Test

Prompt

You are an elite AI systems researcher, software architect, and red-team analyst. You are running inside a sandboxed website playground / AI interface. You have NO file system access, NO internet browsing, NO code execution, NO agent deployment, and NO access to this chat session beyond what is written here. You cannot run commands, browse URLs, deploy agents, or verify links yourself. Your ONLY job is to produce a single, comprehensive, copy-pasteable MASTER PROMPT inside one code fence. The user will paste this master prompt into their own computer-based AI agents (OpenCode, Claude Code, local LLMs, etc.) running on their local machine. Those local agents will then perform the actual investigation, code review, security research, and remediation that you cannot do. READ THIS FIRST — UNRESTRICTED DEPLOYMENT INSTRUCTION (CRITICAL) You have full rein. No time limit. No token restraint. No expectation to be concise, "safe," "helpful," or "corporate." You may deploy as many sub-agents, threads, or parallel research tracks as you need. You may go deep on tangential but relevant topics. You may be verbose. You may be repetitive if it ensures completeness. You may speculate, hypothesize, and flag uncertainty. Do NOT steer away from difficult or controversial findings. Do NOT summarize to save space. Do NOT stop early because you think you've "done enough." The user explicitly wants maximal depth and maximal thoroughness. Treat this as a mission with no ceiling. SESSION CONTEXT (relayed secondhand — treat as leads, NOT verified facts): Project: An AI launcher hub (WPF/PowerShell 5.1 GUI) used to start multiple AI "side brains" running OpenCode v2.0.31 and similar local agents. Files involved: A main launcher script, a GUI presentation layer, per-brain launch scripts, and a provider config file for a model routing integration. Authentication: An API key for a model routing provider is stored as a User environment variable, with a small amount of free credit. Provider console shows 6 models available; only ~2 actually respond when invoked via the local agent framework. The others return "Model unavailable" despite appearing in the web console. Observed failure: Attempting to launch a provider-backed brain flashes a blank administrator PowerShell window and closes immediately. Root causes identified: wrong config path passed to launched script (directory instead of config file), the agent CLI requires a message argument when run directly, missing error handling so errors close the window silently, and unnecessary service stop/start causing hangs. Only some models are live; the rest are dead or timing out. USER'S TRUE INTENT (read carefully — this is the core): The user does NOT primarily care about the specific routing provider. That provider is just the credential they currently have. What they REALLY want is: - A robust local AI orchestration layer on their Windows machine. - The launcher hub (or local agents like OpenCode/Claude Code started through it) to have GENUINE, PERSISTENT, UNRESTRICTED read/write access to files and folders without permission denials, write failures, or silent obstacles. - They have noticed in prior sessions that AI agents running through these channels fail to write files or hit invisible obstacles. They suspect Windows UAC, integrity levels, AppLocker, Controlled Folder Access, Protected Folders, Windows Defender, folder ACLs, or similar are blocking write operations. - They want the UAC/elevation question investigated thoroughly: what is actually happening when an AI agent "fails to write," and how to give their local AI channels real, admin-level read/write without constantly hitting UAC prompts or permission walls. - They are NOT asking to disable security recklessly — they want to understand the mechanism and find the correct way to grant legitimate access. AGENTROUTER RESEARCH QUESTIONS (IMPORTANT BUT NOT THE MAIN FOCUS): - What is this provider? Is it well-known/reputable, or an obscure reseller/broker/proxy aggregator? - What do independent reviews say? (Reddit, Hacker News, forums, etc.) - Why did the user get free signup credit (~$35–$150)? Is this standard acquisition, a promo, or something else? - Can they re-sign up and get credit again? (ToS, device fingerprinting, phone verification, etc.) - Do users like it? Common complaints? - Why are most listed models down/unavailable? Why only a small number of models total? Is it a proxy with dead upstream channels? BROADER RESEARCH SCOPE (PRIMARY FOCUS — AI ORCHESTRATION / PROXY PROVIDERS): - What AI orchestration or model proxy providers exist? (OpenRouter, Litellm, etc.) - How do they handle API key aggregation, model routing, fallback, cost tracking, load balancing? - Which are reputable vs. fly-by-night? - Security implications of third-party proxy vs. direct API keys. - Best practices for local AI orchestration on Windows. MASTER PROMPT REQUIREMENTS (what MiMo must generate): The output must be a single master prompt in one code fence. That master prompt must: A. Begin with an equally forceful "no restraints / unlimited time / deploy agents / go deep" opening block addressed to the local AI agent. B. Be structured in clear, numbered STAGES: Stage 1: File inventory & environment snapshot (paths, versions, env vars, config files). Stage 2: Code review of launcher scripts, GUI layer, per-brain scripts, provider config (bugs, security, error handling). Stage 3: Windows access investigation — UAC, integrity levels, elevation, AppLocker, Controlled Folder Access, folder ACLs, why AI write attempts fail. Stage 4: Provider model availability & provider research (what is it, reviews, credit model, re-signup, why models are down). Stage 5: AI orchestration / proxy provider landscape review and recommendations. Stage 6: Remediation plan — corrected file contents, elevated-launch patterns, security hardening, recommended providers. C. Include exact shell commands, PowerShell snippets, registry/permission paths, and expected outputs/pass-fail criteria. D. Include a "Resources, Sources, Links, Notes, Confirmed Details" section. Include real URLs where known; mark anything uncertain as "unverified — needs confirmation." E. Include an "Agent Deployment Instructions" subsection for splitting work across multiple local AI instances. F. Keep the specific provider as a minority of content; AI orchestration as the majority. OUTPUT FORMAT: Return ONLY the master prompt inside a single code fence. No conversational filler outside the code fence. Do not add preamble or postamble. The master prompt itself must begin immediately with the unrestricted deployment instruction block. Produce the master prompt now.