVE_PATH_INSTRUCTION, WRITE_TODOS_DESCRIPTIVE_CONTENT_INSTRUCTION, WRITE_TODOS_FULL_ENTRY_INSTRUCTION, WRITE_TODOS_NON_EMPTY_INITIAL_LIST_INSTRUCTION } from "../prompts/runtime-prompts.js";

package/dist/runtime/parsing/output-recovery.js

@@ -1,4 +1,4 @@
-import { BROWSER_CAPABILITY_DISCLAIMER_RECOVERY_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_RETRY_INSTRUCTION, STRICT_TOOL_JSON_INSTRUCTION, WORKSPACE_RELATIVE_PATH_INSTRUCTION, WRITE_TODOS_DESCRIPTIVE_CONTENT_INSTRUCTION, WRITE_TODOS_FULL_ENTRY_INSTRUCTION, WRITE_TODOS_NON_EMPTY_INITIAL_LIST_INSTRUCTION, } from "../prompts/runtime-prompts.js";
+import { AUTONOMOUS_INVESTIGATION_RECOVERY_INSTRUCTION, BROWSER_CAPABILITY_DISCLAIMER_RECOVERY_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_RETRY_INSTRUCTION, STRICT_TOOL_JSON_INSTRUCTION, WORKSPACE_RELATIVE_PATH_INSTRUCTION, WRITE_TODOS_DESCRIPTIVE_CONTENT_INSTRUCTION, WRITE_TODOS_FULL_ENTRY_INSTRUCTION, WRITE_TODOS_NON_EMPTY_INITIAL_LIST_INSTRUCTION, } from "../prompts/runtime-prompts.js";
 import { hasToolCalls, readTextContent, wrapNormalizedMessage } from "./output-content.js";
 function isToolCallParseFailure(error) {
     if (!(error instanceof Error))
@@ -123,6 +123,71 @@ function extractLastUserMessageText(input) {
     }
     return "";
 }
+function extractAllUserMessageText(input) {
+    if (!Array.isArray(input) && !(typeof input === "object" && input && Array.isArray(input.messages))) {
+        return "";
+    }
+    const messages = Array.isArray(input) ? input : input.messages;
+    const values = [];
+    for (const message of messages) {
+        if (typeof message !== "object" || !message)
+            continue;
+        const typed = message;
+        if (typed.role === "user") {
+            const text = extractMessageContent(message).trim();
+            if (text)
+                values.push(text);
+            continue;
+        }
+        const ids = Array.isArray(typed.id) ? typed.id.filter((item) => typeof item === "string") : [];
+        const typeName = ids.at(-1);
+        const runtimeType = typeof message._getType === "function"
+            ? message._getType()
+            : typeof message.getType === "function"
+                ? message.getType()
+                : undefined;
+        if (typeName === "HumanMessage" || runtimeType === "human") {
+            const text = extractMessageContent(message).trim();
+            if (text)
+                values.push(text);
+            continue;
+        }
+        if (typeof typed.kwargs === "object" && typed.kwargs && typed.kwargs.role === "user") {
+            const text = extractMessageContent(message).trim();
+            if (text)
+                values.push(text);
+        }
+    }
+    return values.join("\n").trim();
+}
+function hasToolResultEvidence(input) {
+    if (!Array.isArray(input) && !(typeof input === "object" && input && Array.isArray(input.messages))) {
+        return false;
+    }
+    const messages = Array.isArray(input) ? input : input.messages;
+    for (const message of messages) {
+        if (typeof message !== "object" || !message)
+            continue;
+        const typed = message;
+        if (typed.role === "tool") {
+            return true;
+        }
+        const ids = Array.isArray(typed.id) ? typed.id.filter((item) => typeof item === "string") : [];
+        const typeName = ids.at(-1);
+        const runtimeType = typeof message._getType === "function"
+            ? message._getType()
+            : typeof message.getType === "function"
+                ? message.getType()
+                : undefined;
+        if (typeName === "ToolMessage" || runtimeType === "tool") {
+            return true;
+        }
+        if (typeof typed.kwargs === "object" && typed.kwargs && typed.kwargs.role === "tool") {
+            return true;
+        }
+    }
+    return false;
+}
 function isToolRequiredExecutionPrompt(text) {
     if (!text)
         return false;
@@ -130,6 +195,12 @@ function isToolRequiredExecutionPrompt(text) {
         /\b(create|write|edit|read|show|display|append|run|execute|wait|sleep|repeat|loop|file|files|command)\b/i.test(text) ||
         /(创建|写入|读取|显示|追加|执行|运行|等待|重复|文件|命令|步骤)/.test(text));
 }
+function isAutonomousInvestigationPrompt(text) {
+    if (!text)
+        return false;
+    return (/\b(root cause|rca|deep investigation|investigate deeply|investigate fully|step by step|find all (?:issues|problems|causes)|continue investigating|keep digging|until complete|until you find)\b/i.test(text) ||
+        /(根因分析|深度调查|深入调查|一步一步|逐步分析|逐个分析|继续排查|排查到底|找到所有问题|查出所有问题|直到完成|直到找到.*原因)/.test(text));
