v26.7.14

Agent Identity & Execution Discipline Overhaul

• Semantic Intent Router Refactor (reasoning.ts): Upgraded the intent router to utilize a strict CLASSIFICATION MATRIX and Context Hierarchy Rules. This completely eradicates intent misclassification between Web3 and OS workflows, especially during context-switching.
• Agent Identity Compression (promptBuilder.ts): Consolidated verbose [CRITICAL EXECUTION RULES] into 5 high-impact, token-efficient mandates (e.g., OUTPUT RESTRICTION, THINKING GATE, HARD HARDENING) to maximize model attention retention.
• Strict Real-World Facts Guardrail: Introduced mandatory search_web requirements for sports scores, schedules, and real-world news to completely eliminate LLM hallucination on dynamic data.
• Anti-Silent Stop / Interactive Execution Flow (promptBuilder.ts): Replaced standard tool-call enforcements with [INTERACTIVE EXECUTION FLOW]. The agent is now permitted exactly one short conversational sentence before a tool call (both on success and recovery paths), but is strictly forbidden from ending its turn without attaching the corrected tool payload. This cures the "Silent Stop" / "Promise Without Action" bug while maintaining a natural, conversational UX.
• Force Action User Correction (osAgent.ts): Refactored the User Correction Detectors to intercept scoldings with a [CRITICAL INTERCEPT: USER CORRECTION] signal. The LLM is now forced to fetch fresh ground-truth data via tools instead of endlessly apologizing and recycling stale context.

UI/UX & Quality of Life

• Import Project Workflow Redesign: Relocated the "Import Project" button from the main navigation sidebar into the "Workspaces" section header. The action is now represented by an intuitive + icon that perfectly scales with the slightly enlarged 0.85rem section text, achieving a significantly cleaner UI layout while keeping workspace management centralized.

Performance — LLM Response Speed Optimization

Cold Start Latency Fix

• Non-Blocking DeFi Aggregator Discovery (server.ts): aggregatorRegistry.autoDiscover() was previously await-ed synchronously before app.listen(), blocking the entire server startup by 2–5 seconds while it probed external DeFi providers. Refactored to setImmediate() fire-and-forget so the server starts accepting LLM requests immediately after plugins are loaded, with provider discovery happening in the background.
• ML Engine Fetch Timeout (promptBuilder.ts): All 3 network calls to the local Python ML Engine (/memory/rag, /memory/narrative, /skills/list) previously had no timeout. During cold start, when the ML Engine is still booting, these calls would hang for up to 2 minutes waiting for a TCP connection. Added AbortSignal.timeout(1500) to each fetch — if the ML Engine is not ready, calls fail in ≤1.5s and return empty strings gracefully, keeping the first LLM response fast.

Per-Request Latency Optimization

• Parallel Volatile Prompt Parts (promptBuilder.ts): buildEpisodicMemories() and buildNarrativeMemories() were called sequentially (await one, then await the other). Refactored to Promise.all([...]) — both network calls now run concurrently, saving ~300–600ms per request.
• Parallel Narrative + Skills Fetch (promptBuilder.ts): Inside buildNarrativeMemories(), the /memory/narrative and /skills/list fetches were also sequential. Parallelized with Promise.all().
• TTL Cache for Narrative Memory & Skills (promptBuilder.ts): Added a 30-second in-memory TTL cache for narrative memory and skills list. These change only when the user explicitly saves something — re-fetching on every message was wasteful. Cache hit returns instantly with zero network overhead.
• 5-Second Build Cache (promptBuilder.ts): Added a short-lived cache keyed by agentType + userInput[:80]. Prevents double-building the system prompt when the router warm-up and the agent's own getSystemPrompt() call happen within 5 seconds of each other.
• Parallel LLM Router + System Prompt Warm-Up (reasoning.ts): For messages that don't match any keyword (triggering the LLM semantic router), the router call and the OS system prompt pre-build now run simultaneously via Promise.all(). Since 'os' is the most common fallback, its system prompt is pre-warmed into the 5s build cache while the router is deciding — making the router's latency effectively invisible to the user. Applies to both sync and stream paths.
• Static Import for historySanitizer (osAgent.ts, web3Agent.ts): Dynamic require('../utils/historySanitizer') inside function bodies (4 occurrences across 2 files) replaced with static ES import at the top of each file.