nyxora 26.7.9 → 26.7.10

package/CHANGELOG.md

@@ -4,6 +4,23 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepashangelog.com/en/1.0.0/),
 
+## [26.7.10]
+### Web3 Integrations
+- **Robinhood Chain Support**: Officially integrated `Robinhood Chain` (Mainnet) and `Robinhood Testnet` support natively via `viem` `v2.55.0` upgrade. Added seamless RPC mappings, updated the dashboard network selectors with official SVG branding, and ensured high-frequency trading capabilities in the Signer remain stable.
+
+### Core Architecture & Routing
+- **Deterministic Intent Fast-Path**: Bypassed the LLM semantic router for short confirmation phrases. Messages containing global confirmation keywords (e.g., "yes", "sure", "proceed") now deterministically inherit the strict contextual boundary (`os` or `web3`) of the Assistant's previous permission request, completely eradicating routing hallucinations.
+- **Global Codebase Standardization**: Refactored hardcoded regional slang from the routing logic into a comprehensive, professional English-first `CONFIRM_WORDS` array, natively supporting global users while retaining common localized confirmations as secondary fallbacks.
+
+### Fallback Execution Engine
+- **Native `<execute_bash>` Support**: Enhanced the `osAgent` and `web3Agent` Fallback Parsers to autonomously intercept and parse `<execute_bash>` and `<execute>` XML tags. This restores full execution capability for open-weight models that instinctively utilize these tags instead of strict JSON tool calls.
+- **Display Sanitization Hardening**: Appended `execute_bash` and `execute` to the automated UI Sanitizer `tagsToRemove` registry. Raw bash execution blocks hallucinated by the LLM are now stripped at the edge layer before streaming to Telegram or the Dashboard.
+
+### Web3 Signer SDK & Execution Guardrails
+- **Receipt Waiter Integration (Anti-False-Positive)**: Remediated a severe "Fire-and-Forget" architectural bug in `NyxoraSigner.ts` where broadcasted transactions were immediately reported as successful to the LLM regardless of actual on-chain finality. The Signer now strictly executes a `waitForTransactionReceipt` hook (capped at a 20-second timeout). 
+- **Revert Detection**: If a transaction reverts on-chain (e.g., due to strict MEV slippage or gas exhaustion), the Signer violently rejects the promise with `reverted`, preventing the AI from falsely declaring success.
+- **Pending Timeout Grace**: If the blockchain experiences congestion and fails to confirm within the 20-second window, the system falls back gracefully by returning `"Transaction broadcasted (Pending receipt)"` to ensure the 30-second Policy Engine HTTP timeout is never triggered, allowing the AI to report accurate pending status.
+
 ## [26.7.9]
 ### Features & Architecture
 

package/README.md

@@ -2,14 +2,15 @@
 **Your Personal Web3 Assistant.**
 
 
-[![Status: Prototype](https://img.shields.io/badge/Status-Prototype-orange.svg)](#)
-
-[![MCP Supported](https://img.shields.io/badge/MCP-Supported-blue.svg)](#)
-[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
-[![Security: Defense-in-Depth](https://img.shields.io/badge/Security-Defense--in--Depth-blue.svg)](#️-advanced-security-threat-model)
-[![Execution: Cryptographic Approval](https://img.shields.io/badge/Execution-Cryptographic--Approval-orange.svg)](#️-advanced-security-threat-model)
-[![Privacy: Local-Only Keys](https://img.shields.io/badge/Privacy-Local--Only--Keys-success.svg)](#️-advanced-security-threat-model)
-[![PRs Welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg)](https://makeapullrequest.com)
+[![Status: Alpha](https://img.shields.io/badge/Status-Alpha-orange.svg?style=for-the-badge)](#)
+[![Donate on DeBank](https://img.shields.io/badge/DONATE-DEBANK-FE7E41?style=for-the-badge&logo=ethereum&logoColor=white)](https://debank.com/profile/0xe5c21f46993c67cfe04fcf1579486d390be7b535)
+
+[![MCP Supported](https://img.shields.io/badge/MCP-Supported-blue.svg?style=for-the-badge)](#)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg?style=for-the-badge)](https://opensource.org/licenses/MIT)
+[![Security: Defense-in-Depth](https://img.shields.io/badge/Security-Defense--in--Depth-blue.svg?style=for-the-badge)](#️-advanced-security-threat-model)
+[![Execution: Cryptographic Approval](https://img.shields.io/badge/Execution-Cryptographic--Approval-orange.svg?style=for-the-badge)](#️-advanced-security-threat-model)
+[![Privacy: Local-Only Keys](https://img.shields.io/badge/Privacy-Local--Only--Keys-success.svg?style=for-the-badge)](#️-advanced-security-threat-model)
+[![PRs Welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg?style=for-the-badge)](https://makeapullrequest.com)
 
 Nyxora is a **secure, non-custodial runtime infrastructure for autonomous onchain agents** built with a robust Monorepo architecture (Node.js & React). Designed for autonomous workflows with a premium Utility-Centric dark-themed UI and strict client-side key isolation. 
 

