ContextOS v3.0.0

ContextOS 3.0.0 — The context exoskeleton for AI development

ContextOS supports the full AI development lifecycle: understand a project, plan work, inspect evidence, execute, implement, verify, review, and resume. It keeps project state inspectable while reducing repeated token use and main-agent context occupancy.

One OS across the lifecycle

• Evidence and commands: precise source reads, reusable results, and execution receipts.
• Pipelines and changes: serial or parallel batches, scoped edits, verification, and integration.
• Blocks and Chains: architecture connected to the implementation and its dependencies.
• Plans, tasks, and sessions: persistent intent, progress, decisions, and continuity across conversations.
• Micro and CLI: bounded assistance and isolated implementation workers, with the main agent retaining decisions and final review.

App settings and model synchronization

Configure shared models and reasoning levels, then use Use Global for role-level inheritance or choose role-specific settings. Synchronize the models and effort options available from your Micro provider and CLI adapters.

The App has compact settings cards, a single dropdown arrow, improved scrolling, and a responsive Windows layout. An empty key field preserves the existing API key for the same origin; changing origin requires a new key.

Verified checks include 45 macOS Swift tests, the Windows frontend build, and 7 model-catalog tests. Live Micro, Codex, and AGY catalog checks passed, and manual login and invocation tests passed.

Historical measurement

A real plan-dependency task starting at 0dd6431 used fresh conversations and isolated worktrees. Peak main-thread input was 234,174 tokens with native tools, 145,694 with OS + Micro (−37.8%), and 84,681 with OS + Micro + CLI (−63.8%).

Under the specified comparison model, equivalent USD was $1.6230556 native, $1.3157976 OS only, $1.4137650 OS + Micro, and $0.7900901 OS + Micro + CLI. Rates are $2/$0.10/$10 per million uncached-input/cached-input/output tokens, with assistant cost divided by 7. Cached input is included in input; reasoning is included in output.

These are one task's historical results, not a new release benchmark or actual provider invoices. The full workflow used more combined raw tokens. Its 51.3% equivalent-cost reduction is an observed subtotal: D Micro reported usage for 128 of 129 started requests, with one request's usage unknown.

Methodology · Usage JSON · Role costs CSV · Four-arm totals CSV · Recalculation script · SHA-256 hashes

Install with AI

Send your coding assistant:

读取 (https://github.com/yubinbin32-ops/ContextOS/blob/main/setup.md) ，为我安装OS

The installation guide covers installation, host plugins and skills, project registration, Micro API configuration, CLI adapters, authentication, and acceptance checks. Node.js 22+ is required. Micro needs its own API connection; a CLI login alone does not configure it.

English overview · 中文介绍 · Report an issue

Full Changelog: v2.7.1...v3.0.0