RAME Code v0.2.0

RAME Code v0.2.0

Experimental preview of the local-first AI coding agent for Visual Studio Code.

Highlights

• governed, versioned project memory with human authorization;
• persistent Missions with independent Supervised and Autonomous workspace grants;
• host-checked conventional, protected and reinforced verification;
• managed llama-server, plus Ollama and LM Studio/private-LAN model servers;
• optional local-model web research and an experimental, approval-gated Agent terminal;
• direct Meta Model API Chat and Agent, with bounded Mission Advisory that leaves tools under Glimmer control;
• six sanitized REAL/RECORDED evidence packages and a separate reproducible SIMULATED sabotage battery.

Validation

• 759/759 automated tests on Node.js 20 and Node.js 24;
• 44/44 simulated adversarial scenarios: 24 blocked, 17 detected, 3 documented limits;
• GitHub Actions passed on both supported Node versions;
• the attached VSIX was built from commit be1240bec09f7994c1e9c89e05593c36aff0a2c5, installed with VS Code, and identifies itself as be1240b.src.19a005c58df3 / release.

Important limits

This is not a production security boundary. Project scripts and approved terminal commands run with the VS Code user's operating-system permissions; run_command has no OS sandbox. Agent web has one recorded real validation, while Chat and Mission web remain validated only with simulated servers. Three real Meta Spark 1.3 Chat responses are recorded; cloud Agent tool-result handling, Delegated, Mission planning and Mission Advisory remain simulated-provider validations. Vision input and Visual Studio Marketplace distribution are not included yet.

Read the user guide, claims and evidence, and verification guide before use.