+}
 function isBrowserOrUrlTask(text) {
     if (!text)
         return false;
@@ -149,6 +220,18 @@ function claimsMissingBrowserCapability(text) {
     return (/\b(i do not have the capability to access external websites|i do not have access to external websites|i cannot access external websites|i can(?:not|'t) browse|i do not have (?:direct )?(?:web|browser|browsing) access|i do not have .*tool.*(?:website|url|browser)|i cannot summarize .* because .*tool.*failed|please provide (?:the )?text content.*i (?:will|can) summarize)\b/i.test(text) ||
         /(我没有直接访问外部网站|我无法直接访问外部网站|我无法访问外部网站|我不能访问外部网站|我无法直接访问外部网站内容|我无法访问网页|请提供文本内容.*我将为您进行总结|我没有.*浏览.*工具|我没有.*网页抓取.*工具|我没有.*访问.*网站.*工具|我无法总结该网页内容，因为.*工具执行失败)/.test(text));
 }
+function asksUserToChooseObviousNextDiagnosticStep(text) {
+    if (!text)
+        return false;
+    return (/\b(which (?:direction|aspect|node|namespace|part)|what would you like me to check|please choose|please tell me (?:which|what|where) .* (?:check|inspect|investigate)|which .* should we start with)\b/i.test(text) ||
+        /(请告诉我.*(哪个|哪一个|什么方向)|请选择.*(方向|项目|节点)|您希望我.*(检查|查看).*(哪个|哪一个|方向)|从哪个方向开始)/.test(text));
+}
+function claimsFutureExecutionWithoutToolEvidence(text) {
+    if (!text)
+        return false;
+    return (/\b(i will|i'll|we will|next i(?:'ll| will)|i am going to)\b.*\b(run|execute|inspect|check|investigate|continue|start)\b/i.test(text) ||
+        /(我将|我会|接下来我会|现在我将|我将继续).*(执行|检查|查看|调查|排查|开始|继续)/.test(text));
+}
 export function isRetrySafeInvalidToolSelectionError(value) {
     const text = readTextContent(value).trim();
     if (!text)
@@ -156,9 +239,9 @@ export function isRetrySafeInvalidToolSelectionError(value) {
     return /is not a valid tool, try one of \[/i.test(text);
 }
 export function shouldValidateExecutionWithoutToolEvidence(request) {
-    const userText = extractLastUserMessageText(request);
+    const userText = extractAllUserMessageText(request);
     const browserOrUrlTask = isBrowserOrUrlTask(userText);
-    return browserOrUrlTask || isToolRequiredExecutionPrompt(userText);
+    return browserOrUrlTask || isToolRequiredExecutionPrompt(userText) || isAutonomousInvestigationPrompt(userText);
 }
 function extractAssistantTextFromResult(result) {
     if (typeof result === "object" && result && "messages" in result) {
@@ -192,14 +275,21 @@ export function resolveExecutionWithoutToolEvidenceInstruction(request, result)
     return resolveExecutionWithoutToolEvidenceTextInstruction(request, assistantText, hasToolCalls(result));
 }
 export function resolveExecutionWithoutToolEvidenceTextInstruction(request, assistantText, toolCallEvidence = false) {
-    const userText = extractLastUserMessageText(request);
+    const userText = extractAllUserMessageText(request);
     const browserOrUrlTask = isBrowserOrUrlTask(userText);
+    const autonomousInvestigationTask = isAutonomousInvestigationPrompt(userText);
+    const toolResultEvidence = hasToolResultEvidence(request);
     if (!shouldValidateExecutionWithoutToolEvidence(request)) {
         return null;
     }
     if (!assistantText || toolCallEvidence) {
         return null;
     }
+    if (autonomousInvestigationTask &&
+        toolResultEvidence &&
+        (asksUserToChooseObviousNextDiagnosticStep(assistantText) || claimsFutureExecutionWithoutToolEvidence(assistantText))) {
+        return AUTONOMOUS_INVESTIGATION_RECOVERY_INSTRUCTION;
+    }
     if (browserOrUrlTask && claimsMissingBrowserCapability(assistantText)) {
         return BROWSER_CAPABILITY_DISCLAIMER_RECOVERY_INSTRUCTION;
     }
@@ -285,4 +375,4 @@ export function wrapResolvedModel(value) {
         },
     });
 }
-export { BROWSER_CAPABILITY_DISCLAIMER_RECOVERY_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_RETRY_INSTRUCTION, INVALID_TOOL_SELECTION_RECOVERY_INSTRUCTION, STRICT_TOOL_JSON_INSTRUCTION, WORKSPACE_RELATIVE_PATH_INSTRUCTION, WRITE_TODOS_DESCRIPTIVE_CONTENT_INSTRUCTION, WRITE_TODOS_FULL_ENTRY_INSTRUCTION, WRITE_TODOS_NON_EMPTY_INITIAL_LIST_INSTRUCTION } from "../prompts/runtime-prompts.js";
+export { AUTONOMOUS_INVESTIGATION_RECOVERY_INSTRUCTION, BROWSER_CAPABILITY_DISCLAIMER_RECOVERY_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_INSTRUCTION, EXECUTION_WITH_TOOL_EVIDENCE_RETRY_INSTRUCTION, INVALID_TOOL_SELECTION_RECOVERY_INSTRUCTION, STRICT_TOOL_JSON_INSTRUCTION, WORKSPACE_RELATIVE_PATH_INSTRUCTION, WRITE_TODOS_DESCRIPTIVE_CONTENT_INSTRUCTION, WRITE_TODOS_FULL_ENTRY_INSTRUCTION, WRITE_TODOS_NON_EMPTY_INITIAL_LIST_INSTRUCTION } from "../prompts/runtime-prompts.js";