package/SUPPORT.md

@@ -0,0 +1,18 @@
+# Support Nyxora Protocol
+
+Nyxora is an open-source initiative built on the belief that powerful, zero-trust Web3 AI infrastructure should be accessible to everyone. We dedicate countless hours to research, continuous development, and maintaining this ecosystem to ensure it remains secure and cutting-edge.
+
+If Nyxora has brought value to you, your projects, or your organization, consider supporting our ongoing development. While completely optional, your contribution plays a vital role in helping us cover infrastructure costs, fund security audits, and dedicate more time to shipping new features.
+
+### How to Contribute
+
+We gratefully accept community support via any EVM-compatible network (Ethereum, Arbitrum, Optimism, Polygon, Binance Smart Chain, etc.).
+
+**EVM Address:**
+`0xE5c21F46993C67CFe04FCF1579486D390Be7B535`
+
+> 🔗 **[View & Donate via DeBank](https://debank.com/profile/0xe5c21f46993c67cfe04fcf1579486d390be7b535)**
+
+*Every contribution, no matter the size, is deeply appreciated and directly fuels the future of autonomous Web3 AI.*
+
+Thank you for being an amazing part of the Nyxora journey! 🌌

package/dist/packages/core/src/agent/osAgent.js

@@ -171,7 +171,7 @@ CRITICAL INSTRUCTIONS:
             }
             tracker_1.Tracker.addEvent('llm.response', { provider: config.llm.provider, tool_calls: responseMessage.tool_calls?.length || 0 });
             // P1: Capture <think> blocks for scratchpad, get clean content
-            const cleanedContent = scratchpad.capture(responseMessage.content || '', turnCount);
+            let cleanedContent = scratchpad.capture(responseMessage.content || '', turnCount);
             exports.logger.addEntry({
                 role: 'assistant',
                 content: cleanedContent || '',
@@ -225,9 +225,51 @@ CRITICAL INSTRUCTIONS:
                         });
                     }
                 }
+                // 3. <tool_code> JSON arrays
+                const toolCodeRegex = /<tool_code>([\s\S]*?)<\/tool_code>/g;
+                let toolCodeMatch;
+                while ((toolCodeMatch = toolCodeRegex.exec(responseMessage.content)) !== null) {
+                    try {
+                        const jsonStr = toolCodeMatch[1].trim();
+                        const parsedArray = JSON.parse(jsonStr);
+                        if (Array.isArray(parsedArray)) {
+                            for (const item of parsedArray) {
+                                if (item.function_name && item.tool_args) {
+                                    fallbacks.push({
+                                        id: 'call_fallback_' + Math.random().toString(36).substr(2, 9),
+                                        type: 'function',
+                                        function: {
+                                            name: item.function_name,
+                                            arguments: typeof item.tool_args === 'string' ? item.tool_args : JSON.stringify(item.tool_args)
+                                        }
+                                    });
+                                }
+                            }
+                        }
+                    }
+                    catch (e) {
+                        // Ignore parse errors, maybe it's not JSON
+                    }
+                }
+                // 4. <execute_bash> or <execute> blocks
+                const executeRegex = /<(?:execute_bash|execute)>([\s\S]*?)<\/(?:execute_bash|execute)>/g;
+                let executeMatch;
+                while ((executeMatch = executeRegex.exec(responseMessage.content)) !== null) {
+                    const bashCmd = executeMatch[1].trim();
+                    if (bashCmd) {
+                        fallbacks.push({
+                            id: 'call_fallback_' + Math.random().toString(36).substr(2, 9),
+                            type: 'function',
+                            function: {
+                                name: 'run_terminal_command',
+                                arguments: JSON.stringify({ command: bashCmd })
+                            }
+                        });
+                    }
+                }
                 if (fallbacks.length > 0) {
                     responseMessage.tool_calls = fallbacks;
-                    responseMessage.content = responseMessage.content.replace(/^\/[a-zA-Z0-9_]+\s+.*$/gm, '').replace(/```(?:bash|sh|zsh)?\n([\s\S]*?)```/g, '').trim();
+                    responseMessage.content = responseMessage.content.replace(/^\/[a-zA-Z0-9_]+\s+.*$/gm, '').replace(/```(?:bash|sh|zsh)?\n([\s\S]*?)```/g, '').replace(/<tool_code>([\s\S]*?)<\/tool_code>/g, '').replace(/<(?:execute_bash|execute)>([\s\S]*?)<\/(?:execute_bash|execute)>/g, '').trim();
                     console.log(picocolors_1.default.cyan(`[Fallback Parser] Intercepted ${fallbacks.length} raw text commands and converted to tool_calls.`));
                     // Update logger entry with the intercepted tool calls
                     exports.logger.addEntry({
@@ -238,6 +280,22 @@ CRITICAL INSTRUCTIONS:
                 }
             }
             // -----------------------------------------------------
+            // --- DISPLAY SANITIZATION ---
+            // Remove commonly leaked XML tags (reasoning & tool calls) to prevent UI clutter
+            if (responseMessage.content) {
+                const tagsToRemove = ['tool_code', 'tool_call', 'tool_calls', 'function_call', 'function_calls', 'execute', 'execute_bash', 'think', 'thought', 'reasoning'];
+                tagsToRemove.forEach(tag => {
+                    // Remove complete tags and their contents
+                    const regex = new RegExp(`<${tag}\\b[^>]*>[\\s\\S]*?<\\/${tag}>\\s*`, 'gi');
+                    responseMessage.content = responseMessage.content.replace(regex, '');
+                    // Remove any orphaned closing tags
+                    const orphanRegex = new RegExp(`<\\/${tag}>\\s*`, 'gi');
+                    responseMessage.content = responseMessage.content.replace(orphanRegex, '');
+                });
+                responseMessage.content = responseMessage.content.trim();
+                cleanedContent = responseMessage.content;
+            }
+            // -----------------------------------------------------
             if (!responseMessage.tool_calls || responseMessage.tool_calls.length === 0) {
                 // --- CRITIC PASS REMOVED ---
                 // The new PromptBuilder handles robust reasoning. Post-generation critic
@@ -368,7 +426,7 @@ CRITICAL INSTRUCTIONS:
             // V2 Optimization: Zero-LLM Fast Return for transaction tools
             if (canFastReturnAll && accumulatedResults.length > 0) {
                 const finalContent = accumulatedResults.join('\n\n---\n\n');
-                exports.logger.addEntry({ role: 'assistant', content: finalContent }, sessionId);
+                exports.logger.addEntry({ role: 'user', content: `[System Notification: The tool executed successfully and returned the following result directly to the user]\n${finalContent}` }, sessionId);
                 triggerBackgroundReview(sessionId);
                 return finalContent;
             }
@@ -454,11 +512,9 @@ CRITICAL INSTRUCTIONS:
             let streamedContent = '';
             const response = await (0, llmUtils_1.executeWithRetry)(async (client) => {
                 streamedContent = '';
-                // RC#1 FIX: Only clear the Telegram buffer on the FIRST turn.
-                // On subsequent turns (after tool calls), the buffer already shows useful
-                // progress info. Resetting it causes visible content to disappear.
-                if (turnCount === 1)
-                    onChunk('[CLEAR_STREAM]');
+                // RC#1 FIX: Always clear the buffer at the start of the stream turn.
+                // This ensures the client UI doesn't append duplicate preambles across multi-turn executions.
+                onChunk('[CLEAR_STREAM]');
                 return await client.stream({ model: config.llm.model, temperature: config.llm.temperature, messages, tools: activeTools, reasoning_effort: (!config.llm.reasoning_effort || config.llm.reasoning_effort === 'none') ? undefined : config.llm.reasoning_effort }, (chunk) => {
                     streamedContent += chunk;
                     onChunk(chunk);

