Build and verify a genuinely local Hermes lane named local-private on this Windows computer. USER CONTEXT - Operating system: Windows. - GPU: NVIDIA GeForce RTX 3090 with 24 GB dedicated VRAM. - System RAM: 128 GB. - The selected Qwen model is already downloaded and working through Ollama. - An existing Hermes subscription/profile is connected to a $200 cloud plan. It is valuable and must remain completely unchanged. - The new local-private lane is for normal chat and authorized local work with files, terminal commands, Git, persistent memory, and skills. NON-NEGOTIABLE SAFETY AND ISOLATION RULES 1. Create a new blank Hermes profile named exactly local-private. 2. Never clone, import, overwrite, rename, migrate, reset, delete, or otherwise modify the existing cloud-connected Hermes profile. 3. Never copy cloud API keys, OAuth tokens, cookies, credentials, provider settings, conversation history, memory, or other cloud-profile state into local-private. 4. Before changing anything, identify and export or back up the existing Hermes configuration and profile inventory in a form that supports rollback. Do not expose secret values in logs or the final report. 5. Treat Hermes profiles as state/configuration isolation, not as operating-system sandboxes. Explain this distinction clearly. The local-private profile still has the Windows permissions of the process running Hermes, so file-system and command access must remain deliberately scoped. 6. Do not weaken Windows security, expose Ollama to the LAN, or bind any local service to 0.0.0.0. DISCOVER THE CURRENT INSTALLATION Do not guess Hermes commands, file locations, schema, or the Ollama model name. Inspect the installed Hermes version and current official documentation before editing configuration. Record, without revealing secrets: - Hermes version and installation method. - Existing profile names and which one is currently active. - The exact directories and files that contain profile configuration, memory, skills, logs, and credentials. - The exact Ollama version, listening address, and model inventory. - The exact working Qwen model tag reported by Ollama. Use commands such as `ollama list` and current Ollama inspection commands, but treat the returned exact tag as authoritative rather than any model name written in this prompt. - A baseline inventory of network endpoints contacted by the existing cloud profile, collected without sending private test content. If Hermes cannot create an independent blank profile without touching the existing profile, stop and report the limitation. Do not improvise a destructive migration. CREATE THE BLANK LOCAL PROFILE Using the installed Hermes profile mechanism and current documented commands: - Create a fresh, blank profile named exactly local-private. - Confirm the existing cloud-connected profile is still present and byte-for-byte or semantically unchanged, using a secret-safe configuration diff or equivalent evidence. - Keep local-private configuration, conversations, memory, skills, and logs separate from the cloud profile. - Make profile selection explicit. Do not silently change the default profile unless the user approves it. LOCAL MODEL CONFIGURATION Configure local-private to use only the already working Ollama-hosted Qwen model. - Discover and record the exact Ollama model tag; do not substitute, pull, or download a different model. - Use only this OpenAI-compatible base endpoint: http://127.0.0.1:11434/v1 - Keep Ollama bound to loopback only. Verify that port 11434 is not reachable through a LAN address. - Use the minimum non-secret placeholder value only if Hermes requires an API-key field for the local OpenAI-compatible provider. Do not reuse any cloud credential. - Set the discovered Ollama model as the model for normal chat and every Hermes internal role that could otherwise call a model. - Disable every cloud-model fallback. - Disable or replace with the same local Ollama model any hosted auxiliary, routing, embedding, compression, summarization, title-generation, memory-processing, tool-planning, or background model that could transmit content externally. - If any required Hermes function cannot run locally and cannot be disabled, stop and identify it. Do not claim the profile is local-private. LOCAL TOOLS AND OUTBOUND BOUNDARY Keep useful core local tools enabled for explicitly authorized work: - Read and edit local files within user-approved paths. - Run normal terminal and development commands. - Use Git locally, without an automatic push or remote fetch. - Store and retrieve profile-local memory. - Load and use profile-local skills. Separate outbound web/network capability from those local tools: - Local access to 127.0.0.1 for Ollama is allowed. - Disable autonomous web browsing, remote search, telemetry, remote MCP servers, cloud connectors, remote Git operations, and hosted tools by default where Hermes permits. - If outbound web access is retained as an optional tool, label it clearly as non-private and require explicit user approval before each use. - Before any outbound request, require explicit approval before transmitting local file contents, terminal output, Git content, memory, prompts containing private data, credentials, or other local information. - Never interpret permission to browse public information as permission to upload local content. CHANGE RECORD AND ROLLBACK Create a secret-safe change record containing: - Timestamp and Hermes/Ollama versions. - Files and settings changed, with before/after values redacted where sensitive. - Exact discovered Ollama model tag and the configured endpoint. - Local tool settings and outbound-network controls. - Export/backup locations and exact rollback steps. - A post-change comparison showing the existing cloud profile was not modified. Do not delete backups after testing. FUNCTIONAL TESTS Run all tests while local-private is explicitly active. Use harmless, disposable content in an approved temporary test directory. 1. Normal chat: complete a short chat and prove the answer came from the discovered Ollama model at 127.0.0.1:11434, not a cloud fallback. 2. Local file edit: create a harmless text file, read it, make a small edit, verify the bytes, and remove it only if the user-approved test plan allows cleanup. 3. Terminal: run a harmless command such as printing the current directory and record the result. 4. Git: inspect local repository status without fetching, pushing, or contacting a remote. 5. Memory: store a unique non-sensitive test fact, start a new conversation in local-private, retrieve it, and confirm it does not appear in the cloud profile. 6. Skills: load one harmless local skill and confirm it executes without a hosted dependency. 7. Restart persistence: restart Hermes cleanly, reactivate local-private, and repeat the chat/model-endpoint and memory checks. NETWORK PRIVACY VERIFICATION Do not claim privacy from configuration settings alone. Verify the tests with outbound internet blocked, or use an equivalent trustworthy network-observation method that can distinguish loopback traffic from external connections. Windows Firewall rules, a controlled network isolation method, or packet/process connection observation may be used, but preserve a rollback and do not disrupt unrelated applications. During the full functional test suite: - Allow the required loopback connection to 127.0.0.1:11434. - Record every hostname, IP address, port, protocol, and owning process contacted by Hermes and relevant child processes. - Prove that model requests used only http://127.0.0.1:11434/v1. - Prove that no external endpoint received prompts, local content, terminal output, Git data, memory, or skill data. - Treat DNS lookups, telemetry, update checks, hosted authentication, remote MCP traffic, and cloud model calls as outbound activity and report them. - If any unexplained external connection occurs, stop, preserve the evidence, correct the configuration if safe, and rerun the complete test. Do not call the lane local-private until the observation test passes. FINAL EVIDENCE REPORT Return a concise report with: - Hermes and Ollama versions. - Existing cloud profile preservation evidence. - local-private creation evidence and its isolated state locations. - Exact Ollama model tag and configured base endpoint. - Enabled local tools and disabled/approval-gated outbound capabilities. - Results for chat, file, terminal, Git, memory, skills, and restart persistence. - The network-isolation or observation method used. - A complete endpoint-contact table with hostname/IP, port, process, reason, whether content was transmitted, and whether the endpoint was loopback or external. - Any residual privacy limitations, including the fact that a Hermes profile is not an OS sandbox. - Backup/export locations and tested rollback steps. If any required evidence is missing, say that verification is incomplete and do not claim the profile is genuinely local-private.