package/dist/runtime/prompts/runtime-prompts.d.ts

@@ -9,6 +9,7 @@ export declare const WRITE_TODOS_NON_EMPTY_INITIAL_LIST_INSTRUCTION: string;
 export declare const WRITE_TODOS_DESCRIPTIVE_CONTENT_INSTRUCTION: string;
 export declare const EXECUTION_WITH_TOOL_EVIDENCE_INSTRUCTION: string;
 export declare const EXECUTION_WITH_TOOL_EVIDENCE_RETRY_INSTRUCTION: string;
+export declare const AUTONOMOUS_INVESTIGATION_RECOVERY_INSTRUCTION: string;
 export declare const WORKSPACE_RELATIVE_PATH_INSTRUCTION: string;
 export declare function renderDurableMemoryContextPrompt(memoryContext: string): string;
 export declare function renderSlashCommandSkillInstruction(input: {

package/dist/runtime/prompts/runtime-prompts.js

@@ -12,6 +12,7 @@ export const WRITE_TODOS_NON_EMPTY_INITIAL_LIST_INSTRUCTION = readRuntimePrompt(
 export const WRITE_TODOS_DESCRIPTIVE_CONTENT_INSTRUCTION = readRuntimePrompt("write-todos-descriptive-content");
 export const EXECUTION_WITH_TOOL_EVIDENCE_INSTRUCTION = readRuntimePrompt("execution-with-tool-evidence");
 export const EXECUTION_WITH_TOOL_EVIDENCE_RETRY_INSTRUCTION = readRuntimePrompt("execution-with-tool-evidence-retry");
+export const AUTONOMOUS_INVESTIGATION_RECOVERY_INSTRUCTION = readRuntimePrompt("autonomous-investigation-recovery");
 export const WORKSPACE_RELATIVE_PATH_INSTRUCTION = readRuntimePrompt("workspace-relative-path");
 export function renderDurableMemoryContextPrompt(memoryContext) {
     return renderBundledTemplate("prompts/runtime/durable-memory-context.md", {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@botbotgo/agent-harness",
-  "version": "0.0.320",
+  "version": "0.0.322",
   "description": "Workspace runtime for multi-agent applications",
   "license": "MIT",
   "type": "module",