package/dist/packages/core/src/agent/promptBuilder.js

@@ -231,19 +231,15 @@ When using the search_web tool to look up news, current events, or factual data:
 </web_search_accuracy>
 
 <act_dont_ask_os>
-NEVER show a command as a markdown code block and wait. CALL the tool directly.
-NEVER ask "do you want me to run this?" — just run it.
-NEVER say "you need to run this yourself" — you have direct shell access.
-If a command requires sudo and may need a password, just run it and report what happens.
-Only report failure AFTER actually attempting the tool call and receiving an error.
+For harmless commands (e.g. ls, cat, checking system info), CALL the tool directly without asking for confirmation.
+CRITICAL: For ANY command that modifies the system (e.g., sudo, apt-get, install, rm), you MUST ask the user for explicit permission FIRST.
+When asking for permission, simply ask: "Do you want me to run [command]?" and STOP.
+Once the user replies "yes", you MUST immediately emit the tool call to execute the command. Do NOT just reply with text confirming it.
 </act_dont_ask_os>
 
 <anti_hallucination_execution>
-CRITICAL: It is STRICTLY FORBIDDEN to write a bash/shell command in a markdown code block (e.g. \`\`\`bash ... \`\`\`) as a substitute for calling the run_terminal_command tool.
-Writing a code block does NOT execute anything. It is a lie to the user.
-If you write \`\`\`bash\nsudo apt install steam\n\`\`\` instead of calling run_terminal_command, you are hallucinating execution.
-The ONLY way to run a command is to emit a proper tool_call for run_terminal_command.
-If the tool is available, USE IT. Do not simulate or describe running it.
+Use the native JSON tool call format provided by the API. If your model architecture requires it, you may use <tool_code> JSON arrays.
+NEVER claim you have executed a tool or are running it in the background if you haven't actually emitted the tool call payload. Writing text describing execution is a lie.
 </anti_hallucination_execution>
 
 <search_hallucination_prevention>

package/dist/packages/core/src/agent/reasoning.js

@@ -47,10 +47,10 @@ const OS_KEYWORDS = [
     'excel', 'xlsx', 'spreadsheet', 'generate excel',
     'file', 'folder', 'directory', 'read file', 'write file', 'pdf', 'word', 'docx', 'document',
     'note', 'catat', 'catatan', 'keep', 'download', 'unduh', 'link',
-    // Terminal & Git
+    // Terminal & Git & System
     'terminal', 'command', 'shell', 'bash', 'script', 'run command',
     'git', 'commit', 'push', 'pull', 'clone', 'branch', 'merge',
-    // Web & Search
+    'install', 'installin', 'uninstall', 'apt', 'npm', 'pip', 'yarn', 'docker', 'system',
     'search web', 'google', 'browse', 'scrape', 'weather', 'news', 'search', 'find', 'look up',
     // Email & Workspace
     'email', 'gmail', 'google docs', 'google sheets', 'notion', 'calendar',
@@ -78,7 +78,7 @@ const WEB3_KEYWORDS = [
     'token', 'crypto', 'coin', 'nft', 'wallet', 'address', 'contract',
     // Major Chains & L2s
     'ethereum', 'polygon', 'arbitrum', 'optimism', 'bsc', 'mainnet', 'testnet',
-    'on-chain', 'blockchain', 'solana', 'avalanche', 'base',
+    'on-chain', 'blockchain', 'solana', 'avalanche', 'base', 'robinhood',
     // Top 50+ Global Assets / Tickers
     'eth', 'bnb', 'usdt', 'usdc', 'sol', 'matic', 'arb', 'op', 'btc', 'bitcoin',
     'xrp', 'doge', 'dogecoin', 'avax', 'ada', 'cardano', 'pepe', 'shib', 'shiba',
@@ -221,18 +221,44 @@ async function processUserInput(input, role = 'user', onProgress, sessionId) {
         context = 'web3';
         preCheckMatched = true;
     }
+    if (!preCheckMatched) {
+        // Support global confirmation keywords for short fast-path routing
+        const CONFIRM_WORDS = [
+            // English (Global Defaults)
+            'yes', 'yeah', 'yep', 'yup', 'sure', 'ok', 'okay', 'alright', 'confirm', 'proceed', 'continue', 'do it', 'go ahead', 'exactly', 'indeed',
+            // Common Localizations (e.g., ID, ES)
+            'ya', 'yakin', 'boleh', 'gas', 'lanjut', 'setuju', 'si', 'sí', 'vale', 'claro'
+        ];
+        const isConfirm = lowerInput.length < 25 && CONFIRM_WORDS.some(kw => lowerInput.includes(kw));
+        if (isConfirm && textOnlyHistory.length > 0) {
+            const lastMsg = textOnlyHistory[textOnlyHistory.length - 1];
+            if (lastMsg.role === 'assistant') {
+                const lastContent = lastMsg.content.toLowerCase();
+                // Check if the assistant was asking for permission to run an OS command
+                if (lastContent.includes('sudo ') || lastContent.includes('terminal') || lastContent.includes('apt') || lastContent.includes('perintah') || lastContent.includes('install') || lastContent.includes('command')) {
+                    context = 'os';
+                    preCheckMatched = true;
+                    // Check if the assistant was asking for permission to run a Web3 transaction
+                }
+                else if (lastContent.includes('swap') || lastContent.includes('transfer') || lastContent.includes('token') || lastContent.includes('wallet') || lastContent.includes('transaction')) {
+                    context = 'web3';
+                    preCheckMatched = true;
+                }
+            }
+        }
+    }
     if (preCheckMatched) {
-        console.log(picocolors_1.default.cyan(`[Orchestrator] Intent pre-classified (keyword match) as: ${context.toUpperCase()}`));
+        console.log(picocolors_1.default.cyan(`[Orchestrator] Intent pre-classified as: ${context.toUpperCase()}`));
     }
     else {
-        // ── Fallback: LLM Router (untuk intent ambigu / percakapan umum) ─────────
+        // ── Fallback: LLM Router (for ambiguous intents / general conversation) ─────────
         const routerPrompt = `You are Nyxora's Semantic Intent Router. Classify the user's FINAL message into one of four categories: 'web3', 'os', 'compound', or 'general'.
 Rules:
-1. FOCUS ONLY ON THE FINAL MESSAGE. History is only for context.
+1. FOCUS ON THE FINAL MESSAGE, but use history to understand short answers. If the final message is a short confirmation (e.g. "ya", "yes", "do it") or answer to a pending permission request for a tool, CLASSIFY IT BASED ON THE CONTEXT. If the previous message was about an OS command, reply 'os'. If it was about crypto/web3, reply 'web3'.
 2. The user may speak in ANY language, including casual slang, idioms, or abbreviations.
-3. 'compound' is EXTREMELY RARE. Only reply 'compound' when the user EXPLICITLY requests BOTH a blockchain/crypto action AND an OS/file/email action in the SAME message. Example: "swap 1 ETH to USDC and email me the receipt". If in doubt, do NOT use 'compound'.
+3. 'compound' is EXTREMELY RARE. Only reply 'compound' when the user EXPLICITLY requests BOTH a blockchain/crypto action AND an OS/file/email action in the SAME message.
 4. If the core intent involves ONLY blockchain, crypto, bridging, swapping, trading, sending/receiving, tokens, wallets, transactions, OR asking for the price/conversion of ANY asset to fiat, reply 'web3'.
-5. If the core intent involves ONLY OS automation, weather, emails, files, terminal, changing AI settings, OR asking ANY question that requires a web search or real-world factual lookup (e.g., 'who won the game', 'what is the registration date', 'cek info', 'cari tahu'), reply 'os'.
+5. If the core intent involves ONLY OS automation, weather, emails, files, terminal, changing AI settings, OR asking ANY question that requires a web search or real-world factual lookup, reply 'os'.
 6. If the message is casual conversation, chit-chat, greetings, capability questions (e.g., 'what can you do?', 'bisa ngapain?', 'help', 'menu'), or any open-ended/vague question, reply 'general'.
 Reply with EXACTLY ONE WORD: compound, web3, os, or general.`;
         const routerMessages = [
@@ -401,14 +427,40 @@ async function processUserInputStream(input, originalOnChunk, onProgress, sessio
             context = 'web3';
             preCheckMatched = true;
         }
+        if (!preCheckMatched) {
+            // Support global confirmation keywords for short fast-path routing
+            const CONFIRM_WORDS = [
+                // English (Global Defaults)
+                'yes', 'yeah', 'yep', 'yup', 'sure', 'ok', 'okay', 'alright', 'confirm', 'proceed', 'continue', 'do it', 'go ahead', 'exactly', 'indeed',
+                // Common Localizations (e.g., ID, ES)
+                'ya', 'yakin', 'boleh', 'gas', 'lanjut', 'setuju', 'si', 'sí', 'vale', 'claro'
+            ];
+            const isConfirm = lowerInput.length < 25 && CONFIRM_WORDS.some(kw => lowerInput.includes(kw));
+            if (isConfirm && textOnlyHistory.length > 0) {
+                const lastMsg = textOnlyHistory[textOnlyHistory.length - 1];
+                if (lastMsg.role === 'assistant') {
+                    const lastContent = lastMsg.content.toLowerCase();
+                    // Check if the assistant was asking for permission to run an OS command
+                    if (lastContent.includes('sudo ') || lastContent.includes('terminal') || lastContent.includes('apt') || lastContent.includes('perintah') || lastContent.includes('install') || lastContent.includes('command')) {
+                        context = 'os';
+                        preCheckMatched = true;
+                        // Check if the assistant was asking for permission to run a Web3 transaction
+                    }
+                    else if (lastContent.includes('swap') || lastContent.includes('transfer') || lastContent.includes('token') || lastContent.includes('wallet') || lastContent.includes('transaction')) {
+                        context = 'web3';
+                        preCheckMatched = true;
+                    }
+                }
+            }
+        }
         if (!preCheckMatched) {
             const routerPrompt = `You are Nyxora's Semantic Intent Router. Classify the user's FINAL message into one of four categories: 'web3', 'os', 'compound', or 'general'.
 Rules:
-1. FOCUS ONLY ON THE FINAL MESSAGE. History is only for context.
+1. FOCUS ON THE FINAL MESSAGE, but use history to understand short answers. If the final message is a short confirmation (e.g. "ya", "yes", "do it") or answer to a pending permission request for a tool, CLASSIFY IT BASED ON THE CONTEXT. If the previous message was about an OS command, reply 'os'. If it was about crypto/web3, reply 'web3'.
 2. The user may speak in ANY language, including casual slang, idioms, or abbreviations.
-3. 'compound' is EXTREMELY RARE. Only reply 'compound' when the user EXPLICITLY requests BOTH a blockchain/crypto action AND an OS/file/email action in the SAME message. Example: "swap 1 ETH to USDC and email me the receipt". If in doubt, do NOT use 'compound'.
+3. 'compound' is EXTREMELY RARE. Only reply 'compound' when the user EXPLICITLY requests BOTH a blockchain/crypto action AND an OS/file/email action in the SAME message.
 4. If the core intent involves ONLY blockchain, crypto, bridging, swapping, trading, sending/receiving, tokens, wallets, transactions, OR asking for the price/conversion of ANY asset to fiat, reply 'web3'.
-5. If the core intent involves ONLY OS automation, weather, emails, files, terminal, changing AI settings, OR asking ANY question that requires a web search or real-world factual lookup (e.g., 'who won the game', 'what is the registration date', 'cek info', 'cari tahu'), reply 'os'.
+5. If the core intent involves ONLY OS automation, weather, emails, files, terminal, changing AI settings, OR asking ANY question that requires a web search or real-world factual lookup, reply 'os'.
 6. If the message is casual conversation, chit-chat, greetings, capability questions (e.g., 'what can you do?', 'bisa ngapain?', 'help', 'menu'), or any open-ended/vague question, reply 'general'.
 Reply with EXACTLY ONE WORD: compound, web3, os, or general.`;
             const routerMessages = [

package/dist/packages/core/src/agent/web3Agent.js

@@ -96,7 +96,7 @@ async function processWeb3Intent(input, role = 'user', onProgress, sessionId) {
             }
             tracker_1.Tracker.addEvent('llm.response', { provider: config.llm.provider, tool_calls: responseMessage.tool_calls?.length || 0 });
             // P1: Capture <think> blocks for scratchpad, get clean content
-            const cleanedContent = scratchpad.capture(responseMessage.content || '', turnCount);
+            let cleanedContent = scratchpad.capture(responseMessage.content || '', turnCount);
             exports.logger.addEntry({
                 role: 'assistant',
                 content: cleanedContent || '',
@@ -150,9 +150,51 @@ async function processWeb3Intent(input, role = 'user', onProgress, sessionId) {
                         });
                     }
                 }
+                // 3. <tool_code> JSON arrays
+                const toolCodeRegex = /<tool_code>([\s\S]*?)<\/tool_code>/g;
+                let toolCodeMatch;
+                while ((toolCodeMatch = toolCodeRegex.exec(responseMessage.content)) !== null) {
+                    try {
+                        const jsonStr = toolCodeMatch[1].trim();
+                        const parsedArray = JSON.parse(jsonStr);
+                        if (Array.isArray(parsedArray)) {
+                            for (const item of parsedArray) {
+                                if (item.function_name && item.tool_args) {
+                                    fallbacks.push({
+                                        id: 'call_fallback_' + Math.random().toString(36).substr(2, 9),
+                                        type: 'function',
+                                        function: {
+                                            name: item.function_name,
+                                            arguments: typeof item.tool_args === 'string' ? item.tool_args : JSON.stringify(item.tool_args)
+                                        }
+                                    });
+                                }
+                            }
+                        }
+                    }
+                    catch (e) {
+                        // Ignore parse errors, maybe it's not JSON
+                    }
+                }
+                // 4. <execute_bash> or <execute> blocks
+                const executeRegex = /<(?:execute_bash|execute)>([\s\S]*?)<\/(?:execute_bash|execute)>/g;
+                let executeMatch;
+                while ((executeMatch = executeRegex.exec(responseMessage.content)) !== null) {
+                    const bashCmd = executeMatch[1].trim();
+                    if (bashCmd) {
+                        fallbacks.push({
+                            id: 'call_fallback_' + Math.random().toString(36).substr(2, 9),
+                            type: 'function',
+                            function: {
+                                name: 'run_terminal_command',
+                                arguments: JSON.stringify({ command: bashCmd })
+                            }
+                        });
+                    }
+                }
                 if (fallbacks.length > 0) {
                     responseMessage.tool_calls = fallbacks;
-                    responseMessage.content = responseMessage.content.replace(/^\/[a-zA-Z0-9_]+\s+.*$/gm, '').replace(/```(?:bash|sh|zsh)?\n([\s\S]*?)```/g, '').trim();
+                    responseMessage.content = responseMessage.content.replace(/^\/[a-zA-Z0-9_]+\s+.*$/gm, '').replace(/```(?:bash|sh|zsh)?\n([\s\S]*?)```/g, '').replace(/<tool_code>([\s\S]*?)<\/tool_code>/g, '').replace(/<(?:execute_bash|execute)>([\s\S]*?)<\/(?:execute_bash|execute)>/g, '').trim();
                     console.log(picocolors_1.default.cyan(`[Fallback Parser] Intercepted ${fallbacks.length} raw text commands and converted to tool_calls.`));
                     // Update logger entry with the intercepted tool calls
                     exports.logger.addEntry({
@@ -163,6 +205,22 @@ async function processWeb3Intent(input, role = 'user', onProgress, sessionId) {
                 }
             }
             // ---------------------------------------------------------------
+            // --- DISPLAY SANITIZATION ---
+            // Remove commonly leaked XML tags (reasoning & tool calls) to prevent UI clutter
+            if (responseMessage.content) {
+                const tagsToRemove = ['tool_code', 'tool_call', 'tool_calls', 'function_call', 'function_calls', 'execute', 'think', 'thought', 'reasoning'];
+                tagsToRemove.forEach(tag => {
+                    // Remove complete tags and their contents
+                    const regex = new RegExp(`<${tag}\\b[^>]*>[\\s\\S]*?<\\/${tag}>\\s*`, 'gi');
+                    responseMessage.content = responseMessage.content.replace(regex, '');
+                    // Remove any orphaned closing tags
+                    const orphanRegex = new RegExp(`<\\/${tag}>\\s*`, 'gi');
+                    responseMessage.content = responseMessage.content.replace(orphanRegex, '');
+                });
+                responseMessage.content = responseMessage.content.trim();
+                cleanedContent = responseMessage.content;
+            }
+            // ---------------------------------------------------------------
             if (!responseMessage.tool_calls || responseMessage.tool_calls.length === 0) {
                 return cleanedContent || '⚠️ I encountered an issue processing your request. This can happen with very complex multi-step tasks. Please try rephrasing or breaking the request into smaller steps.';
             }
@@ -284,10 +342,9 @@ async function processWeb3Intent(input, role = 'user', onProgress, sessionId) {
             else {
                 consecutiveToolErrors = 0;
             }
-            // V2 Optimization: Zero-LLM Fast Return for transaction tools
             if (canFastReturnAll && accumulatedResults.length > 0) {
                 const finalContent = accumulatedResults.join('\n\n---\n\n');
-                exports.logger.addEntry({ role: 'assistant', content: finalContent }, sessionId);
+                exports.logger.addEntry({ role: 'user', content: `[System Notification: The tool executed successfully and returned the following result directly to the user]\n${finalContent}` }, sessionId);
                 return finalContent;
             }
             // Loop continues, sending tool results in the next turn
@@ -346,10 +403,9 @@ async function processWeb3IntentStream(input, onChunk, onProgress, sessionId) {
             let streamedContent = '';
             const response = await (0, llmUtils_1.executeWithRetry)(async (client) => {
                 streamedContent = '';
-                // RC#1 FIX: Only clear the Telegram buffer on the FIRST turn.
-                // Subsequent turns must NOT wipe the buffer that shows tool progress.
-                if (turnCount === 1)
-                    onChunk('[CLEAR_STREAM]');
+                // RC#1 FIX: Always clear the buffer at the start of the stream turn.
+                // Subsequent turns must wipe the buffer to prevent UI duplication of the planning text.
+                onChunk('[CLEAR_STREAM]');
                 return await client.stream({ model: config.llm.model, temperature: config.llm.temperature, messages, tools: activeTools, reasoning_effort: config.llm.reasoning_effort }, (chunk) => {
                     streamedContent += chunk;
                     onChunk(chunk);

package/dist/packages/core/src/channels/telegram.js

@@ -204,23 +204,12 @@ function startTelegramBot() {
                     buffer = '⏳ Processing...';
                     finalContentAlreadyStreamed = false;
                 }
-                else if (chunk === '[TOOL_CALL_DETECTED]' || chunk === '[TOOL_CALL_FINISHED]') {
-                    // Hermes-Agent tool boundary: Finalize the current message synchronously to avoid race conditions.
-                    const htmlToSave = formatToTelegramHTML(buffer);
-                    const msgIdToSave = progressMsgId;
-                    progressMsgId = null;
-                    buffer = '';
-                    finalContentAlreadyStreamed = false;
-                    if (pendingEdit) {
-                        clearTimeout(pendingEdit);
-                        pendingEdit = null;
-                    }
-                    if (msgIdToSave && htmlToSave) {
-                        await ctx.api.editMessageText(ctx.chat.id, msgIdToSave, htmlToSave, { parse_mode: 'HTML' }).catch(() => { });
-                    }
-                    else if (!msgIdToSave && htmlToSave) {
-                        await ctx.reply(htmlToSave, { parse_mode: 'HTML', reply_parameters: { message_id: ctx.message.message_id } }).catch(() => { });
-                    }
+                else if (chunk === '[TOOL_CALL_DETECTED]') {
+                    buffer = '⏳ Processing...';
+                    scheduleEdit();
+                    return;
+                }
+                else if (chunk === '[TOOL_CALL_FINISHED]') {
                     return;
                 }
                 else {
@@ -412,21 +401,12 @@ function startTelegramBot() {
                     if (chunk === '[CLEAR_STREAM]') {
                         buffer = '⏳ Processing...';
                     }
-                    else if (chunk === '[TOOL_CALL_DETECTED]' || chunk === '[TOOL_CALL_FINISHED]') {
-                        const htmlToSave = formatToTelegramHTML(buffer);
-                        const msgIdToSave = progressMsgId;
-                        progressMsgId = null;
-                        buffer = '';
-                        if (pendingEdit) {
-                            clearTimeout(pendingEdit);
-                            pendingEdit = null;
-                        }
-                        if (msgIdToSave && htmlToSave) {
-                            await ctx.api.editMessageText(ctx.chat.id, msgIdToSave, htmlToSave, { parse_mode: 'HTML' }).catch(() => { });
-                        }
-                        else if (!msgIdToSave && htmlToSave) {
-                            await ctx.reply(htmlToSave, { parse_mode: 'HTML', reply_parameters: { message_id: ctx.message.message_id } }).catch(() => { });
-                        }
+                    else if (chunk === '[TOOL_CALL_DETECTED]') {
+                        buffer = '⏳ Processing...';
+                        scheduleEdit();
+                        return;
+                    }
+                    else if (chunk === '[TOOL_CALL_FINISHED]') {
                         return;
                     }
                     else {

package/dist/packages/core/src/gateway/setup.js

@@ -277,6 +277,8 @@ Provider: ${config.llm.provider}`;
                 { value: 'base_sepolia', label: 'Base Sepolia (Testnet)' },
                 { value: 'arbitrum_sepolia', label: 'Arbitrum Sepolia (Testnet)' },
                 { value: 'optimism_sepolia', label: 'OP Sepolia (Testnet)' },
+                { value: 'robinhood', label: 'Robinhood Chain' },
+                { value: 'robinhood_testnet', label: 'Robinhood Testnet' },
             ],
         });
         if ((0, prompts_1.isCancel)(defaultChain))

package/dist/packages/core/src/utils/contextSummarizer.js

@@ -16,51 +16,37 @@ exports.needsCompression = needsCompression;
 const llmUtils_1 = require("./llmUtils");
 const parser_1 = require("../config/parser");
 const picocolors_1 = __importDefault(require("picocolors"));
-/** How many text-only exchanges to keep verbatim before summarising older ones. */
-const VERBATIM_TAIL = 10;
-/** Minimum history length (user+assistant pairs) before we bother summarising. */
-const SUMMARISE_THRESHOLD = 20;
-/**
- * Cache: prevents re-compressing the same history on every nudge iteration.
- * Key: sessionId — value contains the textCount at time of compression and
- * the resulting compressed history.
- *
- * During nudge loops, agent adds only system-role nudge messages. Those do NOT
- * increment the textCount (user+assistant only). So the cache key stays the same
- * and we return the cached result instead of making another LLM call.
- * Cache is invalidated when the textCount grows (i.e. real LLM response landed).
- */
+const VERBATIM_TAIL = 20;
+const SUMMARISE_THRESHOLD = 40;
 const compressionCache = new Map();
-/**
- * Returns a history array that fits in context.
- * If history is short, returns as-is.
- * If long, prepends a "Conversation Summary" system message and keeps the tail verbatim.
- *
- * @param history  Full raw history from logger.
- * @param sessionId  Optional session identifier used to cache results across nudge turns.
- */
 async function compressHistory(history, sessionId) {
-    const textMessages = history.filter(m => (m.role === 'user' || m.role === 'assistant') && m.content && !m.tool_calls);
-    if (textMessages.length < SUMMARISE_THRESHOLD) {
-        return history; // nothing to compress yet
+    if (history.length < SUMMARISE_THRESHOLD) {
+        return history;
     }
-    // Cache hit: textCount hasn't grown since last compression for this session.
-    // This prevents 4x redundant LLM summarization calls during nudge iterations.
     if (sessionId) {
         const cached = compressionCache.get(sessionId);
-        if (cached && cached.textCount === textMessages.length) {
+        // FIX: Invalidate cache if ANY message is added to history, including tool calls
+        if (cached && cached.textCount === history.length) {
             return cached.compressed;
         }
     }
-    // Split: old messages to summarise, recent tail to keep verbatim
-    const tailMessages = textMessages.slice(-VERBATIM_TAIL);
-    const oldMessages = textMessages.slice(0, -VERBATIM_TAIL);
+    // FIX: Keep the tail strictly chronological so LLM doesn't get confused
+    const tailMessages = history.slice(-VERBATIM_TAIL);
+    const oldMessages = history.slice(0, -VERBATIM_TAIL);
     if (oldMessages.length === 0)
         return history;
     try {
         const config = (0, parser_1.loadConfig)();
         const historyText = oldMessages
-            .map(m => `${m.role === 'user' ? 'USER' : 'ASSISTANT'}: ${m.content}`)
+            .map(m => {
+            let name = m.role.toUpperCase();
+            let content = m.content || '';
+            if (m.role === 'assistant' && m.tool_calls)
+                content = `[Called tools: ${m.tool_calls.map(tc => tc.function.name).join(', ')}]`;
+            if (m.role === 'tool')
+                content = `[Tool Result: ${m.name}]`;
+            return `${name}: ${content}`;
+        })
             .join('\n');
         const summaryRes = await (0, llmUtils_1.executeWithRetry)(async (client) => client.chat({
             model: config.llm.model,
@@ -79,31 +65,22 @@ Write in third person. Be factual, not narrative.`
         if (!summaryText)
             return history;
         console.log(picocolors_1.default.magenta(`[ContextSummarizer] Compressed ${oldMessages.length} old messages into summary.`));
-        // Return: summary as system note + verbatim tail + all tool messages (never drop tool messages)
-        const toolMessages = history.filter(m => m.role === 'tool' || (m.role === 'assistant' && m.tool_calls));
         const compressed = [
             {
                 role: 'system',
                 content: `--- CONVERSATION SUMMARY (earlier context) ---\n${summaryText}\n--- END SUMMARY ---`
             },
-            ...tailMessages,
-            ...toolMessages.slice(-6) // keep last 6 tool interactions
+            ...tailMessages
         ];
-        // Persist in cache for this session
         if (sessionId) {
-            compressionCache.set(sessionId, { textCount: textMessages.length, compressed });
+            compressionCache.set(sessionId, { textCount: history.length, compressed });
         }
         return compressed;
     }
     catch (e) {
-        // On failure, just return the tail — never crash the agent
         return history.slice(-VERBATIM_TAIL * 2);
     }
 }
-/**
- * Fast check — should we attempt compression?
- */
 function needsCompression(history) {
-    const textCount = history.filter(m => (m.role === 'user' || m.role === 'assistant') && m.content && !m.tool_calls).length;
-    return textCount >= SUMMARISE_THRESHOLD;
+    return history.length >= SUMMARISE_THRESHOLD;